Merge branch 'for-4.20' of https://git.kernel.org/pub/scm/linux/kernel/git/broonie...
[sfrench/cifs-2.6.git] / net / rxrpc / conn_client.c
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  */
94 DEFINE_IDR(rxrpc_client_conn_ids);
95 static DEFINE_SPINLOCK(rxrpc_conn_id_lock);
96
97 static 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  */
105 static 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
130 error:
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  */
140 static 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  */
153 void 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  */
172 static struct rxrpc_connection *
173 rxrpc_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
226 error_2:
227         conn->security->clear(conn);
228 error_1:
229         rxrpc_put_client_connection_id(conn);
230 error_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  */
239 static 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
267 mark_dont_reuse:
268         set_bit(RXRPC_CONN_DONT_REUSE, &conn->flags);
269 dont_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  */
279 static int rxrpc_get_client_conn(struct rxrpc_sock *rx,
280                                  struct rxrpc_call *call,
281                                  struct rxrpc_conn_parameters *cp,
282                                  struct sockaddr_rxrpc *srx,
283                                  gfp_t gfp)
284 {
285         struct rxrpc_connection *conn, *candidate = NULL;
286         struct rxrpc_local *local = cp->local;
287         struct rb_node *p, **pp, *parent;
288         long diff;
289         int ret = -ENOMEM;
290
291         _enter("{%d,%lx},", call->debug_id, call->user_call_ID);
292
293         cp->peer = rxrpc_lookup_peer(rx, cp->local, srx, gfp);
294         if (!cp->peer)
295                 goto error;
296
297         call->cong_cwnd = cp->peer->cong_cwnd;
298         if (call->cong_cwnd >= call->cong_ssthresh)
299                 call->cong_mode = RXRPC_CALL_CONGEST_AVOIDANCE;
300         else
301                 call->cong_mode = RXRPC_CALL_SLOW_START;
302
303         /* If the connection is not meant to be exclusive, search the available
304          * connections to see if the connection we want to use already exists.
305          */
306         if (!cp->exclusive) {
307                 _debug("search 1");
308                 spin_lock(&local->client_conns_lock);
309                 p = local->client_conns.rb_node;
310                 while (p) {
311                         conn = rb_entry(p, struct rxrpc_connection, client_node);
312
313 #define cmp(X) ((long)conn->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                                 p = p->rb_left;
321                         } else if (diff > 0) {
322                                 p = p->rb_right;
323                         } else {
324                                 if (rxrpc_may_reuse_conn(conn) &&
325                                     rxrpc_get_connection_maybe(conn))
326                                         goto found_extant_conn;
327                                 /* The connection needs replacing.  It's better
328                                  * to effect that when we have something to
329                                  * replace it with so that we don't have to
330                                  * rebalance the tree twice.
331                                  */
332                                 break;
333                         }
334                 }
335                 spin_unlock(&local->client_conns_lock);
336         }
337
338         /* There wasn't a connection yet or we need an exclusive connection.
339          * We need to create a candidate and then potentially redo the search
340          * in case we're racing with another thread also trying to connect on a
341          * shareable connection.
342          */
343         _debug("new conn");
344         candidate = rxrpc_alloc_client_connection(cp, gfp);
345         if (IS_ERR(candidate)) {
346                 ret = PTR_ERR(candidate);
347                 goto error_peer;
348         }
349
350         /* Add the call to the new connection's waiting list in case we're
351          * going to have to wait for the connection to come live.  It's our
352          * connection, so we want first dibs on the channel slots.  We would
353          * normally have to take channel_lock but we do this before anyone else
354          * can see the connection.
355          */
356         list_add_tail(&call->chan_wait_link, &candidate->waiting_calls);
357
358         if (cp->exclusive) {
359                 call->conn = candidate;
360                 call->security_ix = candidate->security_ix;
361                 call->service_id = candidate->service_id;
362                 _leave(" = 0 [exclusive %d]", candidate->debug_id);
363                 return 0;
364         }
365
366         /* Publish the new connection for userspace to find.  We need to redo
367          * the search before doing this lest we race with someone else adding a
368          * conflicting instance.
