net/rxrpc/conn_client.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /* Client connection-specific management code.
   3  *
   4  * Copyright (C) 2016 Red Hat, Inc. All Rights Reserved.
   5  * Written by David Howells (dhowells@redhat.com)
   6  *
   7  * Client connections need to be cached for a little while after they've made a
   8  * call so as to handle retransmitted DATA packets in case the server didn't
   9  * receive the final ACK or terminating ABORT we sent it.
  10  *
  11  * Client connections can be in one of a number of cache states:
  12  *
  13  *  (1) INACTIVE - The connection is not held in any list and may not have been
  14  *      exposed to the world.  If it has been previously exposed, it was
  15  *      discarded from the idle list after expiring.
  16  *
  17  *  (2) WAITING - The connection is waiting for the number of client conns to
  18  *      drop below the maximum capacity.  Calls may be in progress upon it from
  19  *      when it was active and got culled.
  20  *
  21  *      The connection is on the rxrpc_waiting_client_conns list which is kept
  22  *      in to-be-granted order.  Culled conns with waiters go to the back of
  23  *      the queue just like new conns.
  24  *
  25  *  (3) ACTIVE - The connection has at least one call in progress upon it, it
  26  *      may freely grant available channels to new calls and calls may be
  27  *      waiting on it for channels to become available.
  28  *
  29  *      The connection is on the rxnet->active_client_conns list which is kept
  30  *      in activation order for culling purposes.
  31  *
  32  *      rxrpc_nr_active_client_conns is held incremented also.
  33  *
  34  *  (4) UPGRADE - As for ACTIVE, but only one call may be in progress and is
  35  *      being used to probe for service upgrade.
  36  *
  37  *  (5) CULLED - The connection got summarily culled to try and free up
  38  *      capacity.  Calls currently in progress on the connection are allowed to
  39  *      continue, but new calls will have to wait.  There can be no waiters in
  40  *      this state - the conn would have to go to the WAITING state instead.
  41  *
  42  *  (6) IDLE - The connection has no calls in progress upon it and must have
  43  *      been exposed to the world (ie. the EXPOSED flag must be set).  When it
  44  *      expires, the EXPOSED flag is cleared and the connection transitions to
  45  *      the INACTIVE state.
  46  *
  47  *      The connection is on the rxnet->idle_client_conns list which is kept in
  48  *      order of how soon they'll expire.
  49  *
  50  * There are flags of relevance to the cache:
  51  *
  52  *  (1) EXPOSED - The connection ID got exposed to the world.  If this flag is
  53  *      set, an extra ref is added to the connection preventing it from being
  54  *      reaped when it has no calls outstanding.  This flag is cleared and the
  55  *      ref dropped when a conn is discarded from the idle list.
  56  *
  57  *      This allows us to move terminal call state retransmission to the
  58  *      connection and to discard the call immediately we think it is done
  59  *      with.  It also give us a chance to reuse the connection.
  60  *
  61  *  (2) DONT_REUSE - The connection should be discarded as soon as possible and
  62  *      should not be reused.  This is set when an exclusive connection is used
  63  *      or a call ID counter overflows.
  64  *
  65  * The caching state may only be changed if the cache lock is held.
  66  *
  67  * There are two idle client connection expiry durations.  If the total number
  68  * of connections is below the reap threshold, we use the normal duration; if
  69  * it's above, we use the fast duration.
  70  */
  71
  72 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  73
  74 #include <linux/slab.h>
  75 #include <linux/idr.h>
  76 #include <linux/timer.h>
  77 #include <linux/sched/signal.h>
  78
  79 #include "ar-internal.h"
  80
  81 __read_mostly unsigned int rxrpc_max_client_connections = 1000;
  82 __read_mostly unsigned int rxrpc_reap_client_connections = 900;
  83 __read_mostly unsigned long rxrpc_conn_idle_client_expiry = 2 * 60 * HZ;
  84 __read_mostly unsigned long rxrpc_conn_idle_client_fast_expiry = 2 * HZ;
  85
  86 /*
  87  * We use machine-unique IDs for our client connections.
  88  */
  89 DEFINE_IDR(rxrpc_client_conn_ids);
  90 static DEFINE_SPINLOCK(rxrpc_conn_id_lock);
  91
  92 static void rxrpc_cull_active_client_conns(struct rxrpc_net *);
  93
  94 /*
  95  * Get a connection ID and epoch for a client connection from the global pool.
  96  * The connection struct pointer is then recorded in the idr radix tree.  The
  97  * epoch doesn't change until the client is rebooted (or, at least, unless the
  98  * module is unloaded).
  99  */
 100 static int rxrpc_get_client_connection_id(struct rxrpc_connection *conn,
 101                                           gfp_t gfp)
 102 {
 103         struct rxrpc_net *rxnet = conn->params.local->rxnet;
 104         int id;
 105
 106         _enter("");
 107
 108         idr_preload(gfp);
 109         spin_lock(&rxrpc_conn_id_lock);
 110
 111         id = idr_alloc_cyclic(&rxrpc_client_conn_ids, conn,
 112                               1, 0x40000000, GFP_NOWAIT);
 113         if (id < 0)
 114                 goto error;
 115
 116         spin_unlock(&rxrpc_conn_id_lock);
 117         idr_preload_end();
 118
 119         conn->proto.epoch = rxnet->epoch;
 120         conn->proto.cid = id << RXRPC_CIDSHIFT;
 121         set_bit(RXRPC_CONN_HAS_IDR, &conn->flags);
 122         _leave(" [CID %x]", conn->proto.cid);
 123         return 0;
 124
 125 error:
 126         spin_unlock(&rxrpc_conn_id_lock);
 127         idr_preload_end();
 128         _leave(" = %d", id);
 129         return id;
 130 }
 131
 132 /*
 133  * Release a connection ID for a client connection from the global pool.
 134  */
 135 static void rxrpc_put_client_connection_id(struct rxrpc_connection *conn)
 136 {
 137         if (test_bit(RXRPC_CONN_HAS_IDR, &conn->flags)) {
 138                 spin_lock(&rxrpc_conn_id_lock);
 139                 idr_remove(&rxrpc_client_conn_ids,
 140                            conn->proto.cid >> RXRPC_CIDSHIFT);
 141                 spin_unlock(&rxrpc_conn_id_lock);
 142         }
 143 }
 144
 145 /*
 146  * Destroy the client connection ID tree.
