selftests: kvm: Merge user_msr_test into userspace_msr_exit_test
[linux-2.6-microblaze.git] / net / rxrpc / conn_client.c
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
37 /*
38  * We use machine-unique IDs for our client connections.
39  */
40 DEFINE_IDR(rxrpc_client_conn_ids);
41 static DEFINE_SPINLOCK(rxrpc_conn_id_lock);
42
43 /*
44  * Get a connection ID and epoch for a client connection from the global pool.
45  * The connection struct pointer is then recorded in the idr radix tree.  The
46  * epoch doesn't change until the client is rebooted (or, at least, unless the
47  * module is unloaded).
48  */
49 static int rxrpc_get_client_connection_id(struct rxrpc_connection *conn,
50                                           gfp_t gfp)
51 {
52         struct rxrpc_net *rxnet = conn->params.local->rxnet;
53         int id;
54
55         _enter("");
56
57         idr_preload(gfp);
58         spin_lock(&rxrpc_conn_id_lock);
59
60         id = idr_alloc_cyclic(&rxrpc_client_conn_ids, conn,
61                               1, 0x40000000, GFP_NOWAIT);
62         if (id < 0)
63                 goto error;
64
65         spin_unlock(&rxrpc_conn_id_lock);
66         idr_preload_end();
67
68         conn->proto.epoch = rxnet->epoch;
69         conn->proto.cid = id << RXRPC_CIDSHIFT;
70         set_bit(RXRPC_CONN_HAS_IDR, &conn->flags);
71         _leave(" [CID %x]", conn->proto.cid);
72         return 0;
73
74 error:
75         spin_unlock(&rxrpc_conn_id_lock);
76         idr_preload_end();
77         _leave(" = %d", id);
78         return id;
79 }
80
81 /*
82  * Release a connection ID for a client connection from the global pool.
83  */
84 static void rxrpc_put_client_connection_id(struct rxrpc_connection *conn)
85 {
86         if (test_bit(RXRPC_CONN_HAS_IDR, &conn->flags)) {
87                 spin_lock(&rxrpc_conn_id_lock);
88                 idr_remove(&rxrpc_client_conn_ids,
89                            conn->proto.cid >> RXRPC_CIDSHIFT);
90                 spin_unlock(&rxrpc_conn_id_lock);
91         }
92 }
93
94 /*
95  * Destroy the client connection ID tree.
96  */
97 void rxrpc_destroy_client_conn_ids(void)
98 {
99         struct rxrpc_connection *conn;
100         int id;
101
102         if (!idr_is_empty(&rxrpc_client_conn_ids)) {
103                 idr_for_each_entry(&rxrpc_client_conn_ids, conn, id) {
104                         pr_err("AF_RXRPC: Leaked client conn %p {%d}\n",
105                                conn, atomic_read(&conn->usage));
106                 }
107                 BUG();
108         }
109
110         idr_destroy(&rxrpc_client_conn_ids);
111 }
112
113 /*
114  * Allocate a connection bundle.
115  */
116 static struct rxrpc_bundle *rxrpc_alloc_bundle(struct rxrpc_conn_parameters *cp,
117                                                gfp_t gfp)
118 {
119         struct rxrpc_bundle *bundle;
120
121         bundle = kzalloc(sizeof(*bundle), gfp);
122         if (bundle) {
123                 bundle->params = *cp;
124                 rxrpc_get_peer(bundle->params.peer);
125                 atomic_set(&bundle->usage, 1);
126                 spin_lock_init(&bundle->channel_lock);
127                 INIT_LIST_HEAD(&bundle->waiting_calls);
128         }
129         return bundle;
130 }
131
132 struct rxrpc_bundle *rxrpc_get_bundle(struct rxrpc_bundle *bundle)
133 {
134         atomic_inc(&bundle->usage);
135         return bundle;
136 }
137
138 void rxrpc_put_bundle(struct rxrpc_bundle *bundle)
139 {
140         unsigned int d = bundle->debug_id;
141         unsigned int u = atomic_dec_return(&bundle->usage);
142
143         _debug("PUT B=%x %u", d, u);
144         if (u == 0) {
145                 rxrpc_put_peer(bundle->params.peer);
146                 kfree(bundle);
147         }
148 }
149
150 /*
151  * Allocate a client connection.
