Loading...
1// SPDX-License-Identifier: GPL-2.0-or-later
2/* Client connection-specific management code.
3 *
4 * Copyright (C) 2016, 2020 Red Hat, Inc. All Rights Reserved.
5 * Written by David Howells (dhowells@redhat.com)
6 *
7 * Client connections need to be cached for a little while after they've made a
8 * call so as to handle retransmitted DATA packets in case the server didn't
9 * receive the final ACK or terminating ABORT we sent it.
10 *
11 * There are flags of relevance to the cache:
12 *
13 * (2) DONT_REUSE - The connection should be discarded as soon as possible and
14 * should not be reused. This is set when an exclusive connection is used
15 * or a call ID counter overflows.
16 *
17 * The caching state may only be changed if the cache lock is held.
18 *
19 * There are two idle client connection expiry durations. If the total number
20 * of connections is below the reap threshold, we use the normal duration; if
21 * it's above, we use the fast duration.
22 */
23
24#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
25
26#include <linux/slab.h>
27#include <linux/idr.h>
28#include <linux/timer.h>
29#include <linux/sched/signal.h>
30
31#include "ar-internal.h"
32
33__read_mostly unsigned int rxrpc_reap_client_connections = 900;
34__read_mostly unsigned long rxrpc_conn_idle_client_expiry = 2 * 60 * HZ;
35__read_mostly unsigned long rxrpc_conn_idle_client_fast_expiry = 2 * HZ;
36
37static void rxrpc_activate_bundle(struct rxrpc_bundle *bundle)
38{
39 atomic_inc(&bundle->active);
40}
41
42/*
43 * Release a connection ID for a client connection.
44 */
45static void rxrpc_put_client_connection_id(struct rxrpc_local *local,
46 struct rxrpc_connection *conn)
47{
48 idr_remove(&local->conn_ids, conn->proto.cid >> RXRPC_CIDSHIFT);
49}
50
51/*
52 * Destroy the client connection ID tree.
53 */
54static void rxrpc_destroy_client_conn_ids(struct rxrpc_local *local)
55{
56 struct rxrpc_connection *conn;
57 int id;
58
59 if (!idr_is_empty(&local->conn_ids)) {
60 idr_for_each_entry(&local->conn_ids, conn, id) {
61 pr_err("AF_RXRPC: Leaked client conn %p {%d}\n",
62 conn, refcount_read(&conn->ref));
63 }
64 BUG();
65 }
66
67 idr_destroy(&local->conn_ids);
68}
69
70/*
71 * Allocate a connection bundle.
72 */
73static struct rxrpc_bundle *rxrpc_alloc_bundle(struct rxrpc_call *call,
74 gfp_t gfp)
75{
76 struct rxrpc_bundle *bundle;
77
78 bundle = kzalloc(sizeof(*bundle), gfp);
79 if (bundle) {
80 bundle->local = call->local;
81 bundle->peer = rxrpc_get_peer(call->peer, rxrpc_peer_get_bundle);
82 bundle->key = key_get(call->key);
83 bundle->security = call->security;
84 bundle->exclusive = test_bit(RXRPC_CALL_EXCLUSIVE, &call->flags);
85 bundle->upgrade = test_bit(RXRPC_CALL_UPGRADE, &call->flags);
86 bundle->service_id = call->dest_srx.srx_service;
87 bundle->security_level = call->security_level;
88 refcount_set(&bundle->ref, 1);
89 atomic_set(&bundle->active, 1);
90 INIT_LIST_HEAD(&bundle->waiting_calls);
91 trace_rxrpc_bundle(bundle->debug_id, 1, rxrpc_bundle_new);
92 }
93 return bundle;
94}
95
96struct rxrpc_bundle *rxrpc_get_bundle(struct rxrpc_bundle *bundle,
97 enum rxrpc_bundle_trace why)
98{
99 int r;
100
101 __refcount_inc(&bundle->ref, &r);
102 trace_rxrpc_bundle(bundle->debug_id, r + 1, why);
103 return bundle;
104}
105
106static void rxrpc_free_bundle(struct rxrpc_bundle *bundle)
107{
108 trace_rxrpc_bundle(bundle->debug_id, 1, rxrpc_bundle_free);
109 rxrpc_put_peer(bundle->peer, rxrpc_peer_put_bundle);
110 key_put(bundle->key);
111 kfree(bundle);
112}
113
114void rxrpc_put_bundle(struct rxrpc_bundle *bundle, enum rxrpc_bundle_trace why)
115{
116 unsigned int id;
117 bool dead;
118 int r;
119
120 if (bundle) {
121 id = bundle->debug_id;
122 dead = __refcount_dec_and_test(&bundle->ref, &r);
123 trace_rxrpc_bundle(id, r - 1, why);
124 if (dead)
125 rxrpc_free_bundle(bundle);
126 }
127}
128
129/*
130 * Get rid of outstanding client connection preallocations when a local
131 * endpoint is destroyed.
132 */
133void rxrpc_purge_client_connections(struct rxrpc_local *local)
134{
135 rxrpc_destroy_client_conn_ids(local);
136}
137
138/*
139 * Allocate a client connection.
140 */
141static struct rxrpc_connection *
142rxrpc_alloc_client_connection(struct rxrpc_bundle *bundle)
143{
144 struct rxrpc_connection *conn;
145 struct rxrpc_local *local = bundle->local;
146 struct rxrpc_net *rxnet = local->rxnet;
147 int id;
148
149 _enter("");
150
151 conn = rxrpc_alloc_connection(rxnet, GFP_ATOMIC | __GFP_NOWARN);
152 if (!conn)
153 return ERR_PTR(-ENOMEM);
154
155 id = idr_alloc_cyclic(&local->conn_ids, conn, 1, 0x40000000,
156 GFP_ATOMIC | __GFP_NOWARN);
157 if (id < 0) {
158 kfree(conn);
159 return ERR_PTR(id);
160 }
161
162 refcount_set(&conn->ref, 1);
163 conn->proto.cid = id << RXRPC_CIDSHIFT;
164 conn->proto.epoch = local->rxnet->epoch;
165 conn->out_clientflag = RXRPC_CLIENT_INITIATED;
166 conn->bundle = rxrpc_get_bundle(bundle, rxrpc_bundle_get_client_conn);
167 conn->local = rxrpc_get_local(bundle->local, rxrpc_local_get_client_conn);
168 conn->peer = rxrpc_get_peer(bundle->peer, rxrpc_peer_get_client_conn);
169 conn->key = key_get(bundle->key);
170 conn->security = bundle->security;
171 conn->exclusive = bundle->exclusive;
172 conn->upgrade = bundle->upgrade;
173 conn->orig_service_id = bundle->service_id;
174 conn->security_level = bundle->security_level;
175 conn->state = RXRPC_CONN_CLIENT_UNSECURED;
176 conn->service_id = conn->orig_service_id;
177
178 if (conn->security == &rxrpc_no_security)
179 conn->state = RXRPC_CONN_CLIENT;
180
181 atomic_inc(&rxnet->nr_conns);
182 write_lock(&rxnet->conn_lock);
183 list_add_tail(&conn->proc_link, &rxnet->conn_proc_list);
184 write_unlock(&rxnet->conn_lock);
185
186 rxrpc_see_connection(conn, rxrpc_conn_new_client);
187
188 atomic_inc(&rxnet->nr_client_conns);
189 trace_rxrpc_client(conn, -1, rxrpc_client_alloc);
190 return conn;
191}
192
193/*
194 * Determine if a connection may be reused.
195 */
196static bool rxrpc_may_reuse_conn(struct rxrpc_connection *conn)
197{
198 struct rxrpc_net *rxnet;
199 int id_cursor, id, distance, limit;
200
201 if (!conn)
202 goto dont_reuse;
203
204 rxnet = conn->rxnet;
205 if (test_bit(RXRPC_CONN_DONT_REUSE, &conn->flags))
206 goto dont_reuse;
207
208 if ((conn->state != RXRPC_CONN_CLIENT_UNSECURED &&
209 conn->state != RXRPC_CONN_CLIENT) ||
210 conn->proto.epoch != rxnet->epoch)
211 goto mark_dont_reuse;
212
213 /* The IDR tree gets very expensive on memory if the connection IDs are
214 * widely scattered throughout the number space, so we shall want to
215 * kill off connections that, say, have an ID more than about four
216 * times the maximum number of client conns away from the current
217 * allocation point to try and keep the IDs concentrated.
218 */
219 id_cursor = idr_get_cursor(&conn->local->conn_ids);
220 id = conn->proto.cid >> RXRPC_CIDSHIFT;
221 distance = id - id_cursor;
222 if (distance < 0)
223 distance = -distance;
224 limit = max_t(unsigned long, atomic_read(&rxnet->nr_conns) * 4, 1024);
225 if (distance > limit)
226 goto mark_dont_reuse;
227
228 return true;
229
230mark_dont_reuse:
231 set_bit(RXRPC_CONN_DONT_REUSE, &conn->flags);
232dont_reuse:
233 return false;
234}
235
236/*
237 * Look up the conn bundle that matches the connection parameters, adding it if
238 * it doesn't yet exist.