 147  */
 148 void rxrpc_destroy_client_conn_ids(void)
 149 {
 150         struct rxrpc_connection *conn;
 151         int id;
 152
 153         if (!idr_is_empty(&rxrpc_client_conn_ids)) {
 154                 idr_for_each_entry(&rxrpc_client_conn_ids, conn, id) {
 155                         pr_err("AF_RXRPC: Leaked client conn %p {%d}\n",
 156                                conn, atomic_read(&conn->usage));
 157                 }
 158                 BUG();
 159         }
 160
 161         idr_destroy(&rxrpc_client_conn_ids);
 162 }
 163
 164 /*
 165  * Allocate a client connection.
 166  */
 167 static struct rxrpc_connection *
 168 rxrpc_alloc_client_connection(struct rxrpc_conn_parameters *cp, gfp_t gfp)
 169 {
 170         struct rxrpc_connection *conn;
 171         struct rxrpc_net *rxnet = cp->local->rxnet;
 172         int ret;
 173
 174         _enter("");
 175
 176         conn = rxrpc_alloc_connection(gfp);
 177         if (!conn) {
 178                 _leave(" = -ENOMEM");
 179                 return ERR_PTR(-ENOMEM);
 180         }
 181
 182         atomic_set(&conn->usage, 1);
 183         if (cp->exclusive)
 184                 __set_bit(RXRPC_CONN_DONT_REUSE, &conn->flags);
 185         if (cp->upgrade)
 186                 __set_bit(RXRPC_CONN_PROBING_FOR_UPGRADE, &conn->flags);
 187
 188         conn->params            = *cp;
 189         conn->out_clientflag    = RXRPC_CLIENT_INITIATED;
 190         conn->state             = RXRPC_CONN_CLIENT;
 191         conn->service_id        = cp->service_id;
 192
 193         ret = rxrpc_get_client_connection_id(conn, gfp);
 194         if (ret < 0)
 195                 goto error_0;
 196
 197         ret = rxrpc_init_client_conn_security(conn);
 198         if (ret < 0)
 199                 goto error_1;
 200
 201         ret = conn->security->prime_packet_security(conn);
 202         if (ret < 0)
 203                 goto error_2;
 204
 205         atomic_inc(&rxnet->nr_conns);
 206         write_lock(&rxnet->conn_lock);
 207         list_add_tail(&conn->proc_link, &rxnet->conn_proc_list);
 208         write_unlock(&rxnet->conn_lock);
 209
 210         /* We steal the caller's peer ref. */
 211         cp->peer = NULL;
 212         rxrpc_get_local(conn->params.local);
 213         key_get(conn->params.key);
 214
 215         trace_rxrpc_conn(conn->debug_id, rxrpc_conn_new_client,
 216                          atomic_read(&conn->usage),
 217                          __builtin_return_address(0));
 218         trace_rxrpc_client(conn, -1, rxrpc_client_alloc);
 219         _leave(" = %p", conn);
 220         return conn;
 221
 222 error_2:
 223         conn->security->clear(conn);
 224 error_1:
 225         rxrpc_put_client_connection_id(conn);
 226 error_0:
 227         kfree(conn);
 228         _leave(" = %d", ret);
 229         return ERR_PTR(ret);
 230 }
 231
 232 /*
 233  * Determine if a connection may be reused.
 234  */
 235 static bool rxrpc_may_reuse_conn(struct rxrpc_connection *conn)
 236 {
 237         struct rxrpc_net *rxnet = conn->params.local->rxnet;
 238         int id_cursor, id, distance, limit;
 239
 240         if (test_bit(RXRPC_CONN_DONT_REUSE, &conn->flags))
 241                 goto dont_reuse;
 242
 243         if (conn->proto.epoch != rxnet->epoch)
 244                 goto mark_dont_reuse;
 245
 246         /* The IDR tree gets very expensive on memory if the connection IDs are
 247          * widely scattered throughout the number space, so we shall want to
 248          * kill off connections that, say, have an ID more than about four
 249          * times the maximum number of client conns away from the current
 250          * allocation point to try and keep the IDs concentrated.
 251          */
 252         id_cursor = idr_get_cursor(&rxrpc_client_conn_ids);
 253         id = conn->proto.cid >> RXRPC_CIDSHIFT;
 254         distance = id - id_cursor;
 255         if (distance < 0)
 256                 distance = -distance;
 257         limit = max(rxrpc_max_client_connections * 4, 1024U);
 258         if (distance > limit)
 259                 goto mark_dont_reuse;
 260
 261         return true;
 262
 263 mark_dont_reuse:
 264         set_bit(RXRPC_CONN_DONT_REUSE, &conn->flags);
 265 dont_reuse:
 266         return false;
 267 }
 268
 269 /*
 270  * Create or find a client connection to use for a call.
 271  *
 272  * If we return with a connection, the call will be on its waiting list.  It's
 273  * left to the caller to assign a channel and wake up the call.
 274  */
 275 static int rxrpc_get_client_conn(struct rxrpc_sock *rx,
 276                                  struct rxrpc_call *call,
 277                                  struct rxrpc_conn_parameters *cp,
 278                                  struct sockaddr_rxrpc *srx,
 279                                  gfp_t gfp)
 280 {
 281         struct rxrpc_connection *conn, *candidate = NULL;
 282         struct rxrpc_local *local = cp->local;
 283         struct rb_node *p, **pp, *parent;
 284         long diff;
 285         int ret = -ENOMEM;
 286
 287         _enter("{%d,%lx},", call->debug_id, call->user_call_ID);
 288
 289         cp->peer = rxrpc_lookup_peer(rx, cp->local, srx, gfp);
 290         if (!cp->peer)
 291                 goto error;
 292
 293         call->cong_cwnd = cp->peer->cong_cwnd;
 294         if (call->cong_cwnd >= call->cong_ssthresh)
 295                 call->cong_mode = RXRPC_CALL_CONGEST_AVOIDANCE;
 296         else
 297                 call->cong_mode = RXRPC_CALL_SLOW_START;
 298
 299         /* If the connection is not meant to be exclusive, search the available
 300          * connections to see if the connection we want to use already exists.
