Merge branch 'sched-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel...
[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         write_lock(&rxnet->conn_lock);
211         list_add_tail(&conn->proc_link, &rxnet->conn_proc_list);
212         write_unlock(&rxnet->conn_lock);
213
214         /* We steal the caller's peer ref. */
215         cp->peer = NULL;
216         rxrpc_get_local(conn->params.local);
217         key_get(conn->params.key);
218
219         trace_rxrpc_conn(conn, rxrpc_conn_new_client, atomic_read(&conn->usage),
220                          __builtin_return_address(0));
221         trace_rxrpc_client(conn, -1, rxrpc_client_alloc);
222         _leave(" = %p", conn);
223         return conn;
224
225 error_2:
226         conn->security->clear(conn);
227 error_1:
228         rxrpc_put_client_connection_id(conn);
229 error_0:
230         kfree(conn);
231         _leave(" = %d", ret);
232         return ERR_PTR(ret);
233 }
234
235 /*
236  * Determine if a connection may be reused.
237  */
238 static bool rxrpc_may_reuse_conn(struct rxrpc_connection *conn)
239 {
240         struct rxrpc_net *rxnet = conn->params.local->rxnet;
241         int id_cursor, id, distance, limit;
242
243         if (test_bit(RXRPC_CONN_DONT_REUSE, &conn->flags))
244                 goto dont_reuse;
245
246         if (conn->proto.epoch != rxnet->epoch)
247                 goto mark_dont_reuse;
248
249         /* The IDR tree gets very expensive on memory if the connection IDs are
250          * widely scattered throughout the number space, so we shall want to
251          * kill off connections that, say, have an ID more than about four
252          * times the maximum number of client conns away from the current
253          * allocation point to try and keep the IDs concentrated.
254          */
255         id_cursor = idr_get_cursor(&rxrpc_client_conn_ids);
256         id = conn->proto.cid >> RXRPC_CIDSHIFT;
257         distance = id - id_cursor;
258         if (distance < 0)
259                 distance = -distance;
260         limit = max(rxrpc_max_client_connections * 4, 1024U);
261         if (distance > limit)
262                 goto mark_dont_reuse;
263
264         return true;
265
266 mark_dont_reuse:
267         set_bit(RXRPC_CONN_DONT_REUSE, &conn->flags);
268 dont_reuse:
269         return false;
270 }
271
272 /*
273  * Create or find a client connection to use for a call.
274  *
275  * If we return with a connection, the call will be on its waiting list.  It's
276  * left to the caller to assign a channel and wake up the call.
277  */
278 static int rxrpc_get_client_conn(struct rxrpc_call *call,
279                                  struct rxrpc_conn_parameters *cp,
280                                  struct sockaddr_rxrpc *srx,
281                                  gfp_t gfp)
282 {
283         struct rxrpc_connection *conn, *candidate = NULL;
284         struct rxrpc_local *local = cp->local;
285         struct rb_node *p, **pp, *parent;
286         long diff;
287         int ret = -ENOMEM;
288
289         _enter("{%d,%lx},", call->debug_id, call->user_call_ID);
290
291         cp->peer = rxrpc_lookup_peer(cp->local, srx, gfp);
292         if (!cp->peer)
293                 goto error;
294
295         call->cong_cwnd = cp->peer->cong_cwnd;
296         if (call->cong_cwnd >= call->cong_ssthresh)
297                 call->cong_mode = RXRPC_CALL_CONGEST_AVOIDANCE;
298         else
299                 call->cong_mode = RXRPC_CALL_SLOW_START;
300
301         /* If the connection is not meant to be exclusive, search the available
302          * connections to see if the connection we want to use already exists.
303          */
304         if (!cp->exclusive) {
305                 _debug("search 1");
306                 spin_lock(&local->client_conns_lock);
307                 p = local->client_conns.rb_node;
308                 while (p) {
309                         conn = rb_entry(p, struct rxrpc_connection, client_node);
310
311 #define cmp(X) ((long)conn->params.X - (long)cp->X)
312                         diff = (cmp(peer) ?:
313                                 cmp(key) ?:
314                                 cmp(security_level) ?:
315                                 cmp(upgrade));
316 #undef cmp
317                         if (diff < 0) {
318                                 p = p->rb_left;
319                         } else if (diff > 0) {
320                                 p = p->rb_right;
321                         } else {
322                                 if (rxrpc_may_reuse_conn(conn) &&
323                                     rxrpc_get_connection_maybe(conn))
324                                         goto found_extant_conn;
325                                 /* The connection needs replacing.  It's better
326                                  * to effect that when we have something to
327                                  * replace it with so that we don't have to
328                                  * rebalance the tree twice.
