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, 2020 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  * There are flags of relevance to the cache:
  12  *
  13  *  (2) DONT_REUSE - The connection should be discarded as soon as possible and
  14  *      should not be reused.  This is set when an exclusive connection is used
  15  *      or a call ID counter overflows.
  16  *
  17  * The caching state may only be changed if the cache lock is held.
  18  *
  19  * There are two idle client connection expiry durations.  If the total number
  20  * of connections is below the reap threshold, we use the normal duration; if
  21  * it's above, we use the fast duration.
  22  */
  23
  24 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  25
  26 #include <linux/slab.h>
  27 #include <linux/idr.h>
  28 #include <linux/timer.h>
  29 #include <linux/sched/signal.h>
  30
  31 #include "ar-internal.h"
  32
  33 __read_mostly unsigned int rxrpc_reap_client_connections = 900;
  34 __read_mostly unsigned long rxrpc_conn_idle_client_expiry = 2 * 60 * HZ;
  35 __read_mostly unsigned long rxrpc_conn_idle_client_fast_expiry = 2 * HZ;
  36
  37 /*
  38  * We use machine-unique IDs for our client connections.
  39  */
  40 DEFINE_IDR(rxrpc_client_conn_ids);
  41 static DEFINE_SPINLOCK(rxrpc_conn_id_lock);
  42
  43 /*
  44  * Get a connection ID and epoch for a client connection from the global pool.
  45  * The connection struct pointer is then recorded in the idr radix tree.  The
  46  * epoch doesn't change until the client is rebooted (or, at least, unless the
  47  * module is unloaded).
  48  */
  49 static int rxrpc_get_client_connection_id(struct rxrpc_connection *conn,
  50                                           gfp_t gfp)
  51 {
  52         struct rxrpc_net *rxnet = conn->params.local->rxnet;
  53         int id;
  54
  55         _enter("");
  56
  57         idr_preload(gfp);
  58         spin_lock(&rxrpc_conn_id_lock);
  59
  60         id = idr_alloc_cyclic(&rxrpc_client_conn_ids, conn,
  61                               1, 0x40000000, GFP_NOWAIT);
  62         if (id < 0)
  63                 goto error;
  64
  65         spin_unlock(&rxrpc_conn_id_lock);
  66         idr_preload_end();
  67
  68         conn->proto.epoch = rxnet->epoch;
  69         conn->proto.cid = id << RXRPC_CIDSHIFT;
  70         set_bit(RXRPC_CONN_HAS_IDR, &conn->flags);
  71         _leave(" [CID %x]", conn->proto.cid);
  72         return 0;
  73
  74 error:
  75         spin_unlock(&rxrpc_conn_id_lock);
  76         idr_preload_end();
  77         _leave(" = %d", id);
  78         return id;
  79 }
  80
  81 /*
  82  * Release a connection ID for a client connection from the global pool.
  83  */
  84 static void rxrpc_put_client_connection_id(struct rxrpc_connection *conn)
  85 {
  86         if (test_bit(RXRPC_CONN_HAS_IDR, &conn->flags)) {
  87                 spin_lock(&rxrpc_conn_id_lock);
  88                 idr_remove(&rxrpc_client_conn_ids,
  89                            conn->proto.cid >> RXRPC_CIDSHIFT);
  90                 spin_unlock(&rxrpc_conn_id_lock);
  91         }
  92 }
  93
  94 /*
  95  * Destroy the client connection ID tree.
  96  */
  97 void rxrpc_destroy_client_conn_ids(void)
  98 {
  99         struct rxrpc_connection *conn;
 100         int id;
 101
 102         if (!idr_is_empty(&rxrpc_client_conn_ids)) {
 103                 idr_for_each_entry(&rxrpc_client_conn_ids, conn, id) {
 104                         pr_err("AF_RXRPC: Leaked client conn %p {%d}\n",
 105                                conn, refcount_read(&conn->ref));
 106                 }
 107                 BUG();
 108         }
 109
 110         idr_destroy(&rxrpc_client_conn_ids);
 111 }
 112
 113 /*
 114  * Allocate a connection bundle.
 115  */
 116 static struct rxrpc_bundle *rxrpc_alloc_bundle(struct rxrpc_conn_parameters *cp,
 117                                                gfp_t gfp)
 118 {
 119         struct rxrpc_bundle *bundle;
 120
 121         bundle = kzalloc(sizeof(*bundle), gfp);
 122         if (bundle) {
 123                 bundle->params = *cp;
 124                 rxrpc_get_peer(bundle->params.peer);
 125                 refcount_set(&bundle->ref, 1);
 126                 spin_lock_init(&bundle->channel_lock);
 127                 INIT_LIST_HEAD(&bundle->waiting_calls);
 128         }
 129         return bundle;
 130 }
 131
 132 struct rxrpc_bundle *rxrpc_get_bundle(struct rxrpc_bundle *bundle)
 133 {
 134         refcount_inc(&bundle->ref);
 135         return bundle;
 136 }
 137
 138 static void rxrpc_free_bundle(struct rxrpc_bundle *bundle)
 139 {
 140         rxrpc_put_peer(bundle->params.peer);
 141         kfree(bundle);
 142 }
 143
 144 void rxrpc_put_bundle(struct rxrpc_bundle *bundle)
 145 {
 146         unsigned int d = bundle->debug_id;
 147         bool dead;
 148         int r;
 149
 150         dead = __refcount_dec_and_test(&bundle->ref, &r);
 151
 152         _debug("PUT B=%x %d", d, r);
 153         if (dead)
 154                 rxrpc_free_bundle(bundle);
 155 }
 156
 157 /*
 158  * Allocate a client connection.