 301          */
 302         if (!cp->exclusive) {
 303                 _debug("search 1");
 304                 spin_lock(&local->client_conns_lock);
 305                 p = local->client_conns.rb_node;
 306                 while (p) {
 307                         conn = rb_entry(p, struct rxrpc_connection, client_node);
 308
 309 #define cmp(X) ((long)conn->params.X - (long)cp->X)
 310                         diff = (cmp(peer) ?:
 311                                 cmp(key) ?:
 312                                 cmp(security_level) ?:
 313                                 cmp(upgrade));
 314 #undef cmp
 315                         if (diff < 0) {
 316                                 p = p->rb_left;
 317                         } else if (diff > 0) {
 318                                 p = p->rb_right;
 319                         } else {
 320                                 if (rxrpc_may_reuse_conn(conn) &&
 321                                     rxrpc_get_connection_maybe(conn))
 322                                         goto found_extant_conn;
 323                                 /* The connection needs replacing.  It's better
 324                                  * to effect that when we have something to
 325                                  * replace it with so that we don't have to
 326                                  * rebalance the tree twice.
 327                                  */
 328                                 break;
 329                         }
 330                 }
 331                 spin_unlock(&local->client_conns_lock);
 332         }
 333
 334         /* There wasn't a connection yet or we need an exclusive connection.
 335          * We need to create a candidate and then potentially redo the search
 336          * in case we're racing with another thread also trying to connect on a
 337          * shareable connection.
 338          */
 339         _debug("new conn");
 340         candidate = rxrpc_alloc_client_connection(cp, gfp);
 341         if (IS_ERR(candidate)) {
 342                 ret = PTR_ERR(candidate);
 343                 goto error_peer;
 344         }
 345
 346         /* Add the call to the new connection's waiting list in case we're
 347          * going to have to wait for the connection to come live.  It's our
 348          * connection, so we want first dibs on the channel slots.  We would
 349          * normally have to take channel_lock but we do this before anyone else
 350          * can see the connection.
 351          */
 352         list_add(&call->chan_wait_link, &candidate->waiting_calls);
 353
 354         if (cp->exclusive) {
 355                 call->conn = candidate;
 356                 call->security = candidate->security;
 357                 call->security_ix = candidate->security_ix;
 358                 call->service_id = candidate->service_id;
 359                 _leave(" = 0 [exclusive %d]", candidate->debug_id);
 360                 return 0;
 361         }
 362
 363         /* Publish the new connection for userspace to find.  We need to redo
 364          * the search before doing this lest we race with someone else adding a
 365          * conflicting instance.
 366          */
 367         _debug("search 2");
 368         spin_lock(&local->client_conns_lock);
 369
 370         pp = &local->client_conns.rb_node;
 371         parent = NULL;
 372         while (*pp) {
 373                 parent = *pp;
 374                 conn = rb_entry(parent, struct rxrpc_connection, client_node);
 375
 376 #define cmp(X) ((long)conn->params.X - (long)candidate->params.X)
 377                 diff = (cmp(peer) ?:
 378                         cmp(key) ?:
 379                         cmp(security_level) ?:
 380                         cmp(upgrade));
 381 #undef cmp
 382                 if (diff < 0) {
 383                         pp = &(*pp)->rb_left;
 384                 } else if (diff > 0) {
 385                         pp = &(*pp)->rb_right;
 386                 } else {
 387                         if (rxrpc_may_reuse_conn(conn) &&
 388                             rxrpc_get_connection_maybe(conn))
 389                                 goto found_extant_conn;
 390                         /* The old connection is from an outdated epoch. */
 391                         _debug("replace conn");
 392                         clear_bit(RXRPC_CONN_IN_CLIENT_CONNS, &conn->flags);
 393                         rb_replace_node(&conn->client_node,
 394                                         &candidate->client_node,
 395                                         &local->client_conns);
 396                         trace_rxrpc_client(conn, -1, rxrpc_client_replace);
 397                         goto candidate_published;
 398                 }
 399         }
 400
 401         _debug("new conn");
 402         rb_link_node(&candidate->client_node, parent, pp);
 403         rb_insert_color(&candidate->client_node, &local->client_conns);
 404
 405 candidate_published:
 406         set_bit(RXRPC_CONN_IN_CLIENT_CONNS, &candidate->flags);
 407         call->conn = candidate;
 408         call->security = candidate->security;
 409         call->security_ix = candidate->security_ix;
 410         call->service_id = candidate->service_id;
 411         spin_unlock(&local->client_conns_lock);
 412         _leave(" = 0 [new %d]", candidate->debug_id);
 413         return 0;
 414
 415         /* We come here if we found a suitable connection already in existence.
 416          * Discard any candidate we may have allocated, and try to get a
 417          * channel on this one.
 418          */
 419 found_extant_conn:
 420         _debug("found conn");
 421         spin_unlock(&local->client_conns_lock);
 422
 423         if (candidate) {
 424                 trace_rxrpc_client(candidate, -1, rxrpc_client_duplicate);
 425                 rxrpc_put_connection(candidate);
 426                 candidate = NULL;
 427         }
 428
 429         spin_lock(&conn->channel_lock);
 430         call->conn = conn;
 431         call->security = conn->security;
 432         call->security_ix = conn->security_ix;
 433         call->service_id = conn->service_id;
 434         list_add_tail(&call->chan_wait_link, &conn->waiting_calls);
 435         spin_unlock(&conn->channel_lock);
 436         _leave(" = 0 [extant %d]", conn->debug_id);
 437         return 0;
 438
 439 error_peer:
 440         rxrpc_put_peer(cp->peer);
 441         cp->peer = NULL;
 442 error:
 443         _leave(" = %d", ret);
 444         return ret;
 445 }
 446
 447 /*
 448  * Activate a connection.
 449  */
 450 static void rxrpc_activate_conn(struct rxrpc_net *rxnet,
 451                                 struct rxrpc_connection *conn)
 452 {
 453         if (test_bit(RXRPC_CONN_PROBING_FOR_UPGRADE, &conn->flags)) {
 454                 trace_rxrpc_client(conn, -1, rxrpc_client_to_upgrade);
 455                 conn->cache_state = RXRPC_CONN_CLIENT_UPGRADE;
 456         } else {
 457                 trace_rxrpc_client(conn, -1, rxrpc_client_to_active);
 458                 conn->cache_state = RXRPC_CONN_CLIENT_ACTIVE;
 459         }
 460         rxnet->nr_active_client_conns++;
 461         list_move_tail(&conn->cache_link, &rxnet->active_client_conns);
 462 }
 463
 464 /*
 465  * Attempt to animate a connection for a new call.
