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 static void rxrpc_deactivate_bundle(struct rxrpc_bundle *bundle);
  44
  45 /*
  46  * Get a connection ID and epoch for a client connection from the global pool.
  47  * The connection struct pointer is then recorded in the idr radix tree.  The
  48  * epoch doesn't change until the client is rebooted (or, at least, unless the
  49  * module is unloaded).
  50  */
  51 static int rxrpc_get_client_connection_id(struct rxrpc_connection *conn,
  52                                           gfp_t gfp)
  53 {
  54         struct rxrpc_net *rxnet = conn->rxnet;
  55         int id;
  56
  57         _enter("");
  58
  59         idr_preload(gfp);
  60         spin_lock(&rxrpc_conn_id_lock);
  61
  62         id = idr_alloc_cyclic(&rxrpc_client_conn_ids, conn,
  63                               1, 0x40000000, GFP_NOWAIT);
  64         if (id < 0)
  65                 goto error;
  66
  67         spin_unlock(&rxrpc_conn_id_lock);
  68         idr_preload_end();
  69
  70         conn->proto.epoch = rxnet->epoch;
  71         conn->proto.cid = id << RXRPC_CIDSHIFT;
  72         set_bit(RXRPC_CONN_HAS_IDR, &conn->flags);
  73         _leave(" [CID %x]", conn->proto.cid);
  74         return 0;
  75
  76 error:
  77         spin_unlock(&rxrpc_conn_id_lock);
  78         idr_preload_end();
  79         _leave(" = %d", id);
  80         return id;
  81 }
  82
  83 /*
  84  * Release a connection ID for a client connection from the global pool.
  85  */
  86 static void rxrpc_put_client_connection_id(struct rxrpc_connection *conn)
  87 {
  88         if (test_bit(RXRPC_CONN_HAS_IDR, &conn->flags)) {
  89                 spin_lock(&rxrpc_conn_id_lock);
  90                 idr_remove(&rxrpc_client_conn_ids,
  91                            conn->proto.cid >> RXRPC_CIDSHIFT);
  92                 spin_unlock(&rxrpc_conn_id_lock);
  93         }
  94 }
  95
  96 /*
  97  * Destroy the client connection ID tree.
  98  */
  99 void rxrpc_destroy_client_conn_ids(void)
 100 {
 101         struct rxrpc_connection *conn;
 102         int id;
 103
 104         if (!idr_is_empty(&rxrpc_client_conn_ids)) {
 105                 idr_for_each_entry(&rxrpc_client_conn_ids, conn, id) {
 106                         pr_err("AF_RXRPC: Leaked client conn %p {%d}\n",
 107                                conn, refcount_read(&conn->ref));
 108                 }
 109                 BUG();
 110         }
 111
 112         idr_destroy(&rxrpc_client_conn_ids);
 113 }
 114
 115 /*
 116  * Allocate a connection bundle.
 117  */
 118 static struct rxrpc_bundle *rxrpc_alloc_bundle(struct rxrpc_conn_parameters *cp,
 119                                                gfp_t gfp)
 120 {
 121         struct rxrpc_bundle *bundle;
 122
 123         bundle = kzalloc(sizeof(*bundle), gfp);
 124         if (bundle) {
 125                 bundle->local           = cp->local;
 126                 bundle->peer            = rxrpc_get_peer(cp->peer, rxrpc_peer_get_bundle);
 127                 bundle->key             = cp->key;
 128                 bundle->exclusive       = cp->exclusive;
 129                 bundle->upgrade         = cp->upgrade;
 130                 bundle->service_id      = cp->service_id;
 131                 bundle->security_level  = cp->security_level;
 132                 refcount_set(&bundle->ref, 1);
 133                 atomic_set(&bundle->active, 1);
 134                 spin_lock_init(&bundle->channel_lock);
 135                 INIT_LIST_HEAD(&bundle->waiting_calls);
 136                 trace_rxrpc_bundle(bundle->debug_id, 1, rxrpc_bundle_new);
 137         }
 138         return bundle;
 139 }
 140
 141 struct rxrpc_bundle *rxrpc_get_bundle(struct rxrpc_bundle *bundle,
 142                                       enum rxrpc_bundle_trace why)
 143 {
 144         int r;
 145
 146         __refcount_inc(&bundle->ref, &r);
 147         trace_rxrpc_bundle(bundle->debug_id, r + 1, why);
 148         return bundle;
 149 }
 150
 151 static void rxrpc_free_bundle(struct rxrpc_bundle *bundle)
 152 {
 153         trace_rxrpc_bundle(bundle->debug_id, 1, rxrpc_bundle_free);
 154         rxrpc_put_peer(bundle->peer, rxrpc_peer_put_bundle);
 155         kfree(bundle);
 156 }
 157
 158 void rxrpc_put_bundle(struct rxrpc_bundle *bundle, enum rxrpc_bundle_trace why)
 159 {
 160         unsigned int id = bundle->debug_id;
 161         bool dead;
 162         int r;
 163
 164         dead = __refcount_dec_and_test(&bundle->ref, &r);
 165         trace_rxrpc_bundle(id, r - 1, why);
 166         if (dead)
 167                 rxrpc_free_bundle(bundle);
 168 }
 169
 170 /*
 171  * Allocate a client connection.