239 */
240int rxrpc_look_up_bundle(struct rxrpc_call *call, gfp_t gfp)
241{
242 static atomic_t rxrpc_bundle_id;
243 struct rxrpc_bundle *bundle, *candidate;
244 struct rxrpc_local *local = call->local;
245 struct rb_node *p, **pp, *parent;
246 long diff;
247 bool upgrade = test_bit(RXRPC_CALL_UPGRADE, &call->flags);
248
249 _enter("{%px,%x,%u,%u}",
250 call->peer, key_serial(call->key), call->security_level,
251 upgrade);
252
253 if (test_bit(RXRPC_CALL_EXCLUSIVE, &call->flags)) {
254 call->bundle = rxrpc_alloc_bundle(call, gfp);
255 return call->bundle ? 0 : -ENOMEM;
256 }
257
258 /* First, see if the bundle is already there. */
259 _debug("search 1");
260 spin_lock(&local->client_bundles_lock);
261 p = local->client_bundles.rb_node;
262 while (p) {
263 bundle = rb_entry(p, struct rxrpc_bundle, local_node);
264
265#define cmp(X, Y) ((long)(X) - (long)(Y))
266 diff = (cmp(bundle->peer, call->peer) ?:
267 cmp(bundle->key, call->key) ?:
268 cmp(bundle->security_level, call->security_level) ?:
269 cmp(bundle->upgrade, upgrade));
270#undef cmp
271 if (diff < 0)
272 p = p->rb_left;
273 else if (diff > 0)
274 p = p->rb_right;
275 else
276 goto found_bundle;
277 }
278 spin_unlock(&local->client_bundles_lock);
279 _debug("not found");
280
281 /* It wasn't. We need to add one. */
282 candidate = rxrpc_alloc_bundle(call, gfp);
283 if (!candidate)
284 return -ENOMEM;
285
286 _debug("search 2");
287 spin_lock(&local->client_bundles_lock);
288 pp = &local->client_bundles.rb_node;
289 parent = NULL;
290 while (*pp) {
291 parent = *pp;
292 bundle = rb_entry(parent, struct rxrpc_bundle, local_node);
293
294#define cmp(X, Y) ((long)(X) - (long)(Y))
295 diff = (cmp(bundle->peer, call->peer) ?:
296 cmp(bundle->key, call->key) ?:
297 cmp(bundle->security_level, call->security_level) ?:
298 cmp(bundle->upgrade, upgrade));
299#undef cmp
300 if (diff < 0)
301 pp = &(*pp)->rb_left;
302 else if (diff > 0)
303 pp = &(*pp)->rb_right;
304 else
305 goto found_bundle_free;
306 }
307
308 _debug("new bundle");
309 candidate->debug_id = atomic_inc_return(&rxrpc_bundle_id);
310 rb_link_node(&candidate->local_node, parent, pp);
311 rb_insert_color(&candidate->local_node, &local->client_bundles);
312 call->bundle = rxrpc_get_bundle(candidate, rxrpc_bundle_get_client_call);
313 spin_unlock(&local->client_bundles_lock);
314 _leave(" = B=%u [new]", call->bundle->debug_id);
315 return 0;
316
317found_bundle_free:
318 rxrpc_free_bundle(candidate);
319found_bundle:
320 call->bundle = rxrpc_get_bundle(bundle, rxrpc_bundle_get_client_call);
321 rxrpc_activate_bundle(bundle);
322 spin_unlock(&local->client_bundles_lock);
323 _leave(" = B=%u [found]", call->bundle->debug_id);
324 return 0;
325}
326
327/*
328 * Allocate a new connection and add it into a bundle.
329 */
330static bool rxrpc_add_conn_to_bundle(struct rxrpc_bundle *bundle,
331 unsigned int slot)
332{
333 struct rxrpc_connection *conn, *old;
334 unsigned int shift = slot * RXRPC_MAXCALLS;
335 unsigned int i;
336
337 old = bundle->conns[slot];
338 if (old) {
339 bundle->conns[slot] = NULL;
340 trace_rxrpc_client(old, -1, rxrpc_client_replace);
341 rxrpc_put_connection(old, rxrpc_conn_put_noreuse);
342 }
343
344 conn = rxrpc_alloc_client_connection(bundle);
345 if (IS_ERR(conn)) {
346 bundle->alloc_error = PTR_ERR(conn);
347 return false;
348 }
349
350 rxrpc_activate_bundle(bundle);
351 conn->bundle_shift = shift;
352 bundle->conns[slot] = conn;
353 for (i = 0; i < RXRPC_MAXCALLS; i++)
354 set_bit(shift + i, &bundle->avail_chans);
355 return true;
356}
357
358/*
359 * Add a connection to a bundle if there are no usable connections or we have
360 * connections waiting for extra capacity.
361 */
362static bool rxrpc_bundle_has_space(struct rxrpc_bundle *bundle)
363{
364 int slot = -1, i, usable;
365
366 _enter("");
367
368 bundle->alloc_error = 0;
369
370 /* See if there are any usable connections. */
371 usable = 0;
372 for (i = 0; i < ARRAY_SIZE(bundle->conns); i++) {
373 if (rxrpc_may_reuse_conn(bundle->conns[i]))
374 usable++;
375 else if (slot == -1)
376 slot = i;
377 }
378
379 if (!usable && bundle->upgrade)
380 bundle->try_upgrade = true;
381
382 if (!usable)
383 goto alloc_conn;
384
385 if (!bundle->avail_chans &&
386 !bundle->try_upgrade &&
387 usable < ARRAY_SIZE(bundle->conns))
388 goto alloc_conn;
389
390 _leave("");
391 return usable;
392
393alloc_conn:
394 return slot >= 0 ? rxrpc_add_conn_to_bundle(bundle, slot) : false;
395}
396
397/*
398 * Assign a channel to the call at the front of the queue and wake the call up.
399 * We don't increment the callNumber counter until this number has been exposed
400 * to the world.
401 */
402static void rxrpc_activate_one_channel(struct rxrpc_connection *conn,
403 unsigned int channel)
404{
405 struct rxrpc_channel *chan = &conn->channels[channel];
406 struct rxrpc_bundle *bundle = conn->bundle;
407 struct rxrpc_call *call = list_entry(bundle->waiting_calls.next,
408 struct rxrpc_call, wait_link);
409 u32 call_id = chan->call_counter + 1;
410
411 _enter("C=%x,%u", conn->debug_id, channel);
412
413 list_del_init(&call->wait_link);
414
415 trace_rxrpc_client(conn, channel, rxrpc_client_chan_activate);
416
417 /* Cancel the final ACK on the previous call if it hasn't been sent yet
418 * as the DATA packet will implicitly ACK it.
419 */
420 clear_bit(RXRPC_CONN_FINAL_ACK_0 + channel, &conn->flags);
421 clear_bit(conn->bundle_shift + channel, &bundle->avail_chans);
422
423 rxrpc_see_call(call, rxrpc_call_see_activate_client);
424 call->conn = rxrpc_get_connection(conn, rxrpc_conn_get_activate_call);
425 call->cid = conn->proto.cid | channel;
426 call->call_id = call_id;
427 call->dest_srx.srx_service = conn->service_id;
428 call->cong_ssthresh = call->peer->cong_ssthresh;
429 if (call->cong_cwnd >= call->cong_ssthresh)
430 call->cong_mode = RXRPC_CALL_CONGEST_AVOIDANCE;
431 else
432 call->cong_mode = RXRPC_CALL_SLOW_START;
433
434 chan->call_id = call_id;
435 chan->call_debug_id = call->debug_id;
436 chan->call = call;
437
438 rxrpc_see_call(call, rxrpc_call_see_connected);
439 trace_rxrpc_connect_call(call);
440 call->tx_last_sent = ktime_get_real();
441 rxrpc_start_call_timer(call);
442 rxrpc_set_call_state(call, RXRPC_CALL_CLIENT_SEND_REQUEST);
443 wake_up(&call->waitq);
444}
445
446/*
447 * Remove a connection from the idle list if it's on it.
448 */
449static void rxrpc_unidle_conn(struct rxrpc_connection *conn)
450{
451 if (!list_empty(&conn->cache_link)) {
452 list_del_init(&conn->cache_link);
453 rxrpc_put_connection(conn, rxrpc_conn_put_unidle);
454 }
455}
456
457/*
458 * Assign channels and callNumbers to waiting calls.
459 */
460static void rxrpc_activate_channels(struct rxrpc_bundle *bundle)
461{
462 struct rxrpc_connection *conn;
463 unsigned long avail, mask;
464 unsigned int channel, slot;
465
466 trace_rxrpc_client(NULL, -1, rxrpc_client_activate_chans);
467
468 if (bundle->try_upgrade)
469 mask = 1;
470 else
471 mask = ULONG_MAX;
472
473 while (!list_empty(&bundle->waiting_calls)) {
474 avail = bundle->avail_chans & mask;
475 if (!avail)
476 break;
477 channel = __ffs(avail);
478 clear_bit(channel, &bundle->avail_chans);
479
480 slot = channel / RXRPC_MAXCALLS;
481 conn = bundle->conns[slot];
482 if (!conn)
483 break;
484
485 if (bundle->try_upgrade)
486 set_bit(RXRPC_CONN_PROBING_FOR_UPGRADE, &conn->flags);
487 rxrpc_unidle_conn(conn);
488
489 channel &= (RXRPC_MAXCALLS - 1);
490 conn->act_chans |= 1 << channel;
491 rxrpc_activate_one_channel(conn, channel);
492 }
493}
494
495/*
496 * Connect waiting channels (called from the I/O thread).