 159  */
 160 static struct rxrpc_connection *
 161 rxrpc_alloc_client_connection(struct rxrpc_bundle *bundle, gfp_t gfp)
 162 {
 163         struct rxrpc_connection *conn;
 164         struct rxrpc_net *rxnet = bundle->params.local->rxnet;
 165         int ret;
 166
 167         _enter("");
 168
 169         conn = rxrpc_alloc_connection(gfp);
 170         if (!conn) {
 171                 _leave(" = -ENOMEM");
 172                 return ERR_PTR(-ENOMEM);
 173         }
 174
 175         refcount_set(&conn->ref, 1);
 176         conn->bundle            = bundle;
 177         conn->params            = bundle->params;
 178         conn->out_clientflag    = RXRPC_CLIENT_INITIATED;
 179         conn->state             = RXRPC_CONN_CLIENT;
 180         conn->service_id        = conn->params.service_id;
 181
 182         ret = rxrpc_get_client_connection_id(conn, gfp);
 183         if (ret < 0)
 184                 goto error_0;
 185
 186         ret = rxrpc_init_client_conn_security(conn);
 187         if (ret < 0)
 188                 goto error_1;
 189
 190         atomic_inc(&rxnet->nr_conns);
 191         write_lock(&rxnet->conn_lock);
 192         list_add_tail(&conn->proc_link, &rxnet->conn_proc_list);
 193         write_unlock(&rxnet->conn_lock);
 194
 195         rxrpc_get_bundle(bundle);
 196         rxrpc_get_peer(conn->params.peer);
 197         rxrpc_get_local(conn->params.local);
 198         key_get(conn->params.key);
 199
 200         trace_rxrpc_conn(conn->debug_id, rxrpc_conn_new_client,
 201                          refcount_read(&conn->ref),
 202                          __builtin_return_address(0));
 203
 204         atomic_inc(&rxnet->nr_client_conns);
 205         trace_rxrpc_client(conn, -1, rxrpc_client_alloc);
 206         _leave(" = %p", conn);
 207         return conn;
 208
 209 error_1:
 210         rxrpc_put_client_connection_id(conn);
 211 error_0:
 212         kfree(conn);
 213         _leave(" = %d", ret);
 214         return ERR_PTR(ret);
 215 }
 216
 217 /*
 218  * Determine if a connection may be reused.
 219  */
 220 static bool rxrpc_may_reuse_conn(struct rxrpc_connection *conn)
 221 {
 222         struct rxrpc_net *rxnet;
 223         int id_cursor, id, distance, limit;
 224
 225         if (!conn)
 226                 goto dont_reuse;
 227
 228         rxnet = conn->params.local->rxnet;
 229         if (test_bit(RXRPC_CONN_DONT_REUSE, &conn->flags))
 230                 goto dont_reuse;
 231
 232         if (conn->state != RXRPC_CONN_CLIENT ||
 233             conn->proto.epoch != rxnet->epoch)
 234                 goto mark_dont_reuse;
 235
 236         /* The IDR tree gets very expensive on memory if the connection IDs are
 237          * widely scattered throughout the number space, so we shall want to
 238          * kill off connections that, say, have an ID more than about four
 239          * times the maximum number of client conns away from the current
 240          * allocation point to try and keep the IDs concentrated.
 241          */
 242         id_cursor = idr_get_cursor(&rxrpc_client_conn_ids);
 243         id = conn->proto.cid >> RXRPC_CIDSHIFT;
 244         distance = id - id_cursor;
 245         if (distance < 0)
 246                 distance = -distance;
 247         limit = max_t(unsigned long, atomic_read(&rxnet->nr_conns) * 4, 1024);
 248         if (distance > limit)
 249                 goto mark_dont_reuse;
 250
 251         return true;
 252
 253 mark_dont_reuse:
 254         set_bit(RXRPC_CONN_DONT_REUSE, &conn->flags);
 255 dont_reuse:
 256         return false;
 257 }
 258
 259 /*
 260  * Look up the conn bundle that matches the connection parameters, adding it if
 261  * it doesn't yet exist.
