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 int rxrpc_conn_idle_client_expiry = 2 * 60 * HZ;
  89 __read_mostly unsigned int 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         /* If the connection is not meant to be exclusive, search the available
 296          * connections to see if the connection we want to use already exists.
 297          */
 298         if (!cp->exclusive) {
 299                 _debug("search 1");
 300                 spin_lock(&local->client_conns_lock);
 301                 p = local->client_conns.rb_node;
 302                 while (p) {
 303                         conn = rb_entry(p, struct rxrpc_connection, client_node);
 304
 305 #define cmp(X) ((long)conn->params.X - (long)cp->X)
 306                         diff = (cmp(peer) ?:
 307                                 cmp(key) ?:
 308                                 cmp(security_level) ?:
 309                                 cmp(upgrade));
 310 #undef cmp
 311                         if (diff < 0) {
 312                                 p = p->rb_left;
 313                         } else if (diff > 0) {
 314                                 p = p->rb_right;
 315                         } else {
 316                                 if (rxrpc_may_reuse_conn(conn) &&
 317                                     rxrpc_get_connection_maybe(conn))
 318                                         goto found_extant_conn;
 319                                 /* The connection needs replacing.  It's better
 320                                  * to effect that when we have something to
 321                                  * replace it with so that we don't have to
 322                                  * rebalance the tree twice.
 323                                  */
 324                                 break;
 325                         }
 326                 }
 327                 spin_unlock(&local->client_conns_lock);
 328         }
 329
 330         /* There wasn't a connection yet or we need an exclusive connection.
 331          * We need to create a candidate and then potentially redo the search
 332          * in case we're racing with another thread also trying to connect on a
 333          * shareable connection.
 334          */
 335         _debug("new conn");
 336         candidate = rxrpc_alloc_client_connection(cp, gfp);
 337         if (IS_ERR(candidate)) {
 338                 ret = PTR_ERR(candidate);
 339                 goto error_peer;
 340         }
 341
 342         /* Add the call to the new connection's waiting list in case we're
 343          * going to have to wait for the connection to come live.  It's our
 344          * connection, so we want first dibs on the channel slots.  We would
 345          * normally have to take channel_lock but we do this before anyone else
 346          * can see the connection.
 347          */
 348         list_add_tail(&call->chan_wait_link, &candidate->waiting_calls);
 349
 350         if (cp->exclusive) {
 351                 call->conn = candidate;
 352                 call->security_ix = candidate->security_ix;
 353                 call->service_id = candidate->service_id;
 354                 _leave(" = 0 [exclusive %d]", candidate->debug_id);
 355                 return 0;
 356         }
 357
 358         /* Publish the new connection for userspace to find.  We need to redo
 359          * the search before doing this lest we race with someone else adding a
 360          * conflicting instance.
 361          */
 362         _debug("search 2");
 363         spin_lock(&local->client_conns_lock);
 364
 365         pp = &local->client_conns.rb_node;
 366         parent = NULL;
 367         while (*pp) {
 368                 parent = *pp;
 369                 conn = rb_entry(parent, struct rxrpc_connection, client_node);
 370
 371 #define cmp(X) ((long)conn->params.X - (long)candidate->params.X)
 372                 diff = (cmp(peer) ?:
 373                         cmp(key) ?:
 374                         cmp(security_level) ?:
 375                         cmp(upgrade));
 376 #undef cmp
 377                 if (diff < 0) {
 378                         pp = &(*pp)->rb_left;
 379                 } else if (diff > 0) {
 380                         pp = &(*pp)->rb_right;
 381                 } else {
 382                         if (rxrpc_may_reuse_conn(conn) &&
 383                             rxrpc_get_connection_maybe(conn))
 384                                 goto found_extant_conn;
 385                         /* The old connection is from an outdated epoch. */
 386                         _debug("replace conn");
 387                         clear_bit(RXRPC_CONN_IN_CLIENT_CONNS, &conn->flags);
 388                         rb_replace_node(&conn->client_node,
 389                                         &candidate->client_node,
 390                                         &local->client_conns);
 391                         trace_rxrpc_client(conn, -1, rxrpc_client_replace);
 392                         goto candidate_published;
 393                 }
 394         }
 395
 396         _debug("new conn");
 397         rb_link_node(&candidate->client_node, parent, pp);
 398         rb_insert_color(&candidate->client_node, &local->client_conns);
 399
 400 candidate_published:
 401         set_bit(RXRPC_CONN_IN_CLIENT_CONNS, &candidate->flags);
 402         call->conn = candidate;
 403         call->security_ix = candidate->security_ix;
 404         call->service_id = candidate->service_id;
 405         spin_unlock(&local->client_conns_lock);
 406         _leave(" = 0 [new %d]", candidate->debug_id);
 407         return 0;
 408
 409         /* We come here if we found a suitable connection already in existence.
