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->debug_id, rxrpc_conn_new_client,
 221                          atomic_read(&conn->usage),
 222                          __builtin_return_address(0));
 223         trace_rxrpc_client(conn, -1, rxrpc_client_alloc);
 224         _leave(" = %p", conn);
 225         return conn;
 226
 227 error_2:
 228         conn->security->clear(conn);
 229 error_1:
 230         rxrpc_put_client_connection_id(conn);
 231 error_0:
 232         kfree(conn);
 233         _leave(" = %d", ret);
 234         return ERR_PTR(ret);
 235 }
 236
 237 /*
 238  * Determine if a connection may be reused.
 239  */
 240 static bool rxrpc_may_reuse_conn(struct rxrpc_connection *conn)
 241 {
 242         struct rxrpc_net *rxnet = conn->params.local->rxnet;
 243         int id_cursor, id, distance, limit;
 244
 245         if (test_bit(RXRPC_CONN_DONT_REUSE, &conn->flags))
 246                 goto dont_reuse;
 247
 248         if (conn->proto.epoch != rxnet->epoch)
 249                 goto mark_dont_reuse;
 250
 251         /* The IDR tree gets very expensive on memory if the connection IDs are
 252          * widely scattered throughout the number space, so we shall want to
 253          * kill off connections that, say, have an ID more than about four
 254          * times the maximum number of client conns away from the current
 255          * allocation point to try and keep the IDs concentrated.
 256          */
 257         id_cursor = idr_get_cursor(&rxrpc_client_conn_ids);
 258         id = conn->proto.cid >> RXRPC_CIDSHIFT;
 259         distance = id - id_cursor;
 260         if (distance < 0)
 261                 distance = -distance;
 262         limit = max(rxrpc_max_client_connections * 4, 1024U);
 263         if (distance > limit)
 264                 goto mark_dont_reuse;
 265
 266         return true;
 267
 268 mark_dont_reuse:
 269         set_bit(RXRPC_CONN_DONT_REUSE, &conn->flags);
 270 dont_reuse:
 271         return false;
 272 }
 273
 274 /*
 275  * Create or find a client connection to use for a call.
 276  *
 277  * If we return with a connection, the call will be on its waiting list.  It's
 278  * left to the caller to assign a channel and wake up the call.
 279  */
 280 static int rxrpc_get_client_conn(struct rxrpc_sock *rx,
 281                                  struct rxrpc_call *call,
 282                                  struct rxrpc_conn_parameters *cp,
 283                                  struct sockaddr_rxrpc *srx,
 284                                  gfp_t gfp)
 285 {
 286         struct rxrpc_connection *conn, *candidate = NULL;
 287         struct rxrpc_local *local = cp->local;
 288         struct rb_node *p, **pp, *parent;
 289         long diff;
 290         int ret = -ENOMEM;
 291
 292         _enter("{%d,%lx},", call->debug_id, call->user_call_ID);
 293
 294         cp->peer = rxrpc_lookup_peer(rx, cp->local, srx, gfp);
 295         if (!cp->peer)
 296                 goto error;
 297
 298         call->cong_cwnd = cp->peer->cong_cwnd;
 299         if (call->cong_cwnd >= call->cong_ssthresh)
 300                 call->cong_mode = RXRPC_CALL_CONGEST_AVOIDANCE;
 301         else
 302                 call->cong_mode = RXRPC_CALL_SLOW_START;
 303
 304         /* If the connection is not meant to be exclusive, search the available
 305          * connections to see if the connection we want to use already exists.
 306          */
 307         if (!cp->exclusive) {
 308                 _debug("search 1");
 309                 spin_lock(&local->client_conns_lock);
 310                 p = local->client_conns.rb_node;
 311                 while (p) {
 312                         conn = rb_entry(p, struct rxrpc_connection, client_node);
 313
 314 #define cmp(X) ((long)conn->params.X - (long)cp->X)
 315                         diff = (cmp(peer) ?:
 316                                 cmp(key) ?:
 317                                 cmp(security_level) ?:
 318                                 cmp(upgrade));
 319 #undef cmp
 320                         if (diff < 0) {
 321                                 p = p->rb_left;
 322                         } else if (diff > 0) {
 323                                 p = p->rb_right;
 324                         } else {
 325                                 if (rxrpc_may_reuse_conn(conn) &&
 326                                     rxrpc_get_connection_maybe(conn))
 327                                         goto found_extant_conn;
 328                                 /* The connection needs replacing.  It's better
 329                                  * to effect that when we have something to
 330                                  * replace it with so that we don't have to
 331                                  * rebalance the tree twice.
 332                                  */
 333                                 break;
 334                         }
 335                 }
 336                 spin_unlock(&local->client_conns_lock);
 337         }
 338
 339         /* There wasn't a connection yet or we need an exclusive connection.
 340          * We need to create a candidate and then potentially redo the search
 341          * in case we're racing with another thread also trying to connect on a
 342          * shareable connection.