369          */
370         _debug("search 2");
371         spin_lock(&local->client_conns_lock);
372
373         pp = &local->client_conns.rb_node;
374         parent = NULL;
375         while (*pp) {
376                 parent = *pp;
377                 conn = rb_entry(parent, struct rxrpc_connection, client_node);
378
379 #define cmp(X) ((long)conn->params.X - (long)candidate->params.X)
380                 diff = (cmp(peer) ?:
381                         cmp(key) ?:
382                         cmp(security_level) ?:
383                         cmp(upgrade));
384 #undef cmp
385                 if (diff < 0) {
386                         pp = &(*pp)->rb_left;
387                 } else if (diff > 0) {
388                         pp = &(*pp)->rb_right;
389                 } else {
390                         if (rxrpc_may_reuse_conn(conn) &&
391                             rxrpc_get_connection_maybe(conn))
392                                 goto found_extant_conn;
393                         /* The old connection is from an outdated epoch. */
394                         _debug("replace conn");
395                         clear_bit(RXRPC_CONN_IN_CLIENT_CONNS, &conn->flags);
396                         rb_replace_node(&conn->client_node,
397                                         &candidate->client_node,
398                                         &local->client_conns);
399                         trace_rxrpc_client(conn, -1, rxrpc_client_replace);
400                         goto candidate_published;
401                 }
402         }
403
404         _debug("new conn");
405         rb_link_node(&candidate->client_node, parent, pp);
406         rb_insert_color(&candidate->client_node, &local->client_conns);
407
408 candidate_published:
409         set_bit(RXRPC_CONN_IN_CLIENT_CONNS, &candidate->flags);
410         call->conn = candidate;
411         call->security_ix = candidate->security_ix;
412         call->service_id = candidate->service_id;
413         spin_unlock(&local->client_conns_lock);
414         _leave(" = 0 [new %d]", candidate->debug_id);
415         return 0;
416
417         /* We come here if we found a suitable connection already in existence.
418          * Discard any candidate we may have allocated, and try to get a
419          * channel on this one.
420          */
421 found_extant_conn:
422         _debug("found conn");
423         spin_unlock(&local->client_conns_lock);
424
425         if (candidate) {
426                 trace_rxrpc_client(candidate, -1, rxrpc_client_duplicate);
427                 rxrpc_put_connection(candidate);
428                 candidate = NULL;
429         }
430
431         spin_lock(&conn->channel_lock);
432         call->conn = conn;
433         call->security_ix = conn->security_ix;
434         call->service_id = conn->service_id;
435         list_add(&call->chan_wait_link, &conn->waiting_calls);
436         spin_unlock(&conn->channel_lock);
437         _leave(" = 0 [extant %d]", conn->debug_id);
438         return 0;
439
440 error_peer:
441         rxrpc_put_peer(cp->peer);
442         cp->peer = NULL;
443 error:
444         _leave(" = %d", ret);
445         return ret;
446 }
447
448 /*
449  * Activate a connection.
450  */
451 static void rxrpc_activate_conn(struct rxrpc_net *rxnet,
452                                 struct rxrpc_connection *conn)
453 {
454         if (test_bit(RXRPC_CONN_PROBING_FOR_UPGRADE, &conn->flags)) {
455                 trace_rxrpc_client(conn, -1, rxrpc_client_to_upgrade);
456                 conn->cache_state = RXRPC_CONN_CLIENT_UPGRADE;
457         } else {
458                 trace_rxrpc_client(conn, -1, rxrpc_client_to_active);
459                 conn->cache_state = RXRPC_CONN_CLIENT_ACTIVE;
460         }
461         rxnet->nr_active_client_conns++;
462         list_move_tail(&conn->cache_link, &rxnet->active_client_conns);
463 }
464
465 /*
466  * Attempt to animate a connection for a new call.
467  *
468  * If it's not exclusive, the connection is in the endpoint tree, and we're in
469  * the conn's list of those waiting to grab a channel.  There is, however, a
470  * limit on the number of live connections allowed at any one time, so we may
471  * have to wait for capacity to become available.
472  *
473  * Note that a connection on the waiting queue might *also* have active
474  * channels if it has been culled to make space and then re-requested by a new
475  * call.