 172  */
 173 static struct rxrpc_connection *
 174 rxrpc_alloc_client_connection(struct rxrpc_bundle *bundle, gfp_t gfp)
 175 {
 176         struct rxrpc_connection *conn;
 177         struct rxrpc_net *rxnet = bundle->local->rxnet;
 178         int ret;
 179
 180         _enter("");
 181
 182         conn = rxrpc_alloc_connection(rxnet, gfp);
 183         if (!conn) {
 184                 _leave(" = -ENOMEM");
 185                 return ERR_PTR(-ENOMEM);
 186         }
 187
 188         refcount_set(&conn->ref, 1);
 189         conn->bundle            = bundle;
 190         conn->local             = bundle->local;
 191         conn->peer              = bundle->peer;
 192         conn->key               = bundle->key;
 193         conn->exclusive         = bundle->exclusive;
 194         conn->upgrade           = bundle->upgrade;
 195         conn->orig_service_id   = bundle->service_id;
 196         conn->security_level    = bundle->security_level;
 197         conn->out_clientflag    = RXRPC_CLIENT_INITIATED;
 198         conn->state             = RXRPC_CONN_CLIENT;
 199         conn->service_id        = conn->orig_service_id;
 200
 201         ret = rxrpc_get_client_connection_id(conn, gfp);
 202         if (ret < 0)
 203                 goto error_0;
 204
 205         ret = rxrpc_init_client_conn_security(conn);
 206         if (ret < 0)
 207                 goto error_1;
 208
 209         atomic_inc(&rxnet->nr_conns);
 210         write_lock(&rxnet->conn_lock);
 211         list_add_tail(&conn->proc_link, &rxnet->conn_proc_list);
 212         write_unlock(&rxnet->conn_lock);
 213
 214         rxrpc_get_bundle(bundle, rxrpc_bundle_get_client_conn);
 215         rxrpc_get_peer(conn->peer, rxrpc_peer_get_client_conn);
 216         rxrpc_get_local(conn->local, rxrpc_local_get_client_conn);
 217         key_get(conn->key);
 218
 219         trace_rxrpc_conn(conn->debug_id, refcount_read(&conn->ref),
 220                          rxrpc_conn_new_client);
 221
 222         atomic_inc(&rxnet->nr_client_conns);
 223         trace_rxrpc_client(conn, -1, rxrpc_client_alloc);
 224         _leave(" = %p", conn);
 225         return conn;
 226
 227 error_1:
 228         rxrpc_put_client_connection_id(conn);
 229 error_0:
 230         kfree(conn);
 231         _leave(" = %d", ret);
 232         return ERR_PTR(ret);
 233 }
 234
 235 /*
 236  * Determine if a connection may be reused.
 237  */
 238 static bool rxrpc_may_reuse_conn(struct rxrpc_connection *conn)
 239 {
 240         struct rxrpc_net *rxnet;
 241         int id_cursor, id, distance, limit;
 242
 243         if (!conn)
 244                 goto dont_reuse;
 245
 246         rxnet = conn->rxnet;
 247         if (test_bit(RXRPC_CONN_DONT_REUSE, &conn->flags))
 248                 goto dont_reuse;
 249
 250         if (conn->state != RXRPC_CONN_CLIENT ||
 251             conn->proto.epoch != rxnet->epoch)
 252                 goto mark_dont_reuse;
 253
 254         /* The IDR tree gets very expensive on memory if the connection IDs are
 255          * widely scattered throughout the number space, so we shall want to
 256          * kill off connections that, say, have an ID more than about four
 257          * times the maximum number of client conns away from the current
 258          * allocation point to try and keep the IDs concentrated.
 259          */
 260         id_cursor = idr_get_cursor(&rxrpc_client_conn_ids);
 261         id = conn->proto.cid >> RXRPC_CIDSHIFT;
 262         distance = id - id_cursor;
 263         if (distance < 0)
 264                 distance = -distance;
 265         limit = max_t(unsigned long, atomic_read(&rxnet->nr_conns) * 4, 1024);
 266         if (distance > limit)
 267                 goto mark_dont_reuse;
 268
 269         return true;
 270
 271 mark_dont_reuse:
 272         set_bit(RXRPC_CONN_DONT_REUSE, &conn->flags);
 273 dont_reuse:
 274         return false;
 275 }
 276
 277 /*
 278  * Look up the conn bundle that matches the connection parameters, adding it if
 279  * it doesn't yet exist.