 343          */
 344         _debug("new conn");
 345         candidate = rxrpc_alloc_client_connection(cp, gfp);
 346         if (IS_ERR(candidate)) {
 347                 ret = PTR_ERR(candidate);
 348                 goto error_peer;
 349         }
 350
 351         /* Add the call to the new connection's waiting list in case we're
 352          * going to have to wait for the connection to come live.  It's our
 353          * connection, so we want first dibs on the channel slots.  We would
 354          * normally have to take channel_lock but we do this before anyone else
 355          * can see the connection.
 356          */
 357         list_add(&call->chan_wait_link, &candidate->waiting_calls);
 358
 359         if (cp->exclusive) {
 360                 call->conn = candidate;
 361                 call->security_ix = candidate->security_ix;
 362                 call->service_id = candidate->service_id;
 363                 _leave(" = 0 [exclusive %d]", candidate->debug_id);
 364                 return 0;
 365         }
 366
 367         /* Publish the new connection for userspace to find.  We need to redo
 368          * the search before doing this lest we race with someone else adding a
 369          * conflicting instance.
 370          */
 371         _debug("search 2");
 372         spin_lock(&local->client_conns_lock);
 373
 374         pp = &local->client_conns.rb_node;
 375         parent = NULL;
 376         while (*pp) {
 377                 parent = *pp;
 378                 conn = rb_entry(parent, struct rxrpc_connection, client_node);
 379
 380 #define cmp(X) ((long)conn->params.X - (long)candidate->params.X)
 381                 diff = (cmp(peer) ?:
 382                         cmp(key) ?:
 383                         cmp(security_level) ?:
 384                         cmp(upgrade));
 385 #undef cmp
 386                 if (diff < 0) {
 387                         pp = &(*pp)->rb_left;
 388                 } else if (diff > 0) {
 389                         pp = &(*pp)->rb_right;
 390                 } else {
 391                         if (rxrpc_may_reuse_conn(conn) &&
 392                             rxrpc_get_connection_maybe(conn))
 393                                 goto found_extant_conn;
 394                         /* The old connection is from an outdated epoch. */
 395                         _debug("replace conn");
 396                         clear_bit(RXRPC_CONN_IN_CLIENT_CONNS, &conn->flags);
 397                         rb_replace_node(&conn->client_node,
 398                                         &candidate->client_node,
 399                                         &local->client_conns);
 400                         trace_rxrpc_client(conn, -1, rxrpc_client_replace);
 401                         goto candidate_published;
 402                 }
 403         }
 404
 405         _debug("new conn");
 406         rb_link_node(&candidate->client_node, parent, pp);
 407         rb_insert_color(&candidate->client_node, &local->client_conns);
 408
 409 candidate_published:
 410         set_bit(RXRPC_CONN_IN_CLIENT_CONNS, &candidate->flags);
 411         call->conn = candidate;
 412         call->security_ix = candidate->security_ix;
 413         call->service_id = candidate->service_id;
 414         spin_unlock(&local->client_conns_lock);
 415         _leave(" = 0 [new %d]", candidate->debug_id);
 416         return 0;
 417
 418         /* We come here if we found a suitable connection already in existence.
 419          * Discard any candidate we may have allocated, and try to get a
 420          * channel on this one.
 421          */
 422 found_extant_conn:
 423         _debug("found conn");
 424         spin_unlock(&local->client_conns_lock);
 425
 426         if (candidate) {
 427                 trace_rxrpc_client(candidate, -1, rxrpc_client_duplicate);
 428                 rxrpc_put_connection(candidate);
 429                 candidate = NULL;
 430         }
 431
 432         spin_lock(&conn->channel_lock);
 433         call->conn = conn;
 434         call->security_ix = conn->security_ix;
 435         call->service_id = conn->service_id;
 436         list_add_tail(&call->chan_wait_link, &conn->waiting_calls);
 437         spin_unlock(&conn->channel_lock);
 438         _leave(" = 0 [extant %d]", conn->debug_id);
 439         return 0;
 440
 441 error_peer:
 442         rxrpc_put_peer(cp->peer);
 443         cp->peer = NULL;
 444 error:
 445         _leave(" = %d", ret);
 446         return ret;
 447 }
 448
 449 /*
 450  * Activate a connection.
 451  */
 452 static void rxrpc_activate_conn(struct rxrpc_net *rxnet,
 453                                 struct rxrpc_connection *conn)
 454 {
 455         if (test_bit(RXRPC_CONN_PROBING_FOR_UPGRADE, &conn->flags)) {
 456                 trace_rxrpc_client(conn, -1, rxrpc_client_to_upgrade);
 457                 conn->cache_state = RXRPC_CONN_CLIENT_UPGRADE;
 458         } else {
 459                 trace_rxrpc_client(conn, -1, rxrpc_client_to_active);
 460                 conn->cache_state = RXRPC_CONN_CLIENT_ACTIVE;
 461         }
 462         rxnet->nr_active_client_conns++;
 463         list_move_tail(&conn->cache_link, &rxnet->active_client_conns);
 464 }
 465
 466 /*
 467  * Attempt to animate a connection for a new call.
 468  *
 469  * If it's not exclusive, the connection is in the endpoint tree, and we're in
 470  * the conn's list of those waiting to grab a channel.  There is, however, a
 471  * limit on the number of live connections allowed at any one time, so we may
 472  * have to wait for capacity to become available.