 410          * Discard any candidate we may have allocated, and try to get a
 411          * channel on this one.
 412          */
 413 found_extant_conn:
 414         _debug("found conn");
 415         spin_unlock(&local->client_conns_lock);
 416
 417         if (candidate) {
 418                 trace_rxrpc_client(candidate, -1, rxrpc_client_duplicate);
 419                 rxrpc_put_connection(candidate);
 420                 candidate = NULL;
 421         }
 422
 423         spin_lock(&conn->channel_lock);
 424         call->conn = conn;
 425         call->security_ix = conn->security_ix;
 426         call->service_id = conn->service_id;
 427         list_add(&call->chan_wait_link, &conn->waiting_calls);
 428         spin_unlock(&conn->channel_lock);
 429         _leave(" = 0 [extant %d]", conn->debug_id);
 430         return 0;
 431
 432 error_peer:
 433         rxrpc_put_peer(cp->peer);
 434         cp->peer = NULL;
 435 error:
 436         _leave(" = %d", ret);
 437         return ret;
 438 }
 439
 440 /*
 441  * Activate a connection.
 442  */
 443 static void rxrpc_activate_conn(struct rxrpc_net *rxnet,
 444                                 struct rxrpc_connection *conn)
 445 {
 446         if (test_bit(RXRPC_CONN_PROBING_FOR_UPGRADE, &conn->flags)) {
 447                 trace_rxrpc_client(conn, -1, rxrpc_client_to_upgrade);
 448                 conn->cache_state = RXRPC_CONN_CLIENT_UPGRADE;
 449         } else {
 450                 trace_rxrpc_client(conn, -1, rxrpc_client_to_active);
 451                 conn->cache_state = RXRPC_CONN_CLIENT_ACTIVE;
 452         }
 453         rxnet->nr_active_client_conns++;
 454         list_move_tail(&conn->cache_link, &rxnet->active_client_conns);
 455 }
 456
 457 /*
 458  * Attempt to animate a connection for a new call.
 459  *
 460  * If it's not exclusive, the connection is in the endpoint tree, and we're in
 461  * the conn's list of those waiting to grab a channel.  There is, however, a
 462  * limit on the number of live connections allowed at any one time, so we may
 463  * have to wait for capacity to become available.
 464  *
 465  * Note that a connection on the waiting queue might *also* have active
 466  * channels if it has been culled to make space and then re-requested by a new
 467  * call.
