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