 280  */
 281 static struct rxrpc_bundle *rxrpc_look_up_bundle(struct rxrpc_conn_parameters *cp,
 282                                                  gfp_t gfp)
 283 {
 284         static atomic_t rxrpc_bundle_id;
 285         struct rxrpc_bundle *bundle, *candidate;
 286         struct rxrpc_local *local = cp->local;
 287         struct rb_node *p, **pp, *parent;
 288         long diff;
 289
 290         _enter("{%px,%x,%u,%u}",
 291                cp->peer, key_serial(cp->key), cp->security_level, cp->upgrade);
 292
 293         if (cp->exclusive)
 294                 return rxrpc_alloc_bundle(cp, gfp);
 295
 296         /* First, see if the bundle is already there. */
 297         _debug("search 1");
 298         spin_lock(&local->client_bundles_lock);
 299         p = local->client_bundles.rb_node;
 300         while (p) {
 301                 bundle = rb_entry(p, struct rxrpc_bundle, local_node);
 302
 303 #define cmp(X) ((long)bundle->X - (long)cp->X)
 304                 diff = (cmp(peer) ?:
 305                         cmp(key) ?:
 306                         cmp(security_level) ?:
 307                         cmp(upgrade));
 308 #undef cmp
 309                 if (diff < 0)
 310                         p = p->rb_left;
 311                 else if (diff > 0)
 312                         p = p->rb_right;
 313                 else
 314                         goto found_bundle;
 315         }
 316         spin_unlock(&local->client_bundles_lock);
 317         _debug("not found");
 318
 319         /* It wasn't.  We need to add one. */
 320         candidate = rxrpc_alloc_bundle(cp, gfp);
 321         if (!candidate)
 322                 return NULL;
 323
 324         _debug("search 2");
 325         spin_lock(&local->client_bundles_lock);
 326         pp = &local->client_bundles.rb_node;
 327         parent = NULL;
 328         while (*pp) {
 329                 parent = *pp;
 330                 bundle = rb_entry(parent, struct rxrpc_bundle, local_node);
 331
 332 #define cmp(X) ((long)bundle->X - (long)cp->X)
 333                 diff = (cmp(peer) ?:
 334                         cmp(key) ?:
 335                         cmp(security_level) ?:
 336                         cmp(upgrade));
 337 #undef cmp
 338                 if (diff < 0)
 339                         pp = &(*pp)->rb_left;
 340                 else if (diff > 0)
 341                         pp = &(*pp)->rb_right;
 342                 else
 343                         goto found_bundle_free;
 344         }
 345
 346         _debug("new bundle");
 347         candidate->debug_id = atomic_inc_return(&rxrpc_bundle_id);
 348         rb_link_node(&candidate->local_node, parent, pp);
 349         rb_insert_color(&candidate->local_node, &local->client_bundles);
 350         rxrpc_get_bundle(candidate, rxrpc_bundle_get_client_call);
 351         spin_unlock(&local->client_bundles_lock);
 352         _leave(" = %u [new]", candidate->debug_id);
 353         return candidate;
 354
 355 found_bundle_free:
 356         rxrpc_free_bundle(candidate);
 357 found_bundle:
 358         rxrpc_get_bundle(bundle, rxrpc_bundle_get_client_call);
 359         atomic_inc(&bundle->active);
 360         spin_unlock(&local->client_bundles_lock);
 361         _leave(" = %u [found]", bundle->debug_id);
 362         return bundle;
 363 }
 364
 365 /*
 366  * Create or find a client bundle to use for a call.
 367  *
 368  * If we return with a connection, the call will be on its waiting list.  It's
 369  * left to the caller to assign a channel and wake up the call.
 370  */
 371 static struct rxrpc_bundle *rxrpc_prep_call(struct rxrpc_sock *rx,
 372                                             struct rxrpc_call *call,
 373                                             struct rxrpc_conn_parameters *cp,
 374                                             struct sockaddr_rxrpc *srx,
 375                                             gfp_t gfp)
 376 {
 377         struct rxrpc_bundle *bundle;
 378
 379         _enter("{%d,%lx},", call->debug_id, call->user_call_ID);
 380
 381         cp->peer = rxrpc_lookup_peer(rx, cp->local, srx, gfp);
 382         if (!cp->peer)
 383                 goto error;
 384
 385         call->tx_last_sent = ktime_get_real();
 386         call->cong_ssthresh = cp->peer->cong_ssthresh;
 387         if (call->cong_cwnd >= call->cong_ssthresh)
 388                 call->cong_mode = RXRPC_CALL_CONGEST_AVOIDANCE;
 389         else
 390                 call->cong_mode = RXRPC_CALL_SLOW_START;
 391         if (cp->upgrade)
 392                 __set_bit(RXRPC_CALL_UPGRADE, &call->flags);
 393
 394         /* Find the client connection bundle. */
 395         bundle = rxrpc_look_up_bundle(cp, gfp);
 396         if (!bundle)
 397                 goto error;
 398
 399         /* Get this call queued.  Someone else may activate it whilst we're
 400          * lining up a new connection, but that's fine.