152  */
153 static struct rxrpc_connection *
154 rxrpc_alloc_client_connection(struct rxrpc_bundle *bundle, gfp_t gfp)
155 {
156         struct rxrpc_connection *conn;
157         struct rxrpc_net *rxnet = bundle->params.local->rxnet;
158         int ret;
159
160         _enter("");
161
162         conn = rxrpc_alloc_connection(gfp);
163         if (!conn) {
164                 _leave(" = -ENOMEM");
165                 return ERR_PTR(-ENOMEM);
166         }
167
168         atomic_set(&conn->usage, 1);
169         conn->bundle            = bundle;
170         conn->params            = bundle->params;
171         conn->out_clientflag    = RXRPC_CLIENT_INITIATED;
172         conn->state             = RXRPC_CONN_CLIENT;
173         conn->service_id        = conn->params.service_id;
174
175         ret = rxrpc_get_client_connection_id(conn, gfp);
176         if (ret < 0)
177                 goto error_0;
178
179         ret = rxrpc_init_client_conn_security(conn);
180         if (ret < 0)
181                 goto error_1;
182
183         ret = conn->security->prime_packet_security(conn);
184         if (ret < 0)
185                 goto error_2;
186
187         atomic_inc(&rxnet->nr_conns);
188         write_lock(&rxnet->conn_lock);
189         list_add_tail(&conn->proc_link, &rxnet->conn_proc_list);
190         write_unlock(&rxnet->conn_lock);
191
192         rxrpc_get_bundle(bundle);
193         rxrpc_get_peer(conn->params.peer);
194         rxrpc_get_local(conn->params.local);
195         key_get(conn->params.key);
196
197         trace_rxrpc_conn(conn->debug_id, rxrpc_conn_new_client,
198                          atomic_read(&conn->usage),
199                          __builtin_return_address(0));
200
201         atomic_inc(&rxnet->nr_client_conns);
202         trace_rxrpc_client(conn, -1, rxrpc_client_alloc);
203         _leave(" = %p", conn);
204         return conn;
205
206 error_2:
207         conn->security->clear(conn);
208 error_1:
209         rxrpc_put_client_connection_id(conn);
210 error_0:
211         kfree(conn);
212         _leave(" = %d", ret);
213         return ERR_PTR(ret);
214 }
215
216 /*
217  * Determine if a connection may be reused.
218  */
219 static bool rxrpc_may_reuse_conn(struct rxrpc_connection *conn)
220 {
221         struct rxrpc_net *rxnet;
222         int id_cursor, id, distance, limit;
223
224         if (!conn)
225                 goto dont_reuse;
226
227         rxnet = conn->params.local->rxnet;
228         if (test_bit(RXRPC_CONN_DONT_REUSE, &conn->flags))
229                 goto dont_reuse;
230
231         if (conn->state != RXRPC_CONN_CLIENT ||
232             conn->proto.epoch != rxnet->epoch)
233                 goto mark_dont_reuse;
234
235         /* The IDR tree gets very expensive on memory if the connection IDs are
236          * widely scattered throughout the number space, so we shall want to
237          * kill off connections that, say, have an ID more than about four
238          * times the maximum number of client conns away from the current
239          * allocation point to try and keep the IDs concentrated.
240          */
241         id_cursor = idr_get_cursor(&rxrpc_client_conn_ids);
242         id = conn->proto.cid >> RXRPC_CIDSHIFT;
243         distance = id - id_cursor;
244         if (distance < 0)
245                 distance = -distance;
246         limit = max_t(unsigned long, atomic_read(&rxnet->nr_conns) * 4, 1024);
247         if (distance > limit)
248                 goto mark_dont_reuse;
249
250         return true;
251
252 mark_dont_reuse:
253         set_bit(RXRPC_CONN_DONT_REUSE, &conn->flags);
254 dont_reuse:
255         return false;
256 }
257
258 /*
259  * Look up the conn bundle that matches the connection parameters, adding it if
260  * it doesn't yet exist.
261  */
262 static struct rxrpc_bundle *rxrpc_look_up_bundle(struct rxrpc_conn_parameters *cp,
263                                                  gfp_t gfp)
264 {
265         static atomic_t rxrpc_bundle_id;
266         struct rxrpc_bundle *bundle, *candidate;
267         struct rxrpc_local *local = cp->local;
268         struct rb_node *p, **pp, *parent;
269         long diff;
270
271         _enter("{%px,%x,%u,%u}",
272                cp->peer, key_serial(cp->key), cp->security_level, cp->upgrade);
273
274         if (cp->exclusive)
275                 return rxrpc_alloc_bundle(cp, gfp);
276
277         /* First, see if the bundle is already there. */
278         _debug("search 1");
279         spin_lock(&local->client_bundles_lock);
280         p = local->client_bundles.rb_node;
281         while (p) {
282                 bundle = rb_entry(p, struct rxrpc_bundle, local_node);
283
284 #define cmp(X) ((long)bundle->params.X - (long)cp->X)
285                 diff = (cmp(peer) ?:
286                         cmp(key) ?:
287                         cmp(security_level) ?:
288                         cmp(upgrade));
289 #undef cmp
290                 if (diff < 0)
291                         p = p->rb_left;
292                 else if (diff > 0)
293                         p = p->rb_right;
294                 else
295                         goto found_bundle;
296         }
297         spin_unlock(&local->client_bundles_lock);
298         _debug("not found");
299
300         /* It wasn't.  We need to add one. */
301         candidate = rxrpc_alloc_bundle(cp, gfp);
302         if (!candidate)
303                 return NULL;
304
305         _debug("search 2");
306         spin_lock(&local->client_bundles_lock);
307         pp = &local->client_bundles.rb_node;
308         parent = NULL;
309         while (*pp) {
310                 parent = *pp;
311                 bundle = rb_entry(parent, struct rxrpc_bundle, local_node);
312
313 #define cmp(X) ((long)bundle->params.X - (long)cp->X)
314                 diff = (cmp(peer) ?:
315                         cmp(key) ?:
316                         cmp(security_level) ?:
317                         cmp(upgrade));
318 #undef cmp
319                 if (diff < 0)
320                         pp = &(*pp)->rb_left;
321                 else if (diff > 0)
322                         pp = &(*pp)->rb_right;
323                 else
324                         goto found_bundle_free;
325         }
326
327         _debug("new bundle");
328         candidate->debug_id = atomic_inc_return(&rxrpc_bundle_id);
329         rb_link_node(&candidate->local_node, parent, pp);
330         rb_insert_color(&candidate->local_node, &local->client_bundles);
331         rxrpc_get_bundle(candidate);
332         spin_unlock(&local->client_bundles_lock);
333         _leave(" = %u [new]", candidate->debug_id);
334         return candidate;
335
336 found_bundle_free:
337         kfree(candidate);
338 found_bundle:
339         rxrpc_get_bundle(bundle);
340         spin_unlock(&local->client_bundles_lock);
341         _leave(" = %u [found]", bundle->debug_id);
342         return bundle;
343 }
344
345 /*
346  * Create or find a client bundle to use for a call.