476  */
477 static void rxrpc_animate_client_conn(struct rxrpc_net *rxnet,
478                                       struct rxrpc_connection *conn)
479 {
480         unsigned int nr_conns;
481
482         _enter("%d,%d", conn->debug_id, conn->cache_state);
483
484         if (conn->cache_state == RXRPC_CONN_CLIENT_ACTIVE ||
485             conn->cache_state == RXRPC_CONN_CLIENT_UPGRADE)
486                 goto out;
487
488         spin_lock(&rxnet->client_conn_cache_lock);
489
490         nr_conns = rxnet->nr_client_conns;
491         if (!test_and_set_bit(RXRPC_CONN_COUNTED, &conn->flags)) {
492                 trace_rxrpc_client(conn, -1, rxrpc_client_count);
493                 rxnet->nr_client_conns = nr_conns + 1;
494         }
495
496         switch (conn->cache_state) {
497         case RXRPC_CONN_CLIENT_ACTIVE:
498         case RXRPC_CONN_CLIENT_UPGRADE:
499         case RXRPC_CONN_CLIENT_WAITING:
500                 break;
501
502         case RXRPC_CONN_CLIENT_INACTIVE:
503         case RXRPC_CONN_CLIENT_CULLED:
504         case RXRPC_CONN_CLIENT_IDLE:
505                 if (nr_conns >= rxrpc_max_client_connections)
506                         goto wait_for_capacity;
507                 goto activate_conn;
508
509         default:
510                 BUG();
511         }
512
513 out_unlock:
514         spin_unlock(&rxnet->client_conn_cache_lock);
515 out:
516         _leave(" [%d]", conn->cache_state);
517         return;
518
519 activate_conn:
520         _debug("activate");
521         rxrpc_activate_conn(rxnet, conn);
522         goto out_unlock;
523
524 wait_for_capacity:
525         _debug("wait");
526         trace_rxrpc_client(conn, -1, rxrpc_client_to_waiting);
527         conn->cache_state = RXRPC_CONN_CLIENT_WAITING;
528         list_move_tail(&conn->cache_link, &rxnet->waiting_client_conns);
529         goto out_unlock;
530 }
531
532 /*
533  * Deactivate a channel.
534  */
535 static void rxrpc_deactivate_one_channel(struct rxrpc_connection *conn,
536                                          unsigned int channel)
537 {
538         struct rxrpc_channel *chan = &conn->channels[channel];
539
540         rcu_assign_pointer(chan->call, NULL);
541         conn->active_chans &= ~(1 << channel);
542 }
543
544 /*
545  * Assign a channel to the call at the front of the queue and wake the call up.
546  * We don't increment the callNumber counter until this number has been exposed
547  * to the world.
548  */
549 static void rxrpc_activate_one_channel(struct rxrpc_connection *conn,
550                                        unsigned int channel)
551 {
552         struct rxrpc_channel *chan = &conn->channels[channel];
553         struct rxrpc_call *call = list_entry(conn->waiting_calls.next,
554                                              struct rxrpc_call, chan_wait_link);
555         u32 call_id = chan->call_counter + 1;
556
557         trace_rxrpc_client(conn, channel, rxrpc_client_chan_activate);
558
559         /* Cancel the final ACK on the previous call if it hasn't been sent yet
560          * as the DATA packet will implicitly ACK it.
561          */
562         clear_bit(RXRPC_CONN_FINAL_ACK_0 + channel, &conn->flags);
563
564         write_lock_bh(&call->state_lock);
565         if (!test_bit(RXRPC_CALL_TX_LASTQ, &call->flags))
566                 call->state = RXRPC_CALL_CLIENT_SEND_REQUEST;
567         else
568                 call->state = RXRPC_CALL_CLIENT_AWAIT_REPLY;
569         write_unlock_bh(&call->state_lock);
570
571         rxrpc_see_call(call);
572         list_del_init(&call->chan_wait_link);
573         conn->active_chans |= 1 << channel;
574         call->peer      = rxrpc_get_peer(conn->params.peer);
575         call->cid       = conn->proto.cid | channel;
576         call->call_id   = call_id;
577
578         trace_rxrpc_connect_call(call);
579         _net("CONNECT call %08x:%08x as call %d on conn %d",
580              call->cid, call->call_id, call->debug_id, conn->debug_id);
581
582         /* Paired with the read barrier in rxrpc_wait_for_channel().  This
583          * orders cid and epoch in the connection wrt to call_id without the
584          * need to take the channel_lock.
585          *
586          * We provisionally assign a callNumber at this point, but we don't
587          * confirm it until the call is about to be exposed.
588          *
589          * TODO: Pair with a barrier in the data_ready handler when that looks
590          * at the call ID through a connection channel.
591          */
592         smp_wmb();
593         chan->call_id   = call_id;
594         chan->call_debug_id = call->debug_id;
595         rcu_assign_pointer(chan->call, call);
596         wake_up(&call->waitq);
597 }
598
599 /*
600  * Assign channels and callNumbers to waiting calls with channel_lock
601  * held by caller.