 468  */
 469 static void rxrpc_animate_client_conn(struct rxrpc_net *rxnet,
 470                                       struct rxrpc_connection *conn)
 471 {
 472         unsigned int nr_conns;
 473
 474         _enter("%d,%d", conn->debug_id, conn->cache_state);
 475
 476         if (conn->cache_state == RXRPC_CONN_CLIENT_ACTIVE ||
 477             conn->cache_state == RXRPC_CONN_CLIENT_UPGRADE)
 478                 goto out;
 479
 480         spin_lock(&rxnet->client_conn_cache_lock);
 481
 482         nr_conns = rxnet->nr_client_conns;
 483         if (!test_and_set_bit(RXRPC_CONN_COUNTED, &conn->flags)) {
 484                 trace_rxrpc_client(conn, -1, rxrpc_client_count);
 485                 rxnet->nr_client_conns = nr_conns + 1;
 486         }
 487
 488         switch (conn->cache_state) {
 489         case RXRPC_CONN_CLIENT_ACTIVE:
 490         case RXRPC_CONN_CLIENT_UPGRADE:
 491         case RXRPC_CONN_CLIENT_WAITING:
 492                 break;
 493
 494         case RXRPC_CONN_CLIENT_INACTIVE:
 495         case RXRPC_CONN_CLIENT_CULLED:
 496         case RXRPC_CONN_CLIENT_IDLE:
 497                 if (nr_conns >= rxrpc_max_client_connections)
 498                         goto wait_for_capacity;
 499                 goto activate_conn;
 500
 501         default:
 502                 BUG();
 503         }
 504
 505 out_unlock:
 506         spin_unlock(&rxnet->client_conn_cache_lock);
 507 out:
 508         _leave(" [%d]", conn->cache_state);
 509         return;
 510
 511 activate_conn:
 512         _debug("activate");
 513         rxrpc_activate_conn(rxnet, conn);
 514         goto out_unlock;
 515
 516 wait_for_capacity:
 517         _debug("wait");
 518         trace_rxrpc_client(conn, -1, rxrpc_client_to_waiting);
 519         conn->cache_state = RXRPC_CONN_CLIENT_WAITING;
 520         list_move_tail(&conn->cache_link, &rxnet->waiting_client_conns);
 521         goto out_unlock;
 522 }
 523
 524 /*
 525  * Deactivate a channel.
 526  */
 527 static void rxrpc_deactivate_one_channel(struct rxrpc_connection *conn,
 528                                          unsigned int channel)
 529 {
 530         struct rxrpc_channel *chan = &conn->channels[channel];
 531
 532         rcu_assign_pointer(chan->call, NULL);
 533         conn->active_chans &= ~(1 << channel);
 534 }
 535
 536 /*
 537  * Assign a channel to the call at the front of the queue and wake the call up.
 538  * We don't increment the callNumber counter until this number has been exposed
 539  * to the world.
 540  */
 541 static void rxrpc_activate_one_channel(struct rxrpc_connection *conn,
 542                                        unsigned int channel)
 543 {
 544         struct rxrpc_channel *chan = &conn->channels[channel];
 545         struct rxrpc_call *call = list_entry(conn->waiting_calls.next,
 546                                              struct rxrpc_call, chan_wait_link);
 547         u32 call_id = chan->call_counter + 1;
 548
 549         trace_rxrpc_client(conn, channel, rxrpc_client_chan_activate);
 550
 551         write_lock_bh(&call->state_lock);
 552         call->state = RXRPC_CALL_CLIENT_SEND_REQUEST;
 553         write_unlock_bh(&call->state_lock);
 554
 555         rxrpc_see_call(call);
 556         list_del_init(&call->chan_wait_link);
 557         conn->active_chans |= 1 << channel;
 558         call->peer      = rxrpc_get_peer(conn->params.peer);
 559         call->cid       = conn->proto.cid | channel;
 560         call->call_id   = call_id;
 561
 562         trace_rxrpc_connect_call(call);
 563         _net("CONNECT call %08x:%08x as call %d on conn %d",
 564              call->cid, call->call_id, call->debug_id, conn->debug_id);
 565
 566         /* Paired with the read barrier in rxrpc_wait_for_channel().  This
 567          * orders cid and epoch in the connection wrt to call_id without the
 568          * need to take the channel_lock.
 569          *
 570          * We provisionally assign a callNumber at this point, but we don't
 571          * confirm it until the call is about to be exposed.
 572          *
 573          * TODO: Pair with a barrier in the data_ready handler when that looks
 574          * at the call ID through a connection channel.
 575          */
 576         smp_wmb();
 577         chan->call_id   = call_id;
 578         rcu_assign_pointer(chan->call, call);
 579         wake_up(&call->waitq);
 580 }
 581
 582 /*
 583  * Assign channels and callNumbers to waiting calls with channel_lock
 584  * held by caller.
 585  */
 586 static void rxrpc_activate_channels_locked(struct rxrpc_connection *conn)
 587 {
 588         u8 avail, mask;
 589
 590         switch (conn->cache_state) {
 591         case RXRPC_CONN_CLIENT_ACTIVE:
 592                 mask = RXRPC_ACTIVE_CHANS_MASK;
 593                 break;
 594         case RXRPC_CONN_CLIENT_UPGRADE:
 595                 mask = 0x01;
 596                 break;
 597         default:
 598                 return;
 599         }
 600
 601         while (!list_empty(&conn->waiting_calls) &&
 602                (avail = ~conn->active_chans,
 603                 avail &= mask,
 604                 avail != 0))
 605                 rxrpc_activate_one_channel(conn, __ffs(avail));
 606 }
 607
 608 /*
 609  * Assign channels and callNumbers to waiting calls.