347  *
348  * If we return with a connection, the call will be on its waiting list.  It's
349  * left to the caller to assign a channel and wake up the call.
350  */
351 static struct rxrpc_bundle *rxrpc_prep_call(struct rxrpc_sock *rx,
352                                             struct rxrpc_call *call,
353                                             struct rxrpc_conn_parameters *cp,
354                                             struct sockaddr_rxrpc *srx,
355                                             gfp_t gfp)
356 {
357         struct rxrpc_bundle *bundle;
358
359         _enter("{%d,%lx},", call->debug_id, call->user_call_ID);
360
361         cp->peer = rxrpc_lookup_peer(rx, cp->local, srx, gfp);
362         if (!cp->peer)
363                 goto error;
364
365         call->cong_cwnd = cp->peer->cong_cwnd;
366         if (call->cong_cwnd >= call->cong_ssthresh)
367                 call->cong_mode = RXRPC_CALL_CONGEST_AVOIDANCE;
368         else
369                 call->cong_mode = RXRPC_CALL_SLOW_START;
370         if (cp->upgrade)
371                 __set_bit(RXRPC_CALL_UPGRADE, &call->flags);
372
373         /* Find the client connection bundle. */
374         bundle = rxrpc_look_up_bundle(cp, gfp);
375         if (!bundle)
376                 goto error;
377
378         /* Get this call queued.  Someone else may activate it whilst we're
379          * lining up a new connection, but that's fine.
380          */
381         spin_lock(&bundle->channel_lock);
382         list_add_tail(&call->chan_wait_link, &bundle->waiting_calls);
383         spin_unlock(&bundle->channel_lock);
384
385         _leave(" = [B=%x]", bundle->debug_id);
386         return bundle;
387
388 error:
389         _leave(" = -ENOMEM");
390         return ERR_PTR(-ENOMEM);
391 }
392
393 /*
394  * Allocate a new connection and add it into a bundle.
395  */
396 static void rxrpc_add_conn_to_bundle(struct rxrpc_bundle *bundle, gfp_t gfp)
397         __releases(bundle->channel_lock)
398 {
399         struct rxrpc_connection *candidate = NULL, *old = NULL;
400         bool conflict;
401         int i;
402
403         _enter("");
404
405         conflict = bundle->alloc_conn;
406         if (!conflict)
407                 bundle->alloc_conn = true;
408         spin_unlock(&bundle->channel_lock);
409         if (conflict) {
410                 _leave(" [conf]");
411                 return;
412         }
413
414         candidate = rxrpc_alloc_client_connection(bundle, gfp);
415
416         spin_lock(&bundle->channel_lock);
417         bundle->alloc_conn = false;
418
419         if (IS_ERR(candidate)) {
420                 bundle->alloc_error = PTR_ERR(candidate);
421                 spin_unlock(&bundle->channel_lock);
422                 _leave(" [err %ld]", PTR_ERR(candidate));
423                 return;
424         }
425
426         bundle->alloc_error = 0;
427
428         for (i = 0; i < ARRAY_SIZE(bundle->conns); i++) {
429                 unsigned int shift = i * RXRPC_MAXCALLS;
430                 int j;
431
432                 old = bundle->conns[i];
433                 if (!rxrpc_may_reuse_conn(old)) {
434                         if (old)
435                                 trace_rxrpc_client(old, -1, rxrpc_client_replace);
436                         candidate->bundle_shift = shift;
437                         bundle->conns[i] = candidate;
438                         for (j = 0; j < RXRPC_MAXCALLS; j++)
439                                 set_bit(shift + j, &bundle->avail_chans);
440                         candidate = NULL;
441                         break;
442                 }
443
444                 old = NULL;
445         }
446
447         spin_unlock(&bundle->channel_lock);
448
449         if (candidate) {
450                 _debug("discard C=%x", candidate->debug_id);
451                 trace_rxrpc_client(candidate, -1, rxrpc_client_duplicate);
452                 rxrpc_put_connection(candidate);
453         }
454
455         rxrpc_put_connection(old);
456         _leave("");
457 }
458
459 /*
460  * Add a connection to a bundle if there are no usable connections or we have
461  * connections waiting for extra capacity.