497 */
498void rxrpc_connect_client_calls(struct rxrpc_local *local)
499{
500 struct rxrpc_call *call;
501
502 while ((call = list_first_entry_or_null(&local->new_client_calls,
503 struct rxrpc_call, wait_link))
504 ) {
505 struct rxrpc_bundle *bundle = call->bundle;
506
507 spin_lock(&local->client_call_lock);
508 list_move_tail(&call->wait_link, &bundle->waiting_calls);
509 spin_unlock(&local->client_call_lock);
510
511 if (rxrpc_bundle_has_space(bundle))
512 rxrpc_activate_channels(bundle);
513 }
514}
515
516/*
517 * Note that a call, and thus a connection, is about to be exposed to the
518 * world.
519 */
520void rxrpc_expose_client_call(struct rxrpc_call *call)
521{
522 unsigned int channel = call->cid & RXRPC_CHANNELMASK;
523 struct rxrpc_connection *conn = call->conn;
524 struct rxrpc_channel *chan = &conn->channels[channel];
525
526 if (!test_and_set_bit(RXRPC_CALL_EXPOSED, &call->flags)) {
527 /* Mark the call ID as being used. If the callNumber counter
528 * exceeds ~2 billion, we kill the connection after its
529 * outstanding calls have finished so that the counter doesn't
530 * wrap.
531 */
532 chan->call_counter++;
533 if (chan->call_counter >= INT_MAX)
534 set_bit(RXRPC_CONN_DONT_REUSE, &conn->flags);
535 trace_rxrpc_client(conn, channel, rxrpc_client_exposed);
536
537 spin_lock(&call->peer->lock);
538 hlist_add_head(&call->error_link, &call->peer->error_targets);
539 spin_unlock(&call->peer->lock);
540 }
541}
542
543/*
544 * Set the reap timer.
545 */
546static void rxrpc_set_client_reap_timer(struct rxrpc_local *local)
547{
548 if (!local->kill_all_client_conns) {
549 unsigned long now = jiffies;
550 unsigned long reap_at = now + rxrpc_conn_idle_client_expiry;
551
552 if (local->rxnet->live)
553 timer_reduce(&local->client_conn_reap_timer, reap_at);
554 }
555}
556
557/*
558 * Disconnect a client call.
559 */
560void rxrpc_disconnect_client_call(struct rxrpc_bundle *bundle, struct rxrpc_call *call)
561{
562 struct rxrpc_connection *conn;
563 struct rxrpc_channel *chan = NULL;
564 struct rxrpc_local *local = bundle->local;
565 unsigned int channel;
566 bool may_reuse;
567 u32 cid;
568
569 _enter("c=%x", call->debug_id);
570
571 /* Calls that have never actually been assigned a channel can simply be
572 * discarded.
573 */
574 conn = call->conn;
575 if (!conn) {
576 _debug("call is waiting");
577 ASSERTCMP(call->call_id, ==, 0);
578 ASSERT(!test_bit(RXRPC_CALL_EXPOSED, &call->flags));
579 list_del_init(&call->wait_link);
580 return;
581 }
582
583 cid = call->cid;
584 channel = cid & RXRPC_CHANNELMASK;
585 chan = &conn->channels[channel];
586 trace_rxrpc_client(conn, channel, rxrpc_client_chan_disconnect);
587
588 if (WARN_ON(chan->call != call))
589 return;
590
591 may_reuse = rxrpc_may_reuse_conn(conn);
592
593 /* If a client call was exposed to the world, we save the result for
594 * retransmission.
595 *
596 * We use a barrier here so that the call number and abort code can be
597 * read without needing to take a lock.
598 *
599 * TODO: Make the incoming packet handler check this and handle
600 * terminal retransmission without requiring access to the call.
601 */
602 if (test_bit(RXRPC_CALL_EXPOSED, &call->flags)) {
603 _debug("exposed %u,%u", call->call_id, call->abort_code);
604 __rxrpc_disconnect_call(conn, call);
605
606 if (test_and_clear_bit(RXRPC_CONN_PROBING_FOR_UPGRADE, &conn->flags)) {
607 trace_rxrpc_client(conn, channel, rxrpc_client_to_active);
608 bundle->try_upgrade = false;
609 if (may_reuse)
610 rxrpc_activate_channels(bundle);
611 }
612 }
613
614 /* See if we can pass the channel directly to another call. */
615 if (may_reuse && !list_empty(&bundle->waiting_calls)) {
616 trace_rxrpc_client(conn, channel, rxrpc_client_chan_pass);
617 rxrpc_activate_one_channel(conn, channel);
618 return;
619 }
620
621 /* Schedule the final ACK to be transmitted in a short while so that it
622 * can be skipped if we find a follow-on call. The first DATA packet
623 * of the follow on call will implicitly ACK this call.
624 */
625 if (call->completion == RXRPC_CALL_SUCCEEDED &&
626 test_bit(RXRPC_CALL_EXPOSED, &call->flags)) {
627 unsigned long final_ack_at = jiffies + 2;
628
629 WRITE_ONCE(chan->final_ack_at, final_ack_at);
630 smp_wmb(); /* vs rxrpc_process_delayed_final_acks() */
631 set_bit(RXRPC_CONN_FINAL_ACK_0 + channel, &conn->flags);
632 rxrpc_reduce_conn_timer(conn, final_ack_at);
633 }
634
635 /* Deactivate the channel. */
636 chan->call = NULL;
637 set_bit(conn->bundle_shift + channel, &conn->bundle->avail_chans);
638 conn->act_chans &= ~(1 << channel);
639
640 /* If no channels remain active, then put the connection on the idle
641 * list for a short while. Give it a ref to stop it going away if it
642 * becomes unbundled.
643 */
644 if (!conn->act_chans) {
645 trace_rxrpc_client(conn, channel, rxrpc_client_to_idle);
646 conn->idle_timestamp = jiffies;
647
648 rxrpc_get_connection(conn, rxrpc_conn_get_idle);
649 list_move_tail(&conn->cache_link, &local->idle_client_conns);
650
651 rxrpc_set_client_reap_timer(local);
652 }
653}
654
655/*
656 * Remove a connection from a bundle.
657 */
658static void rxrpc_unbundle_conn(struct rxrpc_connection *conn)
659{
660 struct rxrpc_bundle *bundle = conn->bundle;
661 unsigned int bindex;
662 int i;
663
664 _enter("C=%x", conn->debug_id);
665
666 if (conn->flags & RXRPC_CONN_FINAL_ACK_MASK)
667 rxrpc_process_delayed_final_acks(conn, true);
668
669 bindex = conn->bundle_shift / RXRPC_MAXCALLS;
670 if (bundle->conns[bindex] == conn) {
671 _debug("clear slot %u", bindex);
672 bundle->conns[bindex] = NULL;
673 for (i = 0; i < RXRPC_MAXCALLS; i++)
674 clear_bit(conn->bundle_shift + i, &bundle->avail_chans);
675 rxrpc_put_client_connection_id(bundle->local, conn);
676 rxrpc_deactivate_bundle(bundle);
677 rxrpc_put_connection(conn, rxrpc_conn_put_unbundle);
678 }
679}
680
681/*
682 * Drop the active count on a bundle.
683 */
684void rxrpc_deactivate_bundle(struct rxrpc_bundle *bundle)
685{
686 struct rxrpc_local *local;
687 bool need_put = false;
688
689 if (!bundle)
690 return;
691
692 local = bundle->local;
693 if (atomic_dec_and_lock(&bundle->active, &local->client_bundles_lock)) {
694 if (!bundle->exclusive) {
695 _debug("erase bundle");
696 rb_erase(&bundle->local_node, &local->client_bundles);
697 need_put = true;
698 }
699
700 spin_unlock(&local->client_bundles_lock);
701 if (need_put)
702 rxrpc_put_bundle(bundle, rxrpc_bundle_put_discard);
703 }
704}
705
706/*
707 * Clean up a dead client connection.
708 */
709void rxrpc_kill_client_conn(struct rxrpc_connection *conn)
710{
711 struct rxrpc_local *local = conn->local;
712 struct rxrpc_net *rxnet = local->rxnet;
713
714 _enter("C=%x", conn->debug_id);
715
716 trace_rxrpc_client(conn, -1, rxrpc_client_cleanup);
717 atomic_dec(&rxnet->nr_client_conns);
718
719 rxrpc_put_client_connection_id(local, conn);
720}
721
722/*
723 * Discard expired client connections from the idle list. Each conn in the
724 * idle list has been exposed and holds an extra ref because of that.
725 *
726 * This may be called from conn setup or from a work item so cannot be
727 * considered non-reentrant.
728 */
729void rxrpc_discard_expired_client_conns(struct rxrpc_local *local)
730{
731 struct rxrpc_connection *conn;
732 unsigned long expiry, conn_expires_at, now;
733 unsigned int nr_conns;
734
735 _enter("");
736
737 /* We keep an estimate of what the number of conns ought to be after
738 * we've discarded some so that we don't overdo the discarding.
