Merge tag 'dmaengine-4.19-rc1' of git://git.infradead.org/users/vkoul/slave-dma
[linux-2.6-microblaze.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_call *call,
280                                  struct rxrpc_conn_parameters *cp,
281                                  struct sockaddr_rxrpc *srx,
282                                  gfp_t gfp)
283 {
284         struct rxrpc_connection *conn, *candidate = NULL;
285         struct rxrpc_local *local = cp->local;
286         struct rb_node *p, **pp, *parent;
287         long diff;
288         int ret = -ENOMEM;
289
290         _enter("{%d,%lx},", call->debug_id, call->user_call_ID);
291
292         cp->peer = rxrpc_lookup_peer(cp->local, srx, gfp);
293         if (!cp->peer)
294                 goto error;
295
296         call->cong_cwnd = cp->peer->cong_cwnd;
297         if (call->cong_cwnd >= call->cong_ssthresh)
298                 call->cong_mode = RXRPC_CALL_CONGEST_AVOIDANCE;
299         else
300                 call->cong_mode = RXRPC_CALL_SLOW_START;
301
302         /* If the connection is not meant to be exclusive, search the available
303          * connections to see if the connection we want to use already exists.
304          */
305         if (!cp->exclusive) {
306                 _debug("search 1");
307                 spin_lock(&local->client_conns_lock);
308                 p = local->client_conns.rb_node;
309                 while (p) {
310                         conn = rb_entry(p, struct rxrpc_connection, client_node);
311
312 #define cmp(X) ((long)conn->params.X - (long)cp->X)
313                         diff = (cmp(peer) ?:
314                                 cmp(key) ?:
315                                 cmp(security_level) ?:
316                                 cmp(upgrade));
317 #undef cmp
318                         if (diff < 0) {
319                                 p = p->rb_left;
320                         } else if (diff > 0) {
321                                 p = p->rb_right;
322                         } else {
323                                 if (rxrpc_may_reuse_conn(conn) &&
324                                     rxrpc_get_connection_maybe(conn))
325                                         goto found_extant_conn;
326                                 /* The connection needs replacing.  It's better
327                                  * to effect that when we have something to
328                                  * replace it with so that we don't have to
329                                  * rebalance the tree twice.
330                                  */
331                                 break;
332                         }
333                 }
334                 spin_unlock(&local->client_conns_lock);
335         }
336
337         /* There wasn't a connection yet or we need an exclusive connection.
338          * We need to create a candidate and then potentially redo the search
339          * in case we're racing with another thread also trying to connect on a
340          * shareable connection.
341          */
342         _debug("new conn");
343         candidate = rxrpc_alloc_client_connection(cp, gfp);
344         if (IS_ERR(candidate)) {
345                 ret = PTR_ERR(candidate);
346                 goto error_peer;
347         }
348
349         /* Add the call to the new connection's waiting list in case we're
350          * going to have to wait for the connection to come live.  It's our
351          * connection, so we want first dibs on the channel slots.  We would
352          * normally have to take channel_lock but we do this before anyone else
353          * can see the connection.
354          */
355         list_add_tail(&call->chan_wait_link, &candidate->waiting_calls);
356
357         if (cp->exclusive) {
358                 call->conn = candidate;
359                 call->security_ix = candidate->security_ix;
360                 call->service_id = candidate->service_id;
361                 _leave(" = 0 [exclusive %d]", candidate->debug_id);
362                 return 0;
363         }
364
365         /* Publish the new connection for userspace to find.  We need to redo
366          * the search before doing this lest we race with someone else adding a
367          * conflicting instance.
