?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Merge tag 'hwmon-for-v6.13-rc7' of git://git.kernel.org/pub/scm/linux/kernel/git...
[linux.git] / net / mptcp / protocol.c
blob1b2e7cbb577fc26280f31e58adceb36987112f54
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_states.h>
19 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
20 #include <net/transp_v6.h>
21 #endif
22 #include <net/mptcp.h>
23 #include <net/hotdata.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 WRITE_ONCE(subflow->local_id, 0);
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 ((to->len + from->len) > (sk->sk_rcvbuf >> 3)) ||
140 !skb_try_coalesce(to, from, &fragstolen, &delta))
141 return false;
142
143 pr_debug("colesced seq %llx into %llx new len %d new end seq %llx\n",
144 MPTCP_SKB_CB(from)->map_seq, MPTCP_SKB_CB(to)->map_seq,
145 to->len, MPTCP_SKB_CB(from)->end_seq);
146 MPTCP_SKB_CB(to)->end_seq = MPTCP_SKB_CB(from)->end_seq;
147
148 /* note the fwd memory can reach a negative value after accounting
149 * for the delta, but the later skb free will restore a non
150 * negative one
151 */
152 atomic_add(delta, &sk->sk_rmem_alloc);
153 mptcp_rmem_charge(sk, delta);
154 kfree_skb_partial(from, fragstolen);
155
156 return true;
157 }
158
159 static bool mptcp_ooo_try_coalesce(struct mptcp_sock *msk, struct sk_buff *to,
160 struct sk_buff *from)
161 {
162 if (MPTCP_SKB_CB(from)->map_seq != MPTCP_SKB_CB(to)->end_seq)
163 return false;
164
165 return mptcp_try_coalesce((struct sock *)msk, to, from);
166 }
167
168 static void __mptcp_rmem_reclaim(struct sock *sk, int amount)
169 {
170 amount >>= PAGE_SHIFT;
171 mptcp_rmem_charge(sk, amount << PAGE_SHIFT);
172 __sk_mem_reduce_allocated(sk, amount);
173 }
174
175 static void mptcp_rmem_uncharge(struct sock *sk, int size)
176 {
177 struct mptcp_sock *msk = mptcp_sk(sk);
178 int reclaimable;
179
180 mptcp_rmem_fwd_alloc_add(sk, size);
181 reclaimable = msk->rmem_fwd_alloc - sk_unused_reserved_mem(sk);
182
183 /* see sk_mem_uncharge() for the rationale behind the following schema */
184 if (unlikely(reclaimable >= PAGE_SIZE))
185 __mptcp_rmem_reclaim(sk, reclaimable);
186 }
187
188 static void mptcp_rfree(struct sk_buff *skb)
189 {
190 unsigned int len = skb->truesize;
191 struct sock *sk = skb->sk;
192
193 atomic_sub(len, &sk->sk_rmem_alloc);
194 mptcp_rmem_uncharge(sk, len);
195 }
196
197 void mptcp_set_owner_r(struct sk_buff *skb, struct sock *sk)
198 {
199 skb_orphan(skb);
200 skb->sk = sk;
201 skb->destructor = mptcp_rfree;
202 atomic_add(skb->truesize, &sk->sk_rmem_alloc);
203 mptcp_rmem_charge(sk, skb->truesize);
204 }
205
206 /* "inspired" by tcp_data_queue_ofo(), main differences:
207 * - use mptcp seqs
208 * - don't cope with sacks
209 */
210 static void mptcp_data_queue_ofo(struct mptcp_sock *msk, struct sk_buff *skb)
211 {
212 struct sock *sk = (struct sock *)msk;
213 struct rb_node **p, *parent;
214 u64 seq, end_seq, max_seq;
215 struct sk_buff *skb1;
216
217 seq = MPTCP_SKB_CB(skb)->map_seq;
218 end_seq = MPTCP_SKB_CB(skb)->end_seq;
219 max_seq = atomic64_read(&msk->rcv_wnd_sent);
220
221 pr_debug("msk=%p seq=%llx limit=%llx empty=%d\n", msk, seq, max_seq,
222 RB_EMPTY_ROOT(&msk->out_of_order_queue));
223 if (after64(end_seq, max_seq)) {
224 /* out of window */
225 mptcp_drop(sk, skb);
226 pr_debug("oow by %lld, rcv_wnd_sent %llu\n",
227 (unsigned long long)end_seq - (unsigned long)max_seq,
228 (unsigned long long)atomic64_read(&msk->rcv_wnd_sent));
229 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_NODSSWINDOW);
230 return;
231 }
232
233 p = &msk->out_of_order_queue.rb_node;
234 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUE);
235 if (RB_EMPTY_ROOT(&msk->out_of_order_queue)) {
236 rb_link_node(&skb->rbnode, NULL, p);
237 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
238 msk->ooo_last_skb = skb;
239 goto end;
240 }
241
242 /* with 2 subflows, adding at end of ooo queue is quite likely
243 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
244 */
245 if (mptcp_ooo_try_coalesce(msk, msk->ooo_last_skb, skb)) {
246 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
247 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
248 return;
249 }
250
251 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
252 if (!before64(seq, MPTCP_SKB_CB(msk->ooo_last_skb)->end_seq)) {
253 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
254 parent = &msk->ooo_last_skb->rbnode;
255 p = &parent->rb_right;
256 goto insert;
257 }
258
259 /* Find place to insert this segment. Handle overlaps on the way. */
260 parent = NULL;
261 while (*p) {
262 parent = *p;
263 skb1 = rb_to_skb(parent);
264 if (before64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
265 p = &parent->rb_left;
266 continue;
267 }
268 if (before64(seq, MPTCP_SKB_CB(skb1)->end_seq)) {
269 if (!after64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) {
270 /* All the bits are present. Drop. */
271 mptcp_drop(sk, skb);
272 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
273 return;
274 }
275 if (after64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
276 /* partial overlap:
277 * | skb |
278 * | skb1 |
279 * continue traversing
280 */
281 } else {
282 /* skb's seq == skb1's seq and skb covers skb1.
283 * Replace skb1 with skb.
284 */
285 rb_replace_node(&skb1->rbnode, &skb->rbnode,
286 &msk->out_of_order_queue);
287 mptcp_drop(sk, skb1);
288 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
289 goto merge_right;
290 }
291 } else if (mptcp_ooo_try_coalesce(msk, skb1, skb)) {
292 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
293 return;
294 }
295 p = &parent->rb_right;
296 }
297
298 insert:
299 /* Insert segment into RB tree. */
300 rb_link_node(&skb->rbnode, parent, p);
301 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
302
303 merge_right:
304 /* Remove other segments covered by skb. */
305 while ((skb1 = skb_rb_next(skb)) != NULL) {
306 if (before64(end_seq, MPTCP_SKB_CB(skb1)->end_seq))
307 break;
308 rb_erase(&skb1->rbnode, &msk->out_of_order_queue);
309 mptcp_drop(sk, skb1);
310 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
311 }
312 /* If there is no skb after us, we are the last_skb ! */
313 if (!skb1)
314 msk->ooo_last_skb = skb;
315
316 end:
317 skb_condense(skb);
318 mptcp_set_owner_r(skb, sk);
319 }
320
321 static bool mptcp_rmem_schedule(struct sock *sk, struct sock *ssk, int size)
322 {
323 struct mptcp_sock *msk = mptcp_sk(sk);
324 int amt, amount;
325
326 if (size <= msk->rmem_fwd_alloc)
327 return true;
328
329 size -= msk->rmem_fwd_alloc;
330 amt = sk_mem_pages(size);
331 amount = amt << PAGE_SHIFT;
332 if (!__sk_mem_raise_allocated(sk, size, amt, SK_MEM_RECV))
333 return false;
334
335 mptcp_rmem_fwd_alloc_add(sk, amount);
336 return true;
337 }
338
339 static bool __mptcp_move_skb(struct mptcp_sock *msk, struct sock *ssk,
340 struct sk_buff *skb, unsigned int offset,
341 size_t copy_len)
342 {
343 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
344 struct sock *sk = (struct sock *)msk;
345 struct sk_buff *tail;
346 bool has_rxtstamp;
347
348 __skb_unlink(skb, &ssk->sk_receive_queue);
349
350 skb_ext_reset(skb);
351 skb_orphan(skb);
352
353 /* try to fetch required memory from subflow */
354 if (!mptcp_rmem_schedule(sk, ssk, skb->truesize)) {
355 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RCVPRUNED);
356 goto drop;
357 }
358
359 has_rxtstamp = TCP_SKB_CB(skb)->has_rxtstamp;
360
361 /* the skb map_seq accounts for the skb offset:
362 * mptcp_subflow_get_mapped_dsn() is based on the current tp->copied_seq
363 * value
364 */
365 MPTCP_SKB_CB(skb)->map_seq = mptcp_subflow_get_mapped_dsn(subflow);
366 MPTCP_SKB_CB(skb)->end_seq = MPTCP_SKB_CB(skb)->map_seq + copy_len;
367 MPTCP_SKB_CB(skb)->offset = offset;
368 MPTCP_SKB_CB(skb)->has_rxtstamp = has_rxtstamp;
369
370 if (MPTCP_SKB_CB(skb)->map_seq == msk->ack_seq) {
371 /* in sequence */
372 msk->bytes_received += copy_len;
373 WRITE_ONCE(msk->ack_seq, msk->ack_seq + copy_len);
374 tail = skb_peek_tail(&sk->sk_receive_queue);
375 if (tail && mptcp_try_coalesce(sk, tail, skb))
376 return true;
377
378 mptcp_set_owner_r(skb, sk);
379 __skb_queue_tail(&sk->sk_receive_queue, skb);
380 return true;
381 } else if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) {
382 mptcp_data_queue_ofo(msk, skb);
383 return false;
384 }
385
386 /* old data, keep it simple and drop the whole pkt, sender
387 * will retransmit as needed, if needed.
388 */
389 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
390 drop:
391 mptcp_drop(sk, skb);
392 return false;
393 }
394
395 static void mptcp_stop_rtx_timer(struct sock *sk)
396 {
397 struct inet_connection_sock *icsk = inet_csk(sk);
398
399 sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
400 mptcp_sk(sk)->timer_ival = 0;
401 }
402
403 static void mptcp_close_wake_up(struct sock *sk)
404 {
405 if (sock_flag(sk, SOCK_DEAD))
406 return;
407
408 sk->sk_state_change(sk);
409 if (sk->sk_shutdown == SHUTDOWN_MASK ||
410 sk->sk_state == TCP_CLOSE)
411 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
412 else
413 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
414 }
415
416 /* called under the msk socket lock */
417 static bool mptcp_pending_data_fin_ack(struct sock *sk)
418 {
419 struct mptcp_sock *msk = mptcp_sk(sk);
420
421 return ((1 << sk->sk_state) &
422 (TCPF_FIN_WAIT1 | TCPF_CLOSING | TCPF_LAST_ACK)) &&
423 msk->write_seq == READ_ONCE(msk->snd_una);
424 }
425
426 static void mptcp_check_data_fin_ack(struct sock *sk)
427 {
428 struct mptcp_sock *msk = mptcp_sk(sk);
429
430 /* Look for an acknowledged DATA_FIN */
431 if (mptcp_pending_data_fin_ack(sk)) {
432 WRITE_ONCE(msk->snd_data_fin_enable, 0);
433
434 switch (sk->sk_state) {
435 case TCP_FIN_WAIT1:
436 mptcp_set_state(sk, TCP_FIN_WAIT2);
437 break;
438 case TCP_CLOSING:
439 case TCP_LAST_ACK:
440 mptcp_set_state(sk, TCP_CLOSE);
441 break;
442 }
443
444 mptcp_close_wake_up(sk);
445 }
446 }
447
448 /* can be called with no lock acquired */
449 static bool mptcp_pending_data_fin(struct sock *sk, u64 *seq)
450 {
451 struct mptcp_sock *msk = mptcp_sk(sk);
452
453 if (READ_ONCE(msk->rcv_data_fin) &&
454 ((1 << inet_sk_state_load(sk)) &
455 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2))) {
456 u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq);
457
458 if (READ_ONCE(msk->ack_seq) == rcv_data_fin_seq) {
459 if (seq)
460 *seq = rcv_data_fin_seq;
461
462 return true;
463 }
464 }
465
466 return false;
467 }
468
469 static void mptcp_set_datafin_timeout(struct sock *sk)
470 {
471 struct inet_connection_sock *icsk = inet_csk(sk);
472 u32 retransmits;
473
474 retransmits = min_t(u32, icsk->icsk_retransmits,
475 ilog2(TCP_RTO_MAX / TCP_RTO_MIN));
476
477 mptcp_sk(sk)->timer_ival = TCP_RTO_MIN << retransmits;
478 }
479
480 static void __mptcp_set_timeout(struct sock *sk, long tout)
481 {
482 mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN;
483 }
484
485 static long mptcp_timeout_from_subflow(const struct mptcp_subflow_context *subflow)
486 {
487 const struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
488
489 return inet_csk(ssk)->icsk_pending && !subflow->stale_count ?
490 inet_csk(ssk)->icsk_timeout - jiffies : 0;
491 }
492
493 static void mptcp_set_timeout(struct sock *sk)
494 {
495 struct mptcp_subflow_context *subflow;
496 long tout = 0;
497
498 mptcp_for_each_subflow(mptcp_sk(sk), subflow)
499 tout = max(tout, mptcp_timeout_from_subflow(subflow));
500 __mptcp_set_timeout(sk, tout);
501 }
502
503 static inline bool tcp_can_send_ack(const struct sock *ssk)
504 {
505 return !((1 << inet_sk_state_load(ssk)) &
506 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_TIME_WAIT | TCPF_CLOSE | TCPF_LISTEN));
507 }
508
509 void __mptcp_subflow_send_ack(struct sock *ssk)
510 {
511 if (tcp_can_send_ack(ssk))
512 tcp_send_ack(ssk);
513 }
514
515 static void mptcp_subflow_send_ack(struct sock *ssk)
516 {
517 bool slow;
518
519 slow = lock_sock_fast(ssk);
520 __mptcp_subflow_send_ack(ssk);
521 unlock_sock_fast(ssk, slow);
522 }
523
524 static void mptcp_send_ack(struct mptcp_sock *msk)
525 {
526 struct mptcp_subflow_context *subflow;
527
528 mptcp_for_each_subflow(msk, subflow)
529 mptcp_subflow_send_ack(mptcp_subflow_tcp_sock(subflow));
530 }
531
532 static void mptcp_subflow_cleanup_rbuf(struct sock *ssk, int copied)
533 {
534 bool slow;
535
536 slow = lock_sock_fast(ssk);
537 if (tcp_can_send_ack(ssk))
538 tcp_cleanup_rbuf(ssk, copied);
539 unlock_sock_fast(ssk, slow);
540 }
541
542 static bool mptcp_subflow_could_cleanup(const struct sock *ssk, bool rx_empty)
543 {
544 const struct inet_connection_sock *icsk = inet_csk(ssk);
545 u8 ack_pending = READ_ONCE(icsk->icsk_ack.pending);
546 const struct tcp_sock *tp = tcp_sk(ssk);
547
548 return (ack_pending & ICSK_ACK_SCHED) &&
549 ((READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->rcv_wup) >
550 READ_ONCE(icsk->icsk_ack.rcv_mss)) ||
551 (rx_empty && ack_pending &
552 (ICSK_ACK_PUSHED2 | ICSK_ACK_PUSHED)));
553 }
554
555 static void mptcp_cleanup_rbuf(struct mptcp_sock *msk, int copied)
556 {
557 int old_space = READ_ONCE(msk->old_wspace);
558 struct mptcp_subflow_context *subflow;
559 struct sock *sk = (struct sock *)msk;
560 int space = __mptcp_space(sk);
561 bool cleanup, rx_empty;
562
563 cleanup = (space > 0) && (space >= (old_space << 1)) && copied;
564 rx_empty = !__mptcp_rmem(sk) && copied;
565
566 mptcp_for_each_subflow(msk, subflow) {
567 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
568
569 if (cleanup || mptcp_subflow_could_cleanup(ssk, rx_empty))
570 mptcp_subflow_cleanup_rbuf(ssk, copied);
571 }
572 }
573
574 static bool mptcp_check_data_fin(struct sock *sk)
575 {
576 struct mptcp_sock *msk = mptcp_sk(sk);
577 u64 rcv_data_fin_seq;
578 bool ret = false;
579
580 /* Need to ack a DATA_FIN received from a peer while this side
581 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2.
582 * msk->rcv_data_fin was set when parsing the incoming options
583 * at the subflow level and the msk lock was not held, so this
584 * is the first opportunity to act on the DATA_FIN and change
585 * the msk state.
586 *
587 * If we are caught up to the sequence number of the incoming
588 * DATA_FIN, send the DATA_ACK now and do state transition. If
589 * not caught up, do nothing and let the recv code send DATA_ACK
590 * when catching up.
591 */
592
593 if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) {
594 WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1);
595 WRITE_ONCE(msk->rcv_data_fin, 0);
596
597 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN);
598 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
599
600 switch (sk->sk_state) {
601 case TCP_ESTABLISHED:
602 mptcp_set_state(sk, TCP_CLOSE_WAIT);
603 break;
604 case TCP_FIN_WAIT1:
605 mptcp_set_state(sk, TCP_CLOSING);
606 break;
607 case TCP_FIN_WAIT2:
608 mptcp_set_state(sk, TCP_CLOSE);
609 break;
610 default:
611 /* Other states not expected */
612 WARN_ON_ONCE(1);
613 break;
614 }
615
616 ret = true;
617 if (!__mptcp_check_fallback(msk))
618 mptcp_send_ack(msk);
619 mptcp_close_wake_up(sk);
620 }
621 return ret;
622 }
623
624 static void mptcp_dss_corruption(struct mptcp_sock *msk, struct sock *ssk)
625 {
626 if (READ_ONCE(msk->allow_infinite_fallback)) {
627 MPTCP_INC_STATS(sock_net(ssk),
628 MPTCP_MIB_DSSCORRUPTIONFALLBACK);
629 mptcp_do_fallback(ssk);
630 } else {
631 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_DSSCORRUPTIONRESET);
632 mptcp_subflow_reset(ssk);
633 }
634 }
635
636 static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk,
637 struct sock *ssk,
638 unsigned int *bytes)
639 {
640 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
641 struct sock *sk = (struct sock *)msk;
642 unsigned int moved = 0;
643 bool more_data_avail;
644 struct tcp_sock *tp;
645 bool done = false;
646 int sk_rbuf;
647
648 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
649
650 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
651 int ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
652
653 if (unlikely(ssk_rbuf > sk_rbuf)) {
654 WRITE_ONCE(sk->sk_rcvbuf, ssk_rbuf);
655 sk_rbuf = ssk_rbuf;
656 }
657 }
658
659 pr_debug("msk=%p ssk=%p\n", msk, ssk);
660 tp = tcp_sk(ssk);
661 do {
662 u32 map_remaining, offset;
663 u32 seq = tp->copied_seq;
664 struct sk_buff *skb;
665 bool fin;
666
667 /* try to move as much data as available */
668 map_remaining = subflow->map_data_len -
669 mptcp_subflow_get_map_offset(subflow);
670
671 skb = skb_peek(&ssk->sk_receive_queue);
672 if (!skb) {
673 /* With racing move_skbs_to_msk() and __mptcp_move_skbs(),
674 * a different CPU can have already processed the pending
675 * data, stop here or we can enter an infinite loop
676 */
677 if (!moved)
678 done = true;
679 break;
680 }
681
682 if (__mptcp_check_fallback(msk)) {
683 /* Under fallback skbs have no MPTCP extension and TCP could
684 * collapse them between the dummy map creation and the
685 * current dequeue. Be sure to adjust the map size.