 401          */
 402         spin_lock(&bundle->channel_lock);
 403         list_add_tail(&call->chan_wait_link, &bundle->waiting_calls);
 404         spin_unlock(&bundle->channel_lock);
 405
 406         _leave(" = [B=%x]", bundle->debug_id);
 407         return bundle;
 408
 409 error:
 410         _leave(" = -ENOMEM");
 411         return ERR_PTR(-ENOMEM);
 412 }
 413
 414 /*
 415  * Allocate a new connection and add it into a bundle.
 416  */
 417 static void rxrpc_add_conn_to_bundle(struct rxrpc_bundle *bundle, gfp_t gfp)
 418         __releases(bundle->channel_lock)
 419 {
 420         struct rxrpc_connection *candidate = NULL, *old = NULL;
 421         bool conflict;
 422         int i;
 423
 424         _enter("");
 425
 426         conflict = bundle->alloc_conn;
 427         if (!conflict)
 428                 bundle->alloc_conn = true;
 429         spin_unlock(&bundle->channel_lock);
 430         if (conflict) {
 431                 _leave(" [conf]");
 432                 return;
 433         }
 434
 435         candidate = rxrpc_alloc_client_connection(bundle, gfp);
 436
 437         spin_lock(&bundle->channel_lock);
 438         bundle->alloc_conn = false;
 439
 440         if (IS_ERR(candidate)) {
 441                 bundle->alloc_error = PTR_ERR(candidate);
 442                 spin_unlock(&bundle->channel_lock);
 443                 _leave(" [err %ld]", PTR_ERR(candidate));
 444                 return;
 445         }
 446
 447         bundle->alloc_error = 0;
 448
 449         for (i = 0; i < ARRAY_SIZE(bundle->conns); i++) {
 450                 unsigned int shift = i * RXRPC_MAXCALLS;
 451                 int j;
 452
 453                 old = bundle->conns[i];
 454                 if (!rxrpc_may_reuse_conn(old)) {
 455                         if (old)
 456                                 trace_rxrpc_client(old, -1, rxrpc_client_replace);
 457                         candidate->bundle_shift = shift;
 458                         atomic_inc(&bundle->active);
 459                         bundle->conns[i] = candidate;
 460                         for (j = 0; j < RXRPC_MAXCALLS; j++)
 461                                 set_bit(shift + j, &bundle->avail_chans);
 462                         candidate = NULL;
 463                         break;
 464                 }
 465
 466                 old = NULL;
 467         }
 468
 469         spin_unlock(&bundle->channel_lock);
 470
 471         if (candidate) {
 472                 _debug("discard C=%x", candidate->debug_id);
 473                 trace_rxrpc_client(candidate, -1, rxrpc_client_duplicate);
 474                 rxrpc_put_connection(candidate, rxrpc_conn_put_discard);
 475         }
 476
 477         rxrpc_put_connection(old, rxrpc_conn_put_noreuse);
 478         _leave("");
 479 }
 480
 481 /*
 482  * Add a connection to a bundle if there are no usable connections or we have
 483  * connections waiting for extra capacity.
 484  */
 485 static void rxrpc_maybe_add_conn(struct rxrpc_bundle *bundle, gfp_t gfp)
 486 {
 487         struct rxrpc_call *call;
 488         int i, usable;
 489
 490         _enter("");
 491
 492         spin_lock(&bundle->channel_lock);
 493
 494         /* See if there are any usable connections. */
 495         usable = 0;
 496         for (i = 0; i < ARRAY_SIZE(bundle->conns); i++)
 497                 if (rxrpc_may_reuse_conn(bundle->conns[i]))
 498                         usable++;
 499
 500         if (!usable && !list_empty(&bundle->waiting_calls)) {
 501                 call = list_first_entry(&bundle->waiting_calls,
 502                                         struct rxrpc_call, chan_wait_link);
 503                 if (test_bit(RXRPC_CALL_UPGRADE, &call->flags))
 504                         bundle->try_upgrade = true;
 505         }
 506
 507         if (!usable)
 508                 goto alloc_conn;
 509
 510         if (!bundle->avail_chans &&
 511             !bundle->try_upgrade &&
 512             !list_empty(&bundle->waiting_calls) &&
 513             usable < ARRAY_SIZE(bundle->conns))
 514                 goto alloc_conn;
 515
 516         spin_unlock(&bundle->channel_lock);
 517         _leave("");
 518         return;
 519
 520 alloc_conn:
 521         return rxrpc_add_conn_to_bundle(bundle, gfp);
 522 }
 523
 524 /*
 525  * Assign a channel to the call at the front of the queue and wake the call up.
 526  * We don't increment the callNumber counter until this number has been exposed
 527  * to the world.
 528  */
 529 static void rxrpc_activate_one_channel(struct rxrpc_connection *conn,
 530                                        unsigned int channel)
 531 {
 532         struct rxrpc_channel *chan = &conn->channels[channel];
 533         struct rxrpc_bundle *bundle = conn->bundle;
 534         struct rxrpc_call *call = list_entry(bundle->waiting_calls.next,
 535                                              struct rxrpc_call, chan_wait_link);
 536         u32 call_id = chan->call_counter + 1;
 537
 538         _enter("C=%x,%u", conn->debug_id, channel);
 539
 540         trace_rxrpc_client(conn, channel, rxrpc_client_chan_activate);
 541
 542         /* Cancel the final ACK on the previous call if it hasn't been sent yet
 543          * as the DATA packet will implicitly ACK it.