368          */
369         _debug("search 2");
370         spin_lock(&local->client_conns_lock);
371
372         pp = &local->client_conns.rb_node;
373         parent = NULL;
374         while (*pp) {
375                 parent = *pp;
376                 conn = rb_entry(parent, struct rxrpc_connection, client_node);
377
378 #define cmp(X) ((long)conn->params.X - (long)candidate->params.X)
379                 diff = (cmp(peer) ?:
380                         cmp(key) ?:
381                         cmp(security_level) ?:
382                         cmp(upgrade));
383 #undef cmp
384                 if (diff < 0) {
385                         pp = &(*pp)->rb_left;
386                 } else if (diff > 0) {
387                         pp = &(*pp)->rb_right;
388                 } else {
389                         if (rxrpc_may_reuse_conn(conn) &&
390                             rxrpc_get_connection_maybe(conn))
391                                 goto found_extant_conn;
392                         /* The old connection is from an outdated epoch. */
393                         _debug("replace conn");
394                         clear_bit(RXRPC_CONN_IN_CLIENT_CONNS, &conn->flags);
395                         rb_replace_node(&conn->client_node,
396                                         &candidate->client_node,
397                                         &local->client_conns);
398                         trace_rxrpc_client(conn, -1, rxrpc_client_replace);
399                         goto candidate_published;
400                 }
401         }
402
403         _debug("new conn");
404         rb_link_node(&candidate->client_node, parent, pp);
405         rb_insert_color(&candidate->client_node, &local->client_conns);
406
407 candidate_published:
408         set_bit(RXRPC_CONN_IN_CLIENT_CONNS, &candidate->flags);
409         call->conn = candidate;
410         call->security_ix = candidate->security_ix;
411         call->service_id = candidate->service_id;
412         spin_unlock(&local->client_conns_lock);
413         _leave(" = 0 [new %d]", candidate->debug_id);
414         return 0;
415
416         /* We come here if we found a suitable connection already in existence.
417          * Discard any candidate we may have allocated, and try to get a
418          * channel on this one.
419          */
420 found_extant_conn:
421         _debug("found conn");
422         spin_unlock(&local->client_conns_lock);
423
424         if (candidate) {
425                 trace_rxrpc_client(candidate, -1, rxrpc_client_duplicate);
426                 rxrpc_put_connection(candidate);
427                 candidate = NULL;
428         }
429
430         spin_lock(&conn->channel_lock);
431         call->conn = conn;
432         call->security_ix = conn->security_ix;
433         call->service_id = conn->service_id;
434         list_add(&call->chan_wait_link, &conn->waiting_calls);
435         spin_unlock(&conn->channel_lock);
436         _leave(" = 0 [extant %d]", conn->debug_id);
437         return 0;
438
439 error_peer:
440         rxrpc_put_peer(cp->peer);
441         cp->peer = NULL;
442 error:
443         _leave(" = %d", ret);
444         return ret;
445 }
446
447 /*
448  * Activate a connection.
449  */
450 static void rxrpc_activate_conn(struct rxrpc_net *rxnet,
451                                 struct rxrpc_connection *conn)
452 {
453         if (test_bit(RXRPC_CONN_PROBING_FOR_UPGRADE, &conn->flags)) {
454                 trace_rxrpc_client(conn, -1, rxrpc_client_to_upgrade);
455                 conn->cache_state = RXRPC_CONN_CLIENT_UPGRADE;
456         } else {
457                 trace_rxrpc_client(conn, -1, rxrpc_client_to_active);
458                 conn->cache_state = RXRPC_CONN_CLIENT_ACTIVE;
459         }
460         rxnet->nr_active_client_conns++;
461         list_move_tail(&conn->cache_link, &rxnet->active_client_conns);
462 }
463
464 /*
465  * Attempt to animate a connection for a new call.
466  *
467  * If it's not exclusive, the connection is in the endpoint tree, and we're in
468  * the conn's list of those waiting to grab a channel.  There is, however, a
469  * limit on the number of live connections allowed at any one time, so we may
470  * have to wait for capacity to become available.
471  *
472  * Note that a connection on the waiting queue might *also* have active
473  * channels if it has been culled to make space and then re-requested by a new
474  * call.