462  */
463 static void rxrpc_maybe_add_conn(struct rxrpc_bundle *bundle, gfp_t gfp)
464 {
465         struct rxrpc_call *call;
466         int i, usable;
467
468         _enter("");
469
470         spin_lock(&bundle->channel_lock);
471
472         /* See if there are any usable connections. */
473         usable = 0;
474         for (i = 0; i < ARRAY_SIZE(bundle->conns); i++)
475                 if (rxrpc_may_reuse_conn(bundle->conns[i]))
476                         usable++;
477
478         if (!usable && !list_empty(&bundle->waiting_calls)) {
479                 call = list_first_entry(&bundle->waiting_calls,
480                                         struct rxrpc_call, chan_wait_link);
481                 if (test_bit(RXRPC_CALL_UPGRADE, &call->flags))
482                         bundle->try_upgrade = true;
483         }
484
485         if (!usable)
486                 goto alloc_conn;
487
488         if (!bundle->avail_chans &&
489             !bundle->try_upgrade &&
490             !list_empty(&bundle->waiting_calls) &&
491             usable < ARRAY_SIZE(bundle->conns))
492                 goto alloc_conn;
493
494         spin_unlock(&bundle->channel_lock);
495         _leave("");
496         return;
497
498 alloc_conn:
499         return rxrpc_add_conn_to_bundle(bundle, gfp);
500 }
501
502 /*
503  * Assign a channel to the call at the front of the queue and wake the call up.
504  * We don't increment the callNumber counter until this number has been exposed
505  * to the world.
506  */
507 static void rxrpc_activate_one_channel(struct rxrpc_connection *conn,
508                                        unsigned int channel)
509 {
510         struct rxrpc_channel *chan = &conn->channels[channel];
511         struct rxrpc_bundle *bundle = conn->bundle;
512         struct rxrpc_call *call = list_entry(bundle->waiting_calls.next,
513                                              struct rxrpc_call, chan_wait_link);
514         u32 call_id = chan->call_counter + 1;
515
516         _enter("C=%x,%u", conn->debug_id, channel);
517
518         trace_rxrpc_client(conn, channel, rxrpc_client_chan_activate);
519
520         /* Cancel the final ACK on the previous call if it hasn't been sent yet
521          * as the DATA packet will implicitly ACK it.
522          */
523         clear_bit(RXRPC_CONN_FINAL_ACK_0 + channel, &conn->flags);
524         clear_bit(conn->bundle_shift + channel, &bundle->avail_chans);
525
526         rxrpc_see_call(call);
527         list_del_init(&call->chan_wait_link);
528         call->peer      = rxrpc_get_peer(conn->params.peer);
529         call->conn      = rxrpc_get_connection(conn);
530         call->cid       = conn->proto.cid | channel;
531         call->call_id   = call_id;
532         call->security  = conn->security;
533         call->security_ix = conn->security_ix;
534         call->service_id = conn->service_id;
535
536         trace_rxrpc_connect_call(call);
537         _net("CONNECT call %08x:%08x as call %d on conn %d",
538              call->cid, call->call_id, call->debug_id, conn->debug_id);
539
540         write_lock_bh(&call->state_lock);
541         call->state = RXRPC_CALL_CLIENT_SEND_REQUEST;
542         write_unlock_bh(&call->state_lock);
543
544         /* Paired with the read barrier in rxrpc_connect_call().  This orders
545          * cid and epoch in the connection wrt to call_id without the need to
546          * take the channel_lock.
547          *
548          * We provisionally assign a callNumber at this point, but we don't
549          * confirm it until the call is about to be exposed.
550          *
551          * TODO: Pair with a barrier in the data_ready handler when that looks
552          * at the call ID through a connection channel.
553          */
554         smp_wmb();
555
556         chan->call_id           = call_id;
557         chan->call_debug_id     = call->debug_id;
558         rcu_assign_pointer(chan->call, call);
559         wake_up(&call->waitq);
560 }
561
562 /*
563  * Remove a connection from the idle list if it's on it.
564  */
565 static void rxrpc_unidle_conn(struct rxrpc_bundle *bundle, struct rxrpc_connection *conn)
566 {
567         struct rxrpc_net *rxnet = bundle->params.local->rxnet;
568         bool drop_ref;
569
570         if (!list_empty(&conn->cache_link)) {
571                 drop_ref = false;
572                 spin_lock(&rxnet->client_conn_cache_lock);
573                 if (!list_empty(&conn->cache_link)) {
574                         list_del_init(&conn->cache_link);
575                         drop_ref = true;
576                 }
577                 spin_unlock(&rxnet->client_conn_cache_lock);
578                 if (drop_ref)
579                         rxrpc_put_connection(conn);
580         }
581 }
582
583 /*
584  * Assign channels and callNumbers to waiting calls with channel_lock
585  * held by caller.