 473  *
 474  * Note that a connection on the waiting queue might *also* have active
 475  * channels if it has been culled to make space and then re-requested by a new
 476  * call.
 477  */
 478 static void rxrpc_animate_client_conn(struct rxrpc_net *rxnet,
 479                                       struct rxrpc_connection *conn)
 480 {
 481         unsigned int nr_conns;
 482
 483         _enter("%d,%d", conn->debug_id, conn->cache_state);
 484
 485         if (conn->cache_state == RXRPC_CONN_CLIENT_ACTIVE ||
 486             conn->cache_state == RXRPC_CONN_CLIENT_UPGRADE)
 487                 goto out;
 488
 489         spin_lock(&rxnet->client_conn_cache_lock);
 490
 491         nr_conns = rxnet->nr_client_conns;
 492         if (!test_and_set_bit(RXRPC_CONN_COUNTED, &conn->flags)) {
 493                 trace_rxrpc_client(conn, -1, rxrpc_client_count);
 494                 rxnet->nr_client_conns = nr_conns + 1;
 495         }
 496
 497         switch (conn->cache_state) {
 498         case RXRPC_CONN_CLIENT_ACTIVE:
 499         case RXRPC_CONN_CLIENT_UPGRADE:
 500         case RXRPC_CONN_CLIENT_WAITING:
 501                 break;
 502
 503         case RXRPC_CONN_CLIENT_INACTIVE:
 504         case RXRPC_CONN_CLIENT_CULLED:
 505         case RXRPC_CONN_CLIENT_IDLE:
 506                 if (nr_conns >= rxrpc_max_client_connections)
 507                         goto wait_for_capacity;
 508                 goto activate_conn;
 509
 510         default:
 511                 BUG();
 512         }
 513
 514 out_unlock:
 515         spin_unlock(&rxnet->client_conn_cache_lock);
 516 out:
 517         _leave(" [%d]", conn->cache_state);
 518         return;
 519
 520 activate_conn:
 521         _debug("activate");
 522         rxrpc_activate_conn(rxnet, conn);
 523         goto out_unlock;
 524
 525 wait_for_capacity:
 526         _debug("wait");
 527         trace_rxrpc_client(conn, -1, rxrpc_client_to_waiting);
 528         conn->cache_state = RXRPC_CONN_CLIENT_WAITING;
 529         list_move_tail(&conn->cache_link, &rxnet->waiting_client_conns);
 530         goto out_unlock;
 531 }
 532
 533 /*
 534  * Deactivate a channel.
 535  */
 536 static void rxrpc_deactivate_one_channel(struct rxrpc_connection *conn,
 537                                          unsigned int channel)
 538 {
 539         struct rxrpc_channel *chan = &conn->channels[channel];
 540
 541         rcu_assign_pointer(chan->call, NULL);
 542         conn->active_chans &= ~(1 << channel);
 543 }
 544
 545 /*
 546  * Assign a channel to the call at the front of the queue and wake the call up.
 547  * We don't increment the callNumber counter until this number has been exposed
 548  * to the world.
 549  */
 550 static void rxrpc_activate_one_channel(struct rxrpc_connection *conn,
 551                                        unsigned int channel)
 552 {
 553         struct rxrpc_channel *chan = &conn->channels[channel];
 554         struct rxrpc_call *call = list_entry(conn->waiting_calls.next,
 555                                              struct rxrpc_call, chan_wait_link);
 556         u32 call_id = chan->call_counter + 1;
 557
 558         trace_rxrpc_client(conn, channel, rxrpc_client_chan_activate);
 559
 560         /* Cancel the final ACK on the previous call if it hasn't been sent yet
 561          * as the DATA packet will implicitly ACK it.
 562          */
 563         clear_bit(RXRPC_CONN_FINAL_ACK_0 + channel, &conn->flags);
 564
 565         write_lock_bh(&call->state_lock);
 566         if (!test_bit(RXRPC_CALL_TX_LASTQ, &call->flags))
 567                 call->state = RXRPC_CALL_CLIENT_SEND_REQUEST;
 568         else
 569                 call->state = RXRPC_CALL_CLIENT_AWAIT_REPLY;
 570         write_unlock_bh(&call->state_lock);
 571
 572         rxrpc_see_call(call);
 573         list_del_init(&call->chan_wait_link);
 574         conn->active_chans |= 1 << channel;
 575         call->peer      = rxrpc_get_peer(conn->params.peer);
 576         call->cid       = conn->proto.cid | channel;
 577         call->call_id   = call_id;
 578
 579         trace_rxrpc_connect_call(call);
 580         _net("CONNECT call %08x:%08x as call %d on conn %d",
 581              call->cid, call->call_id, call->debug_id, conn->debug_id);
 582
 583         /* Paired with the read barrier in rxrpc_wait_for_channel().  This
 584          * orders cid and epoch in the connection wrt to call_id without the
 585          * need to take the channel_lock.
 586          *
 587          * We provisionally assign a callNumber at this point, but we don't
 588          * confirm it until the call is about to be exposed.