 544          */
 545         clear_bit(RXRPC_CONN_FINAL_ACK_0 + channel, &conn->flags);
 546         clear_bit(conn->bundle_shift + channel, &bundle->avail_chans);
 547
 548         rxrpc_see_call(call, rxrpc_call_see_activate_client);
 549         list_del_init(&call->chan_wait_link);
 550         call->peer      = rxrpc_get_peer(conn->peer, rxrpc_peer_get_activate_call);
 551         call->conn      = rxrpc_get_connection(conn, rxrpc_conn_get_activate_call);
 552         call->cid       = conn->proto.cid | channel;
 553         call->call_id   = call_id;
 554         call->dest_srx.srx_service = conn->service_id;
 555
 556         trace_rxrpc_connect_call(call);
 557
 558         write_lock(&call->state_lock);
 559         call->state = RXRPC_CALL_CLIENT_SEND_REQUEST;
 560         write_unlock(&call->state_lock);
 561
 562         /* Paired with the read barrier in rxrpc_connect_call().  This orders
 563          * cid and epoch in the connection wrt to call_id without the need to
 564          * take the channel_lock.
 565          *
 566          * We provisionally assign a callNumber at this point, but we don't
 567          * confirm it until the call is about to be exposed.
 568          *
 569          * TODO: Pair with a barrier in the data_ready handler when that looks
 570          * at the call ID through a connection channel.
 571          */
 572         smp_wmb();
 573
 574         chan->call_id           = call_id;
 575         chan->call_debug_id     = call->debug_id;
 576         rcu_assign_pointer(chan->call, call);
 577         wake_up(&call->waitq);
 578 }
 579
 580 /*
 581  * Remove a connection from the idle list if it's on it.
 582  */
 583 static void rxrpc_unidle_conn(struct rxrpc_bundle *bundle, struct rxrpc_connection *conn)
 584 {
 585         struct rxrpc_net *rxnet = bundle->local->rxnet;
 586         bool drop_ref;
 587
 588         if (!list_empty(&conn->cache_link)) {
 589                 drop_ref = false;
 590                 spin_lock(&rxnet->client_conn_cache_lock);
 591                 if (!list_empty(&conn->cache_link)) {
 592                         list_del_init(&conn->cache_link);
 593                         drop_ref = true;
 594                 }
 595                 spin_unlock(&rxnet->client_conn_cache_lock);
 596                 if (drop_ref)
 597                         rxrpc_put_connection(conn, rxrpc_conn_put_unidle);
 598         }
 599 }
 600
 601 /*
 602  * Assign channels and callNumbers to waiting calls with channel_lock
 603  * held by caller.
 604  */
 605 static void rxrpc_activate_channels_locked(struct rxrpc_bundle *bundle)
 606 {
 607         struct rxrpc_connection *conn;
 608         unsigned long avail, mask;
 609         unsigned int channel, slot;
 610
 611         if (bundle->try_upgrade)
 612                 mask = 1;
 613         else
 614                 mask = ULONG_MAX;
 615
 616         while (!list_empty(&bundle->waiting_calls)) {
 617                 avail = bundle->avail_chans & mask;
 618                 if (!avail)
 619                         break;
 620                 channel = __ffs(avail);
 621                 clear_bit(channel, &bundle->avail_chans);
 622
 623                 slot = channel / RXRPC_MAXCALLS;
 624                 conn = bundle->conns[slot];
 625                 if (!conn)
 626                         break;
 627
 628                 if (bundle->try_upgrade)
 629                         set_bit(RXRPC_CONN_PROBING_FOR_UPGRADE, &conn->flags);
 630                 rxrpc_unidle_conn(bundle, conn);
 631
 632                 channel &= (RXRPC_MAXCALLS - 1);
 633                 conn->act_chans |= 1 << channel;
 634                 rxrpc_activate_one_channel(conn, channel);
 635         }
 636 }
 637
 638 /*
 639  * Assign channels and callNumbers to waiting calls.
 640  */
 641 static void rxrpc_activate_channels(struct rxrpc_bundle *bundle)
 642 {
 643         _enter("B=%x", bundle->debug_id);
 644
 645         trace_rxrpc_client(NULL, -1, rxrpc_client_activate_chans);
 646
 647         if (!bundle->avail_chans)
 648                 return;
 649
 650         spin_lock(&bundle->channel_lock);
 651         rxrpc_activate_channels_locked(bundle);
 652         spin_unlock(&bundle->channel_lock);
 653         _leave("");
 654 }
 655
 656 /*
 657  * Wait for a callNumber and a channel to be granted to a call.