475  */
476 static void rxrpc_animate_client_conn(struct rxrpc_net *rxnet,
477                                       struct rxrpc_connection *conn)
478 {
479         unsigned int nr_conns;
480
481         _enter("%d,%d", conn->debug_id, conn->cache_state);
482
483         if (conn->cache_state == RXRPC_CONN_CLIENT_ACTIVE ||
484             conn->cache_state == RXRPC_CONN_CLIENT_UPGRADE)
485                 goto out;
486
487         spin_lock(&rxnet->client_conn_cache_lock);
488
489         nr_conns = rxnet->nr_client_conns;
490         if (!test_and_set_bit(RXRPC_CONN_COUNTED, &conn->flags)) {
491                 trace_rxrpc_client(conn, -1, rxrpc_client_count);
492                 rxnet->nr_client_conns = nr_conns + 1;
493         }
494
495         switch (conn->cache_state) {
496         case RXRPC_CONN_CLIENT_ACTIVE:
497         case RXRPC_CONN_CLIENT_UPGRADE:
498         case RXRPC_CONN_CLIENT_WAITING:
499                 break;
500
501         case RXRPC_CONN_CLIENT_INACTIVE:
502         case RXRPC_CONN_CLIENT_CULLED:
503         case RXRPC_CONN_CLIENT_IDLE:
504                 if (nr_conns >= rxrpc_max_client_connections)
505                         goto wait_for_capacity;
506                 goto activate_conn;
507
508         default:
509                 BUG();
510         }
511
512 out_unlock:
513         spin_unlock(&rxnet->client_conn_cache_lock);
514 out:
515         _leave(" [%d]", conn->cache_state);
516         return;
517
518 activate_conn:
519         _debug("activate");
520         rxrpc_activate_conn(rxnet, conn);
521         goto out_unlock;
522
523 wait_for_capacity:
524         _debug("wait");
525         trace_rxrpc_client(conn, -1, rxrpc_client_to_waiting);
526         conn->cache_state = RXRPC_CONN_CLIENT_WAITING;
527         list_move_tail(&conn->cache_link, &rxnet->waiting_client_conns);
528         goto out_unlock;
529 }
530
531 /*
532  * Deactivate a channel.
533  */
534 static void rxrpc_deactivate_one_channel(struct rxrpc_connection *conn,
535                                          unsigned int channel)
536 {
537         struct rxrpc_channel *chan = &conn->channels[channel];
538
539         rcu_assign_pointer(chan->call, NULL);
540         conn->active_chans &= ~(1 << channel);
541 }
542
543 /*
544  * Assign a channel to the call at the front of the queue and wake the call up.
545  * We don't increment the callNumber counter until this number has been exposed
546  * to the world.
547  */
548 static void rxrpc_activate_one_channel(struct rxrpc_connection *conn,
549                                        unsigned int channel)
550 {
551         struct rxrpc_channel *chan = &conn->channels[channel];
552         struct rxrpc_call *call = list_entry(conn->waiting_calls.next,
553                                              struct rxrpc_call, chan_wait_link);
554         u32 call_id = chan->call_counter + 1;
555
556         trace_rxrpc_client(conn, channel, rxrpc_client_chan_activate);
557
558         /* Cancel the final ACK on the previous call if it hasn't been sent yet
559          * as the DATA packet will implicitly ACK it.
560          */
561         clear_bit(RXRPC_CONN_FINAL_ACK_0 + channel, &conn->flags);
562
563         write_lock_bh(&call->state_lock);
564         if (!test_bit(RXRPC_CALL_TX_LASTQ, &call->flags))
565                 call->state = RXRPC_CALL_CLIENT_SEND_REQUEST;
566         else
567                 call->state = RXRPC_CALL_CLIENT_AWAIT_REPLY;
568         write_unlock_bh(&call->state_lock);
569
570         rxrpc_see_call(call);
571         list_del_init(&call->chan_wait_link);
572         conn->active_chans |= 1 << channel;
573         call->peer      = rxrpc_get_peer(conn->params.peer);
574         call->cid       = conn->proto.cid | channel;
575         call->call_id   = call_id;
576
577         trace_rxrpc_connect_call(call);
578         _net("CONNECT call %08x:%08x as call %d on conn %d",
579              call->cid, call->call_id, call->debug_id, conn->debug_id);
580
581         /* Paired with the read barrier in rxrpc_wait_for_channel().  This
582          * orders cid and epoch in the connection wrt to call_id without the
583          * need to take the channel_lock.
584          *
585          * We provisionally assign a callNumber at this point, but we don't
586          * confirm it until the call is about to be exposed.
587          *
588          * TODO: Pair with a barrier in the data_ready handler when that looks
589          * at the call ID through a connection channel.