 589          *
 590          * TODO: Pair with a barrier in the data_ready handler when that looks
 591          * at the call ID through a connection channel.
 592          */
 593         smp_wmb();
 594         chan->call_id   = call_id;
 595         chan->call_debug_id = call->debug_id;
 596         rcu_assign_pointer(chan->call, call);
 597         wake_up(&call->waitq);
 598 }
 599
 600 /*
 601  * Assign channels and callNumbers to waiting calls with channel_lock
 602  * held by caller.
 603  */
 604 static void rxrpc_activate_channels_locked(struct rxrpc_connection *conn)
 605 {
 606         u8 avail, mask;
 607
 608         switch (conn->cache_state) {
 609         case RXRPC_CONN_CLIENT_ACTIVE:
 610                 mask = RXRPC_ACTIVE_CHANS_MASK;
 611                 break;
 612         case RXRPC_CONN_CLIENT_UPGRADE:
 613                 mask = 0x01;
 614                 break;
 615         default:
 616                 return;
 617         }
 618
 619         while (!list_empty(&conn->waiting_calls) &&
 620                (avail = ~conn->active_chans,
 621                 avail &= mask,
 622                 avail != 0))
 623                 rxrpc_activate_one_channel(conn, __ffs(avail));
 624 }
 625
 626 /*
 627  * Assign channels and callNumbers to waiting calls.
 628  */
 629 static void rxrpc_activate_channels(struct rxrpc_connection *conn)
 630 {
 631         _enter("%d", conn->debug_id);
 632
 633         trace_rxrpc_client(conn, -1, rxrpc_client_activate_chans);
 634
 635         if (conn->active_chans == RXRPC_ACTIVE_CHANS_MASK)
 636                 return;
 637
 638         spin_lock(&conn->channel_lock);
 639         rxrpc_activate_channels_locked(conn);
 640         spin_unlock(&conn->channel_lock);
 641         _leave("");
 642 }
 643
 644 /*
 645  * Wait for a callNumber and a channel to be granted to a call.
 646  */
 647 static int rxrpc_wait_for_channel(struct rxrpc_call *call, gfp_t gfp)
 648 {
 649         int ret = 0;
 650
 651         _enter("%d", call->debug_id);
 652
 653         if (!call->call_id) {
 654                 DECLARE_WAITQUEUE(myself, current);
 655
 656                 if (!gfpflags_allow_blocking(gfp)) {
 657                         ret = -EAGAIN;
 658                         goto out;
 659                 }
 660
 661                 add_wait_queue_exclusive(&call->waitq, &myself);
 662                 for (;;) {
 663                         set_current_state(TASK_INTERRUPTIBLE);
 664                         if (call->call_id)
 665                                 break;
 666                         if (signal_pending(current)) {
 667                                 ret = -ERESTARTSYS;
 668                                 break;
 669                         }
 670                         schedule();
 671                 }
 672                 remove_wait_queue(&call->waitq, &myself);
 673                 __set_current_state(TASK_RUNNING);
 674         }
 675
 676         /* Paired with the write barrier in rxrpc_activate_one_channel(). */
 677         smp_rmb();
 678
 679 out:
 680         _leave(" = %d", ret);
 681         return ret;
 682 }
 683
 684 /*
 685  * find a connection for a call
 686  * - called in process context with IRQs enabled
 687  */
 688 int rxrpc_connect_call(struct rxrpc_sock *rx,
 689                        struct rxrpc_call *call,
 690                        struct rxrpc_conn_parameters *cp,
 691                        struct sockaddr_rxrpc *srx,
 692                        gfp_t gfp)
 693 {
 694         struct rxrpc_net *rxnet = cp->local->rxnet;
 695         int ret;
 696
 697         _enter("{%d,%lx},", call->debug_id, call->user_call_ID);
 698
 699         rxrpc_discard_expired_client_conns(&rxnet->client_conn_reaper);
 700         rxrpc_cull_active_client_conns(rxnet);
 701
 702         ret = rxrpc_get_client_conn(rx, call, cp, srx, gfp);
 703         if (ret < 0)
 704                 goto out;
 705
 706         rxrpc_animate_client_conn(rxnet, call->conn);
 707         rxrpc_activate_channels(call->conn);
 708
 709         ret = rxrpc_wait_for_channel(call, gfp);
 710         if (ret < 0) {
 711                 trace_rxrpc_client(call->conn, ret, rxrpc_client_chan_wait_failed);
 712                 rxrpc_disconnect_client_call(call);
 713                 goto out;
 714         }
 715
 716         spin_lock_bh(&call->conn->params.peer->lock);
 717         hlist_add_head_rcu(&call->error_link,
 718                            &call->conn->params.peer->error_targets);
 719         spin_unlock_bh(&call->conn->params.peer->lock);
 720
 721 out:
 722         _leave(" = %d", ret);
 723         return ret;
 724 }
 725
 726 /*
 727  * Note that a connection is about to be exposed to the world.  Once it is
 728  * exposed, we maintain an extra ref on it that stops it from being summarily
 729  * discarded before it's (a) had a chance to deal with retransmission and (b)
 730  * had a chance at re-use (the per-connection security negotiation is
 731  * expensive).