686 */
687 map_remaining = skb->len;
688 subflow->map_data_len = skb->len;
689 }
690
691 offset = seq - TCP_SKB_CB(skb)->seq;
692 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
693 if (fin) {
694 done = true;
695 seq++;
696 }
697
698 if (offset < skb->len) {
699 size_t len = skb->len - offset;
700
701 if (tp->urg_data)
702 done = true;
703
704 if (__mptcp_move_skb(msk, ssk, skb, offset, len))
705 moved += len;
706 seq += len;
707
708 if (unlikely(map_remaining < len)) {
709 DEBUG_NET_WARN_ON_ONCE(1);
710 mptcp_dss_corruption(msk, ssk);
711 }
712 } else {
713 if (unlikely(!fin)) {
714 DEBUG_NET_WARN_ON_ONCE(1);
715 mptcp_dss_corruption(msk, ssk);
716 }
717
718 sk_eat_skb(ssk, skb);
719 done = true;
720 }
721
722 WRITE_ONCE(tp->copied_seq, seq);
723 more_data_avail = mptcp_subflow_data_available(ssk);
724
725 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf) {
726 done = true;
727 break;
728 }
729 } while (more_data_avail);
730
731 if (moved > 0)
732 msk->last_data_recv = tcp_jiffies32;
733 *bytes += moved;
734 return done;
735 }
736
737 static bool __mptcp_ofo_queue(struct mptcp_sock *msk)
738 {
739 struct sock *sk = (struct sock *)msk;
740 struct sk_buff *skb, *tail;
741 bool moved = false;
742 struct rb_node *p;
743 u64 end_seq;
744
745 p = rb_first(&msk->out_of_order_queue);
746 pr_debug("msk=%p empty=%d\n", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue));
747 while (p) {
748 skb = rb_to_skb(p);
749 if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq))
750 break;
751
752 p = rb_next(p);
753 rb_erase(&skb->rbnode, &msk->out_of_order_queue);
754
755 if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq,
756 msk->ack_seq))) {
757 mptcp_drop(sk, skb);
758 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
759 continue;
760 }
761
762 end_seq = MPTCP_SKB_CB(skb)->end_seq;
763 tail = skb_peek_tail(&sk->sk_receive_queue);
764 if (!tail || !mptcp_ooo_try_coalesce(msk, tail, skb)) {
765 int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
766
767 /* skip overlapping data, if any */
768 pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d\n",
769 MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq,
770 delta);
771 MPTCP_SKB_CB(skb)->offset += delta;
772 MPTCP_SKB_CB(skb)->map_seq += delta;
773 __skb_queue_tail(&sk->sk_receive_queue, skb);
774 }
775 msk->bytes_received += end_seq - msk->ack_seq;
776 WRITE_ONCE(msk->ack_seq, end_seq);
777 moved = true;
778 }
779 return moved;
780 }
781
782 static bool __mptcp_subflow_error_report(struct sock *sk, struct sock *ssk)
783 {
784 int err = sock_error(ssk);
785 int ssk_state;
786
787 if (!err)
788 return false;
789
790 /* only propagate errors on fallen-back sockets or
791 * on MPC connect
792 */
793 if (sk->sk_state != TCP_SYN_SENT && !__mptcp_check_fallback(mptcp_sk(sk)))
794 return false;
795
796 /* We need to propagate only transition to CLOSE state.
797 * Orphaned socket will see such state change via
798 * subflow_sched_work_if_closed() and that path will properly
799 * destroy the msk as needed.
800 */
801 ssk_state = inet_sk_state_load(ssk);
802 if (ssk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DEAD))
803 mptcp_set_state(sk, ssk_state);
804 WRITE_ONCE(sk->sk_err, -err);
805
806 /* This barrier is coupled with smp_rmb() in mptcp_poll() */
807 smp_wmb();
808 sk_error_report(sk);
809 return true;
810 }
811
812 void __mptcp_error_report(struct sock *sk)
813 {
814 struct mptcp_subflow_context *subflow;
815 struct mptcp_sock *msk = mptcp_sk(sk);
816
817 mptcp_for_each_subflow(msk, subflow)
818 if (__mptcp_subflow_error_report(sk, mptcp_subflow_tcp_sock(subflow)))
819 break;
820 }
821
822 /* In most cases we will be able to lock the mptcp socket. If its already
823 * owned, we need to defer to the work queue to avoid ABBA deadlock.
824 */
825 static bool move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk)
826 {
827 struct sock *sk = (struct sock *)msk;
828 unsigned int moved = 0;
829
830 __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
831 __mptcp_ofo_queue(msk);
832 if (unlikely(ssk->sk_err)) {
833 if (!sock_owned_by_user(sk))
834 __mptcp_error_report(sk);
835 else
836 __set_bit(MPTCP_ERROR_REPORT, &msk->cb_flags);
837 }
838
839 /* If the moves have caught up with the DATA_FIN sequence number
840 * it's time to ack the DATA_FIN and change socket state, but
841 * this is not a good place to change state. Let the workqueue
842 * do it.
843 */
844 if (mptcp_pending_data_fin(sk, NULL))
845 mptcp_schedule_work(sk);
846 return moved > 0;
847 }
848
849 void mptcp_data_ready(struct sock *sk, struct sock *ssk)
850 {
851 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
852 struct mptcp_sock *msk = mptcp_sk(sk);
853 int sk_rbuf, ssk_rbuf;
854
855 /* The peer can send data while we are shutting down this
856 * subflow at msk destruction time, but we must avoid enqueuing
857 * more data to the msk receive queue
858 */
859 if (unlikely(subflow->disposable))
860 return;
861
862 ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
863 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
864 if (unlikely(ssk_rbuf > sk_rbuf))
865 sk_rbuf = ssk_rbuf;
866
867 /* over limit? can't append more skbs to msk, Also, no need to wake-up*/
868 if (__mptcp_rmem(sk) > sk_rbuf)
869 return;
870
871 /* Wake-up the reader only for in-sequence data */
872 mptcp_data_lock(sk);
873 if (move_skbs_to_msk(msk, ssk) && mptcp_epollin_ready(sk))
874 sk->sk_data_ready(sk);
875 mptcp_data_unlock(sk);
876 }
877
878 static void mptcp_subflow_joined(struct mptcp_sock *msk, struct sock *ssk)
879 {
880 mptcp_subflow_ctx(ssk)->map_seq = READ_ONCE(msk->ack_seq);
881 WRITE_ONCE(msk->allow_infinite_fallback, false);
882 mptcp_event(MPTCP_EVENT_SUB_ESTABLISHED, msk, ssk, GFP_ATOMIC);
883 }
884
885 static bool __mptcp_finish_join(struct mptcp_sock *msk, struct sock *ssk)
886 {
887 struct sock *sk = (struct sock *)msk;
888
889 if (sk->sk_state != TCP_ESTABLISHED)
890 return false;
891
892 /* attach to msk socket only after we are sure we will deal with it
893 * at close time
894 */
895 if (sk->sk_socket && !ssk->sk_socket)
896 mptcp_sock_graft(ssk, sk->sk_socket);
897
898 mptcp_subflow_ctx(ssk)->subflow_id = msk->subflow_id++;
899 mptcp_sockopt_sync_locked(msk, ssk);
900 mptcp_subflow_joined(msk, ssk);
901 mptcp_stop_tout_timer(sk);
902 __mptcp_propagate_sndbuf(sk, ssk);
903 return true;
904 }
905
906 static void __mptcp_flush_join_list(struct sock *sk, struct list_head *join_list)
907 {
908 struct mptcp_subflow_context *tmp, *subflow;
909 struct mptcp_sock *msk = mptcp_sk(sk);
910
911 list_for_each_entry_safe(subflow, tmp, join_list, node) {
912 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
913 bool slow = lock_sock_fast(ssk);
914
915 list_move_tail(&subflow->node, &msk->conn_list);
916 if (!__mptcp_finish_join(msk, ssk))
917 mptcp_subflow_reset(ssk);
918 unlock_sock_fast(ssk, slow);
919 }
920 }
921
922 static bool mptcp_rtx_timer_pending(struct sock *sk)
923 {
924 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer);
925 }
926
927 static void mptcp_reset_rtx_timer(struct sock *sk)
928 {
929 struct inet_connection_sock *icsk = inet_csk(sk);
930 unsigned long tout;
931
932 /* prevent rescheduling on close */
933 if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE))
934 return;
935
936 tout = mptcp_sk(sk)->timer_ival;
937 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout);
938 }
939
940 bool mptcp_schedule_work(struct sock *sk)
941 {
942 if (inet_sk_state_load(sk) != TCP_CLOSE &&
943 schedule_work(&mptcp_sk(sk)->work)) {
944 /* each subflow already holds a reference to the sk, and the
945 * workqueue is invoked by a subflow, so sk can't go away here.
946 */
947 sock_hold(sk);
948 return true;
949 }
950 return false;
951 }
952
953 static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk)
954 {
955 struct mptcp_subflow_context *subflow;
956
957 msk_owned_by_me(msk);
958
959 mptcp_for_each_subflow(msk, subflow) {
960 if (READ_ONCE(subflow->data_avail))
961 return mptcp_subflow_tcp_sock(subflow);
962 }
963
964 return NULL;
965 }
966
967 static bool mptcp_skb_can_collapse_to(u64 write_seq,
968 const struct sk_buff *skb,
969 const struct mptcp_ext *mpext)
970 {
971 if (!tcp_skb_can_collapse_to(skb))
972 return false;
973
974 /* can collapse only if MPTCP level sequence is in order and this
975 * mapping has not been xmitted yet
976 */
977 return mpext && mpext->data_seq + mpext->data_len == write_seq &&
978 !mpext->frozen;
979 }
980
981 /* we can append data to the given data frag if:
982 * - there is space available in the backing page_frag
983 * - the data frag tail matches the current page_frag free offset
984 * - the data frag end sequence number matches the current write seq
985 */
986 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk,
987 const struct page_frag *pfrag,
988 const struct mptcp_data_frag *df)
989 {
990 return df && pfrag->page == df->page &&
991 pfrag->size - pfrag->offset > 0 &&
992 pfrag->offset == (df->offset + df->data_len) &&
993 df->data_seq + df->data_len == msk->write_seq;
994 }
995
996 static void dfrag_uncharge(struct sock *sk, int len)
997 {
998 sk_mem_uncharge(sk, len);
999 sk_wmem_queued_add(sk, -len);
1000 }
1001
1002 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag)
1003 {
1004 int len = dfrag->data_len + dfrag->overhead;
1005
1006 list_del(&dfrag->list);
1007 dfrag_uncharge(sk, len);
1008 put_page(dfrag->page);
1009 }
1010
1011 /* called under both the msk socket lock and the data lock */
1012 static void __mptcp_clean_una(struct sock *sk)
1013 {
1014 struct mptcp_sock *msk = mptcp_sk(sk);
1015 struct mptcp_data_frag *dtmp, *dfrag;
1016 u64 snd_una;
1017
1018 snd_una = msk->snd_una;
1019 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
1020 if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
1021 break;
1022
1023 if (unlikely(dfrag == msk->first_pending)) {
1024 /* in recovery mode can see ack after the current snd head */
1025 if (WARN_ON_ONCE(!msk->recovery))
1026 break;
1027
1028 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1029 }
1030
1031 dfrag_clear(sk, dfrag);
1032 }
1033
1034 dfrag = mptcp_rtx_head(sk);
1035 if (dfrag && after64(snd_una, dfrag->data_seq)) {
1036 u64 delta = snd_una - dfrag->data_seq;
1037
1038 /* prevent wrap around in recovery mode */
1039 if (unlikely(delta > dfrag->already_sent)) {
1040 if (WARN_ON_ONCE(!msk->recovery))
1041 goto out;
1042 if (WARN_ON_ONCE(delta > dfrag->data_len))
1043 goto out;
1044 dfrag->already_sent += delta - dfrag->already_sent;
1045 }
1046
1047 dfrag->data_seq += delta;
1048 dfrag->offset += delta;
1049 dfrag->data_len -= delta;
1050 dfrag->already_sent -= delta;
1051
1052 dfrag_uncharge(sk, delta);
1053 }
1054
1055 /* all retransmitted data acked, recovery completed */
1056 if (unlikely(msk->recovery) && after64(msk->snd_una, msk->recovery_snd_nxt))
1057 msk->recovery = false;
1058
1059 out:
1060 if (snd_una == msk->snd_nxt && snd_una == msk->write_seq) {
1061 if (mptcp_rtx_timer_pending(sk) && !mptcp_data_fin_enabled(msk))
1062 mptcp_stop_rtx_timer(sk);
1063 } else {
1064 mptcp_reset_rtx_timer(sk);
1065 }
1066
1067 if (mptcp_pending_data_fin_ack(sk))
1068 mptcp_schedule_work(sk);
1069 }
1070
1071 static void __mptcp_clean_una_wakeup(struct sock *sk)
1072 {
1073 lockdep_assert_held_once(&sk->sk_lock.slock);
1074
1075 __mptcp_clean_una(sk);
1076 mptcp_write_space(sk);
1077 }
1078
1079 static void mptcp_clean_una_wakeup(struct sock *sk)
1080 {
1081 mptcp_data_lock(sk);
1082 __mptcp_clean_una_wakeup(sk);
1083 mptcp_data_unlock(sk);
1084 }
1085
1086 static void mptcp_enter_memory_pressure(struct sock *sk)
1087 {
1088 struct mptcp_subflow_context *subflow;
1089 struct mptcp_sock *msk = mptcp_sk(sk);
1090 bool first = true;
1091
1092 mptcp_for_each_subflow(msk, subflow) {
1093 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1094
1095 if (first)
1096 tcp_enter_memory_pressure(ssk);
1097 sk_stream_moderate_sndbuf(ssk);
1098
1099 first = false;
1100 }
1101 __mptcp_sync_sndbuf(sk);
1102 }
1103
1104 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of
1105 * data
1106 */
1107 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
1108 {
1109 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
1110 pfrag, sk->sk_allocation)))
1111 return true;
1112
1113 mptcp_enter_memory_pressure(sk);
1114 return false;
1115 }
1116
1117 static struct mptcp_data_frag *
1118 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag,
1119 int orig_offset)
1120 {
1121 int offset = ALIGN(orig_offset, sizeof(long));
1122 struct mptcp_data_frag *dfrag;
1123
1124 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
1125 dfrag->data_len = 0;
1126 dfrag->data_seq = msk->write_seq;
1127 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
1128 dfrag->offset = offset + sizeof(struct mptcp_data_frag);
1129 dfrag->already_sent = 0;
1130 dfrag->page = pfrag->page;
1131
1132 return dfrag;
1133 }
1134
1135 struct mptcp_sendmsg_info {
1136 int mss_now;
1137 int size_goal;
1138 u16 limit;
1139 u16 sent;
1140 unsigned int flags;
1141 bool data_lock_held;
1142 };
1143
1144 static int mptcp_check_allowed_size(const struct mptcp_sock *msk, struct sock *ssk,
1145 u64 data_seq, int avail_size)
1146 {
1147 u64 window_end = mptcp_wnd_end(msk);
1148 u64 mptcp_snd_wnd;
1149
1150 if (__mptcp_check_fallback(msk))
1151 return avail_size;
1152
1153 mptcp_snd_wnd = window_end - data_seq;
1154 avail_size = min_t(unsigned int, mptcp_snd_wnd, avail_size);
1155
1156 if (unlikely(tcp_sk(ssk)->snd_wnd < mptcp_snd_wnd)) {
1157 tcp_sk(ssk)->snd_wnd = min_t(u64, U32_MAX, mptcp_snd_wnd);
1158 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_SNDWNDSHARED);
1159 }
1160
1161 return avail_size;
1162 }
1163
1164 static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp)
1165 {
1166 struct skb_ext *mpext = __skb_ext_alloc(gfp);
1167
1168 if (!mpext)
1169 return false;
1170 __skb_ext_set(skb, SKB_EXT_MPTCP, mpext);
1171 return true;
1172 }
1173
1174 static struct sk_buff *__mptcp_do_alloc_tx_skb(struct sock *sk, gfp_t gfp)
1175 {
1176 struct sk_buff *skb;
1177
1178 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp);
1179 if (likely(skb)) {
1180 if (likely(__mptcp_add_ext(skb, gfp))) {
1181 skb_reserve(skb, MAX_TCP_HEADER);
1182 skb->ip_summed = CHECKSUM_PARTIAL;
1183 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
1184 return skb;
1185 }
1186 __kfree_skb(skb);
1187 } else {
1188 mptcp_enter_memory_pressure(sk);
1189 }
1190 return NULL;
1191 }
1192
1193 static struct sk_buff *__mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, gfp_t gfp)
1194 {
1195 struct sk_buff *skb;
1196
1197 skb = __mptcp_do_alloc_tx_skb(sk, gfp);
1198 if (!skb)
1199 return NULL;
1200
1201 if (likely(sk_wmem_schedule(ssk, skb->truesize))) {
1202 tcp_skb_entail(ssk, skb);
1203 return skb;
1204 }
1205 tcp_skb_tsorted_anchor_cleanup(skb);
1206 kfree_skb(skb);
1207 return NULL;
1208 }
1209
1210 static struct sk_buff *mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, bool data_lock_held)
1211 {
1212 gfp_t gfp = data_lock_held ? GFP_ATOMIC : sk->sk_allocation;
1213
1214 return __mptcp_alloc_tx_skb(sk, ssk, gfp);
1215 }
1216
1217 /* note: this always recompute the csum on the whole skb, even
1218 * if we just appended a single frag. More status info needed
1219 */
1220 static void mptcp_update_data_checksum(struct sk_buff *skb, int added)
1221 {
1222 struct mptcp_ext *mpext = mptcp_get_ext(skb);
1223 __wsum csum = ~csum_unfold(mpext->csum);
1224 int offset = skb->len - added;
1225
1226 mpext->csum = csum_fold(csum_block_add(csum, skb_checksum(skb, offset, added, 0), offset));
1227 }
1228
1229 static void mptcp_update_infinite_map(struct mptcp_sock *msk,
1230 struct sock *ssk,
1231 struct mptcp_ext *mpext)
1232 {
1233 if (!mpext)
1234 return;
1235
1236 mpext->infinite_map = 1;
1237 mpext->data_len = 0;
1238
1239 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_INFINITEMAPTX);
1240 mptcp_subflow_ctx(ssk)->send_infinite_map = 0;
1241 pr_fallback(msk);
1242 mptcp_do_fallback(ssk);
1243 }
1244
1245 #define MPTCP_MAX_GSO_SIZE (GSO_LEGACY_MAX_SIZE - (MAX_TCP_HEADER + 1))
1246
1247 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
1248 struct mptcp_data_frag *dfrag,
1249 struct mptcp_sendmsg_info *info)
1250 {
1251 u64 data_seq = dfrag->data_seq + info->sent;
1252 int offset = dfrag->offset + info->sent;
1253 struct mptcp_sock *msk = mptcp_sk(sk);
1254 bool zero_window_probe = false;
1255 struct mptcp_ext *mpext = NULL;
1256 bool can_coalesce = false;
1257 bool reuse_skb = true;
1258 struct sk_buff *skb;
1259 size_t copy;
1260 int i;
1261
1262 pr_debug("msk=%p ssk=%p sending dfrag at seq=%llu len=%u already sent=%u\n",
1263 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent);
1264
1265 if (WARN_ON_ONCE(info->sent > info->limit ||
1266 info->limit > dfrag->data_len))
1267 return 0;
1268
1269 if (unlikely(!__tcp_can_send(ssk)))
1270 return -EAGAIN;
1271
1272 /* compute send limit */
1273 if (unlikely(ssk->sk_gso_max_size > MPTCP_MAX_GSO_SIZE))
1274 ssk->sk_gso_max_size = MPTCP_MAX_GSO_SIZE;
1275 info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags);
1276 copy = info->size_goal;
1277
1278 skb = tcp_write_queue_tail(ssk);
1279 if (skb && copy > skb->len) {
1280 /* Limit the write to the size available in the
1281 * current skb, if any, so that we create at most a new skb.