602  */
603 static void rxrpc_activate_channels_locked(struct rxrpc_connection *conn)
604 {
605         u8 avail, mask;
606
607         switch (conn->cache_state) {
608         case RXRPC_CONN_CLIENT_ACTIVE:
609                 mask = RXRPC_ACTIVE_CHANS_MASK;
610                 break;
611         case RXRPC_CONN_CLIENT_UPGRADE:
612                 mask = 0x01;
613                 break;
614         default:
615                 return;
616         }
617
618         while (!list_empty(&conn->waiting_calls) &&
619                (avail = ~conn->active_chans,
620                 avail &= mask,
621                 avail != 0))
622                 rxrpc_activate_one_channel(conn, __ffs(avail));
623 }
624
625 /*
626  * Assign channels and callNumbers to waiting calls.
627  */
628 static void rxrpc_activate_channels(struct rxrpc_connection *conn)
629 {
630         _enter("%d", conn->debug_id);
631
632         trace_rxrpc_client(conn, -1, rxrpc_client_activate_chans);
633
634         if (conn->active_chans == RXRPC_ACTIVE_CHANS_MASK)
635                 return;
636
637         spin_lock(&conn->channel_lock);
638         rxrpc_activate_channels_locked(conn);
639         spin_unlock(&conn->channel_lock);
640         _leave("");
641 }
642
643 /*
644  * Wait for a callNumber and a channel to be granted to a call.
645  */
646 static int rxrpc_wait_for_channel(struct rxrpc_call *call, gfp_t gfp)
647 {
648         int ret = 0;
649
650         _enter("%d", call->debug_id);
651
652         if (!call->call_id) {
653                 DECLARE_WAITQUEUE(myself, current);
654
655                 if (!gfpflags_allow_blocking(gfp)) {
656                         ret = -EAGAIN;
657                         goto out;
658                 }
659
660                 add_wait_queue_exclusive(&call->waitq, &myself);
661                 for (;;) {
662                         set_current_state(TASK_INTERRUPTIBLE);
663                         if (call->call_id)
664                                 break;
665                         if (signal_pending(current)) {
666                                 ret = -ERESTARTSYS;
667                                 break;
668                         }
669                         schedule();
670                 }
671                 remove_wait_queue(&call->waitq, &myself);
672                 __set_current_state(TASK_RUNNING);
673         }
674
675         /* Paired with the write barrier in rxrpc_activate_one_channel(). */
676         smp_rmb();
677
678 out:
679         _leave(" = %d", ret);
680         return ret;
681 }
682
683 /*
684  * find a connection for a call
685  * - called in process context with IRQs enabled
686  */
687 int rxrpc_connect_call(struct rxrpc_sock *rx,
688                        struct rxrpc_call *call,
689                        struct rxrpc_conn_parameters *cp,
690                        struct sockaddr_rxrpc *srx,
691                        gfp_t gfp)
692 {
693         struct rxrpc_net *rxnet = cp->local->rxnet;
694         int ret;
695
696         _enter("{%d,%lx},", call->debug_id, call->user_call_ID);
697
698         rxrpc_discard_expired_client_conns(&rxnet->client_conn_reaper);
699         rxrpc_cull_active_client_conns(rxnet);
700
701         ret = rxrpc_get_client_conn(rx, call, cp, srx, gfp);
702         if (ret < 0)
703                 goto out;
704
705         rxrpc_animate_client_conn(rxnet, call->conn);
706         rxrpc_activate_channels(call->conn);
707
708         ret = rxrpc_wait_for_channel(call, gfp);
709         if (ret < 0) {
710                 rxrpc_disconnect_client_call(call);
711                 goto out;
712         }
713
714         spin_lock_bh(&call->conn->params.peer->lock);
715         hlist_add_head_rcu(&call->error_link,
716                            &call->conn->params.peer->error_targets);
717         spin_unlock_bh(&call->conn->params.peer->lock);
718
719 out:
720         _leave(" = %d", ret);
721         return ret;
722 }
723
724 /*
725  * Note that a connection is about to be exposed to the world.  Once it is
726  * exposed, we maintain an extra ref on it that stops it from being summarily
727  * discarded before it's (a) had a chance to deal with retransmission and (b)
728  * had a chance at re-use (the per-connection security negotiation is
729  * expensive).