 466  *
 467  * If it's not exclusive, the connection is in the endpoint tree, and we're in
 468  * the conn's list of those waiting to grab a channel.  There is, however, a
 469  * limit on the number of live connections allowed at any one time, so we may
 470  * have to wait for capacity to become available.
 471  *
 472  * Note that a connection on the waiting queue might *also* have active
 473  * channels if it has been culled to make space and then re-requested by a new
 474  * call.
 475  */
 476 static void rxrpc_animate_client_conn(struct rxrpc_net *rxnet,
 477                                       struct rxrpc_connection *conn)
 478 {
 479         unsigned int nr_conns;
 480
 481         _enter("%d,%d", conn->debug_id, conn->cache_state);
 482
 483         if (conn->cache_state == RXRPC_CONN_CLIENT_ACTIVE ||
 484             conn->cache_state == RXRPC_CONN_CLIENT_UPGRADE)
 485                 goto out;
 486
 487         spin_lock(&rxnet->client_conn_cache_lock);
 488
 489         nr_conns = rxnet->nr_client_conns;
 490         if (!test_and_set_bit(RXRPC_CONN_COUNTED, &conn->flags)) {
 491                 trace_rxrpc_client(conn, -1, rxrpc_client_count);
 492                 rxnet->nr_client_conns = nr_conns + 1;
 493         }
 494
 495         switch (conn->cache_state) {
 496         case RXRPC_CONN_CLIENT_ACTIVE:
 497         case RXRPC_CONN_CLIENT_UPGRADE:
 498         case RXRPC_CONN_CLIENT_WAITING:
 499                 break;
 500
 501         case RXRPC_CONN_CLIENT_INACTIVE:
 502         case RXRPC_CONN_CLIENT_CULLED:
 503         case RXRPC_CONN_CLIENT_IDLE:
 504                 if (nr_conns >= rxrpc_max_client_connections)
 505                         goto wait_for_capacity;
 506                 goto activate_conn;
 507
 508         default:
 509                 BUG();
 510         }
 511
 512 out_unlock:
 513         spin_unlock(&rxnet->client_conn_cache_lock);
 514 out:
 515         _leave(" [%d]", conn->cache_state);
 516         return;
 517
 518 activate_conn:
 519         _debug("activate");
 520         rxrpc_activate_conn(rxnet, conn);
 521         goto out_unlock;
 522
 523 wait_for_capacity:
 524         _debug("wait");
 525         trace_rxrpc_client(conn, -1, rxrpc_client_to_waiting);
 526         conn->cache_state = RXRPC_CONN_CLIENT_WAITING;
 527         list_move_tail(&conn->cache_link, &rxnet->waiting_client_conns);
 528         goto out_unlock;
 529 }
 530
 531 /*
 532  * Deactivate a channel.
 533  */
 534 static void rxrpc_deactivate_one_channel(struct rxrpc_connection *conn,
 535                                          unsigned int channel)
 536 {
 537         struct rxrpc_channel *chan = &conn->channels[channel];
 538
 539         rcu_assign_pointer(chan->call, NULL);
 540         conn->active_chans &= ~(1 << channel);
 541 }
 542
 543 /*
 544  * Assign a channel to the call at the front of the queue and wake the call up.
 545  * We don't increment the callNumber counter until this number has been exposed
 546  * to the world.
 547  */
 548 static void rxrpc_activate_one_channel(struct rxrpc_connection *conn,
 549                                        unsigned int channel)
 550 {
 551         struct rxrpc_channel *chan = &conn->channels[channel];
 552         struct rxrpc_call *call = list_entry(conn->waiting_calls.next,
 553                                              struct rxrpc_call, chan_wait_link);
 554         u32 call_id = chan->call_counter + 1;
 555
 556         trace_rxrpc_client(conn, channel, rxrpc_client_chan_activate);
 557
 558         /* Cancel the final ACK on the previous call if it hasn't been sent yet
 559          * as the DATA packet will implicitly ACK it.
 560          */
 561         clear_bit(RXRPC_CONN_FINAL_ACK_0 + channel, &conn->flags);
 562
 563         write_lock_bh(&call->state_lock);
 564         call->state = RXRPC_CALL_CLIENT_SEND_REQUEST;
 565         write_unlock_bh(&call->state_lock);
 566
 567         rxrpc_see_call(call);
 568         list_del_init(&call->chan_wait_link);
 569         conn->active_chans |= 1 << channel;
 570         call->peer      = rxrpc_get_peer(conn->params.peer);
 571         call->cid       = conn->proto.cid | channel;
 572         call->call_id   = call_id;
 573
 574         trace_rxrpc_connect_call(call);
 575         _net("CONNECT call %08x:%08x as call %d on conn %d",
 576              call->cid, call->call_id, call->debug_id, conn->debug_id);
 577
 578         /* Paired with the read barrier in rxrpc_wait_for_channel().  This
 579          * orders cid and epoch in the connection wrt to call_id without the
 580          * need to take the channel_lock.
 581          *
 582          * We provisionally assign a callNumber at this point, but we don't
 583          * confirm it until the call is about to be exposed.
 584          *
 585          * TODO: Pair with a barrier in the data_ready handler when that looks
 586          * at the call ID through a connection channel.
 587          */
 588         smp_wmb();
 589         chan->call_id   = call_id;
 590         chan->call_debug_id = call->debug_id;
 591         rcu_assign_pointer(chan->call, call);
 592         wake_up(&call->waitq);
 593 }
 594
 595 /*
 596  * Assign channels and callNumbers to waiting calls with channel_lock
 597  * held by caller.
 598  */
 599 static void rxrpc_activate_channels_locked(struct rxrpc_connection *conn)
 600 {
 601         u8 avail, mask;
 602
 603         switch (conn->cache_state) {
 604         case RXRPC_CONN_CLIENT_ACTIVE:
 605                 mask = RXRPC_ACTIVE_CHANS_MASK;
 606                 break;
 607         case RXRPC_CONN_CLIENT_UPGRADE:
 608                 mask = 0x01;
 609                 break;
 610         default:
 611                 return;
 612         }
 613
 614         while (!list_empty(&conn->waiting_calls) &&
 615                (avail = ~conn->active_chans,
 616                 avail &= mask,
 617                 avail != 0))
 618                 rxrpc_activate_one_channel(conn, __ffs(avail));
 619 }
 620
 621 /*
 622  * Assign channels and callNumbers to waiting calls.