 610  */
 611 static void rxrpc_activate_channels(struct rxrpc_connection *conn)
 612 {
 613         _enter("%d", conn->debug_id);
 614
 615         trace_rxrpc_client(conn, -1, rxrpc_client_activate_chans);
 616
 617         if (conn->active_chans == RXRPC_ACTIVE_CHANS_MASK)
 618                 return;
 619
 620         spin_lock(&conn->channel_lock);
 621         rxrpc_activate_channels_locked(conn);
 622         spin_unlock(&conn->channel_lock);
 623         _leave("");
 624 }
 625
 626 /*
 627  * Wait for a callNumber and a channel to be granted to a call.
 628  */
 629 static int rxrpc_wait_for_channel(struct rxrpc_call *call, gfp_t gfp)
 630 {
 631         int ret = 0;
 632
 633         _enter("%d", call->debug_id);
 634
 635         if (!call->call_id) {
 636                 DECLARE_WAITQUEUE(myself, current);
 637
 638                 if (!gfpflags_allow_blocking(gfp)) {
 639                         ret = -EAGAIN;
 640                         goto out;
 641                 }
 642
 643                 add_wait_queue_exclusive(&call->waitq, &myself);
 644                 for (;;) {
 645                         set_current_state(TASK_INTERRUPTIBLE);
 646                         if (call->call_id)
 647                                 break;
 648                         if (signal_pending(current)) {
 649                                 ret = -ERESTARTSYS;
 650                                 break;
 651                         }
 652                         schedule();
 653                 }
 654                 remove_wait_queue(&call->waitq, &myself);
 655                 __set_current_state(TASK_RUNNING);
 656         }
 657
 658         /* Paired with the write barrier in rxrpc_activate_one_channel(). */
 659         smp_rmb();
 660
 661 out:
 662         _leave(" = %d", ret);
 663         return ret;
 664 }
 665
 666 /*
 667  * find a connection for a call
 668  * - called in process context with IRQs enabled
 669  */
 670 int rxrpc_connect_call(struct rxrpc_call *call,
 671                        struct rxrpc_conn_parameters *cp,
 672                        struct sockaddr_rxrpc *srx,
 673                        gfp_t gfp)
 674 {
 675         struct rxrpc_net *rxnet = cp->local->rxnet;
 676         int ret;
 677
 678         _enter("{%d,%lx},", call->debug_id, call->user_call_ID);
 679
 680         rxrpc_discard_expired_client_conns(&rxnet->client_conn_reaper.work);
 681         rxrpc_cull_active_client_conns(rxnet);
 682
 683         ret = rxrpc_get_client_conn(call, cp, srx, gfp);
 684         if (ret < 0)
 685                 return ret;
 686
 687         rxrpc_animate_client_conn(rxnet, call->conn);
 688         rxrpc_activate_channels(call->conn);
 689
 690         ret = rxrpc_wait_for_channel(call, gfp);
 691         if (ret < 0)
 692                 rxrpc_disconnect_client_call(call);
 693
 694         _leave(" = %d", ret);
 695         return ret;
 696 }
 697
 698 /*
 699  * Note that a connection is about to be exposed to the world.  Once it is
 700  * exposed, we maintain an extra ref on it that stops it from being summarily
 701  * discarded before it's (a) had a chance to deal with retransmission and (b)
 702  * had a chance at re-use (the per-connection security negotiation is
 703  * expensive).
 704  */
 705 static void rxrpc_expose_client_conn(struct rxrpc_connection *conn,
 706                                      unsigned int channel)
 707 {
 708         if (!test_and_set_bit(RXRPC_CONN_EXPOSED, &conn->flags)) {
 709                 trace_rxrpc_client(conn, channel, rxrpc_client_exposed);
 710                 rxrpc_get_connection(conn);
 711         }
 712 }
 713
 714 /*
 715  * Note that a call, and thus a connection, is about to be exposed to the
 716  * world.