 262  */
 263 static struct rxrpc_bundle *rxrpc_look_up_bundle(struct rxrpc_conn_parameters *cp,
 264                                                  gfp_t gfp)
 265 {
 266         static atomic_t rxrpc_bundle_id;
 267         struct rxrpc_bundle *bundle, *candidate;
 268         struct rxrpc_local *local = cp->local;
 269         struct rb_node *p, **pp, *parent;
 270         long diff;
 271
 272         _enter("{%px,%x,%u,%u}",
 273                cp->peer, key_serial(cp->key), cp->security_level, cp->upgrade);
 274
 275         if (cp->exclusive)
 276                 return rxrpc_alloc_bundle(cp, gfp);
 277
 278         /* First, see if the bundle is already there. */
 279         _debug("search 1");
 280         spin_lock(&local->client_bundles_lock);
 281         p = local->client_bundles.rb_node;
 282         while (p) {
 283                 bundle = rb_entry(p, struct rxrpc_bundle, local_node);
 284
 285 #define cmp(X) ((long)bundle->params.X - (long)cp->X)
 286                 diff = (cmp(peer) ?:
 287                         cmp(key) ?:
 288                         cmp(security_level) ?:
 289                         cmp(upgrade));
 290 #undef cmp
 291                 if (diff < 0)
 292                         p = p->rb_left;
 293                 else if (diff > 0)
 294                         p = p->rb_right;
 295                 else
 296                         goto found_bundle;
 297         }
 298         spin_unlock(&local->client_bundles_lock);
 299         _debug("not found");
 300
 301         /* It wasn't.  We need to add one. */
 302         candidate = rxrpc_alloc_bundle(cp, gfp);
 303         if (!candidate)
 304                 return NULL;
 305
 306         _debug("search 2");
 307         spin_lock(&local->client_bundles_lock);
 308         pp = &local->client_bundles.rb_node;
 309         parent = NULL;
 310         while (*pp) {
 311                 parent = *pp;
 312                 bundle = rb_entry(parent, struct rxrpc_bundle, local_node);
 313
 314 #define cmp(X) ((long)bundle->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                         pp = &(*pp)->rb_left;
 322                 else if (diff > 0)
 323                         pp = &(*pp)->rb_right;
 324                 else
 325                         goto found_bundle_free;
 326         }
 327
 328         _debug("new bundle");
 329         candidate->debug_id = atomic_inc_return(&rxrpc_bundle_id);
 330         rb_link_node(&candidate->local_node, parent, pp);
 331         rb_insert_color(&candidate->local_node, &local->client_bundles);
 332         rxrpc_get_bundle(candidate);
 333         spin_unlock(&local->client_bundles_lock);
 334         _leave(" = %u [new]", candidate->debug_id);
 335         return candidate;
 336
 337 found_bundle_free:
 338         rxrpc_free_bundle(candidate);
 339 found_bundle:
 340         rxrpc_get_bundle(bundle);
 341         spin_unlock(&local->client_bundles_lock);
 342         _leave(" = %u [found]", bundle->debug_id);
 343         return bundle;
 344 }
 345
 346 /*
 347  * Create or find a client bundle to use for a call.
 348  *
 349  * If we return with a connection, the call will be on its waiting list.  It's
 350  * left to the caller to assign a channel and wake up the call.
 351  */
 352 static struct rxrpc_bundle *rxrpc_prep_call(struct rxrpc_sock *rx,
 353                                             struct rxrpc_call *call,
 354                                             struct rxrpc_conn_parameters *cp,
 355                                             struct sockaddr_rxrpc *srx,
 356                                             gfp_t gfp)
 357 {
 358         struct rxrpc_bundle *bundle;
 359
 360         _enter("{%d,%lx},", call->debug_id, call->user_call_ID);
 361
 362         cp->peer = rxrpc_lookup_peer(rx, cp->local, srx, gfp);
 363         if (!cp->peer)
 364                 goto error;
 365
 366         call->cong_cwnd = cp->peer->cong_cwnd;
 367         if (call->cong_cwnd >= call->cong_ssthresh)
 368                 call->cong_mode = RXRPC_CALL_CONGEST_AVOIDANCE;
 369         else
 370                 call->cong_mode = RXRPC_CALL_SLOW_START;
 371         if (cp->upgrade)
 372                 __set_bit(RXRPC_CALL_UPGRADE, &call->flags);
 373
 374         /* Find the client connection bundle. */
 375         bundle = rxrpc_look_up_bundle(cp, gfp);
 376         if (!bundle)
 377                 goto error;
 378
 379         /* Get this call queued.  Someone else may activate it whilst we're
 380          * lining up a new connection, but that's fine.
 381          */
 382         spin_lock(&bundle->channel_lock);
 383         list_add_tail(&call->chan_wait_link, &bundle->waiting_calls);
 384         spin_unlock(&bundle->channel_lock);
 385
 386         _leave(" = [B=%x]", bundle->debug_id);
 387         return bundle;
 388
 389 error:
 390         _leave(" = -ENOMEM");
 391         return ERR_PTR(-ENOMEM);
 392 }
 393
 394 /*
 395  * Allocate a new connection and add it into a bundle.
 396  */
 397 static void rxrpc_add_conn_to_bundle(struct rxrpc_bundle *bundle, gfp_t gfp)
 398         __releases(bundle->channel_lock)
 399 {
 400         struct rxrpc_connection *candidate = NULL, *old = NULL;
 401         bool conflict;
 402         int i;
 403
 404         _enter("");
 405
 406         conflict = bundle->alloc_conn;
 407         if (!conflict)
 408                 bundle->alloc_conn = true;
 409         spin_unlock(&bundle->channel_lock);
 410         if (conflict) {
 411                 _leave(" [conf]");
 412                 return;
 413         }
 414
 415         candidate = rxrpc_alloc_client_connection(bundle, gfp);
 416
 417         spin_lock(&bundle->channel_lock);
 418         bundle->alloc_conn = false;
 419
 420         if (IS_ERR(candidate)) {
 421                 bundle->alloc_error = PTR_ERR(candidate);
 422                 spin_unlock(&bundle->channel_lock);
 423                 _leave(" [err %ld]", PTR_ERR(candidate));
 424                 return;
 425         }
 426
 427         bundle->alloc_error = 0;
 428
 429         for (i = 0; i < ARRAY_SIZE(bundle->conns); i++) {
 430                 unsigned int shift = i * RXRPC_MAXCALLS;
 431                 int j;
 432
 433                 old = bundle->conns[i];
 434                 if (!rxrpc_may_reuse_conn(old)) {
 435                         if (old)
 436                                 trace_rxrpc_client(old, -1, rxrpc_client_replace);
 437                         candidate->bundle_shift = shift;
 438                         bundle->conns[i] = candidate;
 439                         for (j = 0; j < RXRPC_MAXCALLS; j++)
 440                                 set_bit(shift + j, &bundle->avail_chans);
 441                         candidate = NULL;
 442                         break;
 443                 }
 444
 445                 old = NULL;
 446         }
 447
 448         spin_unlock(&bundle->channel_lock);
 449
 450         if (candidate) {
 451                 _debug("discard C=%x", candidate->debug_id);
 452                 trace_rxrpc_client(candidate, -1, rxrpc_client_duplicate);
 453                 rxrpc_put_connection(candidate);
 454         }
 455
 456         rxrpc_put_connection(old);
 457         _leave("");
 458 }
 459
 460 /*
 461  * Add a connection to a bundle if there are no usable connections or we have
 462  * connections waiting for extra capacity.