329                                  */
330                                 break;
331                         }
332                 }
333                 spin_unlock(&local->client_conns_lock);
334         }
335
336         /* There wasn't a connection yet or we need an exclusive connection.
337          * We need to create a candidate and then potentially redo the search
338          * in case we're racing with another thread also trying to connect on a
339          * shareable connection.
340          */
341         _debug("new conn");
342         candidate = rxrpc_alloc_client_connection(cp, gfp);
343         if (IS_ERR(candidate)) {
344                 ret = PTR_ERR(candidate);
345                 goto error_peer;
346         }
347
348         /* Add the call to the new connection's waiting list in case we're
349          * going to have to wait for the connection to come live.  It's our
350          * connection, so we want first dibs on the channel slots.  We would
351          * normally have to take channel_lock but we do this before anyone else
352          * can see the connection.
353          */
354         list_add_tail(&call->chan_wait_link, &candidate->waiting_calls);
355
356         if (cp->exclusive) {
357                 call->conn = candidate;
358                 call->security_ix = candidate->security_ix;
359                 call->service_id = candidate->service_id;
360                 _leave(" = 0 [exclusive %d]", candidate->debug_id);
361                 return 0;
362         }
363
364         /* Publish the new connection for userspace to find.  We need to redo
365          * the search before doing this lest we race with someone else adding a
366          * conflicting instance.
367          */
368         _debug("search 2");
369         spin_lock(&local->client_conns_lock);
370
371         pp = &local->client_conns.rb_node;
372         parent = NULL;
373         while (*pp) {
374                 parent = *pp;
375                 conn = rb_entry(parent, struct rxrpc_connection, client_node);
376
377 #define cmp(X) ((long)conn->params.X - (long)candidate->params.X)
378                 diff = (cmp(peer) ?:
379                         cmp(key) ?:
380                         cmp(security_level) ?:
381                         cmp(upgrade));
382 #undef cmp
383                 if (diff < 0) {
384                         pp = &(*pp)->rb_left;
385                 } else if (diff > 0) {
386                         pp = &(*pp)->rb_right;
387                 } else {
388                         if (rxrpc_may_reuse_conn(conn) &&
389                             rxrpc_get_connection_maybe(conn))
390                                 goto found_extant_conn;
391                         /* The old connection is from an outdated epoch. */
392                         _debug("replace conn");
393                         clear_bit(RXRPC_CONN_IN_CLIENT_CONNS, &conn->flags);
394                         rb_replace_node(&conn->client_node,
395                                         &candidate->client_node,
396                                         &local->client_conns);
397                         trace_rxrpc_client(conn, -1, rxrpc_client_replace);
398                         goto candidate_published;
399                 }
400         }
401
402         _debug("new conn");
403         rb_link_node(&candidate->client_node, parent, pp);
404         rb_insert_color(&candidate->client_node, &local->client_conns);
405
406 candidate_published:
407         set_bit(RXRPC_CONN_IN_CLIENT_CONNS, &candidate->flags);
408         call->conn = candidate;
409         call->security_ix = candidate->security_ix;
410         call->service_id = candidate->service_id;
411         spin_unlock(&local->client_conns_lock);
412         _leave(" = 0 [new %d]", candidate->debug_id);
413         return 0;
414
415         /* We come here if we found a suitable connection already in existence.
416          * Discard any candidate we may have allocated, and try to get a
417          * channel on this one.
418          */
419 found_extant_conn:
420         _debug("found conn");
421         spin_unlock(&local->client_conns_lock);
422
423         if (candidate) {
424                 trace_rxrpc_client(candidate, -1, rxrpc_client_duplicate);
425                 rxrpc_put_connection(candidate);
426                 candidate = NULL;
427         }
428
429         spin_lock(&conn->channel_lock);
430         call->conn = conn;
431         call->security_ix = conn->security_ix;
432         call->service_id = conn->service_id;
433         list_add(&call->chan_wait_link, &conn->waiting_calls);
434         spin_unlock(&conn->channel_lock);
435         _leave(" = 0 [extant %d]", conn->debug_id);
436         return 0;
437
438 error_peer:
439         rxrpc_put_peer(cp->peer);
440         cp->peer = NULL;
441 error:
442         _leave(" = %d", ret);
443         return ret;
444 }
445
446 /*
447  * Activate a connection.