1282 * Explicitly tells TCP internals to avoid collapsing on later
1283 * queue management operation, to avoid breaking the ext <->
1284 * SSN association set here
1285 */
1286 mpext = mptcp_get_ext(skb);
1287 if (!mptcp_skb_can_collapse_to(data_seq, skb, mpext)) {
1288 TCP_SKB_CB(skb)->eor = 1;
1289 tcp_mark_push(tcp_sk(ssk), skb);
1290 goto alloc_skb;
1291 }
1292
1293 i = skb_shinfo(skb)->nr_frags;
1294 can_coalesce = skb_can_coalesce(skb, i, dfrag->page, offset);
1295 if (!can_coalesce && i >= READ_ONCE(net_hotdata.sysctl_max_skb_frags)) {
1296 tcp_mark_push(tcp_sk(ssk), skb);
1297 goto alloc_skb;
1298 }
1299
1300 copy -= skb->len;
1301 } else {
1302 alloc_skb:
1303 skb = mptcp_alloc_tx_skb(sk, ssk, info->data_lock_held);
1304 if (!skb)
1305 return -ENOMEM;
1306
1307 i = skb_shinfo(skb)->nr_frags;
1308 reuse_skb = false;
1309 mpext = mptcp_get_ext(skb);
1310 }
1311
1312 /* Zero window and all data acked? Probe. */
1313 copy = mptcp_check_allowed_size(msk, ssk, data_seq, copy);
1314 if (copy == 0) {
1315 u64 snd_una = READ_ONCE(msk->snd_una);
1316
1317 if (snd_una != msk->snd_nxt || tcp_write_queue_tail(ssk)) {
1318 tcp_remove_empty_skb(ssk);
1319 return 0;
1320 }
1321
1322 zero_window_probe = true;
1323 data_seq = snd_una - 1;
1324 copy = 1;
1325 }
1326
1327 copy = min_t(size_t, copy, info->limit - info->sent);
1328 if (!sk_wmem_schedule(ssk, copy)) {
1329 tcp_remove_empty_skb(ssk);
1330 return -ENOMEM;
1331 }
1332
1333 if (can_coalesce) {
1334 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1335 } else {
1336 get_page(dfrag->page);
1337 skb_fill_page_desc(skb, i, dfrag->page, offset, copy);
1338 }
1339
1340 skb->len += copy;
1341 skb->data_len += copy;
1342 skb->truesize += copy;
1343 sk_wmem_queued_add(ssk, copy);
1344 sk_mem_charge(ssk, copy);
1345 WRITE_ONCE(tcp_sk(ssk)->write_seq, tcp_sk(ssk)->write_seq + copy);
1346 TCP_SKB_CB(skb)->end_seq += copy;
1347 tcp_skb_pcount_set(skb, 0);
1348
1349 /* on skb reuse we just need to update the DSS len */
1350 if (reuse_skb) {
1351 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1352 mpext->data_len += copy;
1353 goto out;
1354 }
1355
1356 memset(mpext, 0, sizeof(*mpext));
1357 mpext->data_seq = data_seq;
1358 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
1359 mpext->data_len = copy;
1360 mpext->use_map = 1;
1361 mpext->dsn64 = 1;
1362
1363 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d\n",
1364 mpext->data_seq, mpext->subflow_seq, mpext->data_len,
1365 mpext->dsn64);
1366
1367 if (zero_window_probe) {
1368 mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1369 mpext->frozen = 1;
1370 if (READ_ONCE(msk->csum_enabled))
1371 mptcp_update_data_checksum(skb, copy);
1372 tcp_push_pending_frames(ssk);
1373 return 0;
1374 }
1375 out:
1376 if (READ_ONCE(msk->csum_enabled))
1377 mptcp_update_data_checksum(skb, copy);
1378 if (mptcp_subflow_ctx(ssk)->send_infinite_map)
1379 mptcp_update_infinite_map(msk, ssk, mpext);
1380 trace_mptcp_sendmsg_frag(mpext);
1381 mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1382 return copy;
1383 }
1384
1385 #define MPTCP_SEND_BURST_SIZE ((1 << 16) - \
1386 sizeof(struct tcphdr) - \
1387 MAX_TCP_OPTION_SPACE - \
1388 sizeof(struct ipv6hdr) - \
1389 sizeof(struct frag_hdr))
1390
1391 struct subflow_send_info {
1392 struct sock *ssk;
1393 u64 linger_time;
1394 };
1395
1396 void mptcp_subflow_set_active(struct mptcp_subflow_context *subflow)
1397 {
1398 if (!subflow->stale)
1399 return;
1400
1401 subflow->stale = 0;
1402 MPTCP_INC_STATS(sock_net(mptcp_subflow_tcp_sock(subflow)), MPTCP_MIB_SUBFLOWRECOVER);
1403 }
1404
1405 bool mptcp_subflow_active(struct mptcp_subflow_context *subflow)
1406 {
1407 if (unlikely(subflow->stale)) {
1408 u32 rcv_tstamp = READ_ONCE(tcp_sk(mptcp_subflow_tcp_sock(subflow))->rcv_tstamp);
1409
1410 if (subflow->stale_rcv_tstamp == rcv_tstamp)
1411 return false;
1412
1413 mptcp_subflow_set_active(subflow);
1414 }
1415 return __mptcp_subflow_active(subflow);
1416 }
1417
1418 #define SSK_MODE_ACTIVE 0
1419 #define SSK_MODE_BACKUP 1
1420 #define SSK_MODE_MAX 2
1421
1422 /* implement the mptcp packet scheduler;
1423 * returns the subflow that will transmit the next DSS
1424 * additionally updates the rtx timeout
1425 */
1426 struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk)
1427 {
1428 struct subflow_send_info send_info[SSK_MODE_MAX];
1429 struct mptcp_subflow_context *subflow;
1430 struct sock *sk = (struct sock *)msk;
1431 u32 pace, burst, wmem;
1432 int i, nr_active = 0;
1433 struct sock *ssk;
1434 u64 linger_time;
1435 long tout = 0;
1436
1437 /* pick the subflow with the lower wmem/wspace ratio */
1438 for (i = 0; i < SSK_MODE_MAX; ++i) {
1439 send_info[i].ssk = NULL;
1440 send_info[i].linger_time = -1;
1441 }
1442
1443 mptcp_for_each_subflow(msk, subflow) {
1444 bool backup = subflow->backup || subflow->request_bkup;
1445
1446 trace_mptcp_subflow_get_send(subflow);
1447 ssk = mptcp_subflow_tcp_sock(subflow);
1448 if (!mptcp_subflow_active(subflow))
1449 continue;
1450
1451 tout = max(tout, mptcp_timeout_from_subflow(subflow));
1452 nr_active += !backup;
1453 pace = subflow->avg_pacing_rate;
1454 if (unlikely(!pace)) {
1455 /* init pacing rate from socket */
1456 subflow->avg_pacing_rate = READ_ONCE(ssk->sk_pacing_rate);
1457 pace = subflow->avg_pacing_rate;
1458 if (!pace)
1459 continue;
1460 }
1461
1462 linger_time = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32, pace);
1463 if (linger_time < send_info[backup].linger_time) {
1464 send_info[backup].ssk = ssk;
1465 send_info[backup].linger_time = linger_time;
1466 }
1467 }
1468 __mptcp_set_timeout(sk, tout);
1469
1470 /* pick the best backup if no other subflow is active */
1471 if (!nr_active)
1472 send_info[SSK_MODE_ACTIVE].ssk = send_info[SSK_MODE_BACKUP].ssk;
1473
1474 /* According to the blest algorithm, to avoid HoL blocking for the
1475 * faster flow, we need to:
1476 * - estimate the faster flow linger time
1477 * - use the above to estimate the amount of byte transferred
1478 * by the faster flow
1479 * - check that the amount of queued data is greter than the above,
1480 * otherwise do not use the picked, slower, subflow
1481 * We select the subflow with the shorter estimated time to flush
1482 * the queued mem, which basically ensure the above. We just need
1483 * to check that subflow has a non empty cwin.
1484 */
1485 ssk = send_info[SSK_MODE_ACTIVE].ssk;
1486 if (!ssk || !sk_stream_memory_free(ssk))
1487 return NULL;
1488
1489 burst = min_t(int, MPTCP_SEND_BURST_SIZE, mptcp_wnd_end(msk) - msk->snd_nxt);
1490 wmem = READ_ONCE(ssk->sk_wmem_queued);
1491 if (!burst)
1492 return ssk;
1493
1494 subflow = mptcp_subflow_ctx(ssk);
1495 subflow->avg_pacing_rate = div_u64((u64)subflow->avg_pacing_rate * wmem +
1496 READ_ONCE(ssk->sk_pacing_rate) * burst,
1497 burst + wmem);
1498 msk->snd_burst = burst;
1499 return ssk;
1500 }
1501
1502 static void mptcp_push_release(struct sock *ssk, struct mptcp_sendmsg_info *info)
1503 {
1504 tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal);
1505 release_sock(ssk);
1506 }
1507
1508 static void mptcp_update_post_push(struct mptcp_sock *msk,
1509 struct mptcp_data_frag *dfrag,
1510 u32 sent)
1511 {
1512 u64 snd_nxt_new = dfrag->data_seq;
1513
1514 dfrag->already_sent += sent;
1515
1516 msk->snd_burst -= sent;
1517
1518 snd_nxt_new += dfrag->already_sent;
1519
1520 /* snd_nxt_new can be smaller than snd_nxt in case mptcp
1521 * is recovering after a failover. In that event, this re-sends
1522 * old segments.
1523 *
1524 * Thus compute snd_nxt_new candidate based on
1525 * the dfrag->data_seq that was sent and the data
1526 * that has been handed to the subflow for transmission
1527 * and skip update in case it was old dfrag.
1528 */
1529 if (likely(after64(snd_nxt_new, msk->snd_nxt))) {
1530 msk->bytes_sent += snd_nxt_new - msk->snd_nxt;
1531 WRITE_ONCE(msk->snd_nxt, snd_nxt_new);
1532 }
1533 }
1534
1535 void mptcp_check_and_set_pending(struct sock *sk)
1536 {
1537 if (mptcp_send_head(sk)) {
1538 mptcp_data_lock(sk);
1539 mptcp_sk(sk)->cb_flags |= BIT(MPTCP_PUSH_PENDING);
1540 mptcp_data_unlock(sk);
1541 }
1542 }
1543
1544 static int __subflow_push_pending(struct sock *sk, struct sock *ssk,
1545 struct mptcp_sendmsg_info *info)
1546 {
1547 struct mptcp_sock *msk = mptcp_sk(sk);
1548 struct mptcp_data_frag *dfrag;
1549 int len, copied = 0, err = 0;
1550
1551 while ((dfrag = mptcp_send_head(sk))) {
1552 info->sent = dfrag->already_sent;
1553 info->limit = dfrag->data_len;
1554 len = dfrag->data_len - dfrag->already_sent;
1555 while (len > 0) {
1556 int ret = 0;
1557
1558 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, info);
1559 if (ret <= 0) {
1560 err = copied ? : ret;
1561 goto out;
1562 }
1563
1564 info->sent += ret;
1565 copied += ret;
1566 len -= ret;
1567
1568 mptcp_update_post_push(msk, dfrag, ret);
1569 }
1570 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1571
1572 if (msk->snd_burst <= 0 ||
1573 !sk_stream_memory_free(ssk) ||
1574 !mptcp_subflow_active(mptcp_subflow_ctx(ssk))) {
1575 err = copied;
1576 goto out;
1577 }
1578 mptcp_set_timeout(sk);
1579 }
1580 err = copied;
1581
1582 out:
1583 if (err > 0)
1584 msk->last_data_sent = tcp_jiffies32;
1585 return err;
1586 }
1587
1588 void __mptcp_push_pending(struct sock *sk, unsigned int flags)
1589 {
1590 struct sock *prev_ssk = NULL, *ssk = NULL;
1591 struct mptcp_sock *msk = mptcp_sk(sk);
1592 struct mptcp_sendmsg_info info = {
1593 .flags = flags,
1594 };
1595 bool do_check_data_fin = false;
1596 int push_count = 1;
1597
1598 while (mptcp_send_head(sk) && (push_count > 0)) {
1599 struct mptcp_subflow_context *subflow;
1600 int ret = 0;
1601
1602 if (mptcp_sched_get_send(msk))
1603 break;
1604
1605 push_count = 0;
1606
1607 mptcp_for_each_subflow(msk, subflow) {
1608 if (READ_ONCE(subflow->scheduled)) {
1609 mptcp_subflow_set_scheduled(subflow, false);
1610
1611 prev_ssk = ssk;
1612 ssk = mptcp_subflow_tcp_sock(subflow);
1613 if (ssk != prev_ssk) {
1614 /* First check. If the ssk has changed since
1615 * the last round, release prev_ssk
1616 */
1617 if (prev_ssk)
1618 mptcp_push_release(prev_ssk, &info);
1619
1620 /* Need to lock the new subflow only if different
1621 * from the previous one, otherwise we are still
1622 * helding the relevant lock
1623 */
1624 lock_sock(ssk);
1625 }
1626
1627 push_count++;
1628
1629 ret = __subflow_push_pending(sk, ssk, &info);
1630 if (ret <= 0) {
1631 if (ret != -EAGAIN ||
1632 (1 << ssk->sk_state) &
1633 (TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2 | TCPF_CLOSE))
1634 push_count--;
1635 continue;
1636 }
1637 do_check_data_fin = true;
1638 }
1639 }
1640 }
1641
1642 /* at this point we held the socket lock for the last subflow we used */
1643 if (ssk)
1644 mptcp_push_release(ssk, &info);
1645
1646 /* ensure the rtx timer is running */
1647 if (!mptcp_rtx_timer_pending(sk))
1648 mptcp_reset_rtx_timer(sk);
1649 if (do_check_data_fin)
1650 mptcp_check_send_data_fin(sk);
1651 }
1652
1653 static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk, bool first)
1654 {
1655 struct mptcp_sock *msk = mptcp_sk(sk);
1656 struct mptcp_sendmsg_info info = {
1657 .data_lock_held = true,
1658 };
1659 bool keep_pushing = true;
1660 struct sock *xmit_ssk;
1661 int copied = 0;
1662
1663 info.flags = 0;
1664 while (mptcp_send_head(sk) && keep_pushing) {
1665 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
1666 int ret = 0;
1667
1668 /* check for a different subflow usage only after
1669 * spooling the first chunk of data
1670 */
1671 if (first) {
1672 mptcp_subflow_set_scheduled(subflow, false);
1673 ret = __subflow_push_pending(sk, ssk, &info);
1674 first = false;
1675 if (ret <= 0)
1676 break;
1677 copied += ret;
1678 continue;
1679 }
1680
1681 if (mptcp_sched_get_send(msk))
1682 goto out;
1683
1684 if (READ_ONCE(subflow->scheduled)) {
1685 mptcp_subflow_set_scheduled(subflow, false);
1686 ret = __subflow_push_pending(sk, ssk, &info);
1687 if (ret <= 0)
1688 keep_pushing = false;
1689 copied += ret;
1690 }
1691
1692 mptcp_for_each_subflow(msk, subflow) {
1693 if (READ_ONCE(subflow->scheduled)) {
1694 xmit_ssk = mptcp_subflow_tcp_sock(subflow);
1695 if (xmit_ssk != ssk) {
1696 mptcp_subflow_delegate(subflow,
1697 MPTCP_DELEGATE_SEND);
1698 keep_pushing = false;
1699 }
1700 }
1701 }
1702 }
1703
1704 out:
1705 /* __mptcp_alloc_tx_skb could have released some wmem and we are
1706 * not going to flush it via release_sock()
1707 */
1708 if (copied) {
1709 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
1710 info.size_goal);
1711 if (!mptcp_rtx_timer_pending(sk))
1712 mptcp_reset_rtx_timer(sk);
1713
1714 if (msk->snd_data_fin_enable &&
1715 msk->snd_nxt + 1 == msk->write_seq)
1716 mptcp_schedule_work(sk);
1717 }
1718 }
1719
1720 static int mptcp_disconnect(struct sock *sk, int flags);
1721
1722 static int mptcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1723 size_t len, int *copied_syn)
1724 {
1725 unsigned int saved_flags = msg->msg_flags;
1726 struct mptcp_sock *msk = mptcp_sk(sk);
1727 struct sock *ssk;
1728 int ret;
1729
1730 /* on flags based fastopen the mptcp is supposed to create the
1731 * first subflow right now. Otherwise we are in the defer_connect
1732 * path, and the first subflow must be already present.