590          */
591         smp_wmb();
592         chan->call_id   = call_id;
593         chan->call_debug_id = call->debug_id;
594         rcu_assign_pointer(chan->call, call);
595         wake_up(&call->waitq);
596 }
597
598 /*
599  * Assign channels and callNumbers to waiting calls with channel_lock
600  * held by caller.
601  */
602 static void rxrpc_activate_channels_locked(struct rxrpc_connection *conn)
603 {
604         u8 avail, mask;
605
606         switch (conn->cache_state) {
607         case RXRPC_CONN_CLIENT_ACTIVE:
608                 mask = RXRPC_ACTIVE_CHANS_MASK;
609                 break;
610         case RXRPC_CONN_CLIENT_UPGRADE:
611                 mask = 0x01;
612                 break;
613         default:
614                 return;
615         }
616
617         while (!list_empty(&conn->waiting_calls) &&
618                (avail = ~conn->active_chans,
619                 avail &= mask,
620                 avail != 0))
621                 rxrpc_activate_one_channel(conn, __ffs(avail));
622 }
623
624 /*
625  * Assign channels and callNumbers to waiting calls.
626  */
627 static void rxrpc_activate_channels(struct rxrpc_connection *conn)
628 {
629         _enter("%d", conn->debug_id);
630
631         trace_rxrpc_client(conn, -1, rxrpc_client_activate_chans);
632
633         if (conn->active_chans == RXRPC_ACTIVE_CHANS_MASK)
634                 return;
635
636         spin_lock(&conn->channel_lock);
637         rxrpc_activate_channels_locked(conn);
638         spin_unlock(&conn->channel_lock);
639         _leave("");
640 }
641
642 /*
643  * Wait for a callNumber and a channel to be granted to a call.
644  */
645 static int rxrpc_wait_for_channel(struct rxrpc_call *call, gfp_t gfp)
646 {
647         int ret = 0;
648
649         _enter("%d", call->debug_id);
650
651         if (!call->call_id) {
652                 DECLARE_WAITQUEUE(myself, current);
653
654                 if (!gfpflags_allow_blocking(gfp)) {
655                         ret = -EAGAIN;
656                         goto out;
657                 }
658
659                 add_wait_queue_exclusive(&call->waitq, &myself);
660                 for (;;) {
661                         set_current_state(TASK_INTERRUPTIBLE);
662                         if (call->call_id)
663                                 break;
664                         if (signal_pending(current)) {
665                                 ret = -ERESTARTSYS;
666                                 break;
667                         }
668                         schedule();
669                 }
670                 remove_wait_queue(&call->waitq, &myself);
671                 __set_current_state(TASK_RUNNING);
672         }
673
674         /* Paired with the write barrier in rxrpc_activate_one_channel(). */
675         smp_rmb();
676
677 out:
678         _leave(" = %d", ret);
679         return ret;
680 }
681
682 /*
683  * find a connection for a call
684  * - called in process context with IRQs enabled
685  */
686 int rxrpc_connect_call(struct rxrpc_call *call,
687                        struct rxrpc_conn_parameters *cp,
688                        struct sockaddr_rxrpc *srx,
689                        gfp_t gfp)
690 {
691         struct rxrpc_net *rxnet = cp->local->rxnet;
692         int ret;
693
694         _enter("{%d,%lx},", call->debug_id, call->user_call_ID);
695
696         rxrpc_discard_expired_client_conns(&rxnet->client_conn_reaper);
697         rxrpc_cull_active_client_conns(rxnet);
698
699         ret = rxrpc_get_client_conn(call, cp, srx, gfp);
700         if (ret < 0)
701                 goto out;
702
703         rxrpc_animate_client_conn(rxnet, call->conn);
704         rxrpc_activate_channels(call->conn);
705
706         ret = rxrpc_wait_for_channel(call, gfp);
707         if (ret < 0) {
708                 rxrpc_disconnect_client_call(call);
709                 goto out;
710         }
711
712         spin_lock_bh(&call->conn->params.peer->lock);
713         hlist_add_head(&call->error_link,
714                        &call->conn->params.peer->error_targets);
715         spin_unlock_bh(&call->conn->params.peer->lock);
716
717 out:
718         _leave(" = %d", ret);
719         return ret;
720 }
721
722 /*
723  * Note that a connection is about to be exposed to the world.  Once it is
724  * exposed, we maintain an extra ref on it that stops it from being summarily
725  * discarded before it's (a) had a chance to deal with retransmission and (b)
726  * had a chance at re-use (the per-connection security negotiation is
727  * expensive).