 623  */
 624 static void rxrpc_activate_channels(struct rxrpc_connection *conn)
 625 {
 626         _enter("%d", conn->debug_id);
 627
 628         trace_rxrpc_client(conn, -1, rxrpc_client_activate_chans);
 629
 630         if (conn->active_chans == RXRPC_ACTIVE_CHANS_MASK)
 631                 return;
 632
 633         spin_lock(&conn->channel_lock);
 634         rxrpc_activate_channels_locked(conn);
 635         spin_unlock(&conn->channel_lock);
 636         _leave("");
 637 }
 638
 639 /*
 640  * Wait for a callNumber and a channel to be granted to a call.
 641  */
 642 static int rxrpc_wait_for_channel(struct rxrpc_call *call, gfp_t gfp)
 643 {
 644         int ret = 0;
 645
 646         _enter("%d", call->debug_id);
 647
 648         if (!call->call_id) {
 649                 DECLARE_WAITQUEUE(myself, current);
 650
 651                 if (!gfpflags_allow_blocking(gfp)) {
 652                         ret = -EAGAIN;
 653                         goto out;
 654                 }
 655
 656                 add_wait_queue_exclusive(&call->waitq, &myself);
 657                 for (;;) {
 658                         switch (call->interruptibility) {
 659                         case RXRPC_INTERRUPTIBLE:
 660                         case RXRPC_PREINTERRUPTIBLE:
 661                                 set_current_state(TASK_INTERRUPTIBLE);
 662                                 break;
 663                         case RXRPC_UNINTERRUPTIBLE:
 664                         default:
 665                                 set_current_state(TASK_UNINTERRUPTIBLE);
 666                                 break;
 667                         }
 668                         if (call->call_id)
 669                                 break;
 670                         if ((call->interruptibility == RXRPC_INTERRUPTIBLE ||
 671                              call->interruptibility == RXRPC_PREINTERRUPTIBLE) &&
 672                             signal_pending(current)) {
 673                                 ret = -ERESTARTSYS;
 674                                 break;
 675                         }
 676                         schedule();
 677                 }
 678                 remove_wait_queue(&call->waitq, &myself);
 679                 __set_current_state(TASK_RUNNING);
 680         }
 681
 682         /* Paired with the write barrier in rxrpc_activate_one_channel(). */
 683         smp_rmb();
 684
 685 out:
 686         _leave(" = %d", ret);
 687         return ret;
 688 }
 689
 690 /*
 691  * find a connection for a call
 692  * - called in process context with IRQs enabled
 693  */
 694 int rxrpc_connect_call(struct rxrpc_sock *rx,
 695                        struct rxrpc_call *call,
 696                        struct rxrpc_conn_parameters *cp,
 697                        struct sockaddr_rxrpc *srx,
 698                        gfp_t gfp)
 699 {
 700         struct rxrpc_net *rxnet = cp->local->rxnet;
 701         int ret;
 702
 703         _enter("{%d,%lx},", call->debug_id, call->user_call_ID);
 704
 705         rxrpc_discard_expired_client_conns(&rxnet->client_conn_reaper);
 706         rxrpc_cull_active_client_conns(rxnet);
 707
 708         ret = rxrpc_get_client_conn(rx, call, cp, srx, gfp);
 709         if (ret < 0)
 710                 goto out;
 711
 712         rxrpc_animate_client_conn(rxnet, call->conn);
 713         rxrpc_activate_channels(call->conn);
 714
 715         ret = rxrpc_wait_for_channel(call, gfp);
 716         if (ret < 0) {
 717                 trace_rxrpc_client(call->conn, ret, rxrpc_client_chan_wait_failed);
 718                 rxrpc_disconnect_client_call(call);
 719                 goto out;
 720         }
 721
 722         spin_lock_bh(&call->conn->params.peer->lock);
 723         hlist_add_head_rcu(&call->error_link,
 724                            &call->conn->params.peer->error_targets);
 725         spin_unlock_bh(&call->conn->params.peer->lock);
 726
 727 out:
 728         _leave(" = %d", ret);
 729         return ret;
 730 }
 731
 732 /*
 733  * Note that a connection is about to be exposed to the world.  Once it is
 734  * exposed, we maintain an extra ref on it that stops it from being summarily
 735  * discarded before it's (a) had a chance to deal with retransmission and (b)
 736  * had a chance at re-use (the per-connection security negotiation is
 737  * expensive).
 738  */
 739 static void rxrpc_expose_client_conn(struct rxrpc_connection *conn,
 740                                      unsigned int channel)
 741 {
 742         if (!test_and_set_bit(RXRPC_CONN_EXPOSED, &conn->flags)) {
 743                 trace_rxrpc_client(conn, channel, rxrpc_client_exposed);
 744                 rxrpc_get_connection(conn);
 745         }
 746 }
 747
 748 /*
 749  * Note that a call, and thus a connection, is about to be exposed to the
 750  * world.
 751  */
 752 void rxrpc_expose_client_call(struct rxrpc_call *call)
 753 {
 754         unsigned int channel = call->cid & RXRPC_CHANNELMASK;
 755         struct rxrpc_connection *conn = call->conn;
 756         struct rxrpc_channel *chan = &conn->channels[channel];
 757
 758         if (!test_and_set_bit(RXRPC_CALL_EXPOSED, &call->flags)) {
 759                 /* Mark the call ID as being used.  If the callNumber counter
 760                  * exceeds ~2 billion, we kill the connection after its
 761                  * outstanding calls have finished so that the counter doesn't
 762                  * wrap.
 763                  */
 764                 chan->call_counter++;
 765                 if (chan->call_counter >= INT_MAX)
 766                         set_bit(RXRPC_CONN_DONT_REUSE, &conn->flags);
 767                 rxrpc_expose_client_conn(conn, channel);
 768         }
 769 }
 770
 771 /*
 772  * Set the reap timer.