586  */
587 static void rxrpc_activate_channels_locked(struct rxrpc_bundle *bundle)
588 {
589         struct rxrpc_connection *conn;
590         unsigned long avail, mask;
591         unsigned int channel, slot;
592
593         if (bundle->try_upgrade)
594                 mask = 1;
595         else
596                 mask = ULONG_MAX;
597
598         while (!list_empty(&bundle->waiting_calls)) {
599                 avail = bundle->avail_chans & mask;
600                 if (!avail)
601                         break;
602                 channel = __ffs(avail);
603                 clear_bit(channel, &bundle->avail_chans);
604
605                 slot = channel / RXRPC_MAXCALLS;
606                 conn = bundle->conns[slot];
607                 if (!conn)
608                         break;
609
610                 if (bundle->try_upgrade)
611                         set_bit(RXRPC_CONN_PROBING_FOR_UPGRADE, &conn->flags);
612                 rxrpc_unidle_conn(bundle, conn);
613
614                 channel &= (RXRPC_MAXCALLS - 1);
615                 conn->act_chans |= 1 << channel;
616                 rxrpc_activate_one_channel(conn, channel);
617         }
618 }
619
620 /*
621  * Assign channels and callNumbers to waiting calls.
622  */
623 static void rxrpc_activate_channels(struct rxrpc_bundle *bundle)
624 {
625         _enter("B=%x", bundle->debug_id);
626
627         trace_rxrpc_client(NULL, -1, rxrpc_client_activate_chans);
628
629         if (!bundle->avail_chans)
630                 return;
631
632         spin_lock(&bundle->channel_lock);
633         rxrpc_activate_channels_locked(bundle);
634         spin_unlock(&bundle->channel_lock);
635         _leave("");
636 }
637
638 /*
639  * Wait for a callNumber and a channel to be granted to a call.
640  */
641 static int rxrpc_wait_for_channel(struct rxrpc_bundle *bundle,
642                                   struct rxrpc_call *call, gfp_t gfp)
643 {
644         DECLARE_WAITQUEUE(myself, current);
645         int ret = 0;
646
647         _enter("%d", call->debug_id);
648
649         if (!gfpflags_allow_blocking(gfp)) {
650                 rxrpc_maybe_add_conn(bundle, gfp);
651                 rxrpc_activate_channels(bundle);
652                 ret = bundle->alloc_error ?: -EAGAIN;
653                 goto out;
654         }
655
656         add_wait_queue_exclusive(&call->waitq, &myself);
657         for (;;) {
658                 rxrpc_maybe_add_conn(bundle, gfp);
659                 rxrpc_activate_channels(bundle);
660                 ret = bundle->alloc_error;
661                 if (ret < 0)
662                         break;
663
664                 switch (call->interruptibility) {
665                 case RXRPC_INTERRUPTIBLE:
666                 case RXRPC_PREINTERRUPTIBLE:
667                         set_current_state(TASK_INTERRUPTIBLE);
668                         break;
669                 case RXRPC_UNINTERRUPTIBLE:
670                 default:
671                         set_current_state(TASK_UNINTERRUPTIBLE);
672                         break;
673                 }
674                 if (READ_ONCE(call->state) != RXRPC_CALL_CLIENT_AWAIT_CONN)
675                         break;
676                 if ((call->interruptibility == RXRPC_INTERRUPTIBLE ||
677                      call->interruptibility == RXRPC_PREINTERRUPTIBLE) &&
678                     signal_pending(current)) {
679                         ret = -ERESTARTSYS;
680                         break;
681                 }
682                 schedule();
683         }
684         remove_wait_queue(&call->waitq, &myself);
685         __set_current_state(TASK_RUNNING);
686
687 out:
688         _leave(" = %d", ret);
689         return ret;
690 }
691
692 /*
693  * find a connection for a call
694  * - called in process context with IRQs enabled
695  */
696 int rxrpc_connect_call(struct rxrpc_sock *rx,
697                        struct rxrpc_call *call,
698                        struct rxrpc_conn_parameters *cp,
699                        struct sockaddr_rxrpc *srx,
700                        gfp_t gfp)
701 {
702         struct rxrpc_bundle *bundle;
703         struct rxrpc_net *rxnet = cp->local->rxnet;
704         int ret = 0;
705
706         _enter("{%d,%lx},", call->debug_id, call->user_call_ID);
707
708         rxrpc_discard_expired_client_conns(&rxnet->client_conn_reaper);
709
710         bundle = rxrpc_prep_call(rx, call, cp, srx, gfp);
711         if (IS_ERR(bundle)) {
712                 ret = PTR_ERR(bundle);
713                 goto out;
714         }
715
716         if (call->state == RXRPC_CALL_CLIENT_AWAIT_CONN) {
717                 ret = rxrpc_wait_for_channel(bundle, call, gfp);
718                 if (ret < 0)
719                         goto wait_failed;
720         }
721
722 granted_channel:
723         /* Paired with the write barrier in rxrpc_activate_one_channel(). */
724         smp_rmb();
725
726 out_put_bundle:
727         rxrpc_put_bundle(bundle);
728 out:
729         _leave(" = %d", ret);
730         return ret;
731
732 wait_failed:
733         spin_lock(&bundle->channel_lock);
734         list_del_init(&call->chan_wait_link);
735         spin_unlock(&bundle->channel_lock);
736
737         if (call->state != RXRPC_CALL_CLIENT_AWAIT_CONN) {
738                 ret = 0;
739                 goto granted_channel;
740         }
741
742         trace_rxrpc_client(call->conn, ret, rxrpc_client_chan_wait_failed);
743         rxrpc_set_call_completion(call, RXRPC_CALL_LOCAL_ERROR, 0, ret);
744         rxrpc_disconnect_client_call(bundle, call);
745         goto out_put_bundle;
746 }
747
748 /*
749  * Note that a call, and thus a connection, is about to be exposed to the
750  * world.