 463  */
 464 static void rxrpc_maybe_add_conn(struct rxrpc_bundle *bundle, gfp_t gfp)
 465 {
 466         struct rxrpc_call *call;
 467         int i, usable;
 468
 469         _enter("");
 470
 471         spin_lock(&bundle->channel_lock);
 472
 473         /* See if there are any usable connections. */
 474         usable = 0;
 475         for (i = 0; i < ARRAY_SIZE(bundle->conns); i++)
 476                 if (rxrpc_may_reuse_conn(bundle->conns[i]))
 477                         usable++;
 478
 479         if (!usable && !list_empty(&bundle->waiting_calls)) {
 480                 call = list_first_entry(&bundle->waiting_calls,
 481                                         struct rxrpc_call, chan_wait_link);
 482                 if (test_bit(RXRPC_CALL_UPGRADE, &call->flags))
 483                         bundle->try_upgrade = true;
 484         }
 485
 486         if (!usable)
 487                 goto alloc_conn;
 488
 489         if (!bundle->avail_chans &&
 490             !bundle->try_upgrade &&
 491             !list_empty(&bundle->waiting_calls) &&
 492             usable < ARRAY_SIZE(bundle->conns))
 493                 goto alloc_conn;
 494
 495         spin_unlock(&bundle->channel_lock);
 496         _leave("");
 497         return;
 498
 499 alloc_conn:
 500         return rxrpc_add_conn_to_bundle(bundle, gfp);
 501 }
 502
 503 /*
 504  * Assign a channel to the call at the front of the queue and wake the call up.
 505  * We don't increment the callNumber counter until this number has been exposed
 506  * to the world.
 507  */
 508 static void rxrpc_activate_one_channel(struct rxrpc_connection *conn,
 509                                        unsigned int channel)
 510 {
 511         struct rxrpc_channel *chan = &conn->channels[channel];
 512         struct rxrpc_bundle *bundle = conn->bundle;
 513         struct rxrpc_call *call = list_entry(bundle->waiting_calls.next,
 514                                              struct rxrpc_call, chan_wait_link);
 515         u32 call_id = chan->call_counter + 1;
 516
 517         _enter("C=%x,%u", conn->debug_id, channel);
 518
 519         trace_rxrpc_client(conn, channel, rxrpc_client_chan_activate);
 520
 521         /* Cancel the final ACK on the previous call if it hasn't been sent yet
 522          * as the DATA packet will implicitly ACK it.
 523          */
 524         clear_bit(RXRPC_CONN_FINAL_ACK_0 + channel, &conn->flags);
 525         clear_bit(conn->bundle_shift + channel, &bundle->avail_chans);
 526
 527         rxrpc_see_call(call);
 528         list_del_init(&call->chan_wait_link);
 529         call->peer      = rxrpc_get_peer(conn->params.peer);
 530         call->conn      = rxrpc_get_connection(conn);
 531         call->cid       = conn->proto.cid | channel;
 532         call->call_id   = call_id;
 533         call->security  = conn->security;
 534         call->security_ix = conn->security_ix;
 535         call->service_id = conn->service_id;
 536
 537         trace_rxrpc_connect_call(call);
 538         _net("CONNECT call %08x:%08x as call %d on conn %d",
 539              call->cid, call->call_id, call->debug_id, conn->debug_id);
 540
 541         write_lock_bh(&call->state_lock);
 542         call->state = RXRPC_CALL_CLIENT_SEND_REQUEST;
 543         write_unlock_bh(&call->state_lock);
 544
 545         /* Paired with the read barrier in rxrpc_connect_call().  This orders
 546          * cid and epoch in the connection wrt to call_id without the need to
 547          * take the channel_lock.
 548          *
 549          * We provisionally assign a callNumber at this point, but we don't
 550          * confirm it until the call is about to be exposed.
 551          *
 552          * TODO: Pair with a barrier in the data_ready handler when that looks
 553          * at the call ID through a connection channel.
 554          */
 555         smp_wmb();
 556
 557         chan->call_id           = call_id;
 558         chan->call_debug_id     = call->debug_id;
 559         rcu_assign_pointer(chan->call, call);
 560         wake_up(&call->waitq);
 561 }
 562
 563 /*
 564  * Remove a connection from the idle list if it's on it.