 732  */
 733 static void rxrpc_expose_client_conn(struct rxrpc_connection *conn,
 734                                      unsigned int channel)
 735 {
 736         if (!test_and_set_bit(RXRPC_CONN_EXPOSED, &conn->flags)) {
 737                 trace_rxrpc_client(conn, channel, rxrpc_client_exposed);
 738                 rxrpc_get_connection(conn);
 739         }
 740 }
 741
 742 /*
 743  * Note that a call, and thus a connection, is about to be exposed to the
 744  * world.
 745  */
 746 void rxrpc_expose_client_call(struct rxrpc_call *call)
 747 {
 748         unsigned int channel = call->cid & RXRPC_CHANNELMASK;
 749         struct rxrpc_connection *conn = call->conn;
 750         struct rxrpc_channel *chan = &conn->channels[channel];
 751
 752         if (!test_and_set_bit(RXRPC_CALL_EXPOSED, &call->flags)) {
 753                 /* Mark the call ID as being used.  If the callNumber counter
 754                  * exceeds ~2 billion, we kill the connection after its
 755                  * outstanding calls have finished so that the counter doesn't
 756                  * wrap.
 757                  */
 758                 chan->call_counter++;
 759                 if (chan->call_counter >= INT_MAX)
 760                         set_bit(RXRPC_CONN_DONT_REUSE, &conn->flags);
 761                 rxrpc_expose_client_conn(conn, channel);
 762         }
 763 }
 764
 765 /*
 766  * Set the reap timer.
 767  */
 768 static void rxrpc_set_client_reap_timer(struct rxrpc_net *rxnet)
 769 {
 770         unsigned long now = jiffies;
 771         unsigned long reap_at = now + rxrpc_conn_idle_client_expiry;
 772
 773         if (rxnet->live)
 774                 timer_reduce(&rxnet->client_conn_reap_timer, reap_at);
 775 }
 776
 777 /*
 778  * Disconnect a client call.
 779  */
 780 void rxrpc_disconnect_client_call(struct rxrpc_call *call)
 781 {
 782         struct rxrpc_connection *conn = call->conn;
 783         struct rxrpc_channel *chan = NULL;
 784         struct rxrpc_net *rxnet = conn->params.local->rxnet;
 785         unsigned int channel = -1;
 786         u32 cid;
 787
 788         spin_lock(&conn->channel_lock);
 789         set_bit(RXRPC_CALL_DISCONNECTED, &call->flags);
 790
 791         cid = call->cid;
 792         if (cid) {
 793                 channel = cid & RXRPC_CHANNELMASK;
 794                 chan = &conn->channels[channel];
 795         }
 796         trace_rxrpc_client(conn, channel, rxrpc_client_chan_disconnect);
 797
 798         /* Calls that have never actually been assigned a channel can simply be
 799          * discarded.  If the conn didn't get used either, it will follow
 800          * immediately unless someone else grabs it in the meantime.
 801          */
 802         if (!list_empty(&call->chan_wait_link)) {
 803                 _debug("call is waiting");
 804                 ASSERTCMP(call->call_id, ==, 0);
 805                 ASSERT(!test_bit(RXRPC_CALL_EXPOSED, &call->flags));
 806                 list_del_init(&call->chan_wait_link);
 807
 808                 trace_rxrpc_client(conn, channel, rxrpc_client_chan_unstarted);
 809
 810                 /* We must deactivate or idle the connection if it's now
 811                  * waiting for nothing.
 812                  */
 813                 spin_lock(&rxnet->client_conn_cache_lock);
 814                 if (conn->cache_state == RXRPC_CONN_CLIENT_WAITING &&
 815                     list_empty(&conn->waiting_calls) &&
 816                     !conn->active_chans)
 817                         goto idle_connection;
 818                 goto out;
 819         }
 820
 821         if (rcu_access_pointer(chan->call) != call) {
 822                 spin_unlock(&conn->channel_lock);
 823                 BUG();
 824         }
 825
 826         /* If a client call was exposed to the world, we save the result for
 827          * retransmission.
 828          *
 829          * We use a barrier here so that the call number and abort code can be
 830          * read without needing to take a lock.
 831          *
 832          * TODO: Make the incoming packet handler check this and handle
 833          * terminal retransmission without requiring access to the call.
 834          */
 835         if (test_bit(RXRPC_CALL_EXPOSED, &call->flags)) {
 836                 _debug("exposed %u,%u", call->call_id, call->abort_code);
 837                 __rxrpc_disconnect_call(conn, call);
 838         }
 839
 840         /* See if we can pass the channel directly to another call. */
 841         if (conn->cache_state == RXRPC_CONN_CLIENT_ACTIVE &&
 842             !list_empty(&conn->waiting_calls)) {
 843                 trace_rxrpc_client(conn, channel, rxrpc_client_chan_pass);
 844                 rxrpc_activate_one_channel(conn, channel);
 845                 goto out_2;
 846         }
 847
 848         /* Schedule the final ACK to be transmitted in a short while so that it
 849          * can be skipped if we find a follow-on call.  The first DATA packet
 850          * of the follow on call will implicitly ACK this call.