 658  */
 659 static int rxrpc_wait_for_channel(struct rxrpc_bundle *bundle,
 660                                   struct rxrpc_call *call, gfp_t gfp)
 661 {
 662         DECLARE_WAITQUEUE(myself, current);
 663         int ret = 0;
 664
 665         _enter("%d", call->debug_id);
 666
 667         if (!gfpflags_allow_blocking(gfp)) {
 668                 rxrpc_maybe_add_conn(bundle, gfp);
 669                 rxrpc_activate_channels(bundle);
 670                 ret = bundle->alloc_error ?: -EAGAIN;
 671                 goto out;
 672         }
 673
 674         add_wait_queue_exclusive(&call->waitq, &myself);
 675         for (;;) {
 676                 rxrpc_maybe_add_conn(bundle, gfp);
 677                 rxrpc_activate_channels(bundle);
 678                 ret = bundle->alloc_error;
 679                 if (ret < 0)
 680                         break;
 681
 682                 switch (call->interruptibility) {
 683                 case RXRPC_INTERRUPTIBLE:
 684                 case RXRPC_PREINTERRUPTIBLE:
 685                         set_current_state(TASK_INTERRUPTIBLE);
 686                         break;
 687                 case RXRPC_UNINTERRUPTIBLE:
 688                 default:
 689                         set_current_state(TASK_UNINTERRUPTIBLE);
 690                         break;
 691                 }
 692                 if (READ_ONCE(call->state) != RXRPC_CALL_CLIENT_AWAIT_CONN)
 693                         break;
 694                 if ((call->interruptibility == RXRPC_INTERRUPTIBLE ||
 695                      call->interruptibility == RXRPC_PREINTERRUPTIBLE) &&
 696                     signal_pending(current)) {
 697                         ret = -ERESTARTSYS;
 698                         break;
 699                 }
 700                 schedule();
 701         }
 702         remove_wait_queue(&call->waitq, &myself);
 703         __set_current_state(TASK_RUNNING);
 704
 705 out:
 706         _leave(" = %d", ret);
 707         return ret;
 708 }
 709
 710 /*
 711  * find a connection for a call
 712  * - called in process context with IRQs enabled
 713  */
 714 int rxrpc_connect_call(struct rxrpc_sock *rx,
 715                        struct rxrpc_call *call,
 716                        struct rxrpc_conn_parameters *cp,
 717                        struct sockaddr_rxrpc *srx,
 718                        gfp_t gfp)
 719 {
 720         struct rxrpc_bundle *bundle;
 721         struct rxrpc_net *rxnet = cp->local->rxnet;
 722         int ret = 0;
 723
 724         _enter("{%d,%lx},", call->debug_id, call->user_call_ID);
 725
 726         rxrpc_discard_expired_client_conns(&rxnet->client_conn_reaper);
 727
 728         bundle = rxrpc_prep_call(rx, call, cp, srx, gfp);
 729         if (IS_ERR(bundle)) {
 730                 ret = PTR_ERR(bundle);
 731                 goto out;
 732         }
 733
 734         if (call->state == RXRPC_CALL_CLIENT_AWAIT_CONN) {
 735                 ret = rxrpc_wait_for_channel(bundle, call, gfp);
 736                 if (ret < 0)
 737                         goto wait_failed;
 738         }
 739
 740 granted_channel:
 741         /* Paired with the write barrier in rxrpc_activate_one_channel(). */
 742         smp_rmb();
 743
 744 out_put_bundle:
 745         rxrpc_deactivate_bundle(bundle);
 746         rxrpc_put_bundle(bundle, rxrpc_bundle_get_client_call);
 747 out:
 748         _leave(" = %d", ret);
 749         return ret;
 750
 751 wait_failed:
 752         spin_lock(&bundle->channel_lock);
 753         list_del_init(&call->chan_wait_link);
 754         spin_unlock(&bundle->channel_lock);
 755
 756         if (call->state != RXRPC_CALL_CLIENT_AWAIT_CONN) {
 757                 ret = 0;
 758                 goto granted_channel;
 759         }
 760
 761         trace_rxrpc_client(call->conn, ret, rxrpc_client_chan_wait_failed);
 762         rxrpc_set_call_completion(call, RXRPC_CALL_LOCAL_ERROR, 0, ret);
 763         rxrpc_disconnect_client_call(bundle, call);
 764         goto out_put_bundle;
 765 }
 766
 767 /*
 768  * Note that a call, and thus a connection, is about to be exposed to the
 769  * world.
 770  */
 771 void rxrpc_expose_client_call(struct rxrpc_call *call)
 772 {
 773         unsigned int channel = call->cid & RXRPC_CHANNELMASK;
 774         struct rxrpc_connection *conn = call->conn;
 775         struct rxrpc_channel *chan = &conn->channels[channel];
 776
 777         if (!test_and_set_bit(RXRPC_CALL_EXPOSED, &call->flags)) {
 778                 /* Mark the call ID as being used.  If the callNumber counter
 779                  * exceeds ~2 billion, we kill the connection after its
 780                  * outstanding calls have finished so that the counter doesn't
 781                  * wrap.