730  */
731 static void rxrpc_expose_client_conn(struct rxrpc_connection *conn,
732                                      unsigned int channel)
733 {
734         if (!test_and_set_bit(RXRPC_CONN_EXPOSED, &conn->flags)) {
735                 trace_rxrpc_client(conn, channel, rxrpc_client_exposed);
736                 rxrpc_get_connection(conn);
737         }
738 }
739
740 /*
741  * Note that a call, and thus a connection, is about to be exposed to the
742  * world.
743  */
744 void rxrpc_expose_client_call(struct rxrpc_call *call)
745 {
746         unsigned int channel = call->cid & RXRPC_CHANNELMASK;
747         struct rxrpc_connection *conn = call->conn;
748         struct rxrpc_channel *chan = &conn->channels[channel];
749
750         if (!test_and_set_bit(RXRPC_CALL_EXPOSED, &call->flags)) {
751                 /* Mark the call ID as being used.  If the callNumber counter
752                  * exceeds ~2 billion, we kill the connection after its
753                  * outstanding calls have finished so that the counter doesn't
754                  * wrap.
755                  */
756                 chan->call_counter++;
757                 if (chan->call_counter >= INT_MAX)
758                         set_bit(RXRPC_CONN_DONT_REUSE, &conn->flags);
759                 rxrpc_expose_client_conn(conn, channel);
760         }
761 }
762
763 /*
764  * Set the reap timer.
765  */
766 static void rxrpc_set_client_reap_timer(struct rxrpc_net *rxnet)
767 {
768         unsigned long now = jiffies;
769         unsigned long reap_at = now + rxrpc_conn_idle_client_expiry;
770
771         if (rxnet->live)
772                 timer_reduce(&rxnet->client_conn_reap_timer, reap_at);
773 }
774
775 /*
776  * Disconnect a client call.
777  */
778 void rxrpc_disconnect_client_call(struct rxrpc_call *call)
779 {
780         unsigned int channel = call->cid & RXRPC_CHANNELMASK;
781         struct rxrpc_connection *conn = call->conn;
782         struct rxrpc_channel *chan = &conn->channels[channel];
783         struct rxrpc_net *rxnet = conn->params.local->rxnet;
784
785         trace_rxrpc_client(conn, channel, rxrpc_client_chan_disconnect);
786         call->conn = NULL;
787
788         spin_lock(&conn->channel_lock);
789
790         /* Calls that have never actually been assigned a channel can simply be
791          * discarded.  If the conn didn't get used either, it will follow
792          * immediately unless someone else grabs it in the meantime.
793          */
794         if (!list_empty(&call->chan_wait_link)) {
795                 _debug("call is waiting");
796                 ASSERTCMP(call->call_id, ==, 0);
797                 ASSERT(!test_bit(RXRPC_CALL_EXPOSED, &call->flags));
798                 list_del_init(&call->chan_wait_link);
799
800                 trace_rxrpc_client(conn, channel, rxrpc_client_chan_unstarted);
801
802                 /* We must deactivate or idle the connection if it's now
803                  * waiting for nothing.
804                  */
805                 spin_lock(&rxnet->client_conn_cache_lock);
806                 if (conn->cache_state == RXRPC_CONN_CLIENT_WAITING &&
807                     list_empty(&conn->waiting_calls) &&
808                     !conn->active_chans)
809                         goto idle_connection;
810                 goto out;
811         }
812
813         ASSERTCMP(rcu_access_pointer(chan->call), ==, call);
814
815         /* If a client call was exposed to the world, we save the result for
816          * retransmission.
817          *
818          * We use a barrier here so that the call number and abort code can be
819          * read without needing to take a lock.
820          *
821          * TODO: Make the incoming packet handler check this and handle
822          * terminal retransmission without requiring access to the call.
823          */
824         if (test_bit(RXRPC_CALL_EXPOSED, &call->flags)) {
825                 _debug("exposed %u,%u", call->call_id, call->abort_code);
826                 __rxrpc_disconnect_call(conn, call);
827         }
828
829         /* See if we can pass the channel directly to another call. */
830         if (conn->cache_state == RXRPC_CONN_CLIENT_ACTIVE &&
831             !list_empty(&conn->waiting_calls)) {
832                 trace_rxrpc_client(conn, channel, rxrpc_client_chan_pass);
833                 rxrpc_activate_one_channel(conn, channel);
834                 goto out_2;
835         }
836
837         /* Schedule the final ACK to be transmitted in a short while so that it
838          * can be skipped if we find a follow-on call.  The first DATA packet
839          * of the follow on call will implicitly ACK this call.