 717  */
 718 void rxrpc_expose_client_call(struct rxrpc_call *call)
 719 {
 720         unsigned int channel = call->cid & RXRPC_CHANNELMASK;
 721         struct rxrpc_connection *conn = call->conn;
 722         struct rxrpc_channel *chan = &conn->channels[channel];
 723
 724         if (!test_and_set_bit(RXRPC_CALL_EXPOSED, &call->flags)) {
 725                 /* Mark the call ID as being used.  If the callNumber counter
 726                  * exceeds ~2 billion, we kill the connection after its
 727                  * outstanding calls have finished so that the counter doesn't
 728                  * wrap.
 729                  */
 730                 chan->call_counter++;
 731                 if (chan->call_counter >= INT_MAX)
 732                         set_bit(RXRPC_CONN_DONT_REUSE, &conn->flags);
 733                 rxrpc_expose_client_conn(conn, channel);
 734         }
 735 }
 736
 737 /*
 738  * Disconnect a client call.
 739  */
 740 void rxrpc_disconnect_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         struct rxrpc_net *rxnet = rxrpc_net(sock_net(&call->socket->sk));
 746
 747         trace_rxrpc_client(conn, channel, rxrpc_client_chan_disconnect);
 748         call->conn = NULL;
 749
 750         spin_lock(&conn->channel_lock);
 751
 752         /* Calls that have never actually been assigned a channel can simply be
 753          * discarded.  If the conn didn't get used either, it will follow
 754          * immediately unless someone else grabs it in the meantime.
 755          */
 756         if (!list_empty(&call->chan_wait_link)) {
 757                 _debug("call is waiting");
 758                 ASSERTCMP(call->call_id, ==, 0);
 759                 ASSERT(!test_bit(RXRPC_CALL_EXPOSED, &call->flags));
 760                 list_del_init(&call->chan_wait_link);
 761
 762                 trace_rxrpc_client(conn, channel, rxrpc_client_chan_unstarted);
 763
 764                 /* We must deactivate or idle the connection if it's now
 765                  * waiting for nothing.
 766                  */
 767                 spin_lock(&rxnet->client_conn_cache_lock);
 768                 if (conn->cache_state == RXRPC_CONN_CLIENT_WAITING &&
 769                     list_empty(&conn->waiting_calls) &&
 770                     !conn->active_chans)
 771                         goto idle_connection;
 772                 goto out;
 773         }
 774
 775         ASSERTCMP(rcu_access_pointer(chan->call), ==, call);
 776
 777         /* If a client call was exposed to the world, we save the result for
 778          * retransmission.
 779          *
 780          * We use a barrier here so that the call number and abort code can be
 781          * read without needing to take a lock.
 782          *
 783          * TODO: Make the incoming packet handler check this and handle
 784          * terminal retransmission without requiring access to the call.
 785          */
 786         if (test_bit(RXRPC_CALL_EXPOSED, &call->flags)) {
 787                 _debug("exposed %u,%u", call->call_id, call->abort_code);
 788                 __rxrpc_disconnect_call(conn, call);
 789         }
 790
 791         /* See if we can pass the channel directly to another call. */
 792         if (conn->cache_state == RXRPC_CONN_CLIENT_ACTIVE &&
 793             !list_empty(&conn->waiting_calls)) {
 794                 trace_rxrpc_client(conn, channel, rxrpc_client_chan_pass);
 795                 rxrpc_activate_one_channel(conn, channel);
 796                 goto out_2;
 797         }
 798
 799         /* Things are more complex and we need the cache lock.  We might be
 800          * able to simply idle the conn or it might now be lurking on the wait
 801          * list.  It might even get moved back to the active list whilst we're
 802          * waiting for the lock.