448  */
449 static void rxrpc_activate_conn(struct rxrpc_net *rxnet,
450                                 struct rxrpc_connection *conn)
451 {
452         if (test_bit(RXRPC_CONN_PROBING_FOR_UPGRADE, &conn->flags)) {
453                 trace_rxrpc_client(conn, -1, rxrpc_client_to_upgrade);
454                 conn->cache_state = RXRPC_CONN_CLIENT_UPGRADE;
455         } else {
456                 trace_rxrpc_client(conn, -1, rxrpc_client_to_active);
457                 conn->cache_state = RXRPC_CONN_CLIENT_ACTIVE;
458         }
459         rxnet->nr_active_client_conns++;
460         list_move_tail(&conn->cache_link, &rxnet->active_client_conns);
461 }
462
463 /*
464  * Attempt to animate a connection for a new call.
465  *
466  * If it's not exclusive, the connection is in the endpoint tree, and we're in
467  * the conn's list of those waiting to grab a channel.  There is, however, a
468  * limit on the number of live connections allowed at any one time, so we may
469  * have to wait for capacity to become available.
470  *
471  * Note that a connection on the waiting queue might *also* have active
472  * channels if it has been culled to make space and then re-requested by a new
473  * call.
474  */
475 static void rxrpc_animate_client_conn(struct rxrpc_net *rxnet,
476                                       struct rxrpc_connection *conn)
477 {
478         unsigned int nr_conns;
479
480         _enter("%d,%d", conn->debug_id, conn->cache_state);
481
482         if (conn->cache_state == RXRPC_CONN_CLIENT_ACTIVE ||
483             conn->cache_state == RXRPC_CONN_CLIENT_UPGRADE)
484                 goto out;
485
486         spin_lock(&rxnet->client_conn_cache_lock);
487
488         nr_conns = rxnet->nr_client_conns;
489         if (!test_and_set_bit(RXRPC_CONN_COUNTED, &conn->flags)) {
490                 trace_rxrpc_client(conn, -1, rxrpc_client_count);
491                 rxnet->nr_client_conns = nr_conns + 1;
492         }
493
494         switch (conn->cache_state) {
495         case RXRPC_CONN_CLIENT_ACTIVE:
496         case RXRPC_CONN_CLIENT_UPGRADE:
497         case RXRPC_CONN_CLIENT_WAITING:
498                 break;
499
500         case RXRPC_CONN_CLIENT_INACTIVE:
501         case RXRPC_CONN_CLIENT_CULLED:
502         case RXRPC_CONN_CLIENT_IDLE:
503                 if (nr_conns >= rxrpc_max_client_connections)
504                         goto wait_for_capacity;
505                 goto activate_conn;
506
507         default:
508                 BUG();
509         }
510
511 out_unlock:
512         spin_unlock(&rxnet->client_conn_cache_lock);
513 out:
514         _leave(" [%d]", conn->cache_state);
515         return;
516
517 activate_conn:
518         _debug("activate");
519         rxrpc_activate_conn(rxnet, conn);
520         goto out_unlock;
521
522 wait_for_capacity:
523         _debug("wait");
524         trace_rxrpc_client(conn, -1, rxrpc_client_to_waiting);
525         conn->cache_state = RXRPC_CONN_CLIENT_WAITING;
526         list_move_tail(&conn->cache_link, &rxnet->waiting_client_conns);
527         goto out_unlock;
528 }
529
530 /*
531  * Deactivate a channel.
532  */
533 static void rxrpc_deactivate_one_channel(struct rxrpc_connection *conn,
534                                          unsigned int channel)
535 {
536         struct rxrpc_channel *chan = &conn->channels[channel];
537
538         rcu_assign_pointer(chan->call, NULL);
539         conn->active_chans &= ~(1 << channel);
540 }
541
542 /*
543  * Assign a channel to the call at the front of the queue and wake the call up.
544  * We don't increment the callNumber counter until this number has been exposed
545  * to the world.
546  */
547 static void rxrpc_activate_one_channel(struct rxrpc_connection *conn,
548                                        unsigned int channel)
549 {
550         struct rxrpc_channel *chan = &conn->channels[channel];
551         struct rxrpc_call *call = list_entry(conn->waiting_calls.next,
552                                              struct rxrpc_call, chan_wait_link);
553         u32 call_id = chan->call_counter + 1;
554
555         trace_rxrpc_client(conn, channel, rxrpc_client_chan_activate);
556
557         /* Cancel the final ACK on the previous call if it hasn't been sent yet
558          * as the DATA packet will implicitly ACK it.