1733 * Since the defer_connect flag is cleared after the first succsful
1734 * fastopen attempt, no need to check for additional subflow status.
1735 */
1736 if (msg->msg_flags & MSG_FASTOPEN) {
1737 ssk = __mptcp_nmpc_sk(msk);
1738 if (IS_ERR(ssk))
1739 return PTR_ERR(ssk);
1740 }
1741 if (!msk->first)
1742 return -EINVAL;
1743
1744 ssk = msk->first;
1745
1746 lock_sock(ssk);
1747 msg->msg_flags |= MSG_DONTWAIT;
1748 msk->fastopening = 1;
1749 ret = tcp_sendmsg_fastopen(ssk, msg, copied_syn, len, NULL);
1750 msk->fastopening = 0;
1751 msg->msg_flags = saved_flags;
1752 release_sock(ssk);
1753
1754 /* do the blocking bits of inet_stream_connect outside the ssk socket lock */
1755 if (ret == -EINPROGRESS && !(msg->msg_flags & MSG_DONTWAIT)) {
1756 ret = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1757 msg->msg_namelen, msg->msg_flags, 1);
1758
1759 /* Keep the same behaviour of plain TCP: zero the copied bytes in
1760 * case of any error, except timeout or signal
1761 */
1762 if (ret && ret != -EINPROGRESS && ret != -ERESTARTSYS && ret != -EINTR)
1763 *copied_syn = 0;
1764 } else if (ret && ret != -EINPROGRESS) {
1765 /* The disconnect() op called by tcp_sendmsg_fastopen()/
1766 * __inet_stream_connect() can fail, due to looking check,
1767 * see mptcp_disconnect().
1768 * Attempt it again outside the problematic scope.
1769 */
1770 if (!mptcp_disconnect(sk, 0))
1771 sk->sk_socket->state = SS_UNCONNECTED;
1772 }
1773 inet_clear_bit(DEFER_CONNECT, sk);
1774
1775 return ret;
1776 }
1777
1778 static int do_copy_data_nocache(struct sock *sk, int copy,
1779 struct iov_iter *from, char *to)
1780 {
1781 if (sk->sk_route_caps & NETIF_F_NOCACHE_COPY) {
1782 if (!copy_from_iter_full_nocache(to, copy, from))
1783 return -EFAULT;
1784 } else if (!copy_from_iter_full(to, copy, from)) {
1785 return -EFAULT;
1786 }
1787 return 0;
1788 }
1789
1790 /* open-code sk_stream_memory_free() plus sent limit computation to
1791 * avoid indirect calls in fast-path.
1792 * Called under the msk socket lock, so we can avoid a bunch of ONCE
1793 * annotations.
1794 */
1795 static u32 mptcp_send_limit(const struct sock *sk)
1796 {
1797 const struct mptcp_sock *msk = mptcp_sk(sk);
1798 u32 limit, not_sent;
1799
1800 if (sk->sk_wmem_queued >= READ_ONCE(sk->sk_sndbuf))
1801 return 0;
1802
1803 limit = mptcp_notsent_lowat(sk);
1804 if (limit == UINT_MAX)
1805 return UINT_MAX;
1806
1807 not_sent = msk->write_seq - msk->snd_nxt;
1808 if (not_sent >= limit)
1809 return 0;
1810
1811 return limit - not_sent;
1812 }
1813
1814 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
1815 {
1816 struct mptcp_sock *msk = mptcp_sk(sk);
1817 struct page_frag *pfrag;
1818 size_t copied = 0;
1819 int ret = 0;
1820 long timeo;
1821
1822 /* silently ignore everything else */
1823 msg->msg_flags &= MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL | MSG_FASTOPEN;
1824
1825 lock_sock(sk);
1826
1827 if (unlikely(inet_test_bit(DEFER_CONNECT, sk) ||
1828 msg->msg_flags & MSG_FASTOPEN)) {
1829 int copied_syn = 0;
1830
1831 ret = mptcp_sendmsg_fastopen(sk, msg, len, &copied_syn);
1832 copied += copied_syn;
1833 if (ret == -EINPROGRESS && copied_syn > 0)
1834 goto out;
1835 else if (ret)
1836 goto do_error;
1837 }
1838
1839 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1840
1841 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1842 ret = sk_stream_wait_connect(sk, &timeo);
1843 if (ret)
1844 goto do_error;
1845 }
1846
1847 ret = -EPIPE;
1848 if (unlikely(sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)))
1849 goto do_error;
1850
1851 pfrag = sk_page_frag(sk);
1852
1853 while (msg_data_left(msg)) {
1854 int total_ts, frag_truesize = 0;
1855 struct mptcp_data_frag *dfrag;
1856 bool dfrag_collapsed;
1857 size_t psize, offset;
1858 u32 copy_limit;
1859
1860 /* ensure fitting the notsent_lowat() constraint */
1861 copy_limit = mptcp_send_limit(sk);
1862 if (!copy_limit)
1863 goto wait_for_memory;
1864
1865 /* reuse tail pfrag, if possible, or carve a new one from the
1866 * page allocator
1867 */
1868 dfrag = mptcp_pending_tail(sk);
1869 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
1870 if (!dfrag_collapsed) {
1871 if (!mptcp_page_frag_refill(sk, pfrag))
1872 goto wait_for_memory;
1873
1874 dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset);
1875 frag_truesize = dfrag->overhead;
1876 }
1877
1878 /* we do not bound vs wspace, to allow a single packet.
1879 * memory accounting will prevent execessive memory usage
1880 * anyway
1881 */
1882 offset = dfrag->offset + dfrag->data_len;
1883 psize = pfrag->size - offset;
1884 psize = min_t(size_t, psize, msg_data_left(msg));
1885 psize = min_t(size_t, psize, copy_limit);
1886 total_ts = psize + frag_truesize;
1887
1888 if (!sk_wmem_schedule(sk, total_ts))
1889 goto wait_for_memory;
1890
1891 ret = do_copy_data_nocache(sk, psize, &msg->msg_iter,
1892 page_address(dfrag->page) + offset);
1893 if (ret)
1894 goto do_error;
1895
1896 /* data successfully copied into the write queue */
1897 sk_forward_alloc_add(sk, -total_ts);
1898 copied += psize;
1899 dfrag->data_len += psize;
1900 frag_truesize += psize;
1901 pfrag->offset += frag_truesize;
1902 WRITE_ONCE(msk->write_seq, msk->write_seq + psize);
1903
1904 /* charge data on mptcp pending queue to the msk socket
1905 * Note: we charge such data both to sk and ssk
1906 */
1907 sk_wmem_queued_add(sk, frag_truesize);
1908 if (!dfrag_collapsed) {
1909 get_page(dfrag->page);
1910 list_add_tail(&dfrag->list, &msk->rtx_queue);
1911 if (!msk->first_pending)
1912 WRITE_ONCE(msk->first_pending, dfrag);
1913 }
1914 pr_debug("msk=%p dfrag at seq=%llu len=%u sent=%u new=%d\n", msk,
1915 dfrag->data_seq, dfrag->data_len, dfrag->already_sent,
1916 !dfrag_collapsed);
1917
1918 continue;
1919
1920 wait_for_memory:
1921 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1922 __mptcp_push_pending(sk, msg->msg_flags);
1923 ret = sk_stream_wait_memory(sk, &timeo);
1924 if (ret)
1925 goto do_error;
1926 }
1927
1928 if (copied)
1929 __mptcp_push_pending(sk, msg->msg_flags);
1930
1931 out:
1932 release_sock(sk);
1933 return copied;
1934
1935 do_error:
1936 if (copied)
1937 goto out;
1938
1939 copied = sk_stream_error(sk, msg->msg_flags, ret);
1940 goto out;
1941 }
1942
1943 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied);
1944
1945 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk,
1946 struct msghdr *msg,
1947 size_t len, int flags,
1948 struct scm_timestamping_internal *tss,
1949 int *cmsg_flags)
1950 {
1951 struct sk_buff *skb, *tmp;
1952 int copied = 0;
1953
1954 skb_queue_walk_safe(&msk->receive_queue, skb, tmp) {
1955 u32 offset = MPTCP_SKB_CB(skb)->offset;
1956 u32 data_len = skb->len - offset;
1957 u32 count = min_t(size_t, len - copied, data_len);
1958 int err;
1959
1960 if (!(flags & MSG_TRUNC)) {
1961 err = skb_copy_datagram_msg(skb, offset, msg, count);
1962 if (unlikely(err < 0)) {
1963 if (!copied)
1964 return err;
1965 break;
1966 }
1967 }
1968
1969 if (MPTCP_SKB_CB(skb)->has_rxtstamp) {
1970 tcp_update_recv_tstamps(skb, tss);
1971 *cmsg_flags |= MPTCP_CMSG_TS;
1972 }
1973
1974 copied += count;
1975
1976 if (count < data_len) {
1977 if (!(flags & MSG_PEEK)) {
1978 MPTCP_SKB_CB(skb)->offset += count;
1979 MPTCP_SKB_CB(skb)->map_seq += count;
1980 msk->bytes_consumed += count;
1981 }
1982 break;
1983 }
1984
1985 if (!(flags & MSG_PEEK)) {
1986 /* we will bulk release the skb memory later */
1987 skb->destructor = NULL;
1988 WRITE_ONCE(msk->rmem_released, msk->rmem_released + skb->truesize);
1989 __skb_unlink(skb, &msk->receive_queue);
1990 __kfree_skb(skb);
1991 msk->bytes_consumed += count;
1992 }
1993
1994 if (copied >= len)
1995 break;
1996 }
1997
1998 mptcp_rcv_space_adjust(msk, copied);
1999 return copied;
2000 }
2001
2002 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information.
2003 *
2004 * Only difference: Use highest rtt estimate of the subflows in use.
2005 */
2006 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied)
2007 {
2008 struct mptcp_subflow_context *subflow;
2009 struct sock *sk = (struct sock *)msk;
2010 u8 scaling_ratio = U8_MAX;
2011 u32 time, advmss = 1;
2012 u64 rtt_us, mstamp;
2013
2014 msk_owned_by_me(msk);
2015
2016 if (copied <= 0)
2017 return;
2018
2019 if (!msk->rcvspace_init)
2020 mptcp_rcv_space_init(msk, msk->first);
2021
2022 msk->rcvq_space.copied += copied;
2023
2024 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC);
2025 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time);
2026
2027 rtt_us = msk->rcvq_space.rtt_us;
2028 if (rtt_us && time < (rtt_us >> 3))
2029 return;
2030
2031 rtt_us = 0;
2032 mptcp_for_each_subflow(msk, subflow) {
2033 const struct tcp_sock *tp;
2034 u64 sf_rtt_us;
2035 u32 sf_advmss;
2036
2037 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow));
2038
2039 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us);
2040 sf_advmss = READ_ONCE(tp->advmss);
2041
2042 rtt_us = max(sf_rtt_us, rtt_us);
2043 advmss = max(sf_advmss, advmss);
2044 scaling_ratio = min(tp->scaling_ratio, scaling_ratio);
2045 }
2046
2047 msk->rcvq_space.rtt_us = rtt_us;
2048 msk->scaling_ratio = scaling_ratio;
2049 if (time < (rtt_us >> 3) || rtt_us == 0)
2050 return;
2051
2052 if (msk->rcvq_space.copied <= msk->rcvq_space.space)
2053 goto new_measure;
2054
2055 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf) &&
2056 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
2057 u64 rcvwin, grow;
2058 int rcvbuf;
2059
2060 rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss;
2061
2062 grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space);
2063
2064 do_div(grow, msk->rcvq_space.space);
2065 rcvwin += (grow << 1);
2066
2067 rcvbuf = min_t(u64, mptcp_space_from_win(sk, rcvwin),
2068 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
2069
2070 if (rcvbuf > sk->sk_rcvbuf) {
2071 u32 window_clamp;
2072
2073 window_clamp = mptcp_win_from_space(sk, rcvbuf);
2074 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
2075
2076 /* Make subflows follow along. If we do not do this, we
2077 * get drops at subflow level if skbs can't be moved to
2078 * the mptcp rx queue fast enough (announced rcv_win can
2079 * exceed ssk->sk_rcvbuf).
2080 */
2081 mptcp_for_each_subflow(msk, subflow) {
2082 struct sock *ssk;
2083 bool slow;
2084
2085 ssk = mptcp_subflow_tcp_sock(subflow);
2086 slow = lock_sock_fast(ssk);
2087 WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf);
2088 WRITE_ONCE(tcp_sk(ssk)->window_clamp, window_clamp);
2089 if (tcp_can_send_ack(ssk))
2090 tcp_cleanup_rbuf(ssk, 1);
2091 unlock_sock_fast(ssk, slow);
2092 }
2093 }
2094 }
2095
2096 msk->rcvq_space.space = msk->rcvq_space.copied;
2097 new_measure:
2098 msk->rcvq_space.copied = 0;
2099 msk->rcvq_space.time = mstamp;
2100 }
2101
2102 static void __mptcp_update_rmem(struct sock *sk)
2103 {
2104 struct mptcp_sock *msk = mptcp_sk(sk);
2105
2106 if (!msk->rmem_released)
2107 return;
2108
2109 atomic_sub(msk->rmem_released, &sk->sk_rmem_alloc);
2110 mptcp_rmem_uncharge(sk, msk->rmem_released);
2111 WRITE_ONCE(msk->rmem_released, 0);
2112 }
2113
2114 static void __mptcp_splice_receive_queue(struct sock *sk)
2115 {
2116 struct mptcp_sock *msk = mptcp_sk(sk);
2117
2118 skb_queue_splice_tail_init(&sk->sk_receive_queue, &msk->receive_queue);
2119 }
2120
2121 static bool __mptcp_move_skbs(struct mptcp_sock *msk)
2122 {
2123 struct sock *sk = (struct sock *)msk;
2124 unsigned int moved = 0;
2125 bool ret, done;
2126
2127 do {
2128 struct sock *ssk = mptcp_subflow_recv_lookup(msk);
2129 bool slowpath;
2130
2131 /* we can have data pending in the subflows only if the msk
2132 * receive buffer was full at subflow_data_ready() time,
2133 * that is an unlikely slow path.