739 */
740 nr_conns = atomic_read(&local->rxnet->nr_client_conns);
741
742next:
743 conn = list_first_entry_or_null(&local->idle_client_conns,
744 struct rxrpc_connection, cache_link);
745 if (!conn)
746 return;
747
748 if (!local->kill_all_client_conns) {
749 /* If the number of connections is over the reap limit, we
750 * expedite discard by reducing the expiry timeout. We must,
751 * however, have at least a short grace period to be able to do
752 * final-ACK or ABORT retransmission.
753 */
754 expiry = rxrpc_conn_idle_client_expiry;
755 if (nr_conns > rxrpc_reap_client_connections)
756 expiry = rxrpc_conn_idle_client_fast_expiry;
757 if (conn->local->service_closed)
758 expiry = rxrpc_closed_conn_expiry * HZ;
759
760 conn_expires_at = conn->idle_timestamp + expiry;
761
762 now = READ_ONCE(jiffies);
763 if (time_after(conn_expires_at, now))
764 goto not_yet_expired;
765 }
766
767 atomic_dec(&conn->active);
768 trace_rxrpc_client(conn, -1, rxrpc_client_discard);
769 list_del_init(&conn->cache_link);
770
771 rxrpc_unbundle_conn(conn);
772 /* Drop the ->cache_link ref */
773 rxrpc_put_connection(conn, rxrpc_conn_put_discard_idle);
774
775 nr_conns--;
776 goto next;
777
778not_yet_expired:
779 /* The connection at the front of the queue hasn't yet expired, so
780 * schedule the work item for that point if we discarded something.
781 *
782 * We don't worry if the work item is already scheduled - it can look
783 * after rescheduling itself at a later time. We could cancel it, but
784 * then things get messier.
785 */
786 _debug("not yet");
787 if (!local->kill_all_client_conns)
788 timer_reduce(&local->client_conn_reap_timer, conn_expires_at);
789
790 _leave("");
791}
792
793/*
794 * Clean up the client connections on a local endpoint.
795 */
796void rxrpc_clean_up_local_conns(struct rxrpc_local *local)
797{
798 struct rxrpc_connection *conn;
799
800 _enter("");
801
802 local->kill_all_client_conns = true;
803
804 del_timer_sync(&local->client_conn_reap_timer);
805
806 while ((conn = list_first_entry_or_null(&local->idle_client_conns,
807 struct rxrpc_connection, cache_link))) {
808 list_del_init(&conn->cache_link);
809 atomic_dec(&conn->active);
810 trace_rxrpc_client(conn, -1, rxrpc_client_discard);
811 rxrpc_unbundle_conn(conn);
812 rxrpc_put_connection(conn, rxrpc_conn_put_local_dead);
813 }
814
815 _leave(" [culled]");
816}
1/* Client connection-specific management code.
2 *
3 * Copyright (C) 2016 Red Hat, Inc. All Rights Reserved.
4 * Written by David Howells (dhowells@redhat.com)
5 *
6 * This program is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU General Public Licence
8 * as published by the Free Software Foundation; either version
9 * 2 of the Licence, or (at your option) any later version.
10 *
11 *
12 * Client connections need to be cached for a little while after they've made a
13 * call so as to handle retransmitted DATA packets in case the server didn't
14 * receive the final ACK or terminating ABORT we sent it.
15 *
16 * Client connections can be in one of a number of cache states:
17 *
18 * (1) INACTIVE - The connection is not held in any list and may not have been
19 * exposed to the world. If it has been previously exposed, it was
20 * discarded from the idle list after expiring.
21 *
22 * (2) WAITING - The connection is waiting for the number of client conns to
23 * drop below the maximum capacity. Calls may be in progress upon it from
24 * when it was active and got culled.
25 *
26 * The connection is on the rxrpc_waiting_client_conns list which is kept
27 * in to-be-granted order. Culled conns with waiters go to the back of
28 * the queue just like new conns.
29 *
30 * (3) ACTIVE - The connection has at least one call in progress upon it, it
31 * may freely grant available channels to new calls and calls may be
32 * waiting on it for channels to become available.
33 *
34 * The connection is on the rxnet->active_client_conns list which is kept
35 * in activation order for culling purposes.
36 *
37 * rxrpc_nr_active_client_conns is held incremented also.
38 *
39 * (4) UPGRADE - As for ACTIVE, but only one call may be in progress and is
40 * being used to probe for service upgrade.
41 *
42 * (5) CULLED - The connection got summarily culled to try and free up
43 * capacity. Calls currently in progress on the connection are allowed to
44 * continue, but new calls will have to wait. There can be no waiters in
45 * this state - the conn would have to go to the WAITING state instead.
46 *
47 * (6) IDLE - The connection has no calls in progress upon it and must have
48 * been exposed to the world (ie. the EXPOSED flag must be set). When it
49 * expires, the EXPOSED flag is cleared and the connection transitions to
50 * the INACTIVE state.
51 *
52 * The connection is on the rxnet->idle_client_conns list which is kept in
53 * order of how soon they'll expire.
54 *
55 * There are flags of relevance to the cache:
56 *
57 * (1) EXPOSED - The connection ID got exposed to the world. If this flag is
58 * set, an extra ref is added to the connection preventing it from being
59 * reaped when it has no calls outstanding. This flag is cleared and the
60 * ref dropped when a conn is discarded from the idle list.
61 *
62 * This allows us to move terminal call state retransmission to the
63 * connection and to discard the call immediately we think it is done
64 * with. It also give us a chance to reuse the connection.
65 *
66 * (2) DONT_REUSE - The connection should be discarded as soon as possible and
67 * should not be reused. This is set when an exclusive connection is used
68 * or a call ID counter overflows.
69 *
70 * The caching state may only be changed if the cache lock is held.
71 *
72 * There are two idle client connection expiry durations. If the total number
73 * of connections is below the reap threshold, we use the normal duration; if
74 * it's above, we use the fast duration.
75 */
76
77#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
78
79#include <linux/slab.h>
80#include <linux/idr.h>
81#include <linux/timer.h>
82#include <linux/sched/signal.h>
83
84#include "ar-internal.h"
85
86__read_mostly unsigned int rxrpc_max_client_connections = 1000;
87__read_mostly unsigned int rxrpc_reap_client_connections = 900;
88__read_mostly unsigned long rxrpc_conn_idle_client_expiry = 2 * 60 * HZ;
89__read_mostly unsigned long rxrpc_conn_idle_client_fast_expiry = 2 * HZ;
90
91/*
92 * We use machine-unique IDs for our client connections.
93 */
94DEFINE_IDR(rxrpc_client_conn_ids);
95static DEFINE_SPINLOCK(rxrpc_conn_id_lock);
96
97static void rxrpc_cull_active_client_conns(struct rxrpc_net *);
98
99/*
100 * Get a connection ID and epoch for a client connection from the global pool.
101 * The connection struct pointer is then recorded in the idr radix tree. The
102 * epoch doesn't change until the client is rebooted (or, at least, unless the
103 * module is unloaded).
104 */
105static int rxrpc_get_client_connection_id(struct rxrpc_connection *conn,
106 gfp_t gfp)
107{
108 struct rxrpc_net *rxnet = conn->params.local->rxnet;
109 int id;
110
111 _enter("");
112
113 idr_preload(gfp);
114 spin_lock(&rxrpc_conn_id_lock);
115
116 id = idr_alloc_cyclic(&rxrpc_client_conn_ids, conn,
117 1, 0x40000000, GFP_NOWAIT);
118 if (id < 0)
119 goto error;
120
121 spin_unlock(&rxrpc_conn_id_lock);
122 idr_preload_end();
123
124 conn->proto.epoch = rxnet->epoch;
125 conn->proto.cid = id << RXRPC_CIDSHIFT;
126 set_bit(RXRPC_CONN_HAS_IDR, &conn->flags);
127 _leave(" [CID %x]", conn->proto.cid);
128 return 0;
129
130error:
131 spin_unlock(&rxrpc_conn_id_lock);
132 idr_preload_end();
133 _leave(" = %d", id);
134 return id;
135}
136
137/*
138 * Release a connection ID for a client connection from the global pool.
139 */
140static void rxrpc_put_client_connection_id(struct rxrpc_connection *conn)
141{
142 if (test_bit(RXRPC_CONN_HAS_IDR, &conn->flags)) {
143 spin_lock(&rxrpc_conn_id_lock);
144 idr_remove(&rxrpc_client_conn_ids,
145 conn->proto.cid >> RXRPC_CIDSHIFT);
146 spin_unlock(&rxrpc_conn_id_lock);
147 }
148}
149
150/*
151 * Destroy the client connection ID tree.
152 */
153void rxrpc_destroy_client_conn_ids(void)
154{
155 struct rxrpc_connection *conn;
156 int id;
157
158 if (!idr_is_empty(&rxrpc_client_conn_ids)) {
159 idr_for_each_entry(&rxrpc_client_conn_ids, conn, id) {
160 pr_err("AF_RXRPC: Leaked client conn %p {%d}\n",
161 conn, atomic_read(&conn->usage));
162 }
163 BUG();
164 }
165
166 idr_destroy(&rxrpc_client_conn_ids);
167}
168
169/*
170 * Allocate a client connection.