 851          */
 852         if (call->completion == RXRPC_CALL_SUCCEEDED &&
 853             test_bit(RXRPC_CALL_EXPOSED, &call->flags)) {
 854                 unsigned long final_ack_at = jiffies + 2;
 855
 856                 WRITE_ONCE(chan->final_ack_at, final_ack_at);
 857                 smp_wmb(); /* vs rxrpc_process_delayed_final_acks() */
 858                 set_bit(RXRPC_CONN_FINAL_ACK_0 + channel, &conn->flags);
 859                 rxrpc_reduce_conn_timer(conn, final_ack_at);
 860         }
 861
 862         /* Things are more complex and we need the cache lock.  We might be
 863          * able to simply idle the conn or it might now be lurking on the wait
 864          * list.  It might even get moved back to the active list whilst we're
 865          * waiting for the lock.
 866          */
 867         spin_lock(&rxnet->client_conn_cache_lock);
 868
 869         switch (conn->cache_state) {
 870         case RXRPC_CONN_CLIENT_UPGRADE:
 871                 /* Deal with termination of a service upgrade probe. */
 872                 if (test_bit(RXRPC_CONN_EXPOSED, &conn->flags)) {
 873                         clear_bit(RXRPC_CONN_PROBING_FOR_UPGRADE, &conn->flags);
 874                         trace_rxrpc_client(conn, channel, rxrpc_client_to_active);
 875                         conn->cache_state = RXRPC_CONN_CLIENT_ACTIVE;
 876                         rxrpc_activate_channels_locked(conn);
 877                 }
 878                 /* fall through */
 879         case RXRPC_CONN_CLIENT_ACTIVE:
 880                 if (list_empty(&conn->waiting_calls)) {
 881                         rxrpc_deactivate_one_channel(conn, channel);
 882                         if (!conn->active_chans) {
 883                                 rxnet->nr_active_client_conns--;
 884                                 goto idle_connection;
 885                         }
 886                         goto out;
 887                 }
 888
 889                 trace_rxrpc_client(conn, channel, rxrpc_client_chan_pass);
 890                 rxrpc_activate_one_channel(conn, channel);
 891                 goto out;
 892
 893         case RXRPC_CONN_CLIENT_CULLED:
 894                 rxrpc_deactivate_one_channel(conn, channel);
 895                 ASSERT(list_empty(&conn->waiting_calls));
 896                 if (!conn->active_chans)
 897                         goto idle_connection;
 898                 goto out;
 899
 900         case RXRPC_CONN_CLIENT_WAITING:
 901                 rxrpc_deactivate_one_channel(conn, channel);
 902                 goto out;
 903
 904         default:
 905                 BUG();
 906         }
 907
 908 out:
 909         spin_unlock(&rxnet->client_conn_cache_lock);
 910 out_2:
 911         spin_unlock(&conn->channel_lock);
 912         _leave("");
 913         return;
 914
 915 idle_connection:
 916         /* As no channels remain active, the connection gets deactivated
 917          * immediately or moved to the idle list for a short while.
 918          */
 919         if (test_bit(RXRPC_CONN_EXPOSED, &conn->flags)) {
 920                 trace_rxrpc_client(conn, channel, rxrpc_client_to_idle);
 921                 conn->idle_timestamp = jiffies;
 922                 conn->cache_state = RXRPC_CONN_CLIENT_IDLE;
 923                 list_move_tail(&conn->cache_link, &rxnet->idle_client_conns);
 924                 if (rxnet->idle_client_conns.next == &conn->cache_link &&
 925                     !rxnet->kill_all_client_conns)
 926                         rxrpc_set_client_reap_timer(rxnet);
 927         } else {
 928                 trace_rxrpc_client(conn, channel, rxrpc_client_to_inactive);
 929                 conn->cache_state = RXRPC_CONN_CLIENT_INACTIVE;
 930                 list_del_init(&conn->cache_link);
 931         }
 932         goto out;
 933 }
 934
 935 /*
 936  * Clean up a dead client connection.
 937  */
 938 static struct rxrpc_connection *
 939 rxrpc_put_one_client_conn(struct rxrpc_connection *conn)
 940 {
 941         struct rxrpc_connection *next = NULL;
 942         struct rxrpc_local *local = conn->params.local;
 943         struct rxrpc_net *rxnet = local->rxnet;
 944         unsigned int nr_conns;
 945
 946         trace_rxrpc_client(conn, -1, rxrpc_client_cleanup);
 947
 948         if (test_bit(RXRPC_CONN_IN_CLIENT_CONNS, &conn->flags)) {
 949                 spin_lock(&local->client_conns_lock);
 950                 if (test_and_clear_bit(RXRPC_CONN_IN_CLIENT_CONNS,
 951                                        &conn->flags))
 952                         rb_erase(&conn->client_node, &local->client_conns);
 953                 spin_unlock(&local->client_conns_lock);
 954         }
 955
 956         rxrpc_put_client_connection_id(conn);
 957
 958         ASSERTCMP(conn->cache_state, ==, RXRPC_CONN_CLIENT_INACTIVE);
 959
 960         if (test_bit(RXRPC_CONN_COUNTED, &conn->flags)) {
 961                 trace_rxrpc_client(conn, -1, rxrpc_client_uncount);
 962                 spin_lock(&rxnet->client_conn_cache_lock);
 963                 nr_conns = --rxnet->nr_client_conns;
 964
 965                 if (nr_conns < rxrpc_max_client_connections &&
 966                     !list_empty(&rxnet->waiting_client_conns)) {
 967                         next = list_entry(rxnet->waiting_client_conns.next,
 968                                           struct rxrpc_connection, cache_link);
 969                         rxrpc_get_connection(next);
 970                         rxrpc_activate_conn(rxnet, next);
 971                 }
 972
 973                 spin_unlock(&rxnet->client_conn_cache_lock);
 974         }
 975
 976         rxrpc_kill_connection(conn);
 977         if (next)
 978                 rxrpc_activate_channels(next);
 979
 980         /* We need to get rid of the temporary ref we took upon next, but we
 981          * can't call rxrpc_put_connection() recursively.