 565  */
 566 static void rxrpc_unidle_conn(struct rxrpc_bundle *bundle, struct rxrpc_connection *conn)
 567 {
 568         struct rxrpc_net *rxnet = bundle->params.local->rxnet;
 569         bool drop_ref;
 570
 571         if (!list_empty(&conn->cache_link)) {
 572                 drop_ref = false;
 573                 spin_lock(&rxnet->client_conn_cache_lock);
 574                 if (!list_empty(&conn->cache_link)) {
 575                         list_del_init(&conn->cache_link);
 576                         drop_ref = true;
 577                 }
 578                 spin_unlock(&rxnet->client_conn_cache_lock);
 579                 if (drop_ref)
 580                         rxrpc_put_connection(conn);
 581         }
 582 }
 583
 584 /*
 585  * Assign channels and callNumbers to waiting calls with channel_lock
 586  * held by caller.
 587  */
 588 static void rxrpc_activate_channels_locked(struct rxrpc_bundle *bundle)
 589 {
 590         struct rxrpc_connection *conn;
 591         unsigned long avail, mask;
 592         unsigned int channel, slot;
 593
 594         if (bundle->try_upgrade)
 595                 mask = 1;
 596         else
 597                 mask = ULONG_MAX;
 598
 599         while (!list_empty(&bundle->waiting_calls)) {
 600                 avail = bundle->avail_chans & mask;
 601                 if (!avail)
 602                         break;
 603                 channel = __ffs(avail);
 604                 clear_bit(channel, &bundle->avail_chans);
 605
 606                 slot = channel / RXRPC_MAXCALLS;
 607                 conn = bundle->conns[slot];
 608                 if (!conn)
 609                         break;
 610
 611                 if (bundle->try_upgrade)
 612                         set_bit(RXRPC_CONN_PROBING_FOR_UPGRADE, &conn->flags);
 613                 rxrpc_unidle_conn(bundle, conn);
 614
 615                 channel &= (RXRPC_MAXCALLS - 1);
 616                 conn->act_chans |= 1 << channel;
 617                 rxrpc_activate_one_channel(conn, channel);
 618         }
 619 }
 620
 621 /*
 622  * Assign channels and callNumbers to waiting calls.
 623  */
 624 static void rxrpc_activate_channels(struct rxrpc_bundle *bundle)
 625 {
 626         _enter("B=%x", bundle->debug_id);
 627
 628         trace_rxrpc_client(NULL, -1, rxrpc_client_activate_chans);
 629
 630         if (!bundle->avail_chans)
 631                 return;
 632
 633         spin_lock(&bundle->channel_lock);
 634         rxrpc_activate_channels_locked(bundle);
 635         spin_unlock(&bundle->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_bundle *bundle,
 643                                   struct rxrpc_call *call, gfp_t gfp)
 644 {
 645         DECLARE_WAITQUEUE(myself, current);
 646         int ret = 0;
 647
 648         _enter("%d", call->debug_id);
 649
 650         if (!gfpflags_allow_blocking(gfp)) {
 651                 rxrpc_maybe_add_conn(bundle, gfp);
 652                 rxrpc_activate_channels(bundle);
 653                 ret = bundle->alloc_error ?: -EAGAIN;
 654                 goto out;
 655         }
 656
 657         add_wait_queue_exclusive(&call->waitq, &myself);
 658         for (;;) {
 659                 rxrpc_maybe_add_conn(bundle, gfp);
 660                 rxrpc_activate_channels(bundle);
 661                 ret = bundle->alloc_error;
 662                 if (ret < 0)
 663                         break;
 664
 665                 switch (call->interruptibility) {
 666                 case RXRPC_INTERRUPTIBLE:
 667                 case RXRPC_PREINTERRUPTIBLE:
 668                         set_current_state(TASK_INTERRUPTIBLE);
 669                         break;
 670                 case RXRPC_UNINTERRUPTIBLE:
 671                 default:
 672                         set_current_state(TASK_UNINTERRUPTIBLE);
 673                         break;
 674                 }
 675                 if (READ_ONCE(call->state) != RXRPC_CALL_CLIENT_AWAIT_CONN)
 676                         break;
 677                 if ((call->interruptibility == RXRPC_INTERRUPTIBLE ||
 678                      call->interruptibility == RXRPC_PREINTERRUPTIBLE) &&
 679                     signal_pending(current)) {
 680                         ret = -ERESTARTSYS;
 681                         break;
 682                 }
 683                 schedule();
 684         }
 685         remove_wait_queue(&call->waitq, &myself);
 686         __set_current_state(TASK_RUNNING);
 687
 688 out:
 689         _leave(" = %d", ret);
 690         return ret;
 691 }
 692
 693 /*
 694  * find a connection for a call
 695  * - called in process context with IRQs enabled
 696  */
 697 int rxrpc_connect_call(struct rxrpc_sock *rx,
 698                        struct rxrpc_call *call,
 699                        struct rxrpc_conn_parameters *cp,
 700                        struct sockaddr_rxrpc *srx,
 701                        gfp_t gfp)
 702 {
 703         struct rxrpc_bundle *bundle;
 704         struct rxrpc_net *rxnet = cp->local->rxnet;
 705         int ret = 0;
 706
 707         _enter("{%d,%lx},", call->debug_id, call->user_call_ID);
 708
 709         rxrpc_discard_expired_client_conns(&rxnet->client_conn_reaper);
 710
 711         bundle = rxrpc_prep_call(rx, call, cp, srx, gfp);
 712         if (IS_ERR(bundle)) {
 713                 ret = PTR_ERR(bundle);
 714                 goto out;
 715         }
 716
 717         if (call->state == RXRPC_CALL_CLIENT_AWAIT_CONN) {
 718                 ret = rxrpc_wait_for_channel(bundle, call, gfp);
 719                 if (ret < 0)
 720                         goto wait_failed;
 721         }
 722
 723 granted_channel:
 724         /* Paired with the write barrier in rxrpc_activate_one_channel(). */
 725         smp_rmb();
 726
 727 out_put_bundle:
 728         rxrpc_put_bundle(bundle);
 729 out:
 730         _leave(" = %d", ret);
 731         return ret;
 732
 733 wait_failed:
 734         spin_lock(&bundle->channel_lock);
 735         list_del_init(&call->chan_wait_link);
 736         spin_unlock(&bundle->channel_lock);
 737
 738         if (call->state != RXRPC_CALL_CLIENT_AWAIT_CONN) {
 739                 ret = 0;
 740                 goto granted_channel;
 741         }
 742
 743         trace_rxrpc_client(call->conn, ret, rxrpc_client_chan_wait_failed);
 744         rxrpc_set_call_completion(call, RXRPC_CALL_LOCAL_ERROR, 0, ret);
 745         rxrpc_disconnect_client_call(bundle, call);
 746         goto out_put_bundle;
 747 }
 748
 749 /*
 750  * Note that a call, and thus a connection, is about to be exposed to the
 751  * world.