728  */
729 static void rxrpc_expose_client_conn(struct rxrpc_connection *conn,
730                                      unsigned int channel)
731 {
732         if (!test_and_set_bit(RXRPC_CONN_EXPOSED, &conn->flags)) {
733                 trace_rxrpc_client(conn, channel, rxrpc_client_exposed);
734                 rxrpc_get_connection(conn);
735         }
736 }
737
738 /*
739  * Note that a call, and thus a connection, is about to be exposed to the
740  * world.
741  */
742 void rxrpc_expose_client_call(struct rxrpc_call *call)
743 {
744         unsigned int channel = call->cid & RXRPC_CHANNELMASK;
745         struct rxrpc_connection *conn = call->conn;
746         struct rxrpc_channel *chan = &conn->channels[channel];
747
748         if (!test_and_set_bit(RXRPC_CALL_EXPOSED, &call->flags)) {
749                 /* Mark the call ID as being used.  If the callNumber counter
750                  * exceeds ~2 billion, we kill the connection after its
751                  * outstanding calls have finished so that the counter doesn't
752                  * wrap.
753                  */
754                 chan->call_counter++;
755                 if (chan->call_counter >= INT_MAX)
756                         set_bit(RXRPC_CONN_DONT_REUSE, &conn->flags);
757                 rxrpc_expose_client_conn(conn, channel);
758         }
759 }
760
761 /*
762  * Set the reap timer.
763  */
764 static void rxrpc_set_client_reap_timer(struct rxrpc_net *rxnet)
765 {
766         unsigned long now = jiffies;
767         unsigned long reap_at = now + rxrpc_conn_idle_client_expiry;
768
769         if (rxnet->live)
770                 timer_reduce(&rxnet->client_conn_reap_timer, reap_at);
771 }
772
773 /*
774  * Disconnect a client call.
775  */
776 void rxrpc_disconnect_client_call(struct rxrpc_call *call)
777 {
778         unsigned int channel = call->cid & RXRPC_CHANNELMASK;
779         struct rxrpc_connection *conn = call->conn;
780         struct rxrpc_channel *chan = &conn->channels[channel];
781         struct rxrpc_net *rxnet = conn->params.local->rxnet;
782
783         trace_rxrpc_client(conn, channel, rxrpc_client_chan_disconnect);
784         call->conn = NULL;
785
786         spin_lock(&conn->channel_lock);
787
788         /* Calls that have never actually been assigned a channel can simply be
789          * discarded.  If the conn didn't get used either, it will follow
790          * immediately unless someone else grabs it in the meantime.
791          */
792         if (!list_empty(&call->chan_wait_link)) {
793                 _debug("call is waiting");
794                 ASSERTCMP(call->call_id, ==, 0);
795                 ASSERT(!test_bit(RXRPC_CALL_EXPOSED, &call->flags));
796                 list_del_init(&call->chan_wait_link);
797
798                 trace_rxrpc_client(conn, channel, rxrpc_client_chan_unstarted);
799
800                 /* We must deactivate or idle the connection if it's now
801                  * waiting for nothing.
802                  */
803                 spin_lock(&rxnet->client_conn_cache_lock);
804                 if (conn->cache_state == RXRPC_CONN_CLIENT_WAITING &&
805                     list_empty(&conn->waiting_calls) &&
806                     !conn->active_chans)
807                         goto idle_connection;
808                 goto out;
809         }
810
811         ASSERTCMP(rcu_access_pointer(chan->call), ==, call);
812
813         /* If a client call was exposed to the world, we save the result for
814          * retransmission.
815          *
816          * We use a barrier here so that the call number and abort code can be
817          * read without needing to take a lock.
818          *
819          * TODO: Make the incoming packet handler check this and handle
820          * terminal retransmission without requiring access to the call.