 982          */
 983         return next;
 984 }
 985
 986 /*
 987  * Clean up a dead client connections.
 988  */
 989 void rxrpc_put_client_conn(struct rxrpc_connection *conn)
 990 {
 991         const void *here = __builtin_return_address(0);
 992         unsigned int debug_id = conn->debug_id;
 993         int n;
 994
 995         do {
 996                 n = atomic_dec_return(&conn->usage);
 997                 trace_rxrpc_conn(debug_id, rxrpc_conn_put_client, n, here);
 998                 if (n > 0)
 999                         return;
1000                 ASSERTCMP(n, >=, 0);
1001
1002                 conn = rxrpc_put_one_client_conn(conn);
1003         } while (conn);
1004 }
1005
1006 /*
1007  * Kill the longest-active client connections to make room for new ones.
1008  */
1009 static void rxrpc_cull_active_client_conns(struct rxrpc_net *rxnet)
1010 {
1011         struct rxrpc_connection *conn;
1012         unsigned int nr_conns = rxnet->nr_client_conns;
1013         unsigned int nr_active, limit;
1014
1015         _enter("");
1016
1017         ASSERTCMP(nr_conns, >=, 0);
1018         if (nr_conns < rxrpc_max_client_connections) {
1019                 _leave(" [ok]");
1020                 return;
1021         }
1022         limit = rxrpc_reap_client_connections;
1023
1024         spin_lock(&rxnet->client_conn_cache_lock);
1025         nr_active = rxnet->nr_active_client_conns;
1026
1027         while (nr_active > limit) {
1028                 ASSERT(!list_empty(&rxnet->active_client_conns));
1029                 conn = list_entry(rxnet->active_client_conns.next,
1030                                   struct rxrpc_connection, cache_link);
1031                 ASSERTIFCMP(conn->cache_state != RXRPC_CONN_CLIENT_ACTIVE,
1032                             conn->cache_state, ==, RXRPC_CONN_CLIENT_UPGRADE);
1033
1034                 if (list_empty(&conn->waiting_calls)) {
1035                         trace_rxrpc_client(conn, -1, rxrpc_client_to_culled);
1036                         conn->cache_state = RXRPC_CONN_CLIENT_CULLED;
1037                         list_del_init(&conn->cache_link);
1038                 } else {
1039                         trace_rxrpc_client(conn, -1, rxrpc_client_to_waiting);
1040                         conn->cache_state = RXRPC_CONN_CLIENT_WAITING;
1041                         list_move_tail(&conn->cache_link,
1042                                        &rxnet->waiting_client_conns);
1043                 }
1044
1045                 nr_active--;
1046         }
1047
1048         rxnet->nr_active_client_conns = nr_active;
1049         spin_unlock(&rxnet->client_conn_cache_lock);
1050         ASSERTCMP(nr_active, >=, 0);
1051         _leave(" [culled]");
1052 }
1053
1054 /*
1055  * Discard expired client connections from the idle list.  Each conn in the
1056  * idle list has been exposed and holds an extra ref because of that.
1057  *
1058  * This may be called from conn setup or from a work item so cannot be
1059  * considered non-reentrant.
1060  */
1061 void rxrpc_discard_expired_client_conns(struct work_struct *work)
1062 {
1063         struct rxrpc_connection *conn;
1064         struct rxrpc_net *rxnet =
1065                 container_of(work, struct rxrpc_net, client_conn_reaper);
1066         unsigned long expiry, conn_expires_at, now;
1067         unsigned int nr_conns;
1068
1069         _enter("");
1070
1071         if (list_empty(&rxnet->idle_client_conns)) {
1072                 _leave(" [empty]");
1073                 return;
1074         }
1075
1076         /* Don't double up on the discarding */
1077         if (!spin_trylock(&rxnet->client_conn_discard_lock)) {
1078                 _leave(" [already]");
1079                 return;
1080         }
1081
1082         /* We keep an estimate of what the number of conns ought to be after
1083          * we've discarded some so that we don't overdo the discarding.