559          */
560         clear_bit(RXRPC_CONN_FINAL_ACK_0 + channel, &conn->flags);
561
562         write_lock_bh(&call->state_lock);
563         if (!test_bit(RXRPC_CALL_TX_LASTQ, &call->flags))
564                 call->state = RXRPC_CALL_CLIENT_SEND_REQUEST;
565         else
566                 call->state = RXRPC_CALL_CLIENT_AWAIT_REPLY;
567         write_unlock_bh(&call->state_lock);
568
569         rxrpc_see_call(call);
570         list_del_init(&call->chan_wait_link);
571         conn->active_chans |= 1 << channel;
572         call->peer      = rxrpc_get_peer(conn->params.peer);
573         call->cid       = conn->proto.cid | channel;
574         call->call_id   = call_id;
575
576         trace_rxrpc_connect_call(call);
577         _net("CONNECT call %08x:%08x as call %d on conn %d",
578              call->cid, call->call_id, call->debug_id, conn->debug_id);
579
580         /* Paired with the read barrier in rxrpc_wait_for_channel().  This
581          * orders cid and epoch in the connection wrt to call_id without the
582          * need to take the channel_lock.
583          *
584          * We provisionally assign a callNumber at this point, but we don't
585          * confirm it until the call is about to be exposed.
586          *
587          * TODO: Pair with a barrier in the data_ready handler when that looks
588          * at the call ID through a connection channel.
589          */
590         smp_wmb();
591         chan->call_id   = call_id;
592         rcu_assign_pointer(chan->call, call);
593         wake_up(&call->waitq);
594 }
595
596 /*
597  * Assign channels and callNumbers to waiting calls with channel_lock
598  * held by caller.
599  */
600 static void rxrpc_activate_channels_locked(struct rxrpc_connection *conn)
601 {
602         u8 avail, mask;
603
604         switch (conn->cache_state) {
605         case RXRPC_CONN_CLIENT_ACTIVE:
606                 mask = RXRPC_ACTIVE_CHANS_MASK;
607                 break;
608         case RXRPC_CONN_CLIENT_UPGRADE:
609                 mask = 0x01;
610                 break;
611         default:
612                 return;
613         }
614
615         while (!list_empty(&conn->waiting_calls) &&
616                (avail = ~conn->active_chans,
617                 avail &= mask,
618                 avail != 0))
619                 rxrpc_activate_one_channel(conn, __ffs(avail));
620 }
621
622 /*
623  * Assign channels and callNumbers to waiting calls.
624  */
625 static void rxrpc_activate_channels(struct rxrpc_connection *conn)
626 {
627         _enter("%d", conn->debug_id);
628
629         trace_rxrpc_client(conn, -1, rxrpc_client_activate_chans);
630
631         if (conn->active_chans == RXRPC_ACTIVE_CHANS_MASK)
632                 return;
633
634         spin_lock(&conn->channel_lock);
635         rxrpc_activate_channels_locked(conn);
636         spin_unlock(&conn->channel_lock);
637         _leave("");
638 }
639
640 /*
641  * Wait for a callNumber and a channel to be granted to a call.
642  */
643 static int rxrpc_wait_for_channel(struct rxrpc_call *call, gfp_t gfp)
644 {
645         int ret = 0;
646
647         _enter("%d", call->debug_id);
648
649         if (!call->call_id) {
650                 DECLARE_WAITQUEUE(myself, current);
651
652                 if (!gfpflags_allow_blocking(gfp)) {
653                         ret = -EAGAIN;
654                         goto out;
655                 }
656
657                 add_wait_queue_exclusive(&call->waitq, &myself);
658                 for (;;) {
659                         set_current_state(TASK_INTERRUPTIBLE);
660                         if (call->call_id)
661                                 break;
662                         if (signal_pending(current)) {
663                                 ret = -ERESTARTSYS;
664                                 break;
665                         }
666                         schedule();
667                 }
668                 remove_wait_queue(&call->waitq, &myself);
669                 __set_current_state(TASK_RUNNING);
670         }
671
672         /* Paired with the write barrier in rxrpc_activate_one_channel(). */
673         smp_rmb();
674
675 out:
676         _leave(" = %d", ret);
677         return ret;
678 }
679
680 /*
681  * find a connection for a call
682  * - called in process context with IRQs enabled
683  */
684 int rxrpc_connect_call(struct rxrpc_call *call,
685                        struct rxrpc_conn_parameters *cp,
686                        struct sockaddr_rxrpc *srx,
687                        gfp_t gfp)
688 {
689         struct rxrpc_net *rxnet = cp->local->rxnet;
690         int ret;
691
692         _enter("{%d,%lx},", call->debug_id, call->user_call_ID);
693
694         rxrpc_discard_expired_client_conns(&rxnet->client_conn_reaper);
695         rxrpc_cull_active_client_conns(rxnet);
696
697         ret = rxrpc_get_client_conn(call, cp, srx, gfp);
698         if (ret < 0)
699                 goto out;
700
701         rxrpc_animate_client_conn(rxnet, call->conn);
702         rxrpc_activate_channels(call->conn);
703
704         ret = rxrpc_wait_for_channel(call, gfp);
705         if (ret < 0) {
706                 rxrpc_disconnect_client_call(call);
707                 goto out;
708         }
709
710         spin_lock_bh(&call->conn->params.peer->lock);
711         hlist_add_head(&call->error_link,
712                        &call->conn->params.peer->error_targets);
713         spin_unlock_bh(&call->conn->params.peer->lock);
714
715 out:
716         _leave(" = %d", ret);
717         return ret;
718 }
719
720 /*
721  * Note that a connection is about to be exposed to the world.  Once it is
722  * exposed, we maintain an extra ref on it that stops it from being summarily
723  * discarded before it's (a) had a chance to deal with retransmission and (b)
724  * had a chance at re-use (the per-connection security negotiation is
725  * expensive).