2134 */
2135 if (likely(!ssk))
2136 break;
2137
2138 slowpath = lock_sock_fast(ssk);
2139 mptcp_data_lock(sk);
2140 __mptcp_update_rmem(sk);
2141 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
2142 mptcp_data_unlock(sk);
2143
2144 if (unlikely(ssk->sk_err))
2145 __mptcp_error_report(sk);
2146 unlock_sock_fast(ssk, slowpath);
2147 } while (!done);
2148
2149 /* acquire the data lock only if some input data is pending */
2150 ret = moved > 0;
2151 if (!RB_EMPTY_ROOT(&msk->out_of_order_queue) ||
2152 !skb_queue_empty_lockless(&sk->sk_receive_queue)) {
2153 mptcp_data_lock(sk);
2154 __mptcp_update_rmem(sk);
2155 ret |= __mptcp_ofo_queue(msk);
2156 __mptcp_splice_receive_queue(sk);
2157 mptcp_data_unlock(sk);
2158 }
2159 if (ret)
2160 mptcp_check_data_fin((struct sock *)msk);
2161 return !skb_queue_empty(&msk->receive_queue);
2162 }
2163
2164 static unsigned int mptcp_inq_hint(const struct sock *sk)
2165 {
2166 const struct mptcp_sock *msk = mptcp_sk(sk);
2167 const struct sk_buff *skb;
2168
2169 skb = skb_peek(&msk->receive_queue);
2170 if (skb) {
2171 u64 hint_val = READ_ONCE(msk->ack_seq) - MPTCP_SKB_CB(skb)->map_seq;
2172
2173 if (hint_val >= INT_MAX)
2174 return INT_MAX;
2175
2176 return (unsigned int)hint_val;
2177 }
2178
2179 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
2180 return 1;
2181
2182 return 0;
2183 }
2184
2185 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
2186 int flags, int *addr_len)
2187 {
2188 struct mptcp_sock *msk = mptcp_sk(sk);
2189 struct scm_timestamping_internal tss;
2190 int copied = 0, cmsg_flags = 0;
2191 int target;
2192 long timeo;
2193
2194 /* MSG_ERRQUEUE is really a no-op till we support IP_RECVERR */
2195 if (unlikely(flags & MSG_ERRQUEUE))
2196 return inet_recv_error(sk, msg, len, addr_len);
2197
2198 lock_sock(sk);
2199 if (unlikely(sk->sk_state == TCP_LISTEN)) {
2200 copied = -ENOTCONN;
2201 goto out_err;
2202 }
2203
2204 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
2205
2206 len = min_t(size_t, len, INT_MAX);
2207 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2208
2209 if (unlikely(msk->recvmsg_inq))
2210 cmsg_flags = MPTCP_CMSG_INQ;
2211
2212 while (copied < len) {
2213 int err, bytes_read;
2214
2215 bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied, flags, &tss, &cmsg_flags);
2216 if (unlikely(bytes_read < 0)) {
2217 if (!copied)
2218 copied = bytes_read;
2219 goto out_err;
2220 }
2221
2222 copied += bytes_read;
2223
2224 if (skb_queue_empty(&msk->receive_queue) && __mptcp_move_skbs(msk))
2225 continue;
2226
2227 /* only the MPTCP socket status is relevant here. The exit
2228 * conditions mirror closely tcp_recvmsg()
2229 */
2230 if (copied >= target)
2231 break;
2232
2233 if (copied) {
2234 if (sk->sk_err ||
2235 sk->sk_state == TCP_CLOSE ||
2236 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2237 !timeo ||
2238 signal_pending(current))
2239 break;
2240 } else {
2241 if (sk->sk_err) {
2242 copied = sock_error(sk);
2243 break;
2244 }
2245
2246 if (sk->sk_shutdown & RCV_SHUTDOWN) {
2247 /* race breaker: the shutdown could be after the
2248 * previous receive queue check
2249 */
2250 if (__mptcp_move_skbs(msk))
2251 continue;
2252 break;
2253 }
2254
2255 if (sk->sk_state == TCP_CLOSE) {
2256 copied = -ENOTCONN;
2257 break;
2258 }
2259
2260 if (!timeo) {
2261 copied = -EAGAIN;
2262 break;
2263 }
2264
2265 if (signal_pending(current)) {
2266 copied = sock_intr_errno(timeo);
2267 break;
2268 }
2269 }
2270
2271 pr_debug("block timeout %ld\n", timeo);
2272 mptcp_cleanup_rbuf(msk, copied);
2273 err = sk_wait_data(sk, &timeo, NULL);
2274 if (err < 0) {
2275 err = copied ? : err;
2276 goto out_err;
2277 }
2278 }
2279
2280 mptcp_cleanup_rbuf(msk, copied);
2281
2282 out_err:
2283 if (cmsg_flags && copied >= 0) {
2284 if (cmsg_flags & MPTCP_CMSG_TS)
2285 tcp_recv_timestamp(msg, sk, &tss);
2286
2287 if (cmsg_flags & MPTCP_CMSG_INQ) {
2288 unsigned int inq = mptcp_inq_hint(sk);
2289
2290 put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq);
2291 }
2292 }
2293
2294 pr_debug("msk=%p rx queue empty=%d:%d copied=%d\n",
2295 msk, skb_queue_empty_lockless(&sk->sk_receive_queue),
2296 skb_queue_empty(&msk->receive_queue), copied);
2297
2298 release_sock(sk);
2299 return copied;
2300 }
2301
2302 static void mptcp_retransmit_timer(struct timer_list *t)
2303 {
2304 struct inet_connection_sock *icsk = from_timer(icsk, t,
2305 icsk_retransmit_timer);
2306 struct sock *sk = &icsk->icsk_inet.sk;
2307 struct mptcp_sock *msk = mptcp_sk(sk);
2308
2309 bh_lock_sock(sk);
2310 if (!sock_owned_by_user(sk)) {
2311 /* we need a process context to retransmit */
2312 if (!test_and_set_bit(MPTCP_WORK_RTX, &msk->flags))
2313 mptcp_schedule_work(sk);
2314 } else {
2315 /* delegate our work to tcp_release_cb() */
2316 __set_bit(MPTCP_RETRANSMIT, &msk->cb_flags);
2317 }
2318 bh_unlock_sock(sk);
2319 sock_put(sk);
2320 }
2321
2322 static void mptcp_tout_timer(struct timer_list *t)
2323 {
2324 struct sock *sk = from_timer(sk, t, sk_timer);
2325
2326 mptcp_schedule_work(sk);
2327 sock_put(sk);
2328 }
2329
2330 /* Find an idle subflow. Return NULL if there is unacked data at tcp
2331 * level.
2332 *
2333 * A backup subflow is returned only if that is the only kind available.
2334 */
2335 struct sock *mptcp_subflow_get_retrans(struct mptcp_sock *msk)
2336 {
2337 struct sock *backup = NULL, *pick = NULL;
2338 struct mptcp_subflow_context *subflow;
2339 int min_stale_count = INT_MAX;
2340
2341 mptcp_for_each_subflow(msk, subflow) {
2342 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2343
2344 if (!__mptcp_subflow_active(subflow))
2345 continue;
2346
2347 /* still data outstanding at TCP level? skip this */
2348 if (!tcp_rtx_and_write_queues_empty(ssk)) {
2349 mptcp_pm_subflow_chk_stale(msk, ssk);
2350 min_stale_count = min_t(int, min_stale_count, subflow->stale_count);
2351 continue;
2352 }
2353
2354 if (subflow->backup || subflow->request_bkup) {
2355 if (!backup)
2356 backup = ssk;
2357 continue;
2358 }
2359
2360 if (!pick)
2361 pick = ssk;
2362 }
2363
2364 if (pick)
2365 return pick;
2366
2367 /* use backup only if there are no progresses anywhere */
2368 return min_stale_count > 1 ? backup : NULL;
2369 }
2370
2371 bool __mptcp_retransmit_pending_data(struct sock *sk)
2372 {
2373 struct mptcp_data_frag *cur, *rtx_head;
2374 struct mptcp_sock *msk = mptcp_sk(sk);
2375
2376 if (__mptcp_check_fallback(msk))
2377 return false;
2378
2379 /* the closing socket has some data untransmitted and/or unacked:
2380 * some data in the mptcp rtx queue has not really xmitted yet.
2381 * keep it simple and re-inject the whole mptcp level rtx queue
2382 */
2383 mptcp_data_lock(sk);
2384 __mptcp_clean_una_wakeup(sk);
2385 rtx_head = mptcp_rtx_head(sk);
2386 if (!rtx_head) {
2387 mptcp_data_unlock(sk);
2388 return false;
2389 }
2390
2391 msk->recovery_snd_nxt = msk->snd_nxt;
2392 msk->recovery = true;
2393 mptcp_data_unlock(sk);
2394
2395 msk->first_pending = rtx_head;
2396 msk->snd_burst = 0;
2397
2398 /* be sure to clear the "sent status" on all re-injected fragments */
2399 list_for_each_entry(cur, &msk->rtx_queue, list) {
2400 if (!cur->already_sent)
2401 break;
2402 cur->already_sent = 0;
2403 }
2404
2405 return true;
2406 }
2407
2408 /* flags for __mptcp_close_ssk() */
2409 #define MPTCP_CF_PUSH BIT(1)
2410 #define MPTCP_CF_FASTCLOSE BIT(2)
2411
2412 /* be sure to send a reset only if the caller asked for it, also
2413 * clean completely the subflow status when the subflow reaches
2414 * TCP_CLOSE state
2415 */
2416 static void __mptcp_subflow_disconnect(struct sock *ssk,
2417 struct mptcp_subflow_context *subflow,
2418 unsigned int flags)
2419 {
2420 if (((1 << ssk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) ||
2421 (flags & MPTCP_CF_FASTCLOSE)) {
2422 /* The MPTCP code never wait on the subflow sockets, TCP-level
2423 * disconnect should never fail
2424 */
2425 WARN_ON_ONCE(tcp_disconnect(ssk, 0));
2426 mptcp_subflow_ctx_reset(subflow);
2427 } else {
2428 tcp_shutdown(ssk, SEND_SHUTDOWN);
2429 }
2430 }
2431
2432 /* subflow sockets can be either outgoing (connect) or incoming
2433 * (accept).
2434 *
2435 * Outgoing subflows use in-kernel sockets.
2436 * Incoming subflows do not have their own 'struct socket' allocated,
2437 * so we need to use tcp_close() after detaching them from the mptcp
2438 * parent socket.
2439 */
2440 static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2441 struct mptcp_subflow_context *subflow,
2442 unsigned int flags)
2443 {
2444 struct mptcp_sock *msk = mptcp_sk(sk);
2445 bool dispose_it, need_push = false;
2446
2447 /* If the first subflow moved to a close state before accept, e.g. due
2448 * to an incoming reset or listener shutdown, the subflow socket is
2449 * already deleted by inet_child_forget() and the mptcp socket can't
2450 * survive too.
2451 */
2452 if (msk->in_accept_queue && msk->first == ssk &&
2453 (sock_flag(sk, SOCK_DEAD) || sock_flag(ssk, SOCK_DEAD))) {
2454 /* ensure later check in mptcp_worker() will dispose the msk */
2455 sock_set_flag(sk, SOCK_DEAD);
2456 mptcp_set_close_tout(sk, tcp_jiffies32 - (mptcp_close_timeout(sk) + 1));
2457 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2458 mptcp_subflow_drop_ctx(ssk);
2459 goto out_release;
2460 }
2461
2462 dispose_it = msk->free_first || ssk != msk->first;
2463 if (dispose_it)
2464 list_del(&subflow->node);
2465
2466 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2467
2468 if ((flags & MPTCP_CF_FASTCLOSE) && !__mptcp_check_fallback(msk)) {
2469 /* be sure to force the tcp_close path
2470 * to generate the egress reset
2471 */
2472 ssk->sk_lingertime = 0;
2473 sock_set_flag(ssk, SOCK_LINGER);
2474 subflow->send_fastclose = 1;
2475 }
2476
2477 need_push = (flags & MPTCP_CF_PUSH) && __mptcp_retransmit_pending_data(sk);
2478 if (!dispose_it) {
2479 __mptcp_subflow_disconnect(ssk, subflow, flags);
2480 release_sock(ssk);
2481
2482 goto out;
2483 }
2484
2485 subflow->disposable = 1;
2486
2487 /* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops
2488 * the ssk has been already destroyed, we just need to release the
2489 * reference owned by msk;
2490 */
2491 if (!inet_csk(ssk)->icsk_ulp_ops) {
2492 WARN_ON_ONCE(!sock_flag(ssk, SOCK_DEAD));
2493 kfree_rcu(subflow, rcu);
2494 } else {
2495 /* otherwise tcp will dispose of the ssk and subflow ctx */
2496 __tcp_close(ssk, 0);
2497
2498 /* close acquired an extra ref */
2499 __sock_put(ssk);
2500 }
2501
2502 out_release:
2503 __mptcp_subflow_error_report(sk, ssk);
2504 release_sock(ssk);
2505
2506 sock_put(ssk);
2507
2508 if (ssk == msk->first)
2509 WRITE_ONCE(msk->first, NULL);
2510
2511 out:
2512 __mptcp_sync_sndbuf(sk);
2513 if (need_push)
2514 __mptcp_push_pending(sk, 0);
2515
2516 /* Catch every 'all subflows closed' scenario, including peers silently
2517 * closing them, e.g. due to timeout.
2518 * For established sockets, allow an additional timeout before closing,
2519 * as the protocol can still create more subflows.
2520 */
2521 if (list_is_singular(&msk->conn_list) && msk->first &&
2522 inet_sk_state_load(msk->first) == TCP_CLOSE) {
2523 if (sk->sk_state != TCP_ESTABLISHED ||
2524 msk->in_accept_queue || sock_flag(sk, SOCK_DEAD)) {
2525 mptcp_set_state(sk, TCP_CLOSE);
2526 mptcp_close_wake_up(sk);
2527 } else {
2528 mptcp_start_tout_timer(sk);
2529 }
2530 }
2531 }
2532
2533 void mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2534 struct mptcp_subflow_context *subflow)
2535 {
2536 /* The first subflow can already be closed and still in the list */
2537 if (subflow->close_event_done)
2538 return;
2539
2540 subflow->close_event_done = true;
2541
2542 if (sk->sk_state == TCP_ESTABLISHED)
2543 mptcp_event(MPTCP_EVENT_SUB_CLOSED, mptcp_sk(sk), ssk, GFP_KERNEL);
2544
2545 /* subflow aborted before reaching the fully_established status
2546 * attempt the creation of the next subflow
2547 */
2548 mptcp_pm_subflow_check_next(mptcp_sk(sk), subflow);
2549
2550 __mptcp_close_ssk(sk, ssk, subflow, MPTCP_CF_PUSH);
2551 }
2552
2553 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
2554 {
2555 return 0;
2556 }
2557
2558 static void __mptcp_close_subflow(struct sock *sk)
2559 {
2560 struct mptcp_subflow_context *subflow, *tmp;
2561 struct mptcp_sock *msk = mptcp_sk(sk);
2562
2563 might_sleep();
2564
2565 mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2566 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2567 int ssk_state = inet_sk_state_load(ssk);
2568
2569 if (ssk_state != TCP_CLOSE &&
2570 (ssk_state != TCP_CLOSE_WAIT ||
2571 inet_sk_state_load(sk) != TCP_ESTABLISHED))
2572 continue;
2573
2574 /* 'subflow_data_ready' will re-sched once rx queue is empty */
2575 if (!skb_queue_empty_lockless(&ssk->sk_receive_queue))
2576 continue;
2577
2578 mptcp_close_ssk(sk, ssk, subflow);
2579 }
2580
2581 }
2582
2583 static bool mptcp_close_tout_expired(const struct sock *sk)
2584 {
2585 if (!inet_csk(sk)->icsk_mtup.probe_timestamp ||
2586 sk->sk_state == TCP_CLOSE)
2587 return false;
2588
2589 return time_after32(tcp_jiffies32,
2590 inet_csk(sk)->icsk_mtup.probe_timestamp + mptcp_close_timeout(sk));
2591 }
2592
2593 static void mptcp_check_fastclose(struct mptcp_sock *msk)
2594 {
2595 struct mptcp_subflow_context *subflow, *tmp;
2596 struct sock *sk = (struct sock *)msk;
2597
2598 if (likely(!READ_ONCE(msk->rcv_fastclose)))
2599 return;
2600
2601 mptcp_token_destroy(msk);
2602
2603 mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2604 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2605 bool slow;
2606
2607 slow = lock_sock_fast(tcp_sk);
2608 if (tcp_sk->sk_state != TCP_CLOSE) {
2609 mptcp_send_active_reset_reason(tcp_sk);
2610 tcp_set_state(tcp_sk, TCP_CLOSE);
2611 }
2612 unlock_sock_fast(tcp_sk, slow);
2613 }
2614
2615 /* Mirror the tcp_reset() error propagation */
2616 switch (sk->sk_state) {
2617 case TCP_SYN_SENT:
2618 WRITE_ONCE(sk->sk_err, ECONNREFUSED);
2619 break;
2620 case TCP_CLOSE_WAIT:
2621 WRITE_ONCE(sk->sk_err, EPIPE);
2622 break;
2623 case TCP_CLOSE:
2624 return;
2625 default:
2626 WRITE_ONCE(sk->sk_err, ECONNRESET);
2627 }
2628
2629 mptcp_set_state(sk, TCP_CLOSE);
2630 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
2631 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
2632 set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags);
2633
2634 /* the calling mptcp_worker will properly destroy the socket */
2635 if (sock_flag(sk, SOCK_DEAD))
2636 return;
2637
2638 sk->sk_state_change(sk);
2639 sk_error_report(sk);
2640 }
2641
2642 static void __mptcp_retrans(struct sock *sk)
2643 {
2644 struct mptcp_sock *msk = mptcp_sk(sk);
2645 struct mptcp_subflow_context *subflow;
2646 struct mptcp_sendmsg_info info = {};
2647 struct mptcp_data_frag *dfrag;
2648 struct sock *ssk;
2649 int ret, err;
2650 u16 len = 0;
2651
2652 mptcp_clean_una_wakeup(sk);
2653
2654 /* first check ssk: need to kick "stale" logic */
2655 err = mptcp_sched_get_retrans(msk);
2656 dfrag = mptcp_rtx_head(sk);
2657 if (!dfrag) {
2658 if (mptcp_data_fin_enabled(msk)) {
2659 struct inet_connection_sock *icsk = inet_csk(sk);
2660
2661 icsk->icsk_retransmits++;
2662 mptcp_set_datafin_timeout(sk);
2663 mptcp_send_ack(msk);
2664
2665 goto reset_timer;
2666 }
2667
2668 if (!mptcp_send_head(sk))
2669 return;
2670
2671 goto reset_timer;
2672 }
2673
2674 if (err)
2675 goto reset_timer;
2676
2677 mptcp_for_each_subflow(msk, subflow) {
2678 if (READ_ONCE(subflow->scheduled)) {
2679 u16 copied = 0;
2680