1084          */
1085         nr_conns = rxnet->nr_client_conns;
1086
1087 next:
1088         spin_lock(&rxnet->client_conn_cache_lock);
1089
1090         if (list_empty(&rxnet->idle_client_conns))
1091                 goto out;
1092
1093         conn = list_entry(rxnet->idle_client_conns.next,
1094                           struct rxrpc_connection, cache_link);
1095         ASSERT(test_bit(RXRPC_CONN_EXPOSED, &conn->flags));
1096
1097         if (!rxnet->kill_all_client_conns) {
1098                 /* If the number of connections is over the reap limit, we
1099                  * expedite discard by reducing the expiry timeout.  We must,
1100                  * however, have at least a short grace period to be able to do
1101                  * final-ACK or ABORT retransmission.
1102                  */
1103                 expiry = rxrpc_conn_idle_client_expiry;
1104                 if (nr_conns > rxrpc_reap_client_connections)
1105                         expiry = rxrpc_conn_idle_client_fast_expiry;
1106                 if (conn->params.local->service_closed)
1107                         expiry = rxrpc_closed_conn_expiry * HZ;
1108
1109                 conn_expires_at = conn->idle_timestamp + expiry;
1110
1111                 now = READ_ONCE(jiffies);
1112                 if (time_after(conn_expires_at, now))
1113                         goto not_yet_expired;
1114         }
1115
1116         trace_rxrpc_client(conn, -1, rxrpc_client_discard);
1117         if (!test_and_clear_bit(RXRPC_CONN_EXPOSED, &conn->flags))
1118                 BUG();
1119         conn->cache_state = RXRPC_CONN_CLIENT_INACTIVE;
1120         list_del_init(&conn->cache_link);
1121
1122         spin_unlock(&rxnet->client_conn_cache_lock);
1123
1124         /* When we cleared the EXPOSED flag, we took on responsibility for the
1125          * reference that that had on the usage count.  We deal with that here.
1126          * If someone re-sets the flag and re-gets the ref, that's fine.
1127          */
1128         rxrpc_put_connection(conn);
1129         nr_conns--;
1130         goto next;
1131
1132 not_yet_expired:
1133         /* The connection at the front of the queue hasn't yet expired, so
1134          * schedule the work item for that point if we discarded something.
1135          *
1136          * We don't worry if the work item is already scheduled - it can look
1137          * after rescheduling itself at a later time.  We could cancel it, but
1138          * then things get messier.
1139          */
1140         _debug("not yet");
1141         if (!rxnet->kill_all_client_conns)
1142                 timer_reduce(&rxnet->client_conn_reap_timer,
1143                              conn_expires_at);
1144
1145 out:
1146         spin_unlock(&rxnet->client_conn_cache_lock);
1147         spin_unlock(&rxnet->client_conn_discard_lock);
1148         _leave("");
1149 }
1150
1151 /*
1152  * Preemptively destroy all the client connection records rather than waiting
1153  * for them to time out
1154  */
1155 void rxrpc_destroy_all_client_connections(struct rxrpc_net *rxnet)
1156 {
1157         _enter("");
1158
1159         spin_lock(&rxnet->client_conn_cache_lock);
1160         rxnet->kill_all_client_conns = true;
1161         spin_unlock(&rxnet->client_conn_cache_lock);
1162
1163         del_timer_sync(&rxnet->client_conn_reap_timer);
1164
1165         if (!rxrpc_queue_work(&rxnet->client_conn_reaper))
1166                 _debug("destroy: queue failed");
1167
1168         _leave("");
1169 }
1170
1171 /*
1172  * Clean up the client connections on a local endpoint.
1173  */
1174 void rxrpc_clean_up_local_conns(struct rxrpc_local *local)
1175 {
1176         struct rxrpc_connection *conn, *tmp;
1177         struct rxrpc_net *rxnet = local->rxnet;
1178         unsigned int nr_active;
1179         LIST_HEAD(graveyard);
1180
1181         _enter("");
1182
1183         spin_lock(&rxnet->client_conn_cache_lock);
1184         nr_active = rxnet->nr_active_client_conns;
1185
1186         list_for_each_entry_safe(conn, tmp, &rxnet->idle_client_conns,
1187                                  cache_link) {
1188                 if (conn->params.local == local) {
1189                         ASSERTCMP(conn->cache_state, ==, RXRPC_CONN_CLIENT_IDLE);
1190
1191                         trace_rxrpc_client(conn, -1, rxrpc_client_discard);
1192                         if (!test_and_clear_bit(RXRPC_CONN_EXPOSED, &conn->flags))
1193                                 BUG();
1194                         conn->cache_state = RXRPC_CONN_CLIENT_INACTIVE;
1195                         list_move(&conn->cache_link, &graveyard);
1196                         nr_active--;
1197                 }
1198         }
1199
1200         rxnet->nr_active_client_conns = nr_active;
1201         spin_unlock(&rxnet->client_conn_cache_lock);
1202         ASSERTCMP(nr_active, >=, 0);
1203
1204         while (!list_empty(&graveyard)) {
1205                 conn = list_entry(graveyard.next,
1206                                   struct rxrpc_connection, cache_link);
1207                 list_del_init(&conn->cache_link);
1208
1209                 rxrpc_put_connection(conn);
1210         }
1211
1212         _leave(" [culled]");
1213 }