726  */
727 static void rxrpc_expose_client_conn(struct rxrpc_connection *conn,
728                                      unsigned int channel)
729 {
730         if (!test_and_set_bit(RXRPC_CONN_EXPOSED, &conn->flags)) {
731                 trace_rxrpc_client(conn, channel, rxrpc_client_exposed);
732                 rxrpc_get_connection(conn);
733         }
734 }
735
736 /*
737  * Note that a call, and thus a connection, is about to be exposed to the
738  * world.
739  */
740 void rxrpc_expose_client_call(struct rxrpc_call *call)
741 {
742         unsigned int channel = call->cid & RXRPC_CHANNELMASK;
743         struct rxrpc_connection *conn = call->conn;
744         struct rxrpc_channel *chan = &conn->channels[channel];
745
746         if (!test_and_set_bit(RXRPC_CALL_EXPOSED, &call->flags)) {
747                 /* Mark the call ID as being used.  If the callNumber counter
748                  * exceeds ~2 billion, we kill the connection after its
749                  * outstanding calls have finished so that the counter doesn't
750                  * wrap.
751                  */
752                 chan->call_counter++;
753                 if (chan->call_counter >= INT_MAX)
754                         set_bit(RXRPC_CONN_DONT_REUSE, &conn->flags);
755                 rxrpc_expose_client_conn(conn, channel);
756         }
757 }
758
759 /*
760  * Set the reap timer.
761  */
762 static void rxrpc_set_client_reap_timer(struct rxrpc_net *rxnet)
763 {
764         unsigned long now = jiffies;
765         unsigned long reap_at = now + rxrpc_conn_idle_client_expiry;
766
767         if (rxnet->live)
768                 timer_reduce(&rxnet->client_conn_reap_timer, reap_at);
769 }
770
771 /*
772  * Disconnect a client call.
773  */
774 void rxrpc_disconnect_client_call(struct rxrpc_call *call)
775 {
776         unsigned int channel = call->cid & RXRPC_CHANNELMASK;
777         struct rxrpc_connection *conn = call->conn;
778         struct rxrpc_channel *chan = &conn->channels[channel];
779         struct rxrpc_net *rxnet = rxrpc_net(sock_net(&call->socket->sk));
780
781         trace_rxrpc_client(conn, channel, rxrpc_client_chan_disconnect);
782         call->conn = NULL;
783
784         spin_lock(&conn->channel_lock);
785
786         /* Calls that have never actually been assigned a channel can simply be
787          * discarded.  If the conn didn't get used either, it will follow
788          * immediately unless someone else grabs it in the meantime.
789          */
790         if (!list_empty(&call->chan_wait_link)) {
791                 _debug("call is waiting");
792                 ASSERTCMP(call->call_id, ==, 0);
793                 ASSERT(!test_bit(RXRPC_CALL_EXPOSED, &call->flags));
794                 list_del_init(&call->chan_wait_link);
795
796                 trace_rxrpc_client(conn, channel, rxrpc_client_chan_unstarted);
797
798                 /* We must deactivate or idle the connection if it's now
799                  * waiting for nothing.
800                  */
801                 spin_lock(&rxnet->client_conn_cache_lock);
802                 if (conn->cache_state == RXRPC_CONN_CLIENT_WAITING &&
803                     list_empty(&conn->waiting_calls) &&
804                     !conn->active_chans)
805                         goto idle_connection;
806                 goto out;
807         }
808
809         ASSERTCMP(rcu_access_pointer(chan->call), ==, call);
810
811         /* If a client call was exposed to the world, we save the result for
812          * retransmission.
813          *
814          * We use a barrier here so that the call number and abort code can be
815          * read without needing to take a lock.
816          *
817          * TODO: Make the incoming packet handler check this and handle
818          * terminal retransmission without requiring access to the call.