 803          */
 804         spin_lock(&rxnet->client_conn_cache_lock);
 805
 806         switch (conn->cache_state) {
 807         case RXRPC_CONN_CLIENT_UPGRADE:
 808                 /* Deal with termination of a service upgrade probe. */
 809                 if (test_bit(RXRPC_CONN_EXPOSED, &conn->flags)) {
 810                         clear_bit(RXRPC_CONN_PROBING_FOR_UPGRADE, &conn->flags);
 811                         trace_rxrpc_client(conn, channel, rxrpc_client_to_active);
 812                         conn->cache_state = RXRPC_CONN_CLIENT_ACTIVE;
 813                         rxrpc_activate_channels_locked(conn);
 814                 }
 815                 /* fall through */
 816         case RXRPC_CONN_CLIENT_ACTIVE:
 817                 if (list_empty(&conn->waiting_calls)) {
 818                         rxrpc_deactivate_one_channel(conn, channel);
 819                         if (!conn->active_chans) {
 820                                 rxnet->nr_active_client_conns--;
 821                                 goto idle_connection;
 822                         }
 823                         goto out;
 824                 }
 825
 826                 trace_rxrpc_client(conn, channel, rxrpc_client_chan_pass);
 827                 rxrpc_activate_one_channel(conn, channel);
 828                 goto out;
 829
 830         case RXRPC_CONN_CLIENT_CULLED:
 831                 rxrpc_deactivate_one_channel(conn, channel);
 832                 ASSERT(list_empty(&conn->waiting_calls));
 833                 if (!conn->active_chans)
 834                         goto idle_connection;
 835                 goto out;
 836
 837         case RXRPC_CONN_CLIENT_WAITING:
 838                 rxrpc_deactivate_one_channel(conn, channel);
 839                 goto out;
 840
 841         default:
 842                 BUG();
 843         }
 844
 845 out:
 846         spin_unlock(&rxnet->client_conn_cache_lock);
 847 out_2:
 848         spin_unlock(&conn->channel_lock);
 849         rxrpc_put_connection(conn);
 850         _leave("");
 851         return;
 852
 853 idle_connection:
 854         /* As no channels remain active, the connection gets deactivated
 855          * immediately or moved to the idle list for a short while.
 856          */
 857         if (test_bit(RXRPC_CONN_EXPOSED, &conn->flags)) {
 858                 trace_rxrpc_client(conn, channel, rxrpc_client_to_idle);
 859                 conn->idle_timestamp = jiffies;
 860                 conn->cache_state = RXRPC_CONN_CLIENT_IDLE;
 861                 list_move_tail(&conn->cache_link, &rxnet->idle_client_conns);
 862                 if (rxnet->idle_client_conns.next == &conn->cache_link &&
 863                     !rxnet->kill_all_client_conns)
 864                         queue_delayed_work(rxrpc_workqueue,
 865                                            &rxnet->client_conn_reaper,
 866                                            rxrpc_conn_idle_client_expiry);
 867         } else {
 868                 trace_rxrpc_client(conn, channel, rxrpc_client_to_inactive);
 869                 conn->cache_state = RXRPC_CONN_CLIENT_INACTIVE;
 870                 list_del_init(&conn->cache_link);
 871         }
 872         goto out;
 873 }
 874
 875 /*
 876  * Clean up a dead client connection.
 877  */
 878 static struct rxrpc_connection *
 879 rxrpc_put_one_client_conn(struct rxrpc_connection *conn)
 880 {
 881         struct rxrpc_connection *next = NULL;
 882         struct rxrpc_local *local = conn->params.local;
 883         struct rxrpc_net *rxnet = local->rxnet;
 884         unsigned int nr_conns;
 885
 886         trace_rxrpc_client(conn, -1, rxrpc_client_cleanup);
 887
 888         if (test_bit(RXRPC_CONN_IN_CLIENT_CONNS, &conn->flags)) {
 889                 spin_lock(&local->client_conns_lock);
 890                 if (test_and_clear_bit(RXRPC_CONN_IN_CLIENT_CONNS,
 891                                        &conn->flags))
 892                         rb_erase(&conn->client_node, &local->client_conns);
 893                 spin_unlock(&local->client_conns_lock);
 894         }
 895
 896         rxrpc_put_client_connection_id(conn);
 897
 898         ASSERTCMP(conn->cache_state, ==, RXRPC_CONN_CLIENT_INACTIVE);
 899
 900         if (test_bit(RXRPC_CONN_COUNTED, &conn->flags)) {
 901                 trace_rxrpc_client(conn, -1, rxrpc_client_uncount);
 902                 spin_lock(&rxnet->client_conn_cache_lock);
 903                 nr_conns = --rxnet->nr_client_conns;
 904
 905                 if (nr_conns < rxrpc_max_client_connections &&
 906                     !list_empty(&rxnet->waiting_client_conns)) {
 907                         next = list_entry(rxnet->waiting_client_conns.next,
 908                                           struct rxrpc_connection, cache_link);
 909                         rxrpc_get_connection(next);
 910                         rxrpc_activate_conn(rxnet, next);
 911                 }
 912
 913                 spin_unlock(&rxnet->client_conn_cache_lock);
 914         }
 915
 916         rxrpc_kill_connection(conn);
 917         if (next)
 918                 rxrpc_activate_channels(next);
 919
 920         /* We need to get rid of the temporary ref we took upon next, but we
 921          * can't call rxrpc_put_connection() recursively.