 752  */
 753 void rxrpc_expose_client_call(struct rxrpc_call *call)
 754 {
 755         unsigned int channel = call->cid & RXRPC_CHANNELMASK;
 756         struct rxrpc_connection *conn = call->conn;
 757         struct rxrpc_channel *chan = &conn->channels[channel];
 758
 759         if (!test_and_set_bit(RXRPC_CALL_EXPOSED, &call->flags)) {
 760                 /* Mark the call ID as being used.  If the callNumber counter
 761                  * exceeds ~2 billion, we kill the connection after its
 762                  * outstanding calls have finished so that the counter doesn't
 763                  * wrap.
 764                  */
 765                 chan->call_counter++;
 766                 if (chan->call_counter >= INT_MAX)
 767                         set_bit(RXRPC_CONN_DONT_REUSE, &conn->flags);
 768                 trace_rxrpc_client(conn, channel, rxrpc_client_exposed);
 769         }
 770 }
 771
 772 /*
 773  * Set the reap timer.
 774  */
 775 static void rxrpc_set_client_reap_timer(struct rxrpc_net *rxnet)
 776 {
 777         if (!rxnet->kill_all_client_conns) {
 778                 unsigned long now = jiffies;
 779                 unsigned long reap_at = now + rxrpc_conn_idle_client_expiry;
 780
 781                 if (rxnet->live)
 782                         timer_reduce(&rxnet->client_conn_reap_timer, reap_at);
 783         }
 784 }
 785
 786 /*
 787  * Disconnect a client call.
 788  */
 789 void rxrpc_disconnect_client_call(struct rxrpc_bundle *bundle, struct rxrpc_call *call)
 790 {
 791         struct rxrpc_connection *conn;
 792         struct rxrpc_channel *chan = NULL;
 793         struct rxrpc_net *rxnet = bundle->params.local->rxnet;
 794         unsigned int channel;
 795         bool may_reuse;
 796         u32 cid;
 797
 798         _enter("c=%x", call->debug_id);
 799
 800         spin_lock(&bundle->channel_lock);
 801         set_bit(RXRPC_CALL_DISCONNECTED, &call->flags);
 802
 803         /* Calls that have never actually been assigned a channel can simply be
 804          * discarded.
 805          */
 806         conn = call->conn;
 807         if (!conn) {
 808                 _debug("call is waiting");
 809                 ASSERTCMP(call->call_id, ==, 0);
 810                 ASSERT(!test_bit(RXRPC_CALL_EXPOSED, &call->flags));
 811                 list_del_init(&call->chan_wait_link);
 812                 goto out;
 813         }
 814
 815         cid = call->cid;
 816         channel = cid & RXRPC_CHANNELMASK;
 817         chan = &conn->channels[channel];
 818         trace_rxrpc_client(conn, channel, rxrpc_client_chan_disconnect);
 819
 820         if (rcu_access_pointer(chan->call) != call) {
 821                 spin_unlock(&bundle->channel_lock);
 822                 BUG();
 823         }
 824
 825         may_reuse = rxrpc_may_reuse_conn(conn);
 826
 827         /* If a client call was exposed to the world, we save the result for
 828          * retransmission.
 829          *
 830          * We use a barrier here so that the call number and abort code can be
 831          * read without needing to take a lock.
 832          *
 833          * TODO: Make the incoming packet handler check this and handle
 834          * terminal retransmission without requiring access to the call.
 835          */
 836         if (test_bit(RXRPC_CALL_EXPOSED, &call->flags)) {
 837                 _debug("exposed %u,%u", call->call_id, call->abort_code);
 838                 __rxrpc_disconnect_call(conn, call);
 839
 840                 if (test_and_clear_bit(RXRPC_CONN_PROBING_FOR_UPGRADE, &conn->flags)) {
 841                         trace_rxrpc_client(conn, channel, rxrpc_client_to_active);
 842                         bundle->try_upgrade = false;
 843                         if (may_reuse)
 844                                 rxrpc_activate_channels_locked(bundle);
 845                 }
 846
 847         }
 848
 849         /* See if we can pass the channel directly to another call. */
 850         if (may_reuse && !list_empty(&bundle->waiting_calls)) {
 851                 trace_rxrpc_client(conn, channel, rxrpc_client_chan_pass);
 852                 rxrpc_activate_one_channel(conn, channel);
 853                 goto out;
 854         }
 855
 856         /* Schedule the final ACK to be transmitted in a short while so that it
 857          * can be skipped if we find a follow-on call.  The first DATA packet
 858          * of the follow on call will implicitly ACK this call.