819          */
820         if (test_bit(RXRPC_CALL_EXPOSED, &call->flags)) {
821                 _debug("exposed %u,%u", call->call_id, call->abort_code);
822                 __rxrpc_disconnect_call(conn, call);
823         }
824
825         /* See if we can pass the channel directly to another call. */
826         if (conn->cache_state == RXRPC_CONN_CLIENT_ACTIVE &&
827             !list_empty(&conn->waiting_calls)) {
828                 trace_rxrpc_client(conn, channel, rxrpc_client_chan_pass);
829                 rxrpc_activate_one_channel(conn, channel);
830                 goto out_2;
831         }
832
833         /* Schedule the final ACK to be transmitted in a short while so that it
834          * can be skipped if we find a follow-on call.  The first DATA packet
835          * of the follow on call will implicitly ACK this call.
836          */
837         if (call->completion == RXRPC_CALL_SUCCEEDED &&
838             test_bit(RXRPC_CALL_EXPOSED, &call->flags)) {
839                 unsigned long final_ack_at = jiffies + 2;
840
841                 WRITE_ONCE(chan->final_ack_at, final_ack_at);
842                 smp_wmb(); /* vs rxrpc_process_delayed_final_acks() */
843                 set_bit(RXRPC_CONN_FINAL_ACK_0 + channel, &conn->flags);
844                 rxrpc_reduce_conn_timer(conn, final_ack_at);
845         }
846
847         /* Things are more complex and we need the cache lock.  We might be
848          * able to simply idle the conn or it might now be lurking on the wait
849          * list.  It might even get moved back to the active list whilst we're
850          * waiting for the lock.
851          */
852         spin_lock(&rxnet->client_conn_cache_lock);
853
854         switch (conn->cache_state) {
855         case RXRPC_CONN_CLIENT_UPGRADE:
856                 /* Deal with termination of a service upgrade probe. */
857                 if (test_bit(RXRPC_CONN_EXPOSED, &conn->flags)) {
858                         clear_bit(RXRPC_CONN_PROBING_FOR_UPGRADE, &conn->flags);
859                         trace_rxrpc_client(conn, channel, rxrpc_client_to_active);
860                         conn->cache_state = RXRPC_CONN_CLIENT_ACTIVE;
861                         rxrpc_activate_channels_locked(conn);
862                 }
863                 /* fall through */
864         case RXRPC_CONN_CLIENT_ACTIVE:
865                 if (list_empty(&conn->waiting_calls)) {
866                         rxrpc_deactivate_one_channel(conn, channel);
867                         if (!conn->active_chans) {
868                                 rxnet->nr_active_client_conns--;
869                                 goto idle_connection;
870                         }
871                         goto out;
872                 }
873
874                 trace_rxrpc_client(conn, channel, rxrpc_client_chan_pass);
875                 rxrpc_activate_one_channel(conn, channel);
876                 goto out;
877
878         case RXRPC_CONN_CLIENT_CULLED:
879                 rxrpc_deactivate_one_channel(conn, channel);
880                 ASSERT(list_empty(&conn->waiting_calls));
881                 if (!conn->active_chans)
882                         goto idle_connection;
883                 goto out;
884
885         case RXRPC_CONN_CLIENT_WAITING:
886                 rxrpc_deactivate_one_channel(conn, channel);
887                 goto out;
888
889         default:
890                 BUG();
891         }
892
893 out:
894         spin_unlock(&rxnet->client_conn_cache_lock);
895 out_2:
896         spin_unlock(&conn->channel_lock);
897         rxrpc_put_connection(conn);
898         _leave("");
899         return;
900
901 idle_connection:
902         /* As no channels remain active, the connection gets deactivated
903          * immediately or moved to the idle list for a short while.
904          */
905         if (test_bit(RXRPC_CONN_EXPOSED, &conn->flags)) {
906                 trace_rxrpc_client(conn, channel, rxrpc_client_to_idle);
907                 conn->idle_timestamp = jiffies;
908                 conn->cache_state = RXRPC_CONN_CLIENT_IDLE;
909                 list_move_tail(&conn->cache_link, &rxnet->idle_client_conns);
910                 if (rxnet->idle_client_conns.next == &conn->cache_link &&
911                     !rxnet->kill_all_client_conns)
912                         rxrpc_set_client_reap_timer(rxnet);
913         } else {
914                 trace_rxrpc_client(conn, channel, rxrpc_client_to_inactive);
915                 conn->cache_state = RXRPC_CONN_CLIENT_INACTIVE;
916                 list_del_init(&conn->cache_link);
917         }
918         goto out;
919 }
920
921 /*
922  * Clean up a dead client connection.