2681 mptcp_subflow_set_scheduled(subflow, false);
2682
2683 ssk = mptcp_subflow_tcp_sock(subflow);
2684
2685 lock_sock(ssk);
2686
2687 /* limit retransmission to the bytes already sent on some subflows */
2688 info.sent = 0;
2689 info.limit = READ_ONCE(msk->csum_enabled) ? dfrag->data_len :
2690 dfrag->already_sent;
2691 while (info.sent < info.limit) {
2692 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
2693 if (ret <= 0)
2694 break;
2695
2696 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
2697 copied += ret;
2698 info.sent += ret;
2699 }
2700 if (copied) {
2701 len = max(copied, len);
2702 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
2703 info.size_goal);
2704 WRITE_ONCE(msk->allow_infinite_fallback, false);
2705 }
2706
2707 release_sock(ssk);
2708 }
2709 }
2710
2711 msk->bytes_retrans += len;
2712 dfrag->already_sent = max(dfrag->already_sent, len);
2713
2714 reset_timer:
2715 mptcp_check_and_set_pending(sk);
2716
2717 if (!mptcp_rtx_timer_pending(sk))
2718 mptcp_reset_rtx_timer(sk);
2719 }
2720
2721 /* schedule the timeout timer for the relevant event: either close timeout
2722 * or mp_fail timeout. The close timeout takes precedence on the mp_fail one
2723 */
2724 void mptcp_reset_tout_timer(struct mptcp_sock *msk, unsigned long fail_tout)
2725 {
2726 struct sock *sk = (struct sock *)msk;
2727 unsigned long timeout, close_timeout;
2728
2729 if (!fail_tout && !inet_csk(sk)->icsk_mtup.probe_timestamp)
2730 return;
2731
2732 close_timeout = (unsigned long)inet_csk(sk)->icsk_mtup.probe_timestamp -
2733 tcp_jiffies32 + jiffies + mptcp_close_timeout(sk);
2734
2735 /* the close timeout takes precedence on the fail one, and here at least one of
2736 * them is active
2737 */
2738 timeout = inet_csk(sk)->icsk_mtup.probe_timestamp ? close_timeout : fail_tout;
2739
2740 sk_reset_timer(sk, &sk->sk_timer, timeout);
2741 }
2742
2743 static void mptcp_mp_fail_no_response(struct mptcp_sock *msk)
2744 {
2745 struct sock *ssk = msk->first;
2746 bool slow;
2747
2748 if (!ssk)
2749 return;
2750
2751 pr_debug("MP_FAIL doesn't respond, reset the subflow\n");
2752
2753 slow = lock_sock_fast(ssk);
2754 mptcp_subflow_reset(ssk);
2755 WRITE_ONCE(mptcp_subflow_ctx(ssk)->fail_tout, 0);
2756 unlock_sock_fast(ssk, slow);
2757 }
2758
2759 static void mptcp_do_fastclose(struct sock *sk)
2760 {
2761 struct mptcp_subflow_context *subflow, *tmp;
2762 struct mptcp_sock *msk = mptcp_sk(sk);
2763
2764 mptcp_set_state(sk, TCP_CLOSE);
2765 mptcp_for_each_subflow_safe(msk, subflow, tmp)
2766 __mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow),
2767 subflow, MPTCP_CF_FASTCLOSE);
2768 }
2769
2770 static void mptcp_worker(struct work_struct *work)
2771 {
2772 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
2773 struct sock *sk = (struct sock *)msk;
2774 unsigned long fail_tout;
2775 int state;
2776
2777 lock_sock(sk);
2778 state = sk->sk_state;
2779 if (unlikely((1 << state) & (TCPF_CLOSE | TCPF_LISTEN)))
2780 goto unlock;
2781
2782 mptcp_check_fastclose(msk);
2783
2784 mptcp_pm_nl_work(msk);
2785
2786 mptcp_check_send_data_fin(sk);
2787 mptcp_check_data_fin_ack(sk);
2788 mptcp_check_data_fin(sk);
2789
2790 if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags))
2791 __mptcp_close_subflow(sk);
2792
2793 if (mptcp_close_tout_expired(sk)) {
2794 mptcp_do_fastclose(sk);
2795 mptcp_close_wake_up(sk);
2796 }
2797
2798 if (sock_flag(sk, SOCK_DEAD) && sk->sk_state == TCP_CLOSE) {
2799 __mptcp_destroy_sock(sk);
2800 goto unlock;
2801 }
2802
2803 if (test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
2804 __mptcp_retrans(sk);
2805
2806 fail_tout = msk->first ? READ_ONCE(mptcp_subflow_ctx(msk->first)->fail_tout) : 0;
2807 if (fail_tout && time_after(jiffies, fail_tout))
2808 mptcp_mp_fail_no_response(msk);
2809
2810 unlock:
2811 release_sock(sk);
2812 sock_put(sk);
2813 }
2814
2815 static void __mptcp_init_sock(struct sock *sk)
2816 {
2817 struct mptcp_sock *msk = mptcp_sk(sk);
2818
2819 INIT_LIST_HEAD(&msk->conn_list);
2820 INIT_LIST_HEAD(&msk->join_list);
2821 INIT_LIST_HEAD(&msk->rtx_queue);
2822 INIT_WORK(&msk->work, mptcp_worker);
2823 __skb_queue_head_init(&msk->receive_queue);
2824 msk->out_of_order_queue = RB_ROOT;
2825 msk->first_pending = NULL;
2826 WRITE_ONCE(msk->rmem_fwd_alloc, 0);
2827 WRITE_ONCE(msk->rmem_released, 0);
2828 msk->timer_ival = TCP_RTO_MIN;
2829 msk->scaling_ratio = TCP_DEFAULT_SCALING_RATIO;
2830
2831 WRITE_ONCE(msk->first, NULL);
2832 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
2833 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
2834 WRITE_ONCE(msk->allow_infinite_fallback, true);
2835 msk->recovery = false;
2836 msk->subflow_id = 1;
2837 msk->last_data_sent = tcp_jiffies32;
2838 msk->last_data_recv = tcp_jiffies32;
2839 msk->last_ack_recv = tcp_jiffies32;
2840
2841 mptcp_pm_data_init(msk);
2842
2843 /* re-use the csk retrans timer for MPTCP-level retrans */
2844 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
2845 timer_setup(&sk->sk_timer, mptcp_tout_timer, 0);
2846 }
2847
2848 static void mptcp_ca_reset(struct sock *sk)
2849 {
2850 struct inet_connection_sock *icsk = inet_csk(sk);
2851
2852 tcp_assign_congestion_control(sk);
2853 strscpy(mptcp_sk(sk)->ca_name, icsk->icsk_ca_ops->name,
2854 sizeof(mptcp_sk(sk)->ca_name));
2855
2856 /* no need to keep a reference to the ops, the name will suffice */
2857 tcp_cleanup_congestion_control(sk);
2858 icsk->icsk_ca_ops = NULL;
2859 }
2860
2861 static int mptcp_init_sock(struct sock *sk)
2862 {
2863 struct net *net = sock_net(sk);
2864 int ret;
2865
2866 __mptcp_init_sock(sk);
2867
2868 if (!mptcp_is_enabled(net))
2869 return -ENOPROTOOPT;
2870
2871 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
2872 return -ENOMEM;
2873
2874 rcu_read_lock();
2875 ret = mptcp_init_sched(mptcp_sk(sk),
2876 mptcp_sched_find(mptcp_get_scheduler(net)));
2877 rcu_read_unlock();
2878 if (ret)
2879 return ret;
2880
2881 set_bit(SOCK_CUSTOM_SOCKOPT, &sk->sk_socket->flags);
2882
2883 /* fetch the ca name; do it outside __mptcp_init_sock(), so that clone will
2884 * propagate the correct value
2885 */
2886 mptcp_ca_reset(sk);
2887
2888 sk_sockets_allocated_inc(sk);
2889 sk->sk_rcvbuf = READ_ONCE(net->ipv4.sysctl_tcp_rmem[1]);
2890 sk->sk_sndbuf = READ_ONCE(net->ipv4.sysctl_tcp_wmem[1]);
2891
2892 return 0;
2893 }
2894
2895 static void __mptcp_clear_xmit(struct sock *sk)
2896 {
2897 struct mptcp_sock *msk = mptcp_sk(sk);
2898 struct mptcp_data_frag *dtmp, *dfrag;
2899
2900 WRITE_ONCE(msk->first_pending, NULL);
2901 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
2902 dfrag_clear(sk, dfrag);
2903 }
2904
2905 void mptcp_cancel_work(struct sock *sk)
2906 {
2907 struct mptcp_sock *msk = mptcp_sk(sk);
2908
2909 if (cancel_work_sync(&msk->work))
2910 __sock_put(sk);
2911 }
2912
2913 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
2914 {
2915 lock_sock(ssk);
2916
2917 switch (ssk->sk_state) {
2918 case TCP_LISTEN:
2919 if (!(how & RCV_SHUTDOWN))
2920 break;
2921 fallthrough;
2922 case TCP_SYN_SENT:
2923 WARN_ON_ONCE(tcp_disconnect(ssk, O_NONBLOCK));
2924 break;
2925 default:
2926 if (__mptcp_check_fallback(mptcp_sk(sk))) {
2927 pr_debug("Fallback\n");
2928 ssk->sk_shutdown |= how;
2929 tcp_shutdown(ssk, how);
2930
2931 /* simulate the data_fin ack reception to let the state
2932 * machine move forward
2933 */
2934 WRITE_ONCE(mptcp_sk(sk)->snd_una, mptcp_sk(sk)->snd_nxt);
2935 mptcp_schedule_work(sk);
2936 } else {
2937 pr_debug("Sending DATA_FIN on subflow %p\n", ssk);
2938 tcp_send_ack(ssk);
2939 if (!mptcp_rtx_timer_pending(sk))
2940 mptcp_reset_rtx_timer(sk);
2941 }
2942 break;
2943 }
2944
2945 release_sock(ssk);
2946 }
2947
2948 void mptcp_set_state(struct sock *sk, int state)
2949 {
2950 int oldstate = sk->sk_state;
2951
2952 switch (state) {
2953 case TCP_ESTABLISHED:
2954 if (oldstate != TCP_ESTABLISHED)
2955 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_CURRESTAB);
2956 break;
2957 case TCP_CLOSE_WAIT:
2958 /* Unlike TCP, MPTCP sk would not have the TCP_SYN_RECV state:
2959 * MPTCP "accepted" sockets will be created later on. So no
2960 * transition from TCP_SYN_RECV to TCP_CLOSE_WAIT.
2961 */
2962 break;
2963 default:
2964 if (oldstate == TCP_ESTABLISHED || oldstate == TCP_CLOSE_WAIT)
2965 MPTCP_DEC_STATS(sock_net(sk), MPTCP_MIB_CURRESTAB);
2966 }
2967
2968 inet_sk_state_store(sk, state);
2969 }
2970
2971 static const unsigned char new_state[16] = {
2972 /* current state: new state: action: */
2973 [0 /* (Invalid) */] = TCP_CLOSE,
2974 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2975 [TCP_SYN_SENT] = TCP_CLOSE,
2976 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2977 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2978 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2979 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */
2980 [TCP_CLOSE] = TCP_CLOSE,
2981 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2982 [TCP_LAST_ACK] = TCP_LAST_ACK,
2983 [TCP_LISTEN] = TCP_CLOSE,
2984 [TCP_CLOSING] = TCP_CLOSING,
2985 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2986 };
2987
2988 static int mptcp_close_state(struct sock *sk)
2989 {
2990 int next = (int)new_state[sk->sk_state];
2991 int ns = next & TCP_STATE_MASK;
2992
2993 mptcp_set_state(sk, ns);
2994
2995 return next & TCP_ACTION_FIN;
2996 }
2997
2998 static void mptcp_check_send_data_fin(struct sock *sk)
2999 {
3000 struct mptcp_subflow_context *subflow;
3001 struct mptcp_sock *msk = mptcp_sk(sk);
3002
3003 pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu\n",
3004 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk),
3005 msk->snd_nxt, msk->write_seq);
3006
3007 /* we still need to enqueue subflows or not really shutting down,
3008 * skip this
3009 */
3010 if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq ||
3011 mptcp_send_head(sk))
3012 return;
3013
3014 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
3015
3016 mptcp_for_each_subflow(msk, subflow) {
3017 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
3018
3019 mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN);
3020 }
3021 }
3022
3023 static void __mptcp_wr_shutdown(struct sock *sk)
3024 {
3025 struct mptcp_sock *msk = mptcp_sk(sk);
3026
3027 pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d\n",
3028 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state,
3029 !!mptcp_send_head(sk));
3030
3031 /* will be ignored by fallback sockets */
3032 WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
3033 WRITE_ONCE(msk->snd_data_fin_enable, 1);
3034
3035 mptcp_check_send_data_fin(sk);
3036 }
3037
3038 static void __mptcp_destroy_sock(struct sock *sk)
3039 {
3040 struct mptcp_sock *msk = mptcp_sk(sk);
3041
3042 pr_debug("msk=%p\n", msk);
3043
3044 might_sleep();
3045
3046 mptcp_stop_rtx_timer(sk);
3047 sk_stop_timer(sk, &sk->sk_timer);
3048 msk->pm.status = 0;
3049 mptcp_release_sched(msk);
3050
3051 sk->sk_prot->destroy(sk);
3052
3053 WARN_ON_ONCE(READ_ONCE(msk->rmem_fwd_alloc));
3054 WARN_ON_ONCE(msk->rmem_released);
3055 sk_stream_kill_queues(sk);
3056 xfrm_sk_free_policy(sk);
3057
3058 sock_put(sk);
3059 }
3060
3061 void __mptcp_unaccepted_force_close(struct sock *sk)
3062 {
3063 sock_set_flag(sk, SOCK_DEAD);
3064 mptcp_do_fastclose(sk);
3065 __mptcp_destroy_sock(sk);
3066 }
3067
3068 static __poll_t mptcp_check_readable(struct sock *sk)
3069 {
3070 return mptcp_epollin_ready(sk) ? EPOLLIN | EPOLLRDNORM : 0;
3071 }
3072
3073 static void mptcp_check_listen_stop(struct sock *sk)
3074 {
3075 struct sock *ssk;
3076
3077 if (inet_sk_state_load(sk) != TCP_LISTEN)
3078 return;
3079
3080 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
3081 ssk = mptcp_sk(sk)->first;
3082 if (WARN_ON_ONCE(!ssk || inet_sk_state_load(ssk) != TCP_LISTEN))
3083 return;
3084
3085 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
3086 tcp_set_state(ssk, TCP_CLOSE);
3087 mptcp_subflow_queue_clean(sk, ssk);
3088 inet_csk_listen_stop(ssk);
3089 mptcp_event_pm_listener(ssk, MPTCP_EVENT_LISTENER_CLOSED);
3090 release_sock(ssk);
3091 }
3092
3093 bool __mptcp_close(struct sock *sk, long timeout)
3094 {
3095 struct mptcp_subflow_context *subflow;
3096 struct mptcp_sock *msk = mptcp_sk(sk);
3097 bool do_cancel_work = false;
3098 int subflows_alive = 0;
3099
3100 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
3101
3102 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) {
3103 mptcp_check_listen_stop(sk);
3104 mptcp_set_state(sk, TCP_CLOSE);
3105 goto cleanup;
3106 }
3107
3108 if (mptcp_data_avail(msk) || timeout < 0) {
3109 /* If the msk has read data, or the caller explicitly ask it,
3110 * do the MPTCP equivalent of TCP reset, aka MPTCP fastclose
3111 */
3112 mptcp_do_fastclose(sk);
3113 timeout = 0;
3114 } else if (mptcp_close_state(sk)) {
3115 __mptcp_wr_shutdown(sk);
3116 }
3117
3118 sk_stream_wait_close(sk, timeout);
3119
3120 cleanup:
3121 /* orphan all the subflows */
3122 mptcp_for_each_subflow(msk, subflow) {
3123 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3124 bool slow = lock_sock_fast_nested(ssk);
3125
3126 subflows_alive += ssk->sk_state != TCP_CLOSE;
3127
3128 /* since the close timeout takes precedence on the fail one,
3129 * cancel the latter
3130 */
3131 if (ssk == msk->first)
3132 subflow->fail_tout = 0;
3133
3134 /* detach from the parent socket, but allow data_ready to
3135 * push incoming data into the mptcp stack, to properly ack it
3136 */
3137 ssk->sk_socket = NULL;
3138 ssk->sk_wq = NULL;
3139 unlock_sock_fast(ssk, slow);
3140 }
3141 sock_orphan(sk);
3142
3143 /* all the subflows are closed, only timeout can change the msk
3144 * state, let's not keep resources busy for no reasons
3145 */
3146 if (subflows_alive == 0)
3147 mptcp_set_state(sk, TCP_CLOSE);
3148
3149 sock_hold(sk);
3150 pr_debug("msk=%p state=%d\n", sk, sk->sk_state);
3151 mptcp_pm_connection_closed(msk);
3152
3153 if (sk->sk_state == TCP_CLOSE) {
3154 __mptcp_destroy_sock(sk);
3155 do_cancel_work = true;
3156 } else {
3157 mptcp_start_tout_timer(sk);
3158 }
3159
3160 return do_cancel_work;
3161 }
3162
3163 static void mptcp_close(struct sock *sk, long timeout)
3164 {
3165 bool do_cancel_work;
3166
3167 lock_sock(sk);
3168
3169 do_cancel_work = __mptcp_close(sk, timeout);
3170 release_sock(sk);
3171 if (do_cancel_work)
3172 mptcp_cancel_work(sk);
3173
3174 sock_put(sk);
3175 }
3176
3177 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
3178 {
3179 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3180 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
3181 struct ipv6_pinfo *msk6 = inet6_sk(msk);
3182
3183 msk->sk_v6_daddr = ssk->sk_v6_daddr;
3184 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
3185
3186 if (msk6 && ssk6) {
3187 msk6->saddr = ssk6->saddr;
3188 msk6->flow_label = ssk6->flow_label;
3189 }
3190 #endif
3191
3192 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
3193 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
3194 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
3195 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
3196 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
3197 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
3198 }
3199
3200 static int mptcp_disconnect(struct sock *sk, int flags)
3201 {
3202 struct mptcp_sock *msk = mptcp_sk(sk);
3203
3204 /* We are on the fastopen error path. We can't call straight into the
3205 * subflows cleanup code due to lock nesting (we are already under
3206 * msk->firstsocket lock).