 859          */
 860         if (call->completion == RXRPC_CALL_SUCCEEDED &&
 861             test_bit(RXRPC_CALL_EXPOSED, &call->flags)) {
 862                 unsigned long final_ack_at = jiffies + 2;
 863
 864                 WRITE_ONCE(chan->final_ack_at, final_ack_at);
 865                 smp_wmb(); /* vs rxrpc_process_delayed_final_acks() */
 866                 set_bit(RXRPC_CONN_FINAL_ACK_0 + channel, &conn->flags);
 867                 rxrpc_reduce_conn_timer(conn, final_ack_at);
 868         }
 869
 870         /* Deactivate the channel. */
 871         rcu_assign_pointer(chan->call, NULL);
 872         set_bit(conn->bundle_shift + channel, &conn->bundle->avail_chans);
 873         conn->act_chans &= ~(1 << channel);
 874
 875         /* If no channels remain active, then put the connection on the idle
 876          * list for a short while.  Give it a ref to stop it going away if it
 877          * becomes unbundled.
 878          */
 879         if (!conn->act_chans) {
 880                 trace_rxrpc_client(conn, channel, rxrpc_client_to_idle);
 881                 conn->idle_timestamp = jiffies;
 882
 883                 rxrpc_get_connection(conn);
 884                 spin_lock(&rxnet->client_conn_cache_lock);
 885                 list_move_tail(&conn->cache_link, &rxnet->idle_client_conns);
 886                 spin_unlock(&rxnet->client_conn_cache_lock);
 887
 888                 rxrpc_set_client_reap_timer(rxnet);
 889         }
 890
 891 out:
 892         spin_unlock(&bundle->channel_lock);
 893         _leave("");
 894         return;
 895 }
 896
 897 /*
 898  * Remove a connection from a bundle.
 899  */
 900 static void rxrpc_unbundle_conn(struct rxrpc_connection *conn)
 901 {
 902         struct rxrpc_bundle *bundle = conn->bundle;
 903         struct rxrpc_local *local = bundle->params.local;
 904         unsigned int bindex;
 905         bool need_drop = false, need_put = false;
 906         int i;
 907
 908         _enter("C=%x", conn->debug_id);
 909
 910         if (conn->flags & RXRPC_CONN_FINAL_ACK_MASK)
 911                 rxrpc_process_delayed_final_acks(conn, true);
 912
 913         spin_lock(&bundle->channel_lock);
 914         bindex = conn->bundle_shift / RXRPC_MAXCALLS;
 915         if (bundle->conns[bindex] == conn) {
 916                 _debug("clear slot %u", bindex);
 917                 bundle->conns[bindex] = NULL;
 918                 for (i = 0; i < RXRPC_MAXCALLS; i++)
 919                         clear_bit(conn->bundle_shift + i, &bundle->avail_chans);
 920                 need_drop = true;
 921         }
 922         spin_unlock(&bundle->channel_lock);
 923
 924         /* If there are no more connections, remove the bundle */
 925         if (!bundle->avail_chans) {
 926                 _debug("maybe unbundle");
 927                 spin_lock(&local->client_bundles_lock);
 928
 929                 for (i = 0; i < ARRAY_SIZE(bundle->conns); i++)
 930                         if (bundle->conns[i])
 931                                 break;
 932                 if (i == ARRAY_SIZE(bundle->conns) && !bundle->params.exclusive) {
 933                         _debug("erase bundle");
 934                         rb_erase(&bundle->local_node, &local->client_bundles);
 935                         need_put = true;
 936                 }
 937
 938                 spin_unlock(&local->client_bundles_lock);
 939                 if (need_put)
 940                         rxrpc_put_bundle(bundle);
 941         }
 942
 943         if (need_drop)
 944                 rxrpc_put_connection(conn);
 945         _leave("");
 946 }
 947
 948 /*
 949  * Clean up a dead client connection.
 950  */
 951 static void rxrpc_kill_client_conn(struct rxrpc_connection *conn)
 952 {
 953         struct rxrpc_local *local = conn->params.local;
 954         struct rxrpc_net *rxnet = local->rxnet;
 955
 956         _enter("C=%x", conn->debug_id);
 957
 958         trace_rxrpc_client(conn, -1, rxrpc_client_cleanup);
 959         atomic_dec(&rxnet->nr_client_conns);
 960
 961         rxrpc_put_client_connection_id(conn);
 962         rxrpc_kill_connection(conn);
 963 }
 964
 965 /*
 966  * Clean up a dead client connections.
 967  */
 968 void rxrpc_put_client_conn(struct rxrpc_connection *conn)
 969 {
 970         const void *here = __builtin_return_address(0);
 971         unsigned int debug_id = conn->debug_id;
 972         bool dead;
 973         int r;
 974
 975         dead = __refcount_dec_and_test(&conn->ref, &r);
 976         trace_rxrpc_conn(debug_id, rxrpc_conn_put_client, r - 1, here);
 977         if (dead)
 978                 rxrpc_kill_client_conn(conn);
 979 }
 980
 981 /*
 982  * Discard expired client connections from the idle list.  Each conn in the
 983  * idle list has been exposed and holds an extra ref because of that.