923  */
924 static struct rxrpc_connection *
925 rxrpc_put_one_client_conn(struct rxrpc_connection *conn)
926 {
927         struct rxrpc_connection *next = NULL;
928         struct rxrpc_local *local = conn->params.local;
929         struct rxrpc_net *rxnet = local->rxnet;
930         unsigned int nr_conns;
931
932         trace_rxrpc_client(conn, -1, rxrpc_client_cleanup);
933
934         if (test_bit(RXRPC_CONN_IN_CLIENT_CONNS, &conn->flags)) {
935                 spin_lock(&local->client_conns_lock);
936                 if (test_and_clear_bit(RXRPC_CONN_IN_CLIENT_CONNS,
937                                        &conn->flags))
938                         rb_erase(&conn->client_node, &local->client_conns);
939                 spin_unlock(&local->client_conns_lock);
940         }
941
942         rxrpc_put_client_connection_id(conn);
943
944         ASSERTCMP(conn->cache_state, ==, RXRPC_CONN_CLIENT_INACTIVE);
945
946         if (test_bit(RXRPC_CONN_COUNTED, &conn->flags)) {
947                 trace_rxrpc_client(conn, -1, rxrpc_client_uncount);
948                 spin_lock(&rxnet->client_conn_cache_lock);
949                 nr_conns = --rxnet->nr_client_conns;
950
951                 if (nr_conns < rxrpc_max_client_connections &&
952                     !list_empty(&rxnet->waiting_client_conns)) {
953                         next = list_entry(rxnet->waiting_client_conns.next,
954                                           struct rxrpc_connection, cache_link);
955                         rxrpc_get_connection(next);
956                         rxrpc_activate_conn(rxnet, next);
957                 }
958
959                 spin_unlock(&rxnet->client_conn_cache_lock);
960         }
961
962         rxrpc_kill_connection(conn);
963         if (next)
964                 rxrpc_activate_channels(next);
965
966         /* We need to get rid of the temporary ref we took upon next, but we
967          * can't call rxrpc_put_connection() recursively.
968          */
969         return next;
970 }
971
972 /*
973  * Clean up a dead client connections.
974  */
975 void rxrpc_put_client_conn(struct rxrpc_connection *conn)
976 {
977         const void *here = __builtin_return_address(0);
978         int n;
979
980         do {
981                 n = atomic_dec_return(&conn->usage);
982                 trace_rxrpc_conn(conn, rxrpc_conn_put_client, n, here);
983                 if (n > 0)
984                         return;
985                 ASSERTCMP(n, >=, 0);
986
987                 conn = rxrpc_put_one_client_conn(conn);
988         } while (conn);
989 }
990
991 /*
992  * Kill the longest-active client connections to make room for new ones.
993  */
994 static void rxrpc_cull_active_client_conns(struct rxrpc_net *rxnet)
995 {
996         struct rxrpc_connection *conn;
997         unsigned int nr_conns = rxnet->nr_client_conns;
998         unsigned int nr_active, limit;
999
1000         _enter("");
1001
1002         ASSERTCMP(nr_conns, >=, 0);
1003         if (nr_conns < rxrpc_max_client_connections) {
1004                 _leave(" [ok]");
1005                 return;
1006         }
1007         limit = rxrpc_reap_client_connections;
1008
1009         spin_lock(&rxnet->client_conn_cache_lock);
1010         nr_active = rxnet->nr_active_client_conns;
1011
1012         while (nr_active > limit) {
1013                 ASSERT(!list_empty(&rxnet->active_client_conns));
1014                 conn = list_entry(rxnet->active_client_conns.next,
1015                                   struct rxrpc_connection, cache_link);
1016                 ASSERTIFCMP(conn->cache_state != RXRPC_CONN_CLIENT_ACTIVE,
1017                             conn->cache_state, ==, RXRPC_CONN_CLIENT_UPGRADE);
1018
1019                 if (list_empty(&conn->waiting_calls)) {
1020                         trace_rxrpc_client(conn, -1, rxrpc_client_to_culled);
1021                         conn->cache_state = RXRPC_CONN_CLIENT_CULLED;
1022                         list_del_init(&conn->cache_link);
1023                 } else {
1024                         trace_rxrpc_client(conn, -1, rxrpc_client_to_waiting);
1025                         conn->cache_state = RXRPC_CONN_CLIENT_WAITING;
1026                         list_move_tail(&conn->cache_link,
1027                                        &rxnet->waiting_client_conns);
1028                 }
1029
1030                 nr_active--;
1031         }
1032
1033         rxnet->nr_active_client_conns = nr_active;
1034         spin_unlock(&rxnet->client_conn_cache_lock);
1035         ASSERTCMP(nr_active, >=, 0);
1036         _leave(" [culled]");
1037 }
1038
1039 /*
1040  * Discard expired client connections from the idle list.  Each conn in the
1041  * idle list has been exposed and holds an extra ref because of that.