751  */
752 void rxrpc_expose_client_call(struct rxrpc_call *call)
753 {
754         unsigned int channel = call->cid & RXRPC_CHANNELMASK;
755         struct rxrpc_connection *conn = call->conn;
756         struct rxrpc_channel *chan = &conn->channels[channel];
757
758         if (!test_and_set_bit(RXRPC_CALL_EXPOSED, &call->flags)) {
759                 /* Mark the call ID as being used.  If the callNumber counter
760                  * exceeds ~2 billion, we kill the connection after its
761                  * outstanding calls have finished so that the counter doesn't
762                  * wrap.
763                  */
764                 chan->call_counter++;
765                 if (chan->call_counter >= INT_MAX)
766                         set_bit(RXRPC_CONN_DONT_REUSE, &conn->flags);
767                 trace_rxrpc_client(conn, channel, rxrpc_client_exposed);
768         }
769 }
770
771 /*
772  * Set the reap timer.
773  */
774 static void rxrpc_set_client_reap_timer(struct rxrpc_net *rxnet)
775 {
776         if (!rxnet->kill_all_client_conns) {
777                 unsigned long now = jiffies;
778                 unsigned long reap_at = now + rxrpc_conn_idle_client_expiry;
779
780                 if (rxnet->live)
781                         timer_reduce(&rxnet->client_conn_reap_timer, reap_at);
782         }
783 }
784
785 /*
786  * Disconnect a client call.
787  */
788 void rxrpc_disconnect_client_call(struct rxrpc_bundle *bundle, struct rxrpc_call *call)
789 {
790         struct rxrpc_connection *conn;
791         struct rxrpc_channel *chan = NULL;
792         struct rxrpc_net *rxnet = bundle->params.local->rxnet;
793         unsigned int channel;
794         bool may_reuse;
795         u32 cid;
796
797         _enter("c=%x", call->debug_id);
798
799         spin_lock(&bundle->channel_lock);
800         set_bit(RXRPC_CALL_DISCONNECTED, &call->flags);
801
802         /* Calls that have never actually been assigned a channel can simply be
803          * discarded.
804          */
805         conn = call->conn;
806         if (!conn) {
807                 _debug("call is waiting");
808                 ASSERTCMP(call->call_id, ==, 0);
809                 ASSERT(!test_bit(RXRPC_CALL_EXPOSED, &call->flags));
810                 list_del_init(&call->chan_wait_link);
811                 goto out;
812         }
813
814         cid = call->cid;
815         channel = cid & RXRPC_CHANNELMASK;
816         chan = &conn->channels[channel];
817         trace_rxrpc_client(conn, channel, rxrpc_client_chan_disconnect);
818
819         if (rcu_access_pointer(chan->call) != call) {
820                 spin_unlock(&bundle->channel_lock);
821                 BUG();
822         }
823
824         may_reuse = rxrpc_may_reuse_conn(conn);
825
826         /* If a client call was exposed to the world, we save the result for
827          * retransmission.
828          *
829          * We use a barrier here so that the call number and abort code can be
830          * read without needing to take a lock.
831          *
832          * TODO: Make the incoming packet handler check this and handle
833          * terminal retransmission without requiring access to the call.
834          */
835         if (test_bit(RXRPC_CALL_EXPOSED, &call->flags)) {
836                 _debug("exposed %u,%u", call->call_id, call->abort_code);
837                 __rxrpc_disconnect_call(conn, call);
838
839                 if (test_and_clear_bit(RXRPC_CONN_PROBING_FOR_UPGRADE, &conn->flags)) {
840                         trace_rxrpc_client(conn, channel, rxrpc_client_to_active);
841                         bundle->try_upgrade = false;
842                         if (may_reuse)
843                                 rxrpc_activate_channels_locked(bundle);
844                 }
845
846         }
847
848         /* See if we can pass the channel directly to another call. */
849         if (may_reuse && !list_empty(&bundle->waiting_calls)) {
850                 trace_rxrpc_client(conn, channel, rxrpc_client_chan_pass);
851                 rxrpc_activate_one_channel(conn, channel);
852                 goto out;
853         }
854
855         /* Schedule the final ACK to be transmitted in a short while so that it
856          * can be skipped if we find a follow-on call.  The first DATA packet
857          * of the follow on call will implicitly ACK this call.