171 */
172static struct rxrpc_connection *
173rxrpc_alloc_client_connection(struct rxrpc_conn_parameters *cp, gfp_t gfp)
174{
175 struct rxrpc_connection *conn;
176 struct rxrpc_net *rxnet = cp->local->rxnet;
177 int ret;
178
179 _enter("");
180
181 conn = rxrpc_alloc_connection(gfp);
182 if (!conn) {
183 _leave(" = -ENOMEM");
184 return ERR_PTR(-ENOMEM);
185 }
186
187 atomic_set(&conn->usage, 1);
188 if (cp->exclusive)
189 __set_bit(RXRPC_CONN_DONT_REUSE, &conn->flags);
190 if (cp->upgrade)
191 __set_bit(RXRPC_CONN_PROBING_FOR_UPGRADE, &conn->flags);
192
193 conn->params = *cp;
194 conn->out_clientflag = RXRPC_CLIENT_INITIATED;
195 conn->state = RXRPC_CONN_CLIENT;
196 conn->service_id = cp->service_id;
197
198 ret = rxrpc_get_client_connection_id(conn, gfp);
199 if (ret < 0)
200 goto error_0;
201
202 ret = rxrpc_init_client_conn_security(conn);
203 if (ret < 0)
204 goto error_1;
205
206 ret = conn->security->prime_packet_security(conn);
207 if (ret < 0)
208 goto error_2;
209
210 atomic_inc(&rxnet->nr_conns);
211 write_lock(&rxnet->conn_lock);
212 list_add_tail(&conn->proc_link, &rxnet->conn_proc_list);
213 write_unlock(&rxnet->conn_lock);
214
215 /* We steal the caller's peer ref. */
216 cp->peer = NULL;
217 rxrpc_get_local(conn->params.local);
218 key_get(conn->params.key);
219
220 trace_rxrpc_conn(conn, rxrpc_conn_new_client, atomic_read(&conn->usage),
221 __builtin_return_address(0));
222 trace_rxrpc_client(conn, -1, rxrpc_client_alloc);
223 _leave(" = %p", conn);
224 return conn;
225
226error_2:
227 conn->security->clear(conn);
228error_1:
229 rxrpc_put_client_connection_id(conn);
230error_0:
231 kfree(conn);
232 _leave(" = %d", ret);
233 return ERR_PTR(ret);
234}
235
236/*
237 * Determine if a connection may be reused.
238 */
239static bool rxrpc_may_reuse_conn(struct rxrpc_connection *conn)
240{
241 struct rxrpc_net *rxnet = conn->params.local->rxnet;
242 int id_cursor, id, distance, limit;
243
244 if (test_bit(RXRPC_CONN_DONT_REUSE, &conn->flags))
245 goto dont_reuse;
246
247 if (conn->proto.epoch != rxnet->epoch)
248 goto mark_dont_reuse;
249
250 /* The IDR tree gets very expensive on memory if the connection IDs are
251 * widely scattered throughout the number space, so we shall want to
252 * kill off connections that, say, have an ID more than about four
253 * times the maximum number of client conns away from the current
254 * allocation point to try and keep the IDs concentrated.
255 */
256 id_cursor = idr_get_cursor(&rxrpc_client_conn_ids);
257 id = conn->proto.cid >> RXRPC_CIDSHIFT;
258 distance = id - id_cursor;
259 if (distance < 0)
260 distance = -distance;
261 limit = max(rxrpc_max_client_connections * 4, 1024U);
262 if (distance > limit)
263 goto mark_dont_reuse;
264
265 return true;
266
267mark_dont_reuse:
268 set_bit(RXRPC_CONN_DONT_REUSE, &conn->flags);
269dont_reuse:
270 return false;
271}
272
273/*
274 * Create or find a client connection to use for a call.
275 *
276 * If we return with a connection, the call will be on its waiting list. It's
277 * left to the caller to assign a channel and wake up the call.
278 */
279static int rxrpc_get_client_conn(struct rxrpc_call *call,
280 struct rxrpc_conn_parameters *cp,
281 struct sockaddr_rxrpc *srx,
282 gfp_t gfp)
283{
284 struct rxrpc_connection *conn, *candidate = NULL;
285 struct rxrpc_local *local = cp->local;
286 struct rb_node *p, **pp, *parent;
287 long diff;
288 int ret = -ENOMEM;
289
290 _enter("{%d,%lx},", call->debug_id, call->user_call_ID);
291
292 cp->peer = rxrpc_lookup_peer(cp->local, srx, gfp);
293 if (!cp->peer)
294 goto error;
295
296 call->cong_cwnd = cp->peer->cong_cwnd;
297 if (call->cong_cwnd >= call->cong_ssthresh)
298 call->cong_mode = RXRPC_CALL_CONGEST_AVOIDANCE;
299 else
300 call->cong_mode = RXRPC_CALL_SLOW_START;
301
302 /* If the connection is not meant to be exclusive, search the available
303 * connections to see if the connection we want to use already exists.
304 */
305 if (!cp->exclusive) {
306 _debug("search 1");
307 spin_lock(&local->client_conns_lock);
308 p = local->client_conns.rb_node;
309 while (p) {
310 conn = rb_entry(p, struct rxrpc_connection, client_node);
311
312#define cmp(X) ((long)conn->params.X - (long)cp->X)
313 diff = (cmp(peer) ?:
314 cmp(key) ?:
315 cmp(security_level) ?:
316 cmp(upgrade));
317#undef cmp
318 if (diff < 0) {
319 p = p->rb_left;
320 } else if (diff > 0) {
321 p = p->rb_right;
322 } else {
323 if (rxrpc_may_reuse_conn(conn) &&
324 rxrpc_get_connection_maybe(conn))
325 goto found_extant_conn;
326 /* The connection needs replacing. It's better
327 * to effect that when we have something to
328 * replace it with so that we don't have to
329 * rebalance the tree twice.
330 */
331 break;
332 }
333 }
334 spin_unlock(&local->client_conns_lock);
335 }
336
337 /* There wasn't a connection yet or we need an exclusive connection.
338 * We need to create a candidate and then potentially redo the search
339 * in case we're racing with another thread also trying to connect on a
340 * shareable connection.
341 */
342 _debug("new conn");
343 candidate = rxrpc_alloc_client_connection(cp, gfp);
344 if (IS_ERR(candidate)) {
345 ret = PTR_ERR(candidate);
346 goto error_peer;
347 }
348
349 /* Add the call to the new connection's waiting list in case we're
350 * going to have to wait for the connection to come live. It's our
351 * connection, so we want first dibs on the channel slots. We would
352 * normally have to take channel_lock but we do this before anyone else
353 * can see the connection.
354 */
355 list_add_tail(&call->chan_wait_link, &candidate->waiting_calls);
356
357 if (cp->exclusive) {
358 call->conn = candidate;
359 call->security_ix = candidate->security_ix;
360 call->service_id = candidate->service_id;
361 _leave(" = 0 [exclusive %d]", candidate->debug_id);
362 return 0;
363 }
364
365 /* Publish the new connection for userspace to find. We need to redo
366 * the search before doing this lest we race with someone else adding a
367 * conflicting instance.
368 */
369 _debug("search 2");
370 spin_lock(&local->client_conns_lock);
371
372 pp = &local->client_conns.rb_node;
373 parent = NULL;
374 while (*pp) {
375 parent = *pp;
376 conn = rb_entry(parent, struct rxrpc_connection, client_node);
377
378#define cmp(X) ((long)conn->params.X - (long)candidate->params.X)
379 diff = (cmp(peer) ?:
380 cmp(key) ?:
381 cmp(security_level) ?:
382 cmp(upgrade));
383#undef cmp
384 if (diff < 0) {
385 pp = &(*pp)->rb_left;
386 } else if (diff > 0) {
387 pp = &(*pp)->rb_right;
388 } else {
389 if (rxrpc_may_reuse_conn(conn) &&
390 rxrpc_get_connection_maybe(conn))
391 goto found_extant_conn;
392 /* The old connection is from an outdated epoch. */
393 _debug("replace conn");
394 clear_bit(RXRPC_CONN_IN_CLIENT_CONNS, &conn->flags);
395 rb_replace_node(&conn->client_node,
396 &candidate->client_node,
397 &local->client_conns);
398 trace_rxrpc_client(conn, -1, rxrpc_client_replace);
399 goto candidate_published;
400 }
401 }
402
403 _debug("new conn");
404 rb_link_node(&candidate->client_node, parent, pp);
405 rb_insert_color(&candidate->client_node, &local->client_conns);
406
407candidate_published:
408 set_bit(RXRPC_CONN_IN_CLIENT_CONNS, &candidate->flags);
409 call->conn = candidate;
410 call->security_ix = candidate->security_ix;
411 call->service_id = candidate->service_id;
412 spin_unlock(&local->client_conns_lock);
413 _leave(" = 0 [new %d]", candidate->debug_id);
414 return 0;
415
416 /* We come here if we found a suitable connection already in existence.
417 * Discard any candidate we may have allocated, and try to get a
418 * channel on this one.