1042  *
1043  * This may be called from conn setup or from a work item so cannot be
1044  * considered non-reentrant.
1045  */
1046 void rxrpc_discard_expired_client_conns(struct work_struct *work)
1047 {
1048         struct rxrpc_connection *conn;
1049         struct rxrpc_net *rxnet =
1050                 container_of(work, struct rxrpc_net, client_conn_reaper);
1051         unsigned long expiry, conn_expires_at, now;
1052         unsigned int nr_conns;
1053         bool did_discard = false;
1054
1055         _enter("");
1056
1057         if (list_empty(&rxnet->idle_client_conns)) {
1058                 _leave(" [empty]");
1059                 return;
1060         }
1061
1062         /* Don't double up on the discarding */
1063         if (!spin_trylock(&rxnet->client_conn_discard_lock)) {
1064                 _leave(" [already]");
1065                 return;
1066         }
1067
1068         /* We keep an estimate of what the number of conns ought to be after
1069          * we've discarded some so that we don't overdo the discarding.
1070          */
1071         nr_conns = rxnet->nr_client_conns;
1072
1073 next:
1074         spin_lock(&rxnet->client_conn_cache_lock);
1075
1076         if (list_empty(&rxnet->idle_client_conns))
1077                 goto out;
1078
1079         conn = list_entry(rxnet->idle_client_conns.next,
1080                           struct rxrpc_connection, cache_link);
1081         ASSERT(test_bit(RXRPC_CONN_EXPOSED, &conn->flags));
1082
1083         if (!rxnet->kill_all_client_conns) {
1084                 /* If the number of connections is over the reap limit, we
1085                  * expedite discard by reducing the expiry timeout.  We must,
1086                  * however, have at least a short grace period to be able to do
1087                  * final-ACK or ABORT retransmission.
1088                  */
1089                 expiry = rxrpc_conn_idle_client_expiry;
1090                 if (nr_conns > rxrpc_reap_client_connections)
1091                         expiry = rxrpc_conn_idle_client_fast_expiry;
1092                 if (conn->params.local->service_closed)
1093                         expiry = rxrpc_closed_conn_expiry * HZ;
1094
1095                 conn_expires_at = conn->idle_timestamp + expiry;
1096
1097                 now = READ_ONCE(jiffies);
1098                 if (time_after(conn_expires_at, now))
1099                         goto not_yet_expired;
1100         }
1101
1102         trace_rxrpc_client(conn, -1, rxrpc_client_discard);
1103         if (!test_and_clear_bit(RXRPC_CONN_EXPOSED, &conn->flags))
1104                 BUG();
1105         conn->cache_state = RXRPC_CONN_CLIENT_INACTIVE;
1106         list_del_init(&conn->cache_link);
1107
1108         spin_unlock(&rxnet->client_conn_cache_lock);
1109
1110         /* When we cleared the EXPOSED flag, we took on responsibility for the
1111          * reference that that had on the usage count.  We deal with that here.
1112          * If someone re-sets the flag and re-gets the ref, that's fine.
1113          */
1114         rxrpc_put_connection(conn);
1115         did_discard = true;
1116         nr_conns--;
1117         goto next;
1118
1119 not_yet_expired:
1120         /* The connection at the front of the queue hasn't yet expired, so
1121          * schedule the work item for that point if we discarded something.
1122          *
1123          * We don't worry if the work item is already scheduled - it can look
1124          * after rescheduling itself at a later time.  We could cancel it, but
1125          * then things get messier.
1126          */
1127         _debug("not yet");
1128         if (!rxnet->kill_all_client_conns)
1129                 timer_reduce(&rxnet->client_conn_reap_timer,
1130                              conn_expires_at);
1131
1132 out:
1133         spin_unlock(&rxnet->client_conn_cache_lock);
1134         spin_unlock(&rxnet->client_conn_discard_lock);
1135         _leave("");
1136 }
1137
1138 /*
1139  * Preemptively destroy all the client connection records rather than waiting
1140  * for them to time out
1141  */
1142 void rxrpc_destroy_all_client_connections(struct rxrpc_net *rxnet)
1143 {
1144         _enter("");
1145
1146         spin_lock(&rxnet->client_conn_cache_lock);
1147         rxnet->kill_all_client_conns = true;
1148         spin_unlock(&rxnet->client_conn_cache_lock);
1149
1150         del_timer_sync(&rxnet->client_conn_reap_timer);
1151
1152         if (!rxrpc_queue_work(&rxnet->client_conn_reaper))
1153                 _debug("destroy: queue failed");
1154
1155         _leave("");
1156 }