840          */
841         if (call->completion == RXRPC_CALL_SUCCEEDED &&
842             test_bit(RXRPC_CALL_EXPOSED, &call->flags)) {
843                 unsigned long final_ack_at = jiffies + 2;
844
845                 WRITE_ONCE(chan->final_ack_at, final_ack_at);
846                 smp_wmb(); /* vs rxrpc_process_delayed_final_acks() */
847                 set_bit(RXRPC_CONN_FINAL_ACK_0 + channel, &conn->flags);
848                 rxrpc_reduce_conn_timer(conn, final_ack_at);
849         }
850
851         /* Things are more complex and we need the cache lock.  We might be
852          * able to simply idle the conn or it might now be lurking on the wait
853          * list.  It might even get moved back to the active list whilst we're
854          * waiting for the lock.
855          */
856         spin_lock(&rxnet->client_conn_cache_lock);
857
858         switch (conn->cache_state) {
859         case RXRPC_CONN_CLIENT_UPGRADE:
860                 /* Deal with termination of a service upgrade probe. */
861                 if (test_bit(RXRPC_CONN_EXPOSED, &conn->flags)) {
862                         clear_bit(RXRPC_CONN_PROBING_FOR_UPGRADE, &conn->flags);
863                         trace_rxrpc_client(conn, channel, rxrpc_client_to_active);
864                         conn->cache_state = RXRPC_CONN_CLIENT_ACTIVE;
865                         rxrpc_activate_channels_locked(conn);
866                 }
867                 /* fall through */
868         case RXRPC_CONN_CLIENT_ACTIVE:
869                 if (list_empty(&conn->waiting_calls)) {
870                         rxrpc_deactivate_one_channel(conn, channel);
871                         if (!conn->active_chans) {
872                                 rxnet->nr_active_client_conns--;
873                                 goto idle_connection;
874                         }
875                         goto out;
876                 }
877
878                 trace_rxrpc_client(conn, channel, rxrpc_client_chan_pass);
879                 rxrpc_activate_one_channel(conn, channel);
880                 goto out;
881
882         case RXRPC_CONN_CLIENT_CULLED:
883                 rxrpc_deactivate_one_channel(conn, channel);
884                 ASSERT(list_empty(&conn->waiting_calls));
885                 if (!conn->active_chans)
886                         goto idle_connection;
887                 goto out;
888
889         case RXRPC_CONN_CLIENT_WAITING:
890                 rxrpc_deactivate_one_channel(conn, channel);
891                 goto out;
892
893         default:
894                 BUG();
895         }
896
897 out:
898         spin_unlock(&rxnet->client_conn_cache_lock);
899 out_2:
900         spin_unlock(&conn->channel_lock);
901         rxrpc_put_connection(conn);
902         _leave("");
903         return;
904
905 idle_connection:
906         /* As no channels remain active, the connection gets deactivated
907          * immediately or moved to the idle list for a short while.
908          */
909         if (test_bit(RXRPC_CONN_EXPOSED, &conn->flags)) {
910                 trace_rxrpc_client(conn, channel, rxrpc_client_to_idle);
911                 conn->idle_timestamp = jiffies;
912                 conn->cache_state = RXRPC_CONN_CLIENT_IDLE;
913                 list_move_tail(&conn->cache_link, &rxnet->idle_client_conns);
914                 if (rxnet->idle_client_conns.next == &conn->cache_link &&
915                     !rxnet->kill_all_client_conns)
916                         rxrpc_set_client_reap_timer(rxnet);
917         } else {
918                 trace_rxrpc_client(conn, channel, rxrpc_client_to_inactive);
919                 conn->cache_state = RXRPC_CONN_CLIENT_INACTIVE;
920                 list_del_init(&conn->cache_link);
921         }
922         goto out;
923 }
924
925 /*
926  * Clean up a dead client connection.