419 */
420found_extant_conn:
421 _debug("found conn");
422 spin_unlock(&local->client_conns_lock);
423
424 if (candidate) {
425 trace_rxrpc_client(candidate, -1, rxrpc_client_duplicate);
426 rxrpc_put_connection(candidate);
427 candidate = NULL;
428 }
429
430 spin_lock(&conn->channel_lock);
431 call->conn = conn;
432 call->security_ix = conn->security_ix;
433 call->service_id = conn->service_id;
434 list_add(&call->chan_wait_link, &conn->waiting_calls);
435 spin_unlock(&conn->channel_lock);
436 _leave(" = 0 [extant %d]", conn->debug_id);
437 return 0;
438
439error_peer:
440 rxrpc_put_peer(cp->peer);
441 cp->peer = NULL;
442error:
443 _leave(" = %d", ret);
444 return ret;
445}
446
447/*
448 * Activate a connection.
449 */
450static void rxrpc_activate_conn(struct rxrpc_net *rxnet,
451 struct rxrpc_connection *conn)
452{
453 if (test_bit(RXRPC_CONN_PROBING_FOR_UPGRADE, &conn->flags)) {
454 trace_rxrpc_client(conn, -1, rxrpc_client_to_upgrade);
455 conn->cache_state = RXRPC_CONN_CLIENT_UPGRADE;
456 } else {
457 trace_rxrpc_client(conn, -1, rxrpc_client_to_active);
458 conn->cache_state = RXRPC_CONN_CLIENT_ACTIVE;
459 }
460 rxnet->nr_active_client_conns++;
461 list_move_tail(&conn->cache_link, &rxnet->active_client_conns);
462}
463
464/*
465 * Attempt to animate a connection for a new call.
466 *
467 * If it's not exclusive, the connection is in the endpoint tree, and we're in
468 * the conn's list of those waiting to grab a channel. There is, however, a
469 * limit on the number of live connections allowed at any one time, so we may
470 * have to wait for capacity to become available.
471 *
472 * Note that a connection on the waiting queue might *also* have active
473 * channels if it has been culled to make space and then re-requested by a new
474 * call.
475 */
476static void rxrpc_animate_client_conn(struct rxrpc_net *rxnet,
477 struct rxrpc_connection *conn)
478{
479 unsigned int nr_conns;
480
481 _enter("%d,%d", conn->debug_id, conn->cache_state);
482
483 if (conn->cache_state == RXRPC_CONN_CLIENT_ACTIVE ||
484 conn->cache_state == RXRPC_CONN_CLIENT_UPGRADE)
485 goto out;
486
487 spin_lock(&rxnet->client_conn_cache_lock);
488
489 nr_conns = rxnet->nr_client_conns;
490 if (!test_and_set_bit(RXRPC_CONN_COUNTED, &conn->flags)) {
491 trace_rxrpc_client(conn, -1, rxrpc_client_count);
492 rxnet->nr_client_conns = nr_conns + 1;
493 }
494
495 switch (conn->cache_state) {
496 case RXRPC_CONN_CLIENT_ACTIVE:
497 case RXRPC_CONN_CLIENT_UPGRADE:
498 case RXRPC_CONN_CLIENT_WAITING:
499 break;
500
501 case RXRPC_CONN_CLIENT_INACTIVE:
502 case RXRPC_CONN_CLIENT_CULLED:
503 case RXRPC_CONN_CLIENT_IDLE:
504 if (nr_conns >= rxrpc_max_client_connections)
505 goto wait_for_capacity;
506 goto activate_conn;
507
508 default:
509 BUG();
510 }
511
512out_unlock:
513 spin_unlock(&rxnet->client_conn_cache_lock);
514out:
515 _leave(" [%d]", conn->cache_state);
516 return;
517
518activate_conn:
519 _debug("activate");
520 rxrpc_activate_conn(rxnet, conn);
521 goto out_unlock;
522
523wait_for_capacity:
524 _debug("wait");
525 trace_rxrpc_client(conn, -1, rxrpc_client_to_waiting);
526 conn->cache_state = RXRPC_CONN_CLIENT_WAITING;
527 list_move_tail(&conn->cache_link, &rxnet->waiting_client_conns);
528 goto out_unlock;
529}
530
531/*
532 * Deactivate a channel.
533 */
534static void rxrpc_deactivate_one_channel(struct rxrpc_connection *conn,
535 unsigned int channel)
536{
537 struct rxrpc_channel *chan = &conn->channels[channel];
538
539 rcu_assign_pointer(chan->call, NULL);
540 conn->active_chans &= ~(1 << channel);
541}
542
543/*
544 * Assign a channel to the call at the front of the queue and wake the call up.
545 * We don't increment the callNumber counter until this number has been exposed
546 * to the world.
547 */
548static void rxrpc_activate_one_channel(struct rxrpc_connection *conn,
549 unsigned int channel)
550{
551 struct rxrpc_channel *chan = &conn->channels[channel];
552 struct rxrpc_call *call = list_entry(conn->waiting_calls.next,
553 struct rxrpc_call, chan_wait_link);
554 u32 call_id = chan->call_counter + 1;
555
556 trace_rxrpc_client(conn, channel, rxrpc_client_chan_activate);
557
558 /* Cancel the final ACK on the previous call if it hasn't been sent yet
559 * as the DATA packet will implicitly ACK it.
560 */
561 clear_bit(RXRPC_CONN_FINAL_ACK_0 + channel, &conn->flags);
562
563 write_lock_bh(&call->state_lock);
564 if (!test_bit(RXRPC_CALL_TX_LASTQ, &call->flags))
565 call->state = RXRPC_CALL_CLIENT_SEND_REQUEST;
566 else
567 call->state = RXRPC_CALL_CLIENT_AWAIT_REPLY;
568 write_unlock_bh(&call->state_lock);
569
570 rxrpc_see_call(call);
571 list_del_init(&call->chan_wait_link);
572 conn->active_chans |= 1 << channel;
573 call->peer = rxrpc_get_peer(conn->params.peer);
574 call->cid = conn->proto.cid | channel;
575 call->call_id = call_id;
576
577 trace_rxrpc_connect_call(call);
578 _net("CONNECT call %08x:%08x as call %d on conn %d",
579 call->cid, call->call_id, call->debug_id, conn->debug_id);
580
581 /* Paired with the read barrier in rxrpc_wait_for_channel(). This
582 * orders cid and epoch in the connection wrt to call_id without the
583 * need to take the channel_lock.
584 *
585 * We provisionally assign a callNumber at this point, but we don't
586 * confirm it until the call is about to be exposed.
587 *
588 * TODO: Pair with a barrier in the data_ready handler when that looks
589 * at the call ID through a connection channel.
590 */
591 smp_wmb();
592 chan->call_id = call_id;
593 rcu_assign_pointer(chan->call, call);
594 wake_up(&call->waitq);
595}
596
597/*
598 * Assign channels and callNumbers to waiting calls with channel_lock
599 * held by caller.
600 */
601static void rxrpc_activate_channels_locked(struct rxrpc_connection *conn)
602{
603 u8 avail, mask;
604
605 switch (conn->cache_state) {
606 case RXRPC_CONN_CLIENT_ACTIVE:
607 mask = RXRPC_ACTIVE_CHANS_MASK;
608 break;
609 case RXRPC_CONN_CLIENT_UPGRADE:
610 mask = 0x01;
611 break;
612 default:
613 return;
614 }
615
616 while (!list_empty(&conn->waiting_calls) &&
617 (avail = ~conn->active_chans,
618 avail &= mask,
619 avail != 0))
620 rxrpc_activate_one_channel(conn, __ffs(avail));
621}
622
623/*
624 * Assign channels and callNumbers to waiting calls.
625 */
626static void rxrpc_activate_channels(struct rxrpc_connection *conn)
627{
628 _enter("%d", conn->debug_id);
629
630 trace_rxrpc_client(conn, -1, rxrpc_client_activate_chans);
631
632 if (conn->active_chans == RXRPC_ACTIVE_CHANS_MASK)
633 return;
634
635 spin_lock(&conn->channel_lock);
636 rxrpc_activate_channels_locked(conn);
637 spin_unlock(&conn->channel_lock);
638 _leave("");
639}
640
641/*
642 * Wait for a callNumber and a channel to be granted to a call.