 773  */
 774 static void rxrpc_set_client_reap_timer(struct rxrpc_net *rxnet)
 775 {
 776         unsigned long now = jiffies;
 777         unsigned long reap_at = now + rxrpc_conn_idle_client_expiry;
 778
 779         if (rxnet->live)
 780                 timer_reduce(&rxnet->client_conn_reap_timer, reap_at);
 781 }
 782
 783 /*
 784  * Disconnect a client call.
 785  */
 786 void rxrpc_disconnect_client_call(struct rxrpc_call *call)
 787 {
 788         struct rxrpc_connection *conn = call->conn;
 789         struct rxrpc_channel *chan = NULL;
 790         struct rxrpc_net *rxnet = conn->params.local->rxnet;
 791         unsigned int channel = -1;
 792         u32 cid;
 793
 794         spin_lock(&conn->channel_lock);
 795         set_bit(RXRPC_CALL_DISCONNECTED, &call->flags);
 796
 797         cid = call->cid;
 798         if (cid) {
 799                 channel = cid & RXRPC_CHANNELMASK;
 800                 chan = &conn->channels[channel];
 801         }
 802         trace_rxrpc_client(conn, channel, rxrpc_client_chan_disconnect);
 803
 804         /* Calls that have never actually been assigned a channel can simply be
 805          * discarded.  If the conn didn't get used either, it will follow
 806          * immediately unless someone else grabs it in the meantime.
 807          */
 808         if (!list_empty(&call->chan_wait_link)) {
 809                 _debug("call is waiting");
 810                 ASSERTCMP(call->call_id, ==, 0);
 811                 ASSERT(!test_bit(RXRPC_CALL_EXPOSED, &call->flags));
 812                 list_del_init(&call->chan_wait_link);
 813
 814                 trace_rxrpc_client(conn, channel, rxrpc_client_chan_unstarted);
 815
 816                 /* We must deactivate or idle the connection if it's now
 817                  * waiting for nothing.
 818                  */
 819                 spin_lock(&rxnet->client_conn_cache_lock);
 820                 if (conn->cache_state == RXRPC_CONN_CLIENT_WAITING &&
 821                     list_empty(&conn->waiting_calls) &&
 822                     !conn->active_chans)
 823                         goto idle_connection;
 824                 goto out;
 825         }
 826
 827         if (rcu_access_pointer(chan->call) != call) {
 828                 spin_unlock(&conn->channel_lock);
 829                 BUG();
 830         }
 831
 832         /* If a client call was exposed to the world, we save the result for
 833          * retransmission.
 834          *
 835          * We use a barrier here so that the call number and abort code can be
 836          * read without needing to take a lock.
 837          *
 838          * TODO: Make the incoming packet handler check this and handle
 839          * terminal retransmission without requiring access to the call.
 840          */
 841         if (test_bit(RXRPC_CALL_EXPOSED, &call->flags)) {
 842                 _debug("exposed %u,%u", call->call_id, call->abort_code);
 843                 __rxrpc_disconnect_call(conn, call);
 844         }
 845
 846         /* See if we can pass the channel directly to another call. */
 847         if (conn->cache_state == RXRPC_CONN_CLIENT_ACTIVE &&
 848             !list_empty(&conn->waiting_calls)) {
 849                 trace_rxrpc_client(conn, channel, rxrpc_client_chan_pass);
 850                 rxrpc_activate_one_channel(conn, channel);
 851                 goto out_2;
 852         }
 853
 854         /* Schedule the final ACK to be transmitted in a short while so that it
 855          * can be skipped if we find a follow-on call.  The first DATA packet
 856          * of the follow on call will implicitly ACK this call.
 857          */
 858         if (call->completion == RXRPC_CALL_SUCCEEDED &&
 859             test_bit(RXRPC_CALL_EXPOSED, &call->flags)) {
 860                 unsigned long final_ack_at = jiffies + 2;
 861
 862                 WRITE_ONCE(chan->final_ack_at, final_ack_at);
 863                 smp_wmb(); /* vs rxrpc_process_delayed_final_acks() */
 864                 set_bit(RXRPC_CONN_FINAL_ACK_0 + channel, &conn->flags);
 865                 rxrpc_reduce_conn_timer(conn, final_ack_at);
 866         }
 867
 868         /* Things are more complex and we need the cache lock.  We might be
 869          * able to simply idle the conn or it might now be lurking on the wait
 870          * list.  It might even get moved back to the active list whilst we're
 871          * waiting for the lock.
 872          */
 873         spin_lock(&rxnet->client_conn_cache_lock);
 874
 875         switch (conn->cache_state) {
 876         case RXRPC_CONN_CLIENT_UPGRADE:
 877                 /* Deal with termination of a service upgrade probe. */
 878                 if (test_bit(RXRPC_CONN_EXPOSED, &conn->flags)) {
 879                         clear_bit(RXRPC_CONN_PROBING_FOR_UPGRADE, &conn->flags);
 880                         trace_rxrpc_client(conn, channel, rxrpc_client_to_active);
 881                         conn->cache_state = RXRPC_CONN_CLIENT_ACTIVE;
 882                         rxrpc_activate_channels_locked(conn);
 883                 }
 884                 fallthrough;
 885         case RXRPC_CONN_CLIENT_ACTIVE:
 886                 if (list_empty(&conn->waiting_calls)) {
 887                         rxrpc_deactivate_one_channel(conn, channel);
 888                         if (!conn->active_chans) {
 889                                 rxnet->nr_active_client_conns--;
 890                                 goto idle_connection;
 891                         }
 892                         goto out;
 893                 }
 894
 895                 trace_rxrpc_client(conn, channel, rxrpc_client_chan_pass);
 896                 rxrpc_activate_one_channel(conn, channel);
 897                 goto out;
 898
 899         case RXRPC_CONN_CLIENT_CULLED:
 900                 rxrpc_deactivate_one_channel(conn, channel);
 901                 ASSERT(list_empty(&conn->waiting_calls));
 902                 if (!conn->active_chans)
 903                         goto idle_connection;
 904                 goto out;
 905
 906         case RXRPC_CONN_CLIENT_WAITING:
 907                 rxrpc_deactivate_one_channel(conn, channel);
 908                 goto out;
 909
 910         default:
 911                 BUG();
 912         }
 913
 914 out:
 915         spin_unlock(&rxnet->client_conn_cache_lock);
 916 out_2:
 917         spin_unlock(&conn->channel_lock);
 918         _leave("");
 919         return;
 920
 921 idle_connection:
 922         /* As no channels remain active, the connection gets deactivated
 923          * immediately or moved to the idle list for a short while.