927  */
928 static struct rxrpc_connection *
929 rxrpc_put_one_client_conn(struct rxrpc_connection *conn)
930 {
931         struct rxrpc_connection *next = NULL;
932         struct rxrpc_local *local = conn->params.local;
933         struct rxrpc_net *rxnet = local->rxnet;
934         unsigned int nr_conns;
935
936         trace_rxrpc_client(conn, -1, rxrpc_client_cleanup);
937
938         if (test_bit(RXRPC_CONN_IN_CLIENT_CONNS, &conn->flags)) {
939                 spin_lock(&local->client_conns_lock);
940                 if (test_and_clear_bit(RXRPC_CONN_IN_CLIENT_CONNS,
941                                        &conn->flags))
942                         rb_erase(&conn->client_node, &local->client_conns);
943                 spin_unlock(&local->client_conns_lock);
944         }
945
946         rxrpc_put_client_connection_id(conn);
947
948         ASSERTCMP(conn->cache_state, ==, RXRPC_CONN_CLIENT_INACTIVE);
949
950         if (test_bit(RXRPC_CONN_COUNTED, &conn->flags)) {
951                 trace_rxrpc_client(conn, -1, rxrpc_client_uncount);
952                 spin_lock(&rxnet->client_conn_cache_lock);
953                 nr_conns = --rxnet->nr_client_conns;
954
955                 if (nr_conns < rxrpc_max_client_connections &&
956                     !list_empty(&rxnet->waiting_client_conns)) {
957                         next = list_entry(rxnet->waiting_client_conns.next,
958                                           struct rxrpc_connection, cache_link);
959                         rxrpc_get_connection(next);
960                         rxrpc_activate_conn(rxnet, next);
961                 }
962
963                 spin_unlock(&rxnet->client_conn_cache_lock);
964         }
965
966         rxrpc_kill_connection(conn);
967         if (next)
968                 rxrpc_activate_channels(next);
969
970         /* We need to get rid of the temporary ref we took upon next, but we
971          * can't call rxrpc_put_connection() recursively.
972          */
973         return next;
974 }
975
976 /*
977  * Clean up a dead client connections.
978  */
979 void rxrpc_put_client_conn(struct rxrpc_connection *conn)
980 {
981         const void *here = __builtin_return_address(0);
982         int n;
983
984         do {
985                 n = atomic_dec_return(&conn->usage);
986                 trace_rxrpc_conn(conn, rxrpc_conn_put_client, n, here);
987                 if (n > 0)
988                         return;
989                 ASSERTCMP(n, >=, 0);
990
991                 conn = rxrpc_put_one_client_conn(conn);
992         } while (conn);
993 }
994
995 /*
996  * Kill the longest-active client connections to make room for new ones.
997  */
998 static void rxrpc_cull_active_client_conns(struct rxrpc_net *rxnet)
999 {
1000         struct rxrpc_connection *conn;
1001         unsigned int nr_conns = rxnet->nr_client_conns;
1002         unsigned int nr_active, limit;
1003
1004         _enter("");
1005
1006         ASSERTCMP(nr_conns, >=, 0);
1007         if (nr_conns < rxrpc_max_client_connections) {
1008                 _leave(" [ok]");
1009                 return;
1010         }
1011         limit = rxrpc_reap_client_connections;
1012
1013         spin_lock(&rxnet->client_conn_cache_lock);
1014         nr_active = rxnet->nr_active_client_conns;
1015
1016         while (nr_active > limit) {
1017                 ASSERT(!list_empty(&rxnet->active_client_conns));
1018                 conn = list_entry(rxnet->active_client_conns.next,
1019                                   struct rxrpc_connection, cache_link);
1020                 ASSERTIFCMP(conn->cache_state != RXRPC_CONN_CLIENT_ACTIVE,
1021                             conn->cache_state, ==, RXRPC_CONN_CLIENT_UPGRADE);
1022
1023                 if (list_empty(&conn->waiting_calls)) {
1024                         trace_rxrpc_client(conn, -1, rxrpc_client_to_culled);
1025                         conn->cache_state = RXRPC_CONN_CLIENT_CULLED;
1026                         list_del_init(&conn->cache_link);
1027                 } else {
1028                         trace_rxrpc_client(conn, -1, rxrpc_client_to_waiting);
1029                         conn->cache_state = RXRPC_CONN_CLIENT_WAITING;
1030                         list_move_tail(&conn->cache_link,
1031                                        &rxnet->waiting_client_conns);
1032                 }
1033
1034                 nr_active--;
1035         }
1036
1037         rxnet->nr_active_client_conns = nr_active;
1038         spin_unlock(&rxnet->client_conn_cache_lock);
1039         ASSERTCMP(nr_active, >=, 0);
1040         _leave(" [culled]");
1041 }
1042
1043 /*
1044  * Discard expired client connections from the idle list.  Each conn in the
1045  * idle list has been exposed and holds an extra ref because of that.
1046  *
1047  * This may be called from conn setup or from a work item so cannot be
1048  * considered non-reentrant.