 782                  */
 783                 chan->call_counter++;
 784                 if (chan->call_counter >= INT_MAX)
 785                         set_bit(RXRPC_CONN_DONT_REUSE, &conn->flags);
 786                 trace_rxrpc_client(conn, channel, rxrpc_client_exposed);
 787
 788                 spin_lock(&call->peer->lock);
 789                 hlist_add_head(&call->error_link, &call->peer->error_targets);
 790                 spin_unlock(&call->peer->lock);
 791         }
 792 }
 793
 794 /*
 795  * Set the reap timer.
 796  */
 797 static void rxrpc_set_client_reap_timer(struct rxrpc_net *rxnet)
 798 {
 799         if (!rxnet->kill_all_client_conns) {
 800                 unsigned long now = jiffies;
 801                 unsigned long reap_at = now + rxrpc_conn_idle_client_expiry;
 802
 803                 if (rxnet->live)
 804                         timer_reduce(&rxnet->client_conn_reap_timer, reap_at);
 805         }
 806 }
 807
 808 /*
 809  * Disconnect a client call.
 810  */
 811 void rxrpc_disconnect_client_call(struct rxrpc_bundle *bundle, struct rxrpc_call *call)
 812 {
 813         struct rxrpc_connection *conn;
 814         struct rxrpc_channel *chan = NULL;
 815         struct rxrpc_net *rxnet = bundle->local->rxnet;
 816         unsigned int channel;
 817         bool may_reuse;
 818         u32 cid;
 819
 820         _enter("c=%x", call->debug_id);
 821
 822         spin_lock(&bundle->channel_lock);
 823         set_bit(RXRPC_CALL_DISCONNECTED, &call->flags);
 824
 825         /* Calls that have never actually been assigned a channel can simply be
 826          * discarded.
 827          */
 828         conn = call->conn;
 829         if (!conn) {
 830                 _debug("call is waiting");
 831                 ASSERTCMP(call->call_id, ==, 0);
 832                 ASSERT(!test_bit(RXRPC_CALL_EXPOSED, &call->flags));
 833                 list_del_init(&call->chan_wait_link);
 834                 goto out;
 835         }
 836
 837         cid = call->cid;
 838         channel = cid & RXRPC_CHANNELMASK;
 839         chan = &conn->channels[channel];
 840         trace_rxrpc_client(conn, channel, rxrpc_client_chan_disconnect);
 841
 842         if (rcu_access_pointer(chan->call) != call) {
 843                 spin_unlock(&bundle->channel_lock);
 844                 BUG();
 845         }
 846
 847         may_reuse = rxrpc_may_reuse_conn(conn);
 848
 849         /* If a client call was exposed to the world, we save the result for
 850          * retransmission.
 851          *
 852          * We use a barrier here so that the call number and abort code can be
 853          * read without needing to take a lock.
 854          *
 855          * TODO: Make the incoming packet handler check this and handle
 856          * terminal retransmission without requiring access to the call.
 857          */
 858         if (test_bit(RXRPC_CALL_EXPOSED, &call->flags)) {
 859                 _debug("exposed %u,%u", call->call_id, call->abort_code);
 860                 __rxrpc_disconnect_call(conn, call);
 861
 862                 if (test_and_clear_bit(RXRPC_CONN_PROBING_FOR_UPGRADE, &conn->flags)) {
 863                         trace_rxrpc_client(conn, channel, rxrpc_client_to_active);
 864                         bundle->try_upgrade = false;
 865                         if (may_reuse)
 866                                 rxrpc_activate_channels_locked(bundle);
 867                 }
 868
 869         }
 870
 871         /* See if we can pass the channel directly to another call. */
 872         if (may_reuse && !list_empty(&bundle->waiting_calls)) {
 873                 trace_rxrpc_client(conn, channel, rxrpc_client_chan_pass);
 874                 rxrpc_activate_one_channel(conn, channel);
 875                 goto out;
 876         }
 877
 878         /* Schedule the final ACK to be transmitted in a short while so that it
 879          * can be skipped if we find a follow-on call.  The first DATA packet
 880          * of the follow on call will implicitly ACK this call.
 881          */
 882         if (call->completion == RXRPC_CALL_SUCCEEDED &&
 883             test_bit(RXRPC_CALL_EXPOSED, &call->flags)) {
 884                 unsigned long final_ack_at = jiffies + 2;
 885
 886                 WRITE_ONCE(chan->final_ack_at, final_ack_at);
 887                 smp_wmb(); /* vs rxrpc_process_delayed_final_acks() */
 888                 set_bit(RXRPC_CONN_FINAL_ACK_0 + channel, &conn->flags);
 889                 rxrpc_reduce_conn_timer(conn, final_ack_at);
 890         }
 891
 892         /* Deactivate the channel. */
 893         rcu_assign_pointer(chan->call, NULL);
 894         set_bit(conn->bundle_shift + channel, &conn->bundle->avail_chans);
 895         conn->act_chans &= ~(1 << channel);
 896
 897         /* If no channels remain active, then put the connection on the idle
 898          * list for a short while.  Give it a ref to stop it going away if it
 899          * becomes unbundled.