 924          */
 925         if (test_bit(RXRPC_CONN_EXPOSED, &conn->flags)) {
 926                 trace_rxrpc_client(conn, channel, rxrpc_client_to_idle);
 927                 conn->idle_timestamp = jiffies;
 928                 conn->cache_state = RXRPC_CONN_CLIENT_IDLE;
 929                 list_move_tail(&conn->cache_link, &rxnet->idle_client_conns);
 930                 if (rxnet->idle_client_conns.next == &conn->cache_link &&
 931                     !rxnet->kill_all_client_conns)
 932                         rxrpc_set_client_reap_timer(rxnet);
 933         } else {
 934                 trace_rxrpc_client(conn, channel, rxrpc_client_to_inactive);
 935                 conn->cache_state = RXRPC_CONN_CLIENT_INACTIVE;
 936                 list_del_init(&conn->cache_link);
 937         }
 938         goto out;
 939 }
 940
 941 /*
 942  * Clean up a dead client connection.
 943  */
 944 static struct rxrpc_connection *
 945 rxrpc_put_one_client_conn(struct rxrpc_connection *conn)
 946 {
 947         struct rxrpc_connection *next = NULL;
 948         struct rxrpc_local *local = conn->params.local;
 949         struct rxrpc_net *rxnet = local->rxnet;
 950         unsigned int nr_conns;
 951
 952         trace_rxrpc_client(conn, -1, rxrpc_client_cleanup);
 953
 954         if (test_bit(RXRPC_CONN_IN_CLIENT_CONNS, &conn->flags)) {
 955                 spin_lock(&local->client_conns_lock);
 956                 if (test_and_clear_bit(RXRPC_CONN_IN_CLIENT_CONNS,
 957                                        &conn->flags))
 958                         rb_erase(&conn->client_node, &local->client_conns);
 959                 spin_unlock(&local->client_conns_lock);
 960         }
 961
 962         rxrpc_put_client_connection_id(conn);
 963
 964         ASSERTCMP(conn->cache_state, ==, RXRPC_CONN_CLIENT_INACTIVE);
 965
 966         if (test_bit(RXRPC_CONN_COUNTED, &conn->flags)) {
 967                 trace_rxrpc_client(conn, -1, rxrpc_client_uncount);
 968                 spin_lock(&rxnet->client_conn_cache_lock);
 969                 nr_conns = --rxnet->nr_client_conns;
 970
 971                 if (nr_conns < rxrpc_max_client_connections &&
 972                     !list_empty(&rxnet->waiting_client_conns)) {
 973                         next = list_entry(rxnet->waiting_client_conns.next,
 974                                           struct rxrpc_connection, cache_link);
 975                         rxrpc_get_connection(next);
 976                         rxrpc_activate_conn(rxnet, next);
 977                 }
 978
 979                 spin_unlock(&rxnet->client_conn_cache_lock);
 980         }
 981
 982         rxrpc_kill_connection(conn);
 983         if (next)
 984                 rxrpc_activate_channels(next);
 985
 986         /* We need to get rid of the temporary ref we took upon next, but we
 987          * can't call rxrpc_put_connection() recursively.
 988          */
 989         return next;
 990 }
 991
 992 /*
 993  * Clean up a dead client connections.
 994  */
 995 void rxrpc_put_client_conn(struct rxrpc_connection *conn)
 996 {
 997         const void *here = __builtin_return_address(0);
 998         unsigned int debug_id = conn->debug_id;
 999         int n;
1000
1001         do {
1002                 n = atomic_dec_return(&conn->usage);
1003                 trace_rxrpc_conn(debug_id, rxrpc_conn_put_client, n, here);
1004                 if (n > 0)
1005                         return;
1006                 ASSERTCMP(n, >=, 0);
1007
1008                 conn = rxrpc_put_one_client_conn(conn);
1009         } while (conn);
1010 }
1011
1012 /*
1013  * Kill the longest-active client connections to make room for new ones.
1014  */
1015 static void rxrpc_cull_active_client_conns(struct rxrpc_net *rxnet)
1016 {
1017         struct rxrpc_connection *conn;
1018         unsigned int nr_conns = rxnet->nr_client_conns;
1019         unsigned int nr_active, limit;
1020
1021         _enter("");
1022
1023         ASSERTCMP(nr_conns, >=, 0);
1024         if (nr_conns < rxrpc_max_client_connections) {
1025                 _leave(" [ok]");
1026                 return;
1027         }
1028         limit = rxrpc_reap_client_connections;
1029
1030         spin_lock(&rxnet->client_conn_cache_lock);
1031         nr_active = rxnet->nr_active_client_conns;
1032
1033         while (nr_active > limit) {
1034                 ASSERT(!list_empty(&rxnet->active_client_conns));
1035                 conn = list_entry(rxnet->active_client_conns.next,
1036                                   struct rxrpc_connection, cache_link);
1037                 ASSERTIFCMP(conn->cache_state != RXRPC_CONN_CLIENT_ACTIVE,
1038                             conn->cache_state, ==, RXRPC_CONN_CLIENT_UPGRADE);
1039
1040                 if (list_empty(&conn->waiting_calls)) {
1041                         trace_rxrpc_client(conn, -1, rxrpc_client_to_culled);
1042                         conn->cache_state = RXRPC_CONN_CLIENT_CULLED;
1043                         list_del_init(&conn->cache_link);
1044                 } else {
1045                         trace_rxrpc_client(conn, -1, rxrpc_client_to_waiting);
1046                         conn->cache_state = RXRPC_CONN_CLIENT_WAITING;
1047                         list_move_tail(&conn->cache_link,
1048                                        &rxnet->waiting_client_conns);
1049                 }
1050
1051                 nr_active--;
1052         }
1053
1054         rxnet->nr_active_client_conns = nr_active;
1055         spin_unlock(&rxnet->client_conn_cache_lock);
1056         ASSERTCMP(nr_active, >=, 0);
1057         _leave(" [culled]");
1058 }
1059
1060 /*
1061  * Discard expired client connections from the idle list.  Each conn in the
1062  * idle list has been exposed and holds an extra ref because of that.