3207 */
3208 if (msk->fastopening)
3209 return -EBUSY;
3210
3211 mptcp_check_listen_stop(sk);
3212 mptcp_set_state(sk, TCP_CLOSE);
3213
3214 mptcp_stop_rtx_timer(sk);
3215 mptcp_stop_tout_timer(sk);
3216
3217 mptcp_pm_connection_closed(msk);
3218
3219 /* msk->subflow is still intact, the following will not free the first
3220 * subflow
3221 */
3222 mptcp_destroy_common(msk, MPTCP_CF_FASTCLOSE);
3223 WRITE_ONCE(msk->flags, 0);
3224 msk->cb_flags = 0;
3225 msk->recovery = false;
3226 WRITE_ONCE(msk->can_ack, false);
3227 WRITE_ONCE(msk->fully_established, false);
3228 WRITE_ONCE(msk->rcv_data_fin, false);
3229 WRITE_ONCE(msk->snd_data_fin_enable, false);
3230 WRITE_ONCE(msk->rcv_fastclose, false);
3231 WRITE_ONCE(msk->use_64bit_ack, false);
3232 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
3233 mptcp_pm_data_reset(msk);
3234 mptcp_ca_reset(sk);
3235 msk->bytes_consumed = 0;
3236 msk->bytes_acked = 0;
3237 msk->bytes_received = 0;
3238 msk->bytes_sent = 0;
3239 msk->bytes_retrans = 0;
3240 msk->rcvspace_init = 0;
3241
3242 WRITE_ONCE(sk->sk_shutdown, 0);
3243 sk_error_report(sk);
3244 return 0;
3245 }
3246
3247 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3248 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
3249 {
3250 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
3251
3252 return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
3253 }
3254
3255 static void mptcp_copy_ip6_options(struct sock *newsk, const struct sock *sk)
3256 {
3257 const struct ipv6_pinfo *np = inet6_sk(sk);
3258 struct ipv6_txoptions *opt;
3259 struct ipv6_pinfo *newnp;
3260
3261 newnp = inet6_sk(newsk);
3262
3263 rcu_read_lock();
3264 opt = rcu_dereference(np->opt);
3265 if (opt) {
3266 opt = ipv6_dup_options(newsk, opt);
3267 if (!opt)
3268 net_warn_ratelimited("%s: Failed to copy ip6 options\n", __func__);
3269 }
3270 RCU_INIT_POINTER(newnp->opt, opt);
3271 rcu_read_unlock();
3272 }
3273 #endif
3274
3275 static void mptcp_copy_ip_options(struct sock *newsk, const struct sock *sk)
3276 {
3277 struct ip_options_rcu *inet_opt, *newopt = NULL;
3278 const struct inet_sock *inet = inet_sk(sk);
3279 struct inet_sock *newinet;
3280
3281 newinet = inet_sk(newsk);
3282
3283 rcu_read_lock();
3284 inet_opt = rcu_dereference(inet->inet_opt);
3285 if (inet_opt) {
3286 newopt = sock_kmalloc(newsk, sizeof(*inet_opt) +
3287 inet_opt->opt.optlen, GFP_ATOMIC);
3288 if (newopt)
3289 memcpy(newopt, inet_opt, sizeof(*inet_opt) +
3290 inet_opt->opt.optlen);
3291 else
3292 net_warn_ratelimited("%s: Failed to copy ip options\n", __func__);
3293 }
3294 RCU_INIT_POINTER(newinet->inet_opt, newopt);
3295 rcu_read_unlock();
3296 }
3297
3298 struct sock *mptcp_sk_clone_init(const struct sock *sk,
3299 const struct mptcp_options_received *mp_opt,
3300 struct sock *ssk,
3301 struct request_sock *req)
3302 {
3303 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
3304 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
3305 struct mptcp_subflow_context *subflow;
3306 struct mptcp_sock *msk;
3307
3308 if (!nsk)
3309 return NULL;
3310
3311 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3312 if (nsk->sk_family == AF_INET6)
3313 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
3314 #endif
3315
3316 __mptcp_init_sock(nsk);
3317
3318 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3319 if (nsk->sk_family == AF_INET6)
3320 mptcp_copy_ip6_options(nsk, sk);
3321 else
3322 #endif
3323 mptcp_copy_ip_options(nsk, sk);
3324
3325 msk = mptcp_sk(nsk);
3326 WRITE_ONCE(msk->local_key, subflow_req->local_key);
3327 WRITE_ONCE(msk->token, subflow_req->token);
3328 msk->in_accept_queue = 1;
3329 WRITE_ONCE(msk->fully_established, false);
3330 if (mp_opt->suboptions & OPTION_MPTCP_CSUMREQD)
3331 WRITE_ONCE(msk->csum_enabled, true);
3332
3333 WRITE_ONCE(msk->write_seq, subflow_req->idsn + 1);
3334 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
3335 WRITE_ONCE(msk->snd_una, msk->write_seq);
3336 WRITE_ONCE(msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd);
3337 msk->setsockopt_seq = mptcp_sk(sk)->setsockopt_seq;
3338 mptcp_init_sched(msk, mptcp_sk(sk)->sched);
3339
3340 /* passive msk is created after the first/MPC subflow */
3341 msk->subflow_id = 2;
3342
3343 sock_reset_flag(nsk, SOCK_RCU_FREE);
3344 security_inet_csk_clone(nsk, req);
3345
3346 /* this can't race with mptcp_close(), as the msk is
3347 * not yet exposted to user-space
3348 */
3349 mptcp_set_state(nsk, TCP_ESTABLISHED);
3350
3351 /* The msk maintain a ref to each subflow in the connections list */
3352 WRITE_ONCE(msk->first, ssk);
3353 subflow = mptcp_subflow_ctx(ssk);
3354 list_add(&subflow->node, &msk->conn_list);
3355 sock_hold(ssk);
3356
3357 /* new mpc subflow takes ownership of the newly
3358 * created mptcp socket
3359 */
3360 mptcp_token_accept(subflow_req, msk);
3361
3362 /* set msk addresses early to ensure mptcp_pm_get_local_id()
3363 * uses the correct data
3364 */
3365 mptcp_copy_inaddrs(nsk, ssk);
3366 __mptcp_propagate_sndbuf(nsk, ssk);
3367
3368 mptcp_rcv_space_init(msk, ssk);
3369
3370 if (mp_opt->suboptions & OPTION_MPTCP_MPC_ACK)
3371 __mptcp_subflow_fully_established(msk, subflow, mp_opt);
3372 bh_unlock_sock(nsk);
3373
3374 /* note: the newly allocated socket refcount is 2 now */
3375 return nsk;
3376 }
3377
3378 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
3379 {
3380 const struct tcp_sock *tp = tcp_sk(ssk);
3381
3382 msk->rcvspace_init = 1;
3383 msk->rcvq_space.copied = 0;
3384 msk->rcvq_space.rtt_us = 0;
3385
3386 msk->rcvq_space.time = tp->tcp_mstamp;
3387
3388 /* initial rcv_space offering made to peer */
3389 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
3390 TCP_INIT_CWND * tp->advmss);
3391 if (msk->rcvq_space.space == 0)
3392 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
3393 }
3394
3395 void mptcp_destroy_common(struct mptcp_sock *msk, unsigned int flags)
3396 {
3397 struct mptcp_subflow_context *subflow, *tmp;
3398 struct sock *sk = (struct sock *)msk;
3399
3400 __mptcp_clear_xmit(sk);
3401
3402 /* join list will be eventually flushed (with rst) at sock lock release time */
3403 mptcp_for_each_subflow_safe(msk, subflow, tmp)
3404 __mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow), subflow, flags);
3405
3406 /* move to sk_receive_queue, sk_stream_kill_queues will purge it */
3407 mptcp_data_lock(sk);
3408 skb_queue_splice_tail_init(&msk->receive_queue, &sk->sk_receive_queue);
3409 __skb_queue_purge(&sk->sk_receive_queue);
3410 skb_rbtree_purge(&msk->out_of_order_queue);
3411 mptcp_data_unlock(sk);
3412
3413 /* move all the rx fwd alloc into the sk_mem_reclaim_final in
3414 * inet_sock_destruct() will dispose it
3415 */
3416 sk_forward_alloc_add(sk, msk->rmem_fwd_alloc);
3417 WRITE_ONCE(msk->rmem_fwd_alloc, 0);
3418 mptcp_token_destroy(msk);
3419 mptcp_pm_free_anno_list(msk);
3420 mptcp_free_local_addr_list(msk);
3421 }
3422
3423 static void mptcp_destroy(struct sock *sk)
3424 {
3425 struct mptcp_sock *msk = mptcp_sk(sk);
3426
3427 /* allow the following to close even the initial subflow */
3428 msk->free_first = 1;
3429 mptcp_destroy_common(msk, 0);
3430 sk_sockets_allocated_dec(sk);
3431 }
3432
3433 void __mptcp_data_acked(struct sock *sk)
3434 {
3435 if (!sock_owned_by_user(sk))
3436 __mptcp_clean_una(sk);
3437 else
3438 __set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->cb_flags);
3439 }
3440
3441 void __mptcp_check_push(struct sock *sk, struct sock *ssk)
3442 {
3443 if (!mptcp_send_head(sk))
3444 return;
3445
3446 if (!sock_owned_by_user(sk))
3447 __mptcp_subflow_push_pending(sk, ssk, false);
3448 else
3449 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3450 }
3451
3452 #define MPTCP_FLAGS_PROCESS_CTX_NEED (BIT(MPTCP_PUSH_PENDING) | \
3453 BIT(MPTCP_RETRANSMIT) | \
3454 BIT(MPTCP_FLUSH_JOIN_LIST))
3455
3456 /* processes deferred events and flush wmem */
3457 static void mptcp_release_cb(struct sock *sk)
3458 __must_hold(&sk->sk_lock.slock)
3459 {
3460 struct mptcp_sock *msk = mptcp_sk(sk);
3461
3462 for (;;) {
3463 unsigned long flags = (msk->cb_flags & MPTCP_FLAGS_PROCESS_CTX_NEED);
3464 struct list_head join_list;
3465
3466 if (!flags)
3467 break;
3468
3469 INIT_LIST_HEAD(&join_list);
3470 list_splice_init(&msk->join_list, &join_list);
3471
3472 /* the following actions acquire the subflow socket lock
3473 *
3474 * 1) can't be invoked in atomic scope
3475 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX
3476 * datapath acquires the msk socket spinlock while helding
3477 * the subflow socket lock
3478 */
3479 msk->cb_flags &= ~flags;
3480 spin_unlock_bh(&sk->sk_lock.slock);
3481
3482 if (flags & BIT(MPTCP_FLUSH_JOIN_LIST))
3483 __mptcp_flush_join_list(sk, &join_list);
3484 if (flags & BIT(MPTCP_PUSH_PENDING))
3485 __mptcp_push_pending(sk, 0);
3486 if (flags & BIT(MPTCP_RETRANSMIT))
3487 __mptcp_retrans(sk);
3488
3489 cond_resched();
3490 spin_lock_bh(&sk->sk_lock.slock);
3491 }
3492
3493 if (__test_and_clear_bit(MPTCP_CLEAN_UNA, &msk->cb_flags))
3494 __mptcp_clean_una_wakeup(sk);
3495 if (unlikely(msk->cb_flags)) {
3496 /* be sure to sync the msk state before taking actions
3497 * depending on sk_state (MPTCP_ERROR_REPORT)
3498 * On sk release avoid actions depending on the first subflow
3499 */
3500 if (__test_and_clear_bit(MPTCP_SYNC_STATE, &msk->cb_flags) && msk->first)
3501 __mptcp_sync_state(sk, msk->pending_state);
3502 if (__test_and_clear_bit(MPTCP_ERROR_REPORT, &msk->cb_flags))
3503 __mptcp_error_report(sk);
3504 if (__test_and_clear_bit(MPTCP_SYNC_SNDBUF, &msk->cb_flags))
3505 __mptcp_sync_sndbuf(sk);
3506 }
3507
3508 __mptcp_update_rmem(sk);
3509 }
3510
3511 /* MP_JOIN client subflow must wait for 4th ack before sending any data:
3512 * TCP can't schedule delack timer before the subflow is fully established.
3513 * MPTCP uses the delack timer to do 3rd ack retransmissions
3514 */
3515 static void schedule_3rdack_retransmission(struct sock *ssk)
3516 {
3517 struct inet_connection_sock *icsk = inet_csk(ssk);
3518 struct tcp_sock *tp = tcp_sk(ssk);
3519 unsigned long timeout;
3520
3521 if (READ_ONCE(mptcp_subflow_ctx(ssk)->fully_established))
3522 return;
3523
3524 /* reschedule with a timeout above RTT, as we must look only for drop */
3525 if (tp->srtt_us)
3526 timeout = usecs_to_jiffies(tp->srtt_us >> (3 - 1));
3527 else
3528 timeout = TCP_TIMEOUT_INIT;
3529 timeout += jiffies;
3530
3531 WARN_ON_ONCE(icsk->icsk_ack.pending & ICSK_ACK_TIMER);
3532 smp_store_release(&icsk->icsk_ack.pending,
3533 icsk->icsk_ack.pending | ICSK_ACK_SCHED | ICSK_ACK_TIMER);
3534 icsk->icsk_ack.timeout = timeout;
3535 sk_reset_timer(ssk, &icsk->icsk_delack_timer, timeout);
3536 }
3537
3538 void mptcp_subflow_process_delegated(struct sock *ssk, long status)
3539 {
3540 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3541 struct sock *sk = subflow->conn;
3542
3543 if (status & BIT(MPTCP_DELEGATE_SEND)) {
3544 mptcp_data_lock(sk);
3545 if (!sock_owned_by_user(sk))
3546 __mptcp_subflow_push_pending(sk, ssk, true);
3547 else
3548 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3549 mptcp_data_unlock(sk);
3550 }
3551 if (status & BIT(MPTCP_DELEGATE_SNDBUF)) {
3552 mptcp_data_lock(sk);
3553 if (!sock_owned_by_user(sk))
3554 __mptcp_sync_sndbuf(sk);
3555 else
3556 __set_bit(MPTCP_SYNC_SNDBUF, &mptcp_sk(sk)->cb_flags);
3557 mptcp_data_unlock(sk);
3558 }
3559 if (status & BIT(MPTCP_DELEGATE_ACK))
3560 schedule_3rdack_retransmission(ssk);
3561 }
3562
3563 static int mptcp_hash(struct sock *sk)
3564 {
3565 /* should never be called,
3566 * we hash the TCP subflows not the MPTCP socket
3567 */
3568 WARN_ON_ONCE(1);
3569 return 0;
3570 }
3571
3572 static void mptcp_unhash(struct sock *sk)
3573 {
3574 /* called from sk_common_release(), but nothing to do here */
3575 }
3576
3577 static int mptcp_get_port(struct sock *sk, unsigned short snum)
3578 {
3579 struct mptcp_sock *msk = mptcp_sk(sk);
3580
3581 pr_debug("msk=%p, ssk=%p\n", msk, msk->first);
3582 if (WARN_ON_ONCE(!msk->first))
3583 return -EINVAL;
3584
3585 return inet_csk_get_port(msk->first, snum);
3586 }
3587
3588 void mptcp_finish_connect(struct sock *ssk)
3589 {
3590 struct mptcp_subflow_context *subflow;
3591 struct mptcp_sock *msk;
3592 struct sock *sk;
3593
3594 subflow = mptcp_subflow_ctx(ssk);
3595 sk = subflow->conn;
3596 msk = mptcp_sk(sk);
3597
3598 pr_debug("msk=%p, token=%u\n", sk, subflow->token);
3599
3600 subflow->map_seq = subflow->iasn;
3601 subflow->map_subflow_seq = 1;
3602
3603 /* the socket is not connected yet, no msk/subflow ops can access/race
3604 * accessing the field below
3605 */
3606 WRITE_ONCE(msk->local_key, subflow->local_key);
3607
3608 mptcp_pm_new_connection(msk, ssk, 0);
3609 }
3610
3611 void mptcp_sock_graft(struct sock *sk, struct socket *parent)
3612 {
3613 write_lock_bh(&sk->sk_callback_lock);
3614 rcu_assign_pointer(sk->sk_wq, &parent->wq);
3615 sk_set_socket(sk, parent);
3616 sk->sk_uid = SOCK_INODE(parent)->i_uid;
3617 write_unlock_bh(&sk->sk_callback_lock);
3618 }
3619
3620 bool mptcp_finish_join(struct sock *ssk)
3621 {
3622 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3623 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
3624 struct sock *parent = (void *)msk;
3625 bool ret = true;
3626
3627 pr_debug("msk=%p, subflow=%p\n", msk, subflow);
3628
3629 /* mptcp socket already closing? */
3630 if (!mptcp_is_fully_established(parent)) {
3631 subflow->reset_reason = MPTCP_RST_EMPTCP;
3632 return false;
3633 }
3634
3635 /* active subflow, already present inside the conn_list */
3636 if (!list_empty(&subflow->node)) {
3637 mptcp_subflow_joined(msk, ssk);
3638 mptcp_propagate_sndbuf(parent, ssk);
3639 return true;
3640 }
3641
3642 if (!mptcp_pm_allow_new_subflow(msk))
3643 goto err_prohibited;
3644
3645 /* If we can't acquire msk socket lock here, let the release callback
3646 * handle it
3647 */
3648 mptcp_data_lock(parent);
3649 if (!sock_owned_by_user(parent)) {
3650 ret = __mptcp_finish_join(msk, ssk);
3651 if (ret) {
3652 sock_hold(ssk);
3653 list_add_tail(&subflow->node, &msk->conn_list);
3654 }
3655 } else {
3656 sock_hold(ssk);
3657 list_add_tail(&subflow->node, &msk->join_list);
3658 __set_bit(MPTCP_FLUSH_JOIN_LIST, &msk->cb_flags);
3659 }
3660 mptcp_data_unlock(parent);
3661
3662 if (!ret) {
3663 err_prohibited:
3664 subflow->reset_reason = MPTCP_RST_EPROHIBIT;
3665 return false;
3666 }
3667
3668 return true;
3669 }
3670
3671 static void mptcp_shutdown(struct sock *sk, int how)
3672 {
3673 pr_debug("sk=%p, how=%d\n", sk, how);
3674
3675 if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk))
3676 __mptcp_wr_shutdown(sk);
3677 }
3678
3679 static int mptcp_forward_alloc_get(const struct sock *sk)
3680 {
3681 return READ_ONCE(sk->sk_forward_alloc) +
3682 READ_ONCE(mptcp_sk(sk)->rmem_fwd_alloc);
3683 }
3684
3685 static int mptcp_ioctl_outq(const struct mptcp_sock *msk, u64 v)
3686 {
3687 const struct sock *sk = (void *)msk;
3688 u64 delta;
3689
3690 if (sk->sk_state == TCP_LISTEN)
3691 return -EINVAL;
3692
3693 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
3694 return 0;
3695
3696 delta = msk->write_seq - v;
3697 if (__mptcp_check_fallback(msk) && msk->first) {
3698 struct tcp_sock *tp = tcp_sk(msk->first);
3699
3700 /* the first subflow is disconnected after close - see
3701 * __mptcp_close_ssk(). tcp_disconnect() moves the write_seq
3702 * so ignore that status, too.
3703 */
3704 if (!((1 << msk->first->sk_state) &
3705 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_CLOSE)))
3706 delta += READ_ONCE(tp->write_seq) - tp->snd_una;
3707 }
3708 if (delta > INT_MAX)
3709 delta = INT_MAX;
3710
3711 return (int)delta;
3712 }
3713
3714 static int mptcp_ioctl(struct sock *sk, int cmd, int *karg)
3715 {
3716 struct mptcp_sock *msk = mptcp_sk(sk);
3717 bool slow;
3718
3719 switch (cmd) {
3720 case SIOCINQ:
3721 if (sk->sk_state == TCP_LISTEN)
3722 return -EINVAL;
3723
3724 lock_sock(sk);
3725 __mptcp_move_skbs(msk);
3726 *karg = mptcp_inq_hint(sk);
3727 release_sock(sk);
3728 break;
3729 case SIOCOUTQ:
3730 slow = lock_sock_fast(sk);
3731 *karg = mptcp_ioctl_outq(msk, READ_ONCE(msk->snd_una));
3732 unlock_sock_fast(sk, slow);
3733 break;
3734 case SIOCOUTQNSD:
3735 slow = lock_sock_fast(sk);
3736 *karg = mptcp_ioctl_outq(msk, msk->snd_nxt);
3737 unlock_sock_fast(sk, slow);
3738 break;
3739 default:
3740 return -ENOIOCTLCMD;
3741 }
3742
3743 return 0;
3744 }
3745
3746 static int mptcp_connect(struct sock *sk, struct sockaddr *uaddr, int addr_len)
3747 {
3748 struct mptcp_subflow_context *subflow;
3749 struct mptcp_sock *msk = mptcp_sk(sk);
3750 int err = -EINVAL;
3751 struct sock *ssk;
3752
3753 ssk = __mptcp_nmpc_sk(msk);
3754 if (IS_ERR(ssk))
3755 return PTR_ERR(ssk);
3756
3757 mptcp_set_state(sk, TCP_SYN_SENT);
3758 subflow = mptcp_subflow_ctx(ssk);
3759 #ifdef CONFIG_TCP_MD5SIG
3760 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of
3761 * TCP option space.