858          */
859         if (call->completion == RXRPC_CALL_SUCCEEDED &&
860             test_bit(RXRPC_CALL_EXPOSED, &call->flags)) {
861                 unsigned long final_ack_at = jiffies + 2;
862
863                 WRITE_ONCE(chan->final_ack_at, final_ack_at);
864                 smp_wmb(); /* vs rxrpc_process_delayed_final_acks() */
865                 set_bit(RXRPC_CONN_FINAL_ACK_0 + channel, &conn->flags);
866                 rxrpc_reduce_conn_timer(conn, final_ack_at);
867         }
868
869         /* Deactivate the channel. */
870         rcu_assign_pointer(chan->call, NULL);
871         set_bit(conn->bundle_shift + channel, &conn->bundle->avail_chans);
872         conn->act_chans &= ~(1 << channel);
873
874         /* If no channels remain active, then put the connection on the idle
875          * list for a short while.  Give it a ref to stop it going away if it
876          * becomes unbundled.
877          */
878         if (!conn->act_chans) {
879                 trace_rxrpc_client(conn, channel, rxrpc_client_to_idle);
880                 conn->idle_timestamp = jiffies;
881
882                 rxrpc_get_connection(conn);
883                 spin_lock(&rxnet->client_conn_cache_lock);
884                 list_move_tail(&conn->cache_link, &rxnet->idle_client_conns);
885                 spin_unlock(&rxnet->client_conn_cache_lock);
886
887                 rxrpc_set_client_reap_timer(rxnet);
888         }
889
890 out:
891         spin_unlock(&bundle->channel_lock);
892         _leave("");
893         return;
894 }
895
896 /*
897  * Remove a connection from a bundle.
898  */
899 static void rxrpc_unbundle_conn(struct rxrpc_connection *conn)
900 {
901         struct rxrpc_bundle *bundle = conn->bundle;
902         struct rxrpc_local *local = bundle->params.local;
903         unsigned int bindex;
904         bool need_drop = false, need_put = false;
905         int i;
906
907         _enter("C=%x", conn->debug_id);
908
909         if (conn->flags & RXRPC_CONN_FINAL_ACK_MASK)
910                 rxrpc_process_delayed_final_acks(conn, true);
911
912         spin_lock(&bundle->channel_lock);
913         bindex = conn->bundle_shift / RXRPC_MAXCALLS;
914         if (bundle->conns[bindex] == conn) {
915                 _debug("clear slot %u", bindex);
916                 bundle->conns[bindex] = NULL;
917                 for (i = 0; i < RXRPC_MAXCALLS; i++)
918                         clear_bit(conn->bundle_shift + i, &bundle->avail_chans);
919                 need_drop = true;
920         }
921         spin_unlock(&bundle->channel_lock);
922
923         /* If there are no more connections, remove the bundle */
924         if (!bundle->avail_chans) {
925                 _debug("maybe unbundle");
926                 spin_lock(&local->client_bundles_lock);
927
928                 for (i = 0; i < ARRAY_SIZE(bundle->conns); i++)
929                         if (bundle->conns[i])
930                                 break;
931                 if (i == ARRAY_SIZE(bundle->conns) && !bundle->params.exclusive) {
932                         _debug("erase bundle");
933                         rb_erase(&bundle->local_node, &local->client_bundles);
934                         need_put = true;
935                 }
936
937                 spin_unlock(&local->client_bundles_lock);
938                 if (need_put)
939                         rxrpc_put_bundle(bundle);
940         }
941
942         if (need_drop)
943                 rxrpc_put_connection(conn);
944         _leave("");
945 }
946
947 /*
948  * Clean up a dead client connection.
949  */
950 static void rxrpc_kill_client_conn(struct rxrpc_connection *conn)
951 {
952         struct rxrpc_local *local = conn->params.local;
953         struct rxrpc_net *rxnet = local->rxnet;
954
955         _enter("C=%x", conn->debug_id);
956
957         trace_rxrpc_client(conn, -1, rxrpc_client_cleanup);
958         atomic_dec(&rxnet->nr_client_conns);
959
960         rxrpc_put_client_connection_id(conn);
961         rxrpc_kill_connection(conn);
962 }
963
964 /*
965  * Clean up a dead client connections.
966  */
967 void rxrpc_put_client_conn(struct rxrpc_connection *conn)
968 {
969         const void *here = __builtin_return_address(0);
970         unsigned int debug_id = conn->debug_id;
971         int n;
972
973         n = atomic_dec_return(&conn->usage);
974         trace_rxrpc_conn(debug_id, rxrpc_conn_put_client, n, here);
975         if (n <= 0) {
976                 ASSERTCMP(n, >=, 0);
977                 rxrpc_kill_client_conn(conn);
978         }
979 }
980
981 /*
982  * Discard expired client connections from the idle list.  Each conn in the
983  * idle list has been exposed and holds an extra ref because of that.
984  *
985  * This may be called from conn setup or from a work item so cannot be
986  * considered non-reentrant.