821          */
822         if (test_bit(RXRPC_CALL_EXPOSED, &call->flags)) {
823                 _debug("exposed %u,%u", call->call_id, call->abort_code);
824                 __rxrpc_disconnect_call(conn, call);
825         }
826
827         /* See if we can pass the channel directly to another call. */
828         if (conn->cache_state == RXRPC_CONN_CLIENT_ACTIVE &&
829             !list_empty(&conn->waiting_calls)) {
830                 trace_rxrpc_client(conn, channel, rxrpc_client_chan_pass);
831                 rxrpc_activate_one_channel(conn, channel);
832                 goto out_2;
833         }
834
835         /* Schedule the final ACK to be transmitted in a short while so that it
836          * can be skipped if we find a follow-on call.  The first DATA packet
837          * of the follow on call will implicitly ACK this call.
838          */
839         if (call->completion == RXRPC_CALL_SUCCEEDED &&
840             test_bit(RXRPC_CALL_EXPOSED, &call->flags)) {
841                 unsigned long final_ack_at = jiffies + 2;
842
843                 WRITE_ONCE(chan->final_ack_at, final_ack_at);
844                 smp_wmb(); /* vs rxrpc_process_delayed_final_acks() */
845                 set_bit(RXRPC_CONN_FINAL_ACK_0 + channel, &conn->flags);
846                 rxrpc_reduce_conn_timer(conn, final_ack_at);
847         }
848
849         /* Things are more complex and we need the cache lock.  We might be
850          * able to simply idle the conn or it might now be lurking on the wait
851          * list.  It might even get moved back to the active list whilst we're
852          * waiting for the lock.
853          */
854         spin_lock(&rxnet->client_conn_cache_lock);
855
856         switch (conn->cache_state) {
857         case RXRPC_CONN_CLIENT_UPGRADE:
858                 /* Deal with termination of a service upgrade probe. */
859                 if (test_bit(RXRPC_CONN_EXPOSED, &conn->flags)) {
860                         clear_bit(RXRPC_CONN_PROBING_FOR_UPGRADE, &conn->flags);
861                         trace_rxrpc_client(conn, channel, rxrpc_client_to_active);
862                         conn->cache_state = RXRPC_CONN_CLIENT_ACTIVE;
863                         rxrpc_activate_channels_locked(conn);
864                 }
865                 /* fall through */
866         case RXRPC_CONN_CLIENT_ACTIVE:
867                 if (list_empty(&conn->waiting_calls)) {
868                         rxrpc_deactivate_one_channel(conn, channel);
869                         if (!conn->active_chans) {
870                                 rxnet->nr_active_client_conns--;
871                                 goto idle_connection;
872                         }
873                         goto out;
874                 }
875
876                 trace_rxrpc_client(conn, channel, rxrpc_client_chan_pass);
877                 rxrpc_activate_one_channel(conn, channel);
878                 goto out;
879
880         case RXRPC_CONN_CLIENT_CULLED:
881                 rxrpc_deactivate_one_channel(conn, channel);
882                 ASSERT(list_empty(&conn->waiting_calls));
883                 if (!conn->active_chans)
884                         goto idle_connection;
885                 goto out;
886
887         case RXRPC_CONN_CLIENT_WAITING:
888                 rxrpc_deactivate_one_channel(conn, channel);
889                 goto out;
890
891         default:
892                 BUG();
893         }
894
895 out:
896         spin_unlock(&rxnet->client_conn_cache_lock);
897 out_2:
898         spin_unlock(&conn->channel_lock);
899         rxrpc_put_connection(conn);
900         _leave("");
901         return;
902
903 idle_connection:
904         /* As no channels remain active, the connection gets deactivated
905          * immediately or moved to the idle list for a short while.
906          */
907         if (test_bit(RXRPC_CONN_EXPOSED, &conn->flags)) {
908                 trace_rxrpc_client(conn, channel, rxrpc_client_to_idle);
909                 conn->idle_timestamp = jiffies;
910                 conn->cache_state = RXRPC_CONN_CLIENT_IDLE;
911                 list_move_tail(&conn->cache_link, &rxnet->idle_client_conns);
912                 if (rxnet->idle_client_conns.next == &conn->cache_link &&
913                     !rxnet->kill_all_client_conns)
914                         rxrpc_set_client_reap_timer(rxnet);
915         } else {
916                 trace_rxrpc_client(conn, channel, rxrpc_client_to_inactive);
917                 conn->cache_state = RXRPC_CONN_CLIENT_INACTIVE;
918                 list_del_init(&conn->cache_link);
919         }
920         goto out;
921 }
922
923 /*
924  * Clean up a dead client connection.