643 */
644static int rxrpc_wait_for_channel(struct rxrpc_call *call, gfp_t gfp)
645{
646 int ret = 0;
647
648 _enter("%d", call->debug_id);
649
650 if (!call->call_id) {
651 DECLARE_WAITQUEUE(myself, current);
652
653 if (!gfpflags_allow_blocking(gfp)) {
654 ret = -EAGAIN;
655 goto out;
656 }
657
658 add_wait_queue_exclusive(&call->waitq, &myself);
659 for (;;) {
660 set_current_state(TASK_INTERRUPTIBLE);
661 if (call->call_id)
662 break;
663 if (signal_pending(current)) {
664 ret = -ERESTARTSYS;
665 break;
666 }
667 schedule();
668 }
669 remove_wait_queue(&call->waitq, &myself);
670 __set_current_state(TASK_RUNNING);
671 }
672
673 /* Paired with the write barrier in rxrpc_activate_one_channel(). */
674 smp_rmb();
675
676out:
677 _leave(" = %d", ret);
678 return ret;
679}
680
681/*
682 * find a connection for a call
683 * - called in process context with IRQs enabled
684 */
685int rxrpc_connect_call(struct rxrpc_call *call,
686 struct rxrpc_conn_parameters *cp,
687 struct sockaddr_rxrpc *srx,
688 gfp_t gfp)
689{
690 struct rxrpc_net *rxnet = cp->local->rxnet;
691 int ret;
692
693 _enter("{%d,%lx},", call->debug_id, call->user_call_ID);
694
695 rxrpc_discard_expired_client_conns(&rxnet->client_conn_reaper);
696 rxrpc_cull_active_client_conns(rxnet);
697
698 ret = rxrpc_get_client_conn(call, cp, srx, gfp);
699 if (ret < 0)
700 goto out;
701
702 rxrpc_animate_client_conn(rxnet, call->conn);
703 rxrpc_activate_channels(call->conn);
704
705 ret = rxrpc_wait_for_channel(call, gfp);
706 if (ret < 0) {
707 rxrpc_disconnect_client_call(call);
708 goto out;
709 }
710
711 spin_lock_bh(&call->conn->params.peer->lock);
712 hlist_add_head(&call->error_link,
713 &call->conn->params.peer->error_targets);
714 spin_unlock_bh(&call->conn->params.peer->lock);
715
716out:
717 _leave(" = %d", ret);
718 return ret;
719}
720
721/*
722 * Note that a connection is about to be exposed to the world. Once it is
723 * exposed, we maintain an extra ref on it that stops it from being summarily
724 * discarded before it's (a) had a chance to deal with retransmission and (b)
725 * had a chance at re-use (the per-connection security negotiation is
726 * expensive).
727 */
728static void rxrpc_expose_client_conn(struct rxrpc_connection *conn,
729 unsigned int channel)
730{
731 if (!test_and_set_bit(RXRPC_CONN_EXPOSED, &conn->flags)) {
732 trace_rxrpc_client(conn, channel, rxrpc_client_exposed);
733 rxrpc_get_connection(conn);
734 }
735}
736
737/*
738 * Note that a call, and thus a connection, is about to be exposed to the
739 * world.
740 */
741void rxrpc_expose_client_call(struct rxrpc_call *call)
742{
743 unsigned int channel = call->cid & RXRPC_CHANNELMASK;
744 struct rxrpc_connection *conn = call->conn;
745 struct rxrpc_channel *chan = &conn->channels[channel];
746
747 if (!test_and_set_bit(RXRPC_CALL_EXPOSED, &call->flags)) {
748 /* Mark the call ID as being used. If the callNumber counter
749 * exceeds ~2 billion, we kill the connection after its
750 * outstanding calls have finished so that the counter doesn't
751 * wrap.
752 */
753 chan->call_counter++;
754 if (chan->call_counter >= INT_MAX)
755 set_bit(RXRPC_CONN_DONT_REUSE, &conn->flags);
756 rxrpc_expose_client_conn(conn, channel);
757 }
758}
759
760/*
761 * Set the reap timer.
762 */
763static void rxrpc_set_client_reap_timer(struct rxrpc_net *rxnet)
764{
765 unsigned long now = jiffies;
766 unsigned long reap_at = now + rxrpc_conn_idle_client_expiry;
767
768 if (rxnet->live)
769 timer_reduce(&rxnet->client_conn_reap_timer, reap_at);
770}
771
772/*
773 * Disconnect a client call.
774 */
775void rxrpc_disconnect_client_call(struct rxrpc_call *call)
776{
777 unsigned int channel = call->cid & RXRPC_CHANNELMASK;
778 struct rxrpc_connection *conn = call->conn;
779 struct rxrpc_channel *chan = &conn->channels[channel];
780 struct rxrpc_net *rxnet = conn->params.local->rxnet;
781
782 trace_rxrpc_client(conn, channel, rxrpc_client_chan_disconnect);
783 call->conn = NULL;
784
785 spin_lock(&conn->channel_lock);
786
787 /* Calls that have never actually been assigned a channel can simply be
788 * discarded. If the conn didn't get used either, it will follow
789 * immediately unless someone else grabs it in the meantime.
790 */
791 if (!list_empty(&call->chan_wait_link)) {
792 _debug("call is waiting");
793 ASSERTCMP(call->call_id, ==, 0);
794 ASSERT(!test_bit(RXRPC_CALL_EXPOSED, &call->flags));
795 list_del_init(&call->chan_wait_link);
796
797 trace_rxrpc_client(conn, channel, rxrpc_client_chan_unstarted);
798
799 /* We must deactivate or idle the connection if it's now
800 * waiting for nothing.
801 */
802 spin_lock(&rxnet->client_conn_cache_lock);
803 if (conn->cache_state == RXRPC_CONN_CLIENT_WAITING &&
804 list_empty(&conn->waiting_calls) &&
805 !conn->active_chans)
806 goto idle_connection;
807 goto out;
808 }
809
810 ASSERTCMP(rcu_access_pointer(chan->call), ==, call);
811
812 /* If a client call was exposed to the world, we save the result for
813 * retransmission.
814 *
815 * We use a barrier here so that the call number and abort code can be
816 * read without needing to take a lock.
817 *
818 * TODO: Make the incoming packet handler check this and handle
819 * terminal retransmission without requiring access to the call.
820 */
821 if (test_bit(RXRPC_CALL_EXPOSED, &call->flags)) {
822 _debug("exposed %u,%u", call->call_id, call->abort_code);
823 __rxrpc_disconnect_call(conn, call);
824 }
825
826 /* See if we can pass the channel directly to another call. */
827 if (conn->cache_state == RXRPC_CONN_CLIENT_ACTIVE &&
828 !list_empty(&conn->waiting_calls)) {
829 trace_rxrpc_client(conn, channel, rxrpc_client_chan_pass);
830 rxrpc_activate_one_channel(conn, channel);
831 goto out_2;
832 }
833
834 /* Schedule the final ACK to be transmitted in a short while so that it
835 * can be skipped if we find a follow-on call. The first DATA packet
836 * of the follow on call will implicitly ACK this call.
837 */
838 if (call->completion == RXRPC_CALL_SUCCEEDED &&
839 test_bit(RXRPC_CALL_EXPOSED, &call->flags)) {
840 unsigned long final_ack_at = jiffies + 2;
841
842 WRITE_ONCE(chan->final_ack_at, final_ack_at);
843 smp_wmb(); /* vs rxrpc_process_delayed_final_acks() */
844 set_bit(RXRPC_CONN_FINAL_ACK_0 + channel, &conn->flags);
845 rxrpc_reduce_conn_timer(conn, final_ack_at);
846 }
847
848 /* Things are more complex and we need the cache lock. We might be
849 * able to simply idle the conn or it might now be lurking on the wait
850 * list. It might even get moved back to the active list whilst we're
851 * waiting for the lock.
852 */
853 spin_lock(&rxnet->client_conn_cache_lock);
854
855 switch (conn->cache_state) {
856 case RXRPC_CONN_CLIENT_UPGRADE:
857 /* Deal with termination of a service upgrade probe. */
858 if (test_bit(RXRPC_CONN_EXPOSED, &conn->flags)) {
859 clear_bit(RXRPC_CONN_PROBING_FOR_UPGRADE, &conn->flags);
860 trace_rxrpc_client(conn, channel, rxrpc_client_to_active);
861 conn->cache_state = RXRPC_CONN_CLIENT_ACTIVE;
862 rxrpc_activate_channels_locked(conn);
863 }
864 /* fall through */
865 case RXRPC_CONN_CLIENT_ACTIVE:
866 if (list_empty(&conn->waiting_calls)) {
867 rxrpc_deactivate_one_channel(conn, channel);
868 if (!conn->active_chans) {
869 rxnet->nr_active_client_conns--;
870 goto idle_connection;
871 }
872 goto out;
873 }
874
875 trace_rxrpc_client(conn, channel, rxrpc_client_chan_pass);
876 rxrpc_activate_one_channel(conn, channel);
877 goto out;
878
879 case RXRPC_CONN_CLIENT_CULLED:
880 rxrpc_deactivate_one_channel(conn, channel);
881 ASSERT(list_empty(&conn->waiting_calls));
882 if (!conn->active_chans)
883 goto idle_connection;
884 goto out;
885
886 case RXRPC_CONN_CLIENT_WAITING:
887 rxrpc_deactivate_one_channel(conn, channel);
888 goto out;
889
890 default:
891 BUG();
892 }
893
894out:
895 spin_unlock(&rxnet->client_conn_cache_lock);
896out_2:
897 spin_unlock(&conn->channel_lock);
898 rxrpc_put_connection(conn);
899 _leave("");
900 return;
901
902idle_connection:
903 /* As no channels remain active, the connection gets deactivated
904 * immediately or moved to the idle list for a short while.
905 */
906 if (test_bit(RXRPC_CONN_EXPOSED, &conn->flags)) {
907 trace_rxrpc_client(conn, channel, rxrpc_client_to_idle);
908 conn->idle_timestamp = jiffies;
909 conn->cache_state = RXRPC_CONN_CLIENT_IDLE;
910 list_move_tail(&conn->cache_link, &rxnet->idle_client_conns);
911 if (rxnet->idle_client_conns.next == &conn->cache_link &&
912 !rxnet->kill_all_client_conns)
913 rxrpc_set_client_reap_timer(rxnet);
914 } else {
915 trace_rxrpc_client(conn, channel, rxrpc_client_to_inactive);
916 conn->cache_state = RXRPC_CONN_CLIENT_INACTIVE;
917 list_del_init(&conn->cache_link);
918 }
919 goto out;
920}
921
922/*
923 * Clean up a dead client connection.