3762 */
3763 if (rcu_access_pointer(tcp_sk(ssk)->md5sig_info))
3764 mptcp_subflow_early_fallback(msk, subflow);
3765 #endif
3766 if (subflow->request_mptcp) {
3767 if (mptcp_active_should_disable(sk)) {
3768 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_MPCAPABLEACTIVEDISABLED);
3769 mptcp_subflow_early_fallback(msk, subflow);
3770 } else if (mptcp_token_new_connect(ssk) < 0) {
3771 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_TOKENFALLBACKINIT);
3772 mptcp_subflow_early_fallback(msk, subflow);
3773 }
3774 }
3775
3776 WRITE_ONCE(msk->write_seq, subflow->idsn);
3777 WRITE_ONCE(msk->snd_nxt, subflow->idsn);
3778 WRITE_ONCE(msk->snd_una, subflow->idsn);
3779 if (likely(!__mptcp_check_fallback(msk)))
3780 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEACTIVE);
3781
3782 /* if reaching here via the fastopen/sendmsg path, the caller already
3783 * acquired the subflow socket lock, too.
3784 */
3785 if (!msk->fastopening)
3786 lock_sock(ssk);
3787
3788 /* the following mirrors closely a very small chunk of code from
3789 * __inet_stream_connect()
3790 */
3791 if (ssk->sk_state != TCP_CLOSE)
3792 goto out;
3793
3794 if (BPF_CGROUP_PRE_CONNECT_ENABLED(ssk)) {
3795 err = ssk->sk_prot->pre_connect(ssk, uaddr, addr_len);
3796 if (err)
3797 goto out;
3798 }
3799
3800 err = ssk->sk_prot->connect(ssk, uaddr, addr_len);
3801 if (err < 0)
3802 goto out;
3803
3804 inet_assign_bit(DEFER_CONNECT, sk, inet_test_bit(DEFER_CONNECT, ssk));
3805
3806 out:
3807 if (!msk->fastopening)
3808 release_sock(ssk);
3809
3810 /* on successful connect, the msk state will be moved to established by
3811 * subflow_finish_connect()
3812 */
3813 if (unlikely(err)) {
3814 /* avoid leaving a dangling token in an unconnected socket */
3815 mptcp_token_destroy(msk);
3816 mptcp_set_state(sk, TCP_CLOSE);
3817 return err;
3818 }
3819
3820 mptcp_copy_inaddrs(sk, ssk);
3821 return 0;
3822 }
3823
3824 static struct proto mptcp_prot = {
3825 .name = "MPTCP",
3826 .owner = THIS_MODULE,
3827 .init = mptcp_init_sock,
3828 .connect = mptcp_connect,
3829 .disconnect = mptcp_disconnect,
3830 .close = mptcp_close,
3831 .setsockopt = mptcp_setsockopt,
3832 .getsockopt = mptcp_getsockopt,
3833 .shutdown = mptcp_shutdown,
3834 .destroy = mptcp_destroy,
3835 .sendmsg = mptcp_sendmsg,
3836 .ioctl = mptcp_ioctl,
3837 .recvmsg = mptcp_recvmsg,
3838 .release_cb = mptcp_release_cb,
3839 .hash = mptcp_hash,
3840 .unhash = mptcp_unhash,
3841 .get_port = mptcp_get_port,
3842 .forward_alloc_get = mptcp_forward_alloc_get,
3843 .stream_memory_free = mptcp_stream_memory_free,
3844 .sockets_allocated = &mptcp_sockets_allocated,
3845
3846 .memory_allocated = &tcp_memory_allocated,
3847 .per_cpu_fw_alloc = &tcp_memory_per_cpu_fw_alloc,
3848
3849 .memory_pressure = &tcp_memory_pressure,
3850 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
3851 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem),
3852 .sysctl_mem = sysctl_tcp_mem,
3853 .obj_size = sizeof(struct mptcp_sock),
3854 .slab_flags = SLAB_TYPESAFE_BY_RCU,
3855 .no_autobind = true,
3856 };
3857
3858 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
3859 {
3860 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3861 struct sock *ssk, *sk = sock->sk;
3862 int err = -EINVAL;
3863
3864 lock_sock(sk);
3865 ssk = __mptcp_nmpc_sk(msk);
3866 if (IS_ERR(ssk)) {
3867 err = PTR_ERR(ssk);
3868 goto unlock;
3869 }
3870
3871 if (sk->sk_family == AF_INET)
3872 err = inet_bind_sk(ssk, uaddr, addr_len);
3873 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3874 else if (sk->sk_family == AF_INET6)
3875 err = inet6_bind_sk(ssk, uaddr, addr_len);
3876 #endif
3877 if (!err)
3878 mptcp_copy_inaddrs(sk, ssk);
3879
3880 unlock:
3881 release_sock(sk);
3882 return err;
3883 }
3884
3885 static int mptcp_listen(struct socket *sock, int backlog)
3886 {
3887 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3888 struct sock *sk = sock->sk;
3889 struct sock *ssk;
3890 int err;
3891
3892 pr_debug("msk=%p\n", msk);
3893
3894 lock_sock(sk);
3895
3896 err = -EINVAL;
3897 if (sock->state != SS_UNCONNECTED || sock->type != SOCK_STREAM)
3898 goto unlock;
3899
3900 ssk = __mptcp_nmpc_sk(msk);
3901 if (IS_ERR(ssk)) {
3902 err = PTR_ERR(ssk);
3903 goto unlock;
3904 }
3905
3906 mptcp_set_state(sk, TCP_LISTEN);
3907 sock_set_flag(sk, SOCK_RCU_FREE);
3908
3909 lock_sock(ssk);
3910 err = __inet_listen_sk(ssk, backlog);
3911 release_sock(ssk);
3912 mptcp_set_state(sk, inet_sk_state_load(ssk));
3913
3914 if (!err) {
3915 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
3916 mptcp_copy_inaddrs(sk, ssk);
3917 mptcp_event_pm_listener(ssk, MPTCP_EVENT_LISTENER_CREATED);
3918 }
3919
3920 unlock:
3921 release_sock(sk);
3922 return err;
3923 }
3924
3925 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
3926 struct proto_accept_arg *arg)
3927 {
3928 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3929 struct sock *ssk, *newsk;
3930
3931 pr_debug("msk=%p\n", msk);
3932
3933 /* Buggy applications can call accept on socket states other then LISTEN
3934 * but no need to allocate the first subflow just to error out.
3935 */
3936 ssk = READ_ONCE(msk->first);
3937 if (!ssk)
3938 return -EINVAL;
3939
3940 pr_debug("ssk=%p, listener=%p\n", ssk, mptcp_subflow_ctx(ssk));
3941 newsk = inet_csk_accept(ssk, arg);
3942 if (!newsk)
3943 return arg->err;
3944
3945 pr_debug("newsk=%p, subflow is mptcp=%d\n", newsk, sk_is_mptcp(newsk));
3946 if (sk_is_mptcp(newsk)) {
3947 struct mptcp_subflow_context *subflow;
3948 struct sock *new_mptcp_sock;
3949
3950 subflow = mptcp_subflow_ctx(newsk);
3951 new_mptcp_sock = subflow->conn;
3952
3953 /* is_mptcp should be false if subflow->conn is missing, see
3954 * subflow_syn_recv_sock()
3955 */
3956 if (WARN_ON_ONCE(!new_mptcp_sock)) {
3957 tcp_sk(newsk)->is_mptcp = 0;
3958 goto tcpfallback;
3959 }
3960
3961 newsk = new_mptcp_sock;
3962 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
3963
3964 newsk->sk_kern_sock = arg->kern;
3965 lock_sock(newsk);
3966 __inet_accept(sock, newsock, newsk);
3967
3968 set_bit(SOCK_CUSTOM_SOCKOPT, &newsock->flags);
3969 msk = mptcp_sk(newsk);
3970 msk->in_accept_queue = 0;
3971
3972 /* set ssk->sk_socket of accept()ed flows to mptcp socket.
3973 * This is needed so NOSPACE flag can be set from tcp stack.
3974 */
3975 mptcp_for_each_subflow(msk, subflow) {
3976 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3977
3978 if (!ssk->sk_socket)
3979 mptcp_sock_graft(ssk, newsock);
3980 }
3981
3982 /* Do late cleanup for the first subflow as necessary. Also
3983 * deal with bad peers not doing a complete shutdown.
3984 */
3985 if (unlikely(inet_sk_state_load(msk->first) == TCP_CLOSE)) {
3986 __mptcp_close_ssk(newsk, msk->first,
3987 mptcp_subflow_ctx(msk->first), 0);
3988 if (unlikely(list_is_singular(&msk->conn_list)))
3989 mptcp_set_state(newsk, TCP_CLOSE);
3990 }
3991 } else {
3992 tcpfallback:
3993 newsk->sk_kern_sock = arg->kern;
3994 lock_sock(newsk);
3995 __inet_accept(sock, newsock, newsk);
3996 /* we are being invoked after accepting a non-mp-capable
3997 * flow: sk is a tcp_sk, not an mptcp one.
3998 *
3999 * Hand the socket over to tcp so all further socket ops
4000 * bypass mptcp.
4001 */
4002 WRITE_ONCE(newsock->sk->sk_socket->ops,
4003 mptcp_fallback_tcp_ops(newsock->sk));
4004 }
4005 release_sock(newsk);
4006
4007 return 0;
4008 }
4009
4010 static __poll_t mptcp_check_writeable(struct mptcp_sock *msk)
4011 {
4012 struct sock *sk = (struct sock *)msk;
4013
4014 if (__mptcp_stream_is_writeable(sk, 1))
4015 return EPOLLOUT | EPOLLWRNORM;
4016
4017 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
4018 smp_mb__after_atomic(); /* NOSPACE is changed by mptcp_write_space() */
4019 if (__mptcp_stream_is_writeable(sk, 1))
4020 return EPOLLOUT | EPOLLWRNORM;
4021
4022 return 0;
4023 }
4024
4025 static __poll_t mptcp_poll(struct file *file, struct socket *sock,
4026 struct poll_table_struct *wait)
4027 {
4028 struct sock *sk = sock->sk;
4029 struct mptcp_sock *msk;
4030 __poll_t mask = 0;
4031 u8 shutdown;
4032 int state;
4033
4034 msk = mptcp_sk(sk);
4035 sock_poll_wait(file, sock, wait);
4036
4037 state = inet_sk_state_load(sk);
4038 pr_debug("msk=%p state=%d flags=%lx\n", msk, state, msk->flags);
4039 if (state == TCP_LISTEN) {
4040 struct sock *ssk = READ_ONCE(msk->first);
4041
4042 if (WARN_ON_ONCE(!ssk))
4043 return 0;
4044
4045 return inet_csk_listen_poll(ssk);
4046 }
4047
4048 shutdown = READ_ONCE(sk->sk_shutdown);
4049 if (shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
4050 mask |= EPOLLHUP;
4051 if (shutdown & RCV_SHUTDOWN)
4052 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
4053
4054 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
4055 mask |= mptcp_check_readable(sk);
4056 if (shutdown & SEND_SHUTDOWN)
4057 mask |= EPOLLOUT | EPOLLWRNORM;
4058 else
4059 mask |= mptcp_check_writeable(msk);
4060 } else if (state == TCP_SYN_SENT &&
4061 inet_test_bit(DEFER_CONNECT, sk)) {
4062 /* cf tcp_poll() note about TFO */
4063 mask |= EPOLLOUT | EPOLLWRNORM;
4064 }
4065
4066 /* This barrier is coupled with smp_wmb() in __mptcp_error_report() */
4067 smp_rmb();
4068 if (READ_ONCE(sk->sk_err))
4069 mask |= EPOLLERR;
4070
4071 return mask;
4072 }
4073
4074 static const struct proto_ops mptcp_stream_ops = {
4075 .family = PF_INET,
4076 .owner = THIS_MODULE,
4077 .release = inet_release,
4078 .bind = mptcp_bind,
4079 .connect = inet_stream_connect,
4080 .socketpair = sock_no_socketpair,
4081 .accept = mptcp_stream_accept,
4082 .getname = inet_getname,
4083 .poll = mptcp_poll,
4084 .ioctl = inet_ioctl,
4085 .gettstamp = sock_gettstamp,
4086 .listen = mptcp_listen,
4087 .shutdown = inet_shutdown,
4088 .setsockopt = sock_common_setsockopt,
4089 .getsockopt = sock_common_getsockopt,
4090 .sendmsg = inet_sendmsg,
4091 .recvmsg = inet_recvmsg,
4092 .mmap = sock_no_mmap,
4093 .set_rcvlowat = mptcp_set_rcvlowat,
4094 };
4095
4096 static struct inet_protosw mptcp_protosw = {
4097 .type = SOCK_STREAM,
4098 .protocol = IPPROTO_MPTCP,
4099 .prot = &mptcp_prot,
4100 .ops = &mptcp_stream_ops,
4101 .flags = INET_PROTOSW_ICSK,
4102 };
4103
4104 static int mptcp_napi_poll(struct napi_struct *napi, int budget)
4105 {
4106 struct mptcp_delegated_action *delegated;
4107 struct mptcp_subflow_context *subflow;
4108 int work_done = 0;
4109
4110 delegated = container_of(napi, struct mptcp_delegated_action, napi);
4111 while ((subflow = mptcp_subflow_delegated_next(delegated)) != NULL) {
4112 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
4113
4114 bh_lock_sock_nested(ssk);
4115 if (!sock_owned_by_user(ssk)) {
4116 mptcp_subflow_process_delegated(ssk, xchg(&subflow->delegated_status, 0));
4117 } else {
4118 /* tcp_release_cb_override already processed
4119 * the action or will do at next release_sock().
4120 * In both case must dequeue the subflow here - on the same
4121 * CPU that scheduled it.
4122 */
4123 smp_wmb();
4124 clear_bit(MPTCP_DELEGATE_SCHEDULED, &subflow->delegated_status);
4125 }
4126 bh_unlock_sock(ssk);
4127 sock_put(ssk);
4128
4129 if (++work_done == budget)
4130 return budget;
4131 }
4132
4133 /* always provide a 0 'work_done' argument, so that napi_complete_done
4134 * will not try accessing the NULL napi->dev ptr
4135 */
4136 napi_complete_done(napi, 0);
4137 return work_done;
4138 }
4139
4140 void __init mptcp_proto_init(void)
4141 {
4142 struct mptcp_delegated_action *delegated;
4143 int cpu;
4144
4145 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
4146
4147 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
4148 panic("Failed to allocate MPTCP pcpu counter\n");
4149
4150 init_dummy_netdev(&mptcp_napi_dev);
4151 for_each_possible_cpu(cpu) {
4152 delegated = per_cpu_ptr(&mptcp_delegated_actions, cpu);
4153 INIT_LIST_HEAD(&delegated->head);
4154 netif_napi_add_tx(&mptcp_napi_dev, &delegated->napi,
4155 mptcp_napi_poll);
4156 napi_enable(&delegated->napi);
4157 }
4158
4159 mptcp_subflow_init();
4160 mptcp_pm_init();
4161 mptcp_sched_init();
4162 mptcp_token_init();
4163
4164 if (proto_register(&mptcp_prot, 1) != 0)
4165 panic("Failed to register MPTCP proto.\n");
4166
4167 inet_register_protosw(&mptcp_protosw);
4168
4169 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
4170 }
4171
4172 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
4173 static const struct proto_ops mptcp_v6_stream_ops = {
4174 .family = PF_INET6,
4175 .owner = THIS_MODULE,
4176 .release = inet6_release,
4177 .bind = mptcp_bind,
4178 .connect = inet_stream_connect,
4179 .socketpair = sock_no_socketpair,
4180 .accept = mptcp_stream_accept,
4181 .getname = inet6_getname,
4182 .poll = mptcp_poll,
4183 .ioctl = inet6_ioctl,
4184 .gettstamp = sock_gettstamp,
4185 .listen = mptcp_listen,
4186 .shutdown = inet_shutdown,
4187 .setsockopt = sock_common_setsockopt,
4188 .getsockopt = sock_common_getsockopt,
4189 .sendmsg = inet6_sendmsg,
4190 .recvmsg = inet6_recvmsg,
4191 .mmap = sock_no_mmap,
4192 #ifdef CONFIG_COMPAT
4193 .compat_ioctl = inet6_compat_ioctl,
4194 #endif
4195 .set_rcvlowat = mptcp_set_rcvlowat,
4196 };
4197
4198 static struct proto mptcp_v6_prot;
4199
4200 static struct inet_protosw mptcp_v6_protosw = {
4201 .type = SOCK_STREAM,
4202 .protocol = IPPROTO_MPTCP,
4203 .prot = &mptcp_v6_prot,
4204 .ops = &mptcp_v6_stream_ops,
4205 .flags = INET_PROTOSW_ICSK,
4206 };
4207
4208 int __init mptcp_proto_v6_init(void)
4209 {
4210 int err;
4211
4212 mptcp_v6_prot = mptcp_prot;
4213 strscpy(mptcp_v6_prot.name, "MPTCPv6", sizeof(mptcp_v6_prot.name));
4214 mptcp_v6_prot.slab = NULL;
4215 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
4216 mptcp_v6_prot.ipv6_pinfo_offset = offsetof(struct mptcp6_sock, np);
4217
4218 err = proto_register(&mptcp_v6_prot, 1);
4219 if (err)
4220 return err;
4221
4222 err = inet6_register_protosw(&mptcp_v6_protosw);
4223 if (err)
4224 proto_unregister(&mptcp_v6_prot);
4225
4226 return err;
4227 }
4228 #endif
Linux Kernel