 922          */
 923         return next;
 924 }
 925
 926 /*
 927  * Clean up a dead client connections.
 928  */
 929 void rxrpc_put_client_conn(struct rxrpc_connection *conn)
 930 {
 931         const void *here = __builtin_return_address(0);
 932         int n;
 933
 934         do {
 935                 n = atomic_dec_return(&conn->usage);
 936                 trace_rxrpc_conn(conn, rxrpc_conn_put_client, n, here);
 937                 if (n > 0)
 938                         return;
 939                 ASSERTCMP(n, >=, 0);
 940
 941                 conn = rxrpc_put_one_client_conn(conn);
 942         } while (conn);
 943 }
 944
 945 /*
 946  * Kill the longest-active client connections to make room for new ones.
 947  */
 948 static void rxrpc_cull_active_client_conns(struct rxrpc_net *rxnet)
 949 {
 950         struct rxrpc_connection *conn;
 951         unsigned int nr_conns = rxnet->nr_client_conns;
 952         unsigned int nr_active, limit;
 953
 954         _enter("");
 955
 956         ASSERTCMP(nr_conns, >=, 0);
 957         if (nr_conns < rxrpc_max_client_connections) {
 958                 _leave(" [ok]");
 959                 return;
 960         }
 961         limit = rxrpc_reap_client_connections;
 962
 963         spin_lock(&rxnet->client_conn_cache_lock);
 964         nr_active = rxnet->nr_active_client_conns;
 965
 966         while (nr_active > limit) {
 967                 ASSERT(!list_empty(&rxnet->active_client_conns));
 968                 conn = list_entry(rxnet->active_client_conns.next,
 969                                   struct rxrpc_connection, cache_link);
 970                 ASSERTIFCMP(conn->cache_state != RXRPC_CONN_CLIENT_ACTIVE,
 971                             conn->cache_state, ==, RXRPC_CONN_CLIENT_UPGRADE);
 972
 973                 if (list_empty(&conn->waiting_calls)) {
 974                         trace_rxrpc_client(conn, -1, rxrpc_client_to_culled);
 975                         conn->cache_state = RXRPC_CONN_CLIENT_CULLED;
 976                         list_del_init(&conn->cache_link);
 977                 } else {
 978                         trace_rxrpc_client(conn, -1, rxrpc_client_to_waiting);
 979                         conn->cache_state = RXRPC_CONN_CLIENT_WAITING;
 980                         list_move_tail(&conn->cache_link,
 981                                        &rxnet->waiting_client_conns);
 982                 }
 983
 984                 nr_active--;
 985         }
 986
 987         rxnet->nr_active_client_conns = nr_active;
 988         spin_unlock(&rxnet->client_conn_cache_lock);
 989         ASSERTCMP(nr_active, >=, 0);
 990         _leave(" [culled]");
 991 }
 992
 993 /*
 994  * Discard expired client connections from the idle list.  Each conn in the
 995  * idle list has been exposed and holds an extra ref because of that.
 996  *
 997  * This may be called from conn setup or from a work item so cannot be
 998  * considered non-reentrant.