924 */
925static struct rxrpc_connection *
926rxrpc_put_one_client_conn(struct rxrpc_connection *conn)
927{
928 struct rxrpc_connection *next = NULL;
929 struct rxrpc_local *local = conn->params.local;
930 struct rxrpc_net *rxnet = local->rxnet;
931 unsigned int nr_conns;
932
933 trace_rxrpc_client(conn, -1, rxrpc_client_cleanup);
934
935 if (test_bit(RXRPC_CONN_IN_CLIENT_CONNS, &conn->flags)) {
936 spin_lock(&local->client_conns_lock);
937 if (test_and_clear_bit(RXRPC_CONN_IN_CLIENT_CONNS,
938 &conn->flags))
939 rb_erase(&conn->client_node, &local->client_conns);
940 spin_unlock(&local->client_conns_lock);
941 }
942
943 rxrpc_put_client_connection_id(conn);
944
945 ASSERTCMP(conn->cache_state, ==, RXRPC_CONN_CLIENT_INACTIVE);
946
947 if (test_bit(RXRPC_CONN_COUNTED, &conn->flags)) {
948 trace_rxrpc_client(conn, -1, rxrpc_client_uncount);
949 spin_lock(&rxnet->client_conn_cache_lock);
950 nr_conns = --rxnet->nr_client_conns;
951
952 if (nr_conns < rxrpc_max_client_connections &&
953 !list_empty(&rxnet->waiting_client_conns)) {
954 next = list_entry(rxnet->waiting_client_conns.next,
955 struct rxrpc_connection, cache_link);
956 rxrpc_get_connection(next);
957 rxrpc_activate_conn(rxnet, next);
958 }
959
960 spin_unlock(&rxnet->client_conn_cache_lock);
961 }
962
963 rxrpc_kill_connection(conn);
964 if (next)
965 rxrpc_activate_channels(next);
966
967 /* We need to get rid of the temporary ref we took upon next, but we
968 * can't call rxrpc_put_connection() recursively.
969 */
970 return next;
971}
972
973/*
974 * Clean up a dead client connections.
975 */
976void rxrpc_put_client_conn(struct rxrpc_connection *conn)
977{
978 const void *here = __builtin_return_address(0);
979 int n;
980
981 do {
982 n = atomic_dec_return(&conn->usage);
983 trace_rxrpc_conn(conn, rxrpc_conn_put_client, n, here);
984 if (n > 0)
985 return;
986 ASSERTCMP(n, >=, 0);
987
988 conn = rxrpc_put_one_client_conn(conn);
989 } while (conn);
990}
991
992/*
993 * Kill the longest-active client connections to make room for new ones.
994 */
995static void rxrpc_cull_active_client_conns(struct rxrpc_net *rxnet)
996{
997 struct rxrpc_connection *conn;
998 unsigned int nr_conns = rxnet->nr_client_conns;
999 unsigned int nr_active, limit;
1000
1001 _enter("");
1002
1003 ASSERTCMP(nr_conns, >=, 0);
1004 if (nr_conns < rxrpc_max_client_connections) {
1005 _leave(" [ok]");
1006 return;
1007 }
1008 limit = rxrpc_reap_client_connections;
1009
1010 spin_lock(&rxnet->client_conn_cache_lock);
1011 nr_active = rxnet->nr_active_client_conns;
1012
1013 while (nr_active > limit) {
1014 ASSERT(!list_empty(&rxnet->active_client_conns));
1015 conn = list_entry(rxnet->active_client_conns.next,
1016 struct rxrpc_connection, cache_link);
1017 ASSERTIFCMP(conn->cache_state != RXRPC_CONN_CLIENT_ACTIVE,
1018 conn->cache_state, ==, RXRPC_CONN_CLIENT_UPGRADE);
1019
1020 if (list_empty(&conn->waiting_calls)) {
1021 trace_rxrpc_client(conn, -1, rxrpc_client_to_culled);
1022 conn->cache_state = RXRPC_CONN_CLIENT_CULLED;
1023 list_del_init(&conn->cache_link);
1024 } else {
1025 trace_rxrpc_client(conn, -1, rxrpc_client_to_waiting);
1026 conn->cache_state = RXRPC_CONN_CLIENT_WAITING;
1027 list_move_tail(&conn->cache_link,
1028 &rxnet->waiting_client_conns);
1029 }
1030
1031 nr_active--;
1032 }
1033
1034 rxnet->nr_active_client_conns = nr_active;
1035 spin_unlock(&rxnet->client_conn_cache_lock);
1036 ASSERTCMP(nr_active, >=, 0);
1037 _leave(" [culled]");
1038}
1039
1040/*
1041 * Discard expired client connections from the idle list. Each conn in the
1042 * idle list has been exposed and holds an extra ref because of that.
1043 *
1044 * This may be called from conn setup or from a work item so cannot be
1045 * considered non-reentrant.
1046 */
1047void rxrpc_discard_expired_client_conns(struct work_struct *work)
1048{
1049 struct rxrpc_connection *conn;
1050 struct rxrpc_net *rxnet =
1051 container_of(work, struct rxrpc_net, client_conn_reaper);
1052 unsigned long expiry, conn_expires_at, now;
1053 unsigned int nr_conns;
1054 bool did_discard = false;
1055
1056 _enter("");
1057
1058 if (list_empty(&rxnet->idle_client_conns)) {
1059 _leave(" [empty]");
1060 return;
1061 }
1062
1063 /* Don't double up on the discarding */
1064 if (!spin_trylock(&rxnet->client_conn_discard_lock)) {
1065 _leave(" [already]");
1066 return;
1067 }
1068
1069 /* We keep an estimate of what the number of conns ought to be after
1070 * we've discarded some so that we don't overdo the discarding.
1071 */
1072 nr_conns = rxnet->nr_client_conns;
1073
1074next:
1075 spin_lock(&rxnet->client_conn_cache_lock);
1076
1077 if (list_empty(&rxnet->idle_client_conns))
1078 goto out;
1079
1080 conn = list_entry(rxnet->idle_client_conns.next,
1081 struct rxrpc_connection, cache_link);
1082 ASSERT(test_bit(RXRPC_CONN_EXPOSED, &conn->flags));
1083
1084 if (!rxnet->kill_all_client_conns) {
1085 /* If the number of connections is over the reap limit, we
1086 * expedite discard by reducing the expiry timeout. We must,
1087 * however, have at least a short grace period to be able to do
1088 * final-ACK or ABORT retransmission.
1089 */
1090 expiry = rxrpc_conn_idle_client_expiry;
1091 if (nr_conns > rxrpc_reap_client_connections)
1092 expiry = rxrpc_conn_idle_client_fast_expiry;
1093 if (conn->params.local->service_closed)
1094 expiry = rxrpc_closed_conn_expiry * HZ;
1095
1096 conn_expires_at = conn->idle_timestamp + expiry;
1097
1098 now = READ_ONCE(jiffies);
1099 if (time_after(conn_expires_at, now))
1100 goto not_yet_expired;
1101 }
1102
1103 trace_rxrpc_client(conn, -1, rxrpc_client_discard);
1104 if (!test_and_clear_bit(RXRPC_CONN_EXPOSED, &conn->flags))
1105 BUG();
1106 conn->cache_state = RXRPC_CONN_CLIENT_INACTIVE;
1107 list_del_init(&conn->cache_link);
1108
1109 spin_unlock(&rxnet->client_conn_cache_lock);
1110
1111 /* When we cleared the EXPOSED flag, we took on responsibility for the
1112 * reference that that had on the usage count. We deal with that here.
1113 * If someone re-sets the flag and re-gets the ref, that's fine.
1114 */
1115 rxrpc_put_connection(conn);
1116 did_discard = true;
1117 nr_conns--;
1118 goto next;
1119
1120not_yet_expired:
1121 /* The connection at the front of the queue hasn't yet expired, so
1122 * schedule the work item for that point if we discarded something.
1123 *
1124 * We don't worry if the work item is already scheduled - it can look
1125 * after rescheduling itself at a later time. We could cancel it, but
1126 * then things get messier.
1127 */
1128 _debug("not yet");
1129 if (!rxnet->kill_all_client_conns)
1130 timer_reduce(&rxnet->client_conn_reap_timer,
1131 conn_expires_at);
1132
1133out:
1134 spin_unlock(&rxnet->client_conn_cache_lock);
1135 spin_unlock(&rxnet->client_conn_discard_lock);
1136 _leave("");
1137}
1138
1139/*
1140 * Preemptively destroy all the client connection records rather than waiting
1141 * for them to time out
1142 */
1143void rxrpc_destroy_all_client_connections(struct rxrpc_net *rxnet)
1144{
1145 _enter("");
1146
1147 spin_lock(&rxnet->client_conn_cache_lock);
1148 rxnet->kill_all_client_conns = true;
1149 spin_unlock(&rxnet->client_conn_cache_lock);
1150
1151 del_timer_sync(&rxnet->client_conn_reap_timer);
1152
1153 if (!rxrpc_queue_work(&rxnet->client_conn_reaper))
1154 _debug("destroy: queue failed");
1155
1156 _leave("");
1157}