 984  *
 985  * This may be called from conn setup or from a work item so cannot be
 986  * considered non-reentrant.
 987  */
 988 void rxrpc_discard_expired_client_conns(struct work_struct *work)
 989 {
 990         struct rxrpc_connection *conn;
 991         struct rxrpc_net *rxnet =
 992                 container_of(work, struct rxrpc_net, client_conn_reaper);
 993         unsigned long expiry, conn_expires_at, now;
 994         unsigned int nr_conns;
 995
 996         _enter("");
 997
 998         if (list_empty(&rxnet->idle_client_conns)) {
 999                 _leave(" [empty]");
1000                 return;
1001         }
1002
1003         /* Don't double up on the discarding */
1004         if (!spin_trylock(&rxnet->client_conn_discard_lock)) {
1005                 _leave(" [already]");
1006                 return;
1007         }
1008
1009         /* We keep an estimate of what the number of conns ought to be after
1010          * we've discarded some so that we don't overdo the discarding.
1011          */
1012         nr_conns = atomic_read(&rxnet->nr_client_conns);
1013
1014 next:
1015         spin_lock(&rxnet->client_conn_cache_lock);
1016
1017         if (list_empty(&rxnet->idle_client_conns))
1018                 goto out;
1019
1020         conn = list_entry(rxnet->idle_client_conns.next,
1021                           struct rxrpc_connection, cache_link);
1022
1023         if (!rxnet->kill_all_client_conns) {
1024                 /* If the number of connections is over the reap limit, we
1025                  * expedite discard by reducing the expiry timeout.  We must,
1026                  * however, have at least a short grace period to be able to do
1027                  * final-ACK or ABORT retransmission.
1028                  */
1029                 expiry = rxrpc_conn_idle_client_expiry;
1030                 if (nr_conns > rxrpc_reap_client_connections)
1031                         expiry = rxrpc_conn_idle_client_fast_expiry;
1032                 if (conn->params.local->service_closed)
1033                         expiry = rxrpc_closed_conn_expiry * HZ;
1034
1035                 conn_expires_at = conn->idle_timestamp + expiry;
1036
1037                 now = READ_ONCE(jiffies);
1038                 if (time_after(conn_expires_at, now))
1039                         goto not_yet_expired;
1040         }
1041
1042         trace_rxrpc_client(conn, -1, rxrpc_client_discard);
1043         list_del_init(&conn->cache_link);
1044
1045         spin_unlock(&rxnet->client_conn_cache_lock);
1046
1047         rxrpc_unbundle_conn(conn);
1048         rxrpc_put_connection(conn); /* Drop the ->cache_link ref */
1049
1050         nr_conns--;
1051         goto next;
1052
1053 not_yet_expired:
1054         /* The connection at the front of the queue hasn't yet expired, so
1055          * schedule the work item for that point if we discarded something.
1056          *
1057          * We don't worry if the work item is already scheduled - it can look
1058          * after rescheduling itself at a later time.  We could cancel it, but
1059          * then things get messier.
1060          */
1061         _debug("not yet");
1062         if (!rxnet->kill_all_client_conns)
1063                 timer_reduce(&rxnet->client_conn_reap_timer, conn_expires_at);
1064
1065 out:
1066         spin_unlock(&rxnet->client_conn_cache_lock);
1067         spin_unlock(&rxnet->client_conn_discard_lock);
1068         _leave("");
1069 }
1070
1071 /*
1072  * Preemptively destroy all the client connection records rather than waiting
1073  * for them to time out
1074  */
1075 void rxrpc_destroy_all_client_connections(struct rxrpc_net *rxnet)
1076 {
1077         _enter("");
1078
1079         spin_lock(&rxnet->client_conn_cache_lock);
1080         rxnet->kill_all_client_conns = true;
1081         spin_unlock(&rxnet->client_conn_cache_lock);
1082
1083         del_timer_sync(&rxnet->client_conn_reap_timer);
1084
1085         if (!rxrpc_queue_work(&rxnet->client_conn_reaper))
1086                 _debug("destroy: queue failed");
1087
1088         _leave("");
1089 }
1090
1091 /*
1092  * Clean up the client connections on a local endpoint.
1093  */
1094 void rxrpc_clean_up_local_conns(struct rxrpc_local *local)
1095 {
1096         struct rxrpc_connection *conn, *tmp;
1097         struct rxrpc_net *rxnet = local->rxnet;
1098         LIST_HEAD(graveyard);
1099
1100         _enter("");
1101
1102         spin_lock(&rxnet->client_conn_cache_lock);
1103
1104         list_for_each_entry_safe(conn, tmp, &rxnet->idle_client_conns,
1105                                  cache_link) {
1106                 if (conn->params.local == local) {
1107                         trace_rxrpc_client(conn, -1, rxrpc_client_discard);
1108                         list_move(&conn->cache_link, &graveyard);
1109                 }
1110         }
1111
1112         spin_unlock(&rxnet->client_conn_cache_lock);
1113
1114         while (!list_empty(&graveyard)) {
1115                 conn = list_entry(graveyard.next,
1116                                   struct rxrpc_connection, cache_link);
1117                 list_del_init(&conn->cache_link);
1118                 rxrpc_unbundle_conn(conn);
1119                 rxrpc_put_connection(conn);
1120         }
1121
1122         _leave(" [culled]");
1123 }