925  */
926 static struct rxrpc_connection *
927 rxrpc_put_one_client_conn(struct rxrpc_connection *conn)
928 {
929         struct rxrpc_connection *next = NULL;
930         struct rxrpc_local *local = conn->params.local;
931         struct rxrpc_net *rxnet = local->rxnet;
932         unsigned int nr_conns;
933
934         trace_rxrpc_client(conn, -1, rxrpc_client_cleanup);
935
936         if (test_bit(RXRPC_CONN_IN_CLIENT_CONNS, &conn->flags)) {
937                 spin_lock(&local->client_conns_lock);
938                 if (test_and_clear_bit(RXRPC_CONN_IN_CLIENT_CONNS,
939                                        &conn->flags))
940                         rb_erase(&conn->client_node, &local->client_conns);
941                 spin_unlock(&local->client_conns_lock);
942         }
943
944         rxrpc_put_client_connection_id(conn);
945
946         ASSERTCMP(conn->cache_state, ==, RXRPC_CONN_CLIENT_INACTIVE);
947
948         if (test_bit(RXRPC_CONN_COUNTED, &conn->flags)) {
949                 trace_rxrpc_client(conn, -1, rxrpc_client_uncount);
950                 spin_lock(&rxnet->client_conn_cache_lock);
951                 nr_conns = --rxnet->nr_client_conns;
952
953                 if (nr_conns < rxrpc_max_client_connections &&
954                     !list_empty(&rxnet->waiting_client_conns)) {
955                         next = list_entry(rxnet->waiting_client_conns.next,
956                                           struct rxrpc_connection, cache_link);
957                         rxrpc_get_connection(next);
958                         rxrpc_activate_conn(rxnet, next);
959                 }
960
961                 spin_unlock(&rxnet->client_conn_cache_lock);
962         }
963
964         rxrpc_kill_connection(conn);
965         if (next)
966                 rxrpc_activate_channels(next);
967
968         /* We need to get rid of the temporary ref we took upon next, but we
969          * can't call rxrpc_put_connection() recursively.
970          */
971         return next;
972 }
973
974 /*
975  * Clean up a dead client connections.
976  */
977 void rxrpc_put_client_conn(struct rxrpc_connection *conn)
978 {
979         const void *here = __builtin_return_address(0);
980         int n;
981
982         do {
983                 n = atomic_dec_return(&conn->usage);
984                 trace_rxrpc_conn(conn, rxrpc_conn_put_client, n, here);
985                 if (n > 0)
986                         return;
987                 ASSERTCMP(n, >=, 0);
988
989                 conn = rxrpc_put_one_client_conn(conn);
990         } while (conn);
991 }
992
993 /*
994  * Kill the longest-active client connections to make room for new ones.