1049  */
1050 void rxrpc_discard_expired_client_conns(struct work_struct *work)
1051 {
1052         struct rxrpc_connection *conn;
1053         struct rxrpc_net *rxnet =
1054                 container_of(work, struct rxrpc_net, client_conn_reaper);
1055         unsigned long expiry, conn_expires_at, now;
1056         unsigned int nr_conns;
1057
1058         _enter("");
1059
1060         if (list_empty(&rxnet->idle_client_conns)) {
1061                 _leave(" [empty]");
1062                 return;
1063         }
1064
1065         /* Don't double up on the discarding */
1066         if (!spin_trylock(&rxnet->client_conn_discard_lock)) {
1067                 _leave(" [already]");
1068                 return;
1069         }
1070
1071         /* We keep an estimate of what the number of conns ought to be after
1072          * we've discarded some so that we don't overdo the discarding.
1073          */
1074         nr_conns = rxnet->nr_client_conns;
1075
1076 next:
1077         spin_lock(&rxnet->client_conn_cache_lock);
1078
1079         if (list_empty(&rxnet->idle_client_conns))
1080                 goto out;
1081
1082         conn = list_entry(rxnet->idle_client_conns.next,
1083                           struct rxrpc_connection, cache_link);
1084         ASSERT(test_bit(RXRPC_CONN_EXPOSED, &conn->flags));
1085
1086         if (!rxnet->kill_all_client_conns) {
1087                 /* If the number of connections is over the reap limit, we
1088                  * expedite discard by reducing the expiry timeout.  We must,
1089                  * however, have at least a short grace period to be able to do
1090                  * final-ACK or ABORT retransmission.
1091                  */
1092                 expiry = rxrpc_conn_idle_client_expiry;
1093                 if (nr_conns > rxrpc_reap_client_connections)
1094                         expiry = rxrpc_conn_idle_client_fast_expiry;
1095                 if (conn->params.local->service_closed)
1096                         expiry = rxrpc_closed_conn_expiry * HZ;
1097
1098                 conn_expires_at = conn->idle_timestamp + expiry;
1099
1100                 now = READ_ONCE(jiffies);
1101                 if (time_after(conn_expires_at, now))
1102                         goto not_yet_expired;
1103         }
1104
1105         trace_rxrpc_client(conn, -1, rxrpc_client_discard);
1106         if (!test_and_clear_bit(RXRPC_CONN_EXPOSED, &conn->flags))
1107                 BUG();
1108         conn->cache_state = RXRPC_CONN_CLIENT_INACTIVE;
1109         list_del_init(&conn->cache_link);
1110
1111         spin_unlock(&rxnet->client_conn_cache_lock);
1112
1113         /* When we cleared the EXPOSED flag, we took on responsibility for the
1114          * reference that that had on the usage count.  We deal with that here.
1115          * If someone re-sets the flag and re-gets the ref, that's fine.
1116          */
1117         rxrpc_put_connection(conn);
1118         nr_conns--;
1119         goto next;
1120
1121 not_yet_expired:
1122         /* The connection at the front of the queue hasn't yet expired, so
1123          * schedule the work item for that point if we discarded something.
1124          *
1125          * We don't worry if the work item is already scheduled - it can look
1126          * after rescheduling itself at a later time.  We could cancel it, but
1127          * then things get messier.
1128          */
1129         _debug("not yet");
1130         if (!rxnet->kill_all_client_conns)
1131                 timer_reduce(&rxnet->client_conn_reap_timer,
1132                              conn_expires_at);
1133
1134 out:
1135         spin_unlock(&rxnet->client_conn_cache_lock);
1136         spin_unlock(&rxnet->client_conn_discard_lock);
1137         _leave("");
1138 }
1139
1140 /*
1141  * Preemptively destroy all the client connection records rather than waiting
1142  * for them to time out
1143  */
1144 void rxrpc_destroy_all_client_connections(struct rxrpc_net *rxnet)
1145 {
1146         _enter("");
1147
1148         spin_lock(&rxnet->client_conn_cache_lock);
1149         rxnet->kill_all_client_conns = true;
1150         spin_unlock(&rxnet->client_conn_cache_lock);
1151
1152         del_timer_sync(&rxnet->client_conn_reap_timer);
1153
1154         if (!rxrpc_queue_work(&rxnet->client_conn_reaper))
1155                 _debug("destroy: queue failed");
1156
1157         _leave("");
1158 }