 999  */
1000 void rxrpc_discard_expired_client_conns(struct work_struct *work)
1001 {
1002         struct rxrpc_connection *conn;
1003         struct rxrpc_net *rxnet =
1004                 container_of(to_delayed_work(work),
1005                              struct rxrpc_net, client_conn_reaper);
1006         unsigned long expiry, conn_expires_at, now;
1007         unsigned int nr_conns;
1008         bool did_discard = false;
1009
1010         _enter("");
1011
1012         if (list_empty(&rxnet->idle_client_conns)) {
1013                 _leave(" [empty]");
1014                 return;
1015         }
1016
1017         /* Don't double up on the discarding */
1018         if (!spin_trylock(&rxnet->client_conn_discard_lock)) {
1019                 _leave(" [already]");
1020                 return;
1021         }
1022
1023         /* We keep an estimate of what the number of conns ought to be after
1024          * we've discarded some so that we don't overdo the discarding.
1025          */
1026         nr_conns = rxnet->nr_client_conns;
1027
1028 next:
1029         spin_lock(&rxnet->client_conn_cache_lock);
1030
1031         if (list_empty(&rxnet->idle_client_conns))
1032                 goto out;
1033
1034         conn = list_entry(rxnet->idle_client_conns.next,
1035                           struct rxrpc_connection, cache_link);
1036         ASSERT(test_bit(RXRPC_CONN_EXPOSED, &conn->flags));
1037
1038         if (!rxnet->kill_all_client_conns) {
1039                 /* If the number of connections is over the reap limit, we
1040                  * expedite discard by reducing the expiry timeout.  We must,
1041                  * however, have at least a short grace period to be able to do
1042                  * final-ACK or ABORT retransmission.
1043                  */
1044                 expiry = rxrpc_conn_idle_client_expiry;
1045                 if (nr_conns > rxrpc_reap_client_connections)
1046                         expiry = rxrpc_conn_idle_client_fast_expiry;
1047
1048                 conn_expires_at = conn->idle_timestamp + expiry;
1049
1050                 now = READ_ONCE(jiffies);
1051                 if (time_after(conn_expires_at, now))
1052                         goto not_yet_expired;
1053         }
1054
1055         trace_rxrpc_client(conn, -1, rxrpc_client_discard);
1056         if (!test_and_clear_bit(RXRPC_CONN_EXPOSED, &conn->flags))
1057                 BUG();
1058         conn->cache_state = RXRPC_CONN_CLIENT_INACTIVE;
1059         list_del_init(&conn->cache_link);
1060
1061         spin_unlock(&rxnet->client_conn_cache_lock);
1062
1063         /* When we cleared the EXPOSED flag, we took on responsibility for the
1064          * reference that that had on the usage count.  We deal with that here.
1065          * If someone re-sets the flag and re-gets the ref, that's fine.
1066          */
1067         rxrpc_put_connection(conn);
1068         did_discard = true;
1069         nr_conns--;
1070         goto next;
1071
1072 not_yet_expired:
1073         /* The connection at the front of the queue hasn't yet expired, so
1074          * schedule the work item for that point if we discarded something.
1075          *
1076          * We don't worry if the work item is already scheduled - it can look
1077          * after rescheduling itself at a later time.  We could cancel it, but
1078          * then things get messier.
1079          */
1080         _debug("not yet");
1081         if (!rxnet->kill_all_client_conns)
1082                 queue_delayed_work(rxrpc_workqueue,
1083                                    &rxnet->client_conn_reaper,
1084                                    conn_expires_at - now);
1085
1086 out:
1087         spin_unlock(&rxnet->client_conn_cache_lock);
1088         spin_unlock(&rxnet->client_conn_discard_lock);
1089         _leave("");
1090 }
1091
1092 /*
1093  * Preemptively destroy all the client connection records rather than waiting
1094  * for them to time out
1095  */
1096 void rxrpc_destroy_all_client_connections(struct rxrpc_net *rxnet)
1097 {
1098         _enter("");
1099
1100         spin_lock(&rxnet->client_conn_cache_lock);
1101         rxnet->kill_all_client_conns = true;
1102         spin_unlock(&rxnet->client_conn_cache_lock);
1103
1104         cancel_delayed_work(&rxnet->client_conn_reaper);
1105
1106         if (!queue_delayed_work(rxrpc_workqueue, &rxnet->client_conn_reaper, 0))
1107                 _debug("destroy: queue failed");
1108
1109         _leave("");
1110 }