995  */
996 static void rxrpc_cull_active_client_conns(struct rxrpc_net *rxnet)
997 {
998         struct rxrpc_connection *conn;
999         unsigned int nr_conns = rxnet->nr_client_conns;
1000         unsigned int nr_active, limit;
1001
1002         _enter("");
1003
1004         ASSERTCMP(nr_conns, >=, 0);
1005         if (nr_conns < rxrpc_max_client_connections) {
1006                 _leave(" [ok]");
1007                 return;
1008         }
1009         limit = rxrpc_reap_client_connections;
1010
1011         spin_lock(&rxnet->client_conn_cache_lock);
1012         nr_active = rxnet->nr_active_client_conns;
1013
1014         while (nr_active > limit) {
1015                 ASSERT(!list_empty(&rxnet->active_client_conns));
1016                 conn = list_entry(rxnet->active_client_conns.next,
1017                                   struct rxrpc_connection, cache_link);
1018                 ASSERTIFCMP(conn->cache_state != RXRPC_CONN_CLIENT_ACTIVE,
1019                             conn->cache_state, ==, RXRPC_CONN_CLIENT_UPGRADE);
1020
1021                 if (list_empty(&conn->waiting_calls)) {
1022                         trace_rxrpc_client(conn, -1, rxrpc_client_to_culled);
1023                         conn->cache_state = RXRPC_CONN_CLIENT_CULLED;
1024                         list_del_init(&conn->cache_link);
1025                 } else {
1026                         trace_rxrpc_client(conn, -1, rxrpc_client_to_waiting);
1027                         conn->cache_state = RXRPC_CONN_CLIENT_WAITING;
1028                         list_move_tail(&conn->cache_link,
1029                                        &rxnet->waiting_client_conns);
1030                 }
1031
1032                 nr_active--;
1033         }
1034
1035         rxnet->nr_active_client_conns = nr_active;
1036         spin_unlock(&rxnet->client_conn_cache_lock);
1037         ASSERTCMP(nr_active, >=, 0);
1038         _leave(" [culled]");
1039 }
1040
1041 /*
1042  * Discard expired client connections from the idle list.  Each conn in the
1043  * idle list has been exposed and holds an extra ref because of that.
1044  *
1045  * This may be called from conn setup or from a work item so cannot be
1046  * considered non-reentrant.
1047  */
1048 void rxrpc_discard_expired_client_conns(struct work_struct *work)
1049 {
1050         struct rxrpc_connection *conn;
1051         struct rxrpc_net *rxnet =
1052                 container_of(work, struct rxrpc_net, client_conn_reaper);
1053         unsigned long expiry, conn_expires_at, now;
1054         unsigned int nr_conns;
1055
1056         _enter("");
1057
1058         if (list_empty(&rxnet->idle_client_conns)) {
1059                 _leave(" [empty]");
1060                 return;
1061         }
1062
1063         /* Don't double up on the discarding */
1064         if (!spin_trylock(&rxnet->client_conn_discard_lock)) {
1065                 _leave(" [already]");
1066                 return;
1067         }
1068
1069         /* We keep an estimate of what the number of conns ought to be after
1070          * we've discarded some so that we don't overdo the discarding.
1071          */
1072         nr_conns = rxnet->nr_client_conns;
1073
1074 next:
1075         spin_lock(&rxnet->client_conn_cache_lock);
1076
1077         if (list_empty(&rxnet->idle_client_conns))
1078                 goto out;
1079
1080         conn = list_entry(rxnet->idle_client_conns.next,
1081                           struct rxrpc_connection, cache_link);
1082         ASSERT(test_bit(RXRPC_CONN_EXPOSED, &conn->flags));
1083
1084         if (!rxnet->kill_all_client_conns) {
1085                 /* If the number of connections is over the reap limit, we
1086                  * expedite discard by reducing the expiry timeout.  We must,
1087                  * however, have at least a short grace period to be able to do
1088                  * final-ACK or ABORT retransmission.
1089                  */
1090                 expiry = rxrpc_conn_idle_client_expiry;
1091                 if (nr_conns > rxrpc_reap_client_connections)
1092                         expiry = rxrpc_conn_idle_client_fast_expiry;
1093                 if (conn->params.local->service_closed)
1094                         expiry = rxrpc_closed_conn_expiry * HZ;
1095
1096                 conn_expires_at = conn->idle_timestamp + expiry;
1097
1098                 now = READ_ONCE(jiffies);
1099                 if (time_after(conn_expires_at, now))
1100                         goto not_yet_expired;
1101         }
1102
1103         trace_rxrpc_client(conn, -1, rxrpc_client_discard);
1104         if (!test_and_clear_bit(RXRPC_CONN_EXPOSED, &conn->flags))
1105                 BUG();
1106         conn->cache_state = RXRPC_CONN_CLIENT_INACTIVE;
1107         list_del_init(&conn->cache_link);
1108
1109         spin_unlock(&rxnet->client_conn_cache_lock);
1110
1111         /* When we cleared the EXPOSED flag, we took on responsibility for the
1112          * reference that that had on the usage count.  We deal with that here.
1113          * If someone re-sets the flag and re-gets the ref, that's fine.
1114          */
1115         rxrpc_put_connection(conn);
1116         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 }