1063  *
1064  * This may be called from conn setup or from a work item so cannot be
1065  * considered non-reentrant.
1066  */
1067 void rxrpc_discard_expired_client_conns(struct work_struct *work)
1068 {
1069         struct rxrpc_connection *conn;
1070         struct rxrpc_net *rxnet =
1071                 container_of(work, struct rxrpc_net, client_conn_reaper);
1072         unsigned long expiry, conn_expires_at, now;
1073         unsigned int nr_conns;
1074
1075         _enter("");
1076
1077         if (list_empty(&rxnet->idle_client_conns)) {
1078                 _leave(" [empty]");
1079                 return;
1080         }
1081
1082         /* Don't double up on the discarding */
1083         if (!spin_trylock(&rxnet->client_conn_discard_lock)) {
1084                 _leave(" [already]");
1085                 return;
1086         }
1087
1088         /* We keep an estimate of what the number of conns ought to be after
1089          * we've discarded some so that we don't overdo the discarding.
1090          */
1091         nr_conns = rxnet->nr_client_conns;
1092
1093 next:
1094         spin_lock(&rxnet->client_conn_cache_lock);
1095
1096         if (list_empty(&rxnet->idle_client_conns))
1097                 goto out;
1098
1099         conn = list_entry(rxnet->idle_client_conns.next,
1100                           struct rxrpc_connection, cache_link);
1101         ASSERT(test_bit(RXRPC_CONN_EXPOSED, &conn->flags));
1102
1103         if (!rxnet->kill_all_client_conns) {
1104                 /* If the number of connections is over the reap limit, we
1105                  * expedite discard by reducing the expiry timeout.  We must,
1106                  * however, have at least a short grace period to be able to do
1107                  * final-ACK or ABORT retransmission.
1108                  */
1109                 expiry = rxrpc_conn_idle_client_expiry;
1110                 if (nr_conns > rxrpc_reap_client_connections)
1111                         expiry = rxrpc_conn_idle_client_fast_expiry;
1112                 if (conn->params.local->service_closed)
1113                         expiry = rxrpc_closed_conn_expiry * HZ;
1114
1115                 conn_expires_at = conn->idle_timestamp + expiry;
1116
1117                 now = READ_ONCE(jiffies);
1118                 if (time_after(conn_expires_at, now))
1119                         goto not_yet_expired;
1120         }
1121
1122         trace_rxrpc_client(conn, -1, rxrpc_client_discard);
1123         if (!test_and_clear_bit(RXRPC_CONN_EXPOSED, &conn->flags))
1124                 BUG();
1125         conn->cache_state = RXRPC_CONN_CLIENT_INACTIVE;
1126         list_del_init(&conn->cache_link);
1127
1128         spin_unlock(&rxnet->client_conn_cache_lock);
1129
1130         /* When we cleared the EXPOSED flag, we took on responsibility for the
1131          * reference that that had on the usage count.  We deal with that here.
1132          * If someone re-sets the flag and re-gets the ref, that's fine.
1133          */
1134         rxrpc_put_connection(conn);
1135         nr_conns--;
1136         goto next;
1137
1138 not_yet_expired:
1139         /* The connection at the front of the queue hasn't yet expired, so
1140          * schedule the work item for that point if we discarded something.
1141          *
1142          * We don't worry if the work item is already scheduled - it can look
1143          * after rescheduling itself at a later time.  We could cancel it, but
1144          * then things get messier.
1145          */
1146         _debug("not yet");
1147         if (!rxnet->kill_all_client_conns)
1148                 timer_reduce(&rxnet->client_conn_reap_timer,
1149                              conn_expires_at);
1150
1151 out:
1152         spin_unlock(&rxnet->client_conn_cache_lock);
1153         spin_unlock(&rxnet->client_conn_discard_lock);
1154         _leave("");
1155 }
1156
1157 /*
1158  * Preemptively destroy all the client connection records rather than waiting
1159  * for them to time out
1160  */
1161 void rxrpc_destroy_all_client_connections(struct rxrpc_net *rxnet)
1162 {
1163         _enter("");
1164
1165         spin_lock(&rxnet->client_conn_cache_lock);
1166         rxnet->kill_all_client_conns = true;
1167         spin_unlock(&rxnet->client_conn_cache_lock);
1168
1169         del_timer_sync(&rxnet->client_conn_reap_timer);
1170
1171         if (!rxrpc_queue_work(&rxnet->client_conn_reaper))
1172                 _debug("destroy: queue failed");
1173
1174         _leave("");
1175 }
1176
1177 /*
1178  * Clean up the client connections on a local endpoint.
1179  */
1180 void rxrpc_clean_up_local_conns(struct rxrpc_local *local)
1181 {
1182         struct rxrpc_connection *conn, *tmp;
1183         struct rxrpc_net *rxnet = local->rxnet;
1184         unsigned int nr_active;
1185         LIST_HEAD(graveyard);
1186
1187         _enter("");
1188
1189         spin_lock(&rxnet->client_conn_cache_lock);
1190         nr_active = rxnet->nr_active_client_conns;
1191
1192         list_for_each_entry_safe(conn, tmp, &rxnet->idle_client_conns,
1193                                  cache_link) {
1194                 if (conn->params.local == local) {
1195                         ASSERTCMP(conn->cache_state, ==, RXRPC_CONN_CLIENT_IDLE);
1196
1197                         trace_rxrpc_client(conn, -1, rxrpc_client_discard);
1198                         if (!test_and_clear_bit(RXRPC_CONN_EXPOSED, &conn->flags))
1199                                 BUG();
1200                         conn->cache_state = RXRPC_CONN_CLIENT_INACTIVE;
1201                         list_move(&conn->cache_link, &graveyard);
1202                         nr_active--;
1203                 }
1204         }
1205
1206         rxnet->nr_active_client_conns = nr_active;
1207         spin_unlock(&rxnet->client_conn_cache_lock);
1208         ASSERTCMP(nr_active, >=, 0);
1209
1210         while (!list_empty(&graveyard)) {
1211                 conn = list_entry(graveyard.next,
1212                                   struct rxrpc_connection, cache_link);
1213                 list_del_init(&conn->cache_link);
1214
1215                 rxrpc_put_connection(conn);
1216         }
1217
1218         _leave(" [culled]");
1219 }