987  */
988 void rxrpc_discard_expired_client_conns(struct work_struct *work)
989 {
990         struct rxrpc_connection *conn;
991         struct rxrpc_net *rxnet =
992                 container_of(work, struct rxrpc_net, client_conn_reaper);
993         unsigned long expiry, conn_expires_at, now;
994         unsigned int nr_conns;
995
996         _enter("");
997
998         if (list_empty(&rxnet->idle_client_conns)) {
999                 _leave(" [empty]");
1000                 return;
1001         }
1002
1003         /* Don't double up on the discarding */
1004         if (!spin_trylock(&rxnet->client_conn_discard_lock)) {
1005                 _leave(" [already]");
1006                 return;
1007         }
1008
1009         /* We keep an estimate of what the number of conns ought to be after
1010          * we've discarded some so that we don't overdo the discarding.
1011          */
1012         nr_conns = atomic_read(&rxnet->nr_client_conns);
1013
1014 next:
1015         spin_lock(&rxnet->client_conn_cache_lock);
1016
1017         if (list_empty(&rxnet->idle_client_conns))
1018                 goto out;
1019
1020         conn = list_entry(rxnet->idle_client_conns.next,
1021                           struct rxrpc_connection, cache_link);
1022
1023         if (!rxnet->kill_all_client_conns) {
1024                 /* If the number of connections is over the reap limit, we
1025                  * expedite discard by reducing the expiry timeout.  We must,
1026                  * however, have at least a short grace period to be able to do
1027                  * final-ACK or ABORT retransmission.
1028                  */
1029                 expiry = rxrpc_conn_idle_client_expiry;
1030                 if (nr_conns > rxrpc_reap_client_connections)
1031                         expiry = rxrpc_conn_idle_client_fast_expiry;
1032                 if (conn->params.local->service_closed)
1033                         expiry = rxrpc_closed_conn_expiry * HZ;
1034
1035                 conn_expires_at = conn->idle_timestamp + expiry;
1036
1037                 now = READ_ONCE(jiffies);
1038                 if (time_after(conn_expires_at, now))
1039                         goto not_yet_expired;
1040         }
1041
1042         trace_rxrpc_client(conn, -1, rxrpc_client_discard);
1043         list_del_init(&conn->cache_link);
1044
1045         spin_unlock(&rxnet->client_conn_cache_lock);
1046
1047         rxrpc_unbundle_conn(conn);
1048         rxrpc_put_connection(conn); /* Drop the ->cache_link ref */
1049
1050         nr_conns--;
1051         goto next;
1052
1053 not_yet_expired:
1054         /* The connection at the front of the queue hasn't yet expired, so
1055          * schedule the work item for that point if we discarded something.
1056          *
1057          * We don't worry if the work item is already scheduled - it can look
1058          * after rescheduling itself at a later time.  We could cancel it, but
1059          * then things get messier.
1060          */
1061         _debug("not yet");
1062         if (!rxnet->kill_all_client_conns)
1063                 timer_reduce(&rxnet->client_conn_reap_timer, conn_expires_at);
1064
1065 out:
1066         spin_unlock(&rxnet->client_conn_cache_lock);
1067         spin_unlock(&rxnet->client_conn_discard_lock);
1068         _leave("");
1069 }
1070
1071 /*
1072  * Preemptively destroy all the client connection records rather than waiting
1073  * for them to time out
1074  */
1075 void rxrpc_destroy_all_client_connections(struct rxrpc_net *rxnet)
1076 {
1077         _enter("");
1078
1079         spin_lock(&rxnet->client_conn_cache_lock);
1080         rxnet->kill_all_client_conns = true;
1081         spin_unlock(&rxnet->client_conn_cache_lock);
1082
1083         del_timer_sync(&rxnet->client_conn_reap_timer);
1084
1085         if (!rxrpc_queue_work(&rxnet->client_conn_reaper))
1086                 _debug("destroy: queue failed");
1087
1088         _leave("");
1089 }
1090
1091 /*
1092  * Clean up the client connections on a local endpoint.
1093  */
1094 void rxrpc_clean_up_local_conns(struct rxrpc_local *local)
1095 {
1096         struct rxrpc_connection *conn, *tmp;
1097         struct rxrpc_net *rxnet = local->rxnet;
1098         LIST_HEAD(graveyard);
1099
1100         _enter("");
1101
1102         spin_lock(&rxnet->client_conn_cache_lock);
1103
1104         list_for_each_entry_safe(conn, tmp, &rxnet->idle_client_conns,
1105                                  cache_link) {
1106                 if (conn->params.local == local) {
1107                         trace_rxrpc_client(conn, -1, rxrpc_client_discard);
1108                         list_move(&conn->cache_link, &graveyard);
1109                 }
1110         }
1111
1112         spin_unlock(&rxnet->client_conn_cache_lock);
1113
1114         while (!list_empty(&graveyard)) {
1115                 conn = list_entry(graveyard.next,
1116                                   struct rxrpc_connection, cache_link);
1117                 list_del_init(&conn->cache_link);
1118                 rxrpc_unbundle_conn(conn);
1119                 rxrpc_put_connection(conn);
1120         }
1121
1122         _leave(" [culled]");
1123 }