 900          */
 901         if (!conn->act_chans) {
 902                 trace_rxrpc_client(conn, channel, rxrpc_client_to_idle);
 903                 conn->idle_timestamp = jiffies;
 904
 905                 rxrpc_get_connection(conn, rxrpc_conn_get_idle);
 906                 spin_lock(&rxnet->client_conn_cache_lock);
 907                 list_move_tail(&conn->cache_link, &rxnet->idle_client_conns);
 908                 spin_unlock(&rxnet->client_conn_cache_lock);
 909
 910                 rxrpc_set_client_reap_timer(rxnet);
 911         }
 912
 913 out:
 914         spin_unlock(&bundle->channel_lock);
 915         _leave("");
 916         return;
 917 }
 918
 919 /*
 920  * Remove a connection from a bundle.
 921  */
 922 static void rxrpc_unbundle_conn(struct rxrpc_connection *conn)
 923 {
 924         struct rxrpc_bundle *bundle = conn->bundle;
 925         unsigned int bindex;
 926         bool need_drop = false;
 927         int i;
 928
 929         _enter("C=%x", conn->debug_id);
 930
 931         if (conn->flags & RXRPC_CONN_FINAL_ACK_MASK)
 932                 rxrpc_process_delayed_final_acks(conn, true);
 933
 934         spin_lock(&bundle->channel_lock);
 935         bindex = conn->bundle_shift / RXRPC_MAXCALLS;
 936         if (bundle->conns[bindex] == conn) {
 937                 _debug("clear slot %u", bindex);
 938                 bundle->conns[bindex] = NULL;
 939                 for (i = 0; i < RXRPC_MAXCALLS; i++)
 940                         clear_bit(conn->bundle_shift + i, &bundle->avail_chans);
 941                 need_drop = true;
 942         }
 943         spin_unlock(&bundle->channel_lock);
 944
 945         if (need_drop) {
 946                 rxrpc_deactivate_bundle(bundle);
 947                 rxrpc_put_connection(conn, rxrpc_conn_put_unbundle);
 948         }
 949 }
 950
 951 /*
 952  * Drop the active count on a bundle.
 953  */
 954 static void rxrpc_deactivate_bundle(struct rxrpc_bundle *bundle)
 955 {
 956         struct rxrpc_local *local = bundle->local;
 957         bool need_put = false;
 958
 959         if (atomic_dec_and_lock(&bundle->active, &local->client_bundles_lock)) {
 960                 if (!bundle->exclusive) {
 961                         _debug("erase bundle");
 962                         rb_erase(&bundle->local_node, &local->client_bundles);
 963                         need_put = true;
 964                 }
 965
 966                 spin_unlock(&local->client_bundles_lock);
 967                 if (need_put)
 968                         rxrpc_put_bundle(bundle, rxrpc_bundle_put_discard);
 969         }
 970 }
 971
 972 /*
 973  * Clean up a dead client connection.
 974  */
 975 void rxrpc_kill_client_conn(struct rxrpc_connection *conn)
 976 {
 977         struct rxrpc_local *local = conn->local;
 978         struct rxrpc_net *rxnet = local->rxnet;
 979
 980         _enter("C=%x", conn->debug_id);
 981
 982         trace_rxrpc_client(conn, -1, rxrpc_client_cleanup);
 983         atomic_dec(&rxnet->nr_client_conns);
 984
 985         rxrpc_put_client_connection_id(conn);
 986 }
 987
 988 /*
 989  * Discard expired client connections from the idle list.  Each conn in the
 990  * idle list has been exposed and holds an extra ref because of that.
 991  *
 992  * This may be called from conn setup or from a work item so cannot be
 993  * considered non-reentrant.
 994  */
 995 void rxrpc_discard_expired_client_conns(struct work_struct *work)
 996 {
 997         struct rxrpc_connection *conn;
 998         struct rxrpc_net *rxnet =
 999                 container_of(work, struct rxrpc_net, client_conn_reaper);
1000         unsigned long expiry, conn_expires_at, now;
1001         unsigned int nr_conns;
1002
1003         _enter("");
1004
1005         if (list_empty(&rxnet->idle_client_conns)) {
1006                 _leave(" [empty]");
1007                 return;
1008         }
1009
1010         /* Don't double up on the discarding */
1011         if (!mutex_trylock(&rxnet->client_conn_discard_lock)) {
1012                 _leave(" [already]");
1013                 return;
1014         }
1015
1016         /* We keep an estimate of what the number of conns ought to be after
1017          * we've discarded some so that we don't overdo the discarding.
1018          */
1019         nr_conns = atomic_read(&rxnet->nr_client_conns);
1020
1021 next:
1022         spin_lock(&rxnet->client_conn_cache_lock);
1023
1024         if (list_empty(&rxnet->idle_client_conns))
1025                 goto out;
1026
1027         conn = list_entry(rxnet->idle_client_conns.next,
1028                           struct rxrpc_connection, cache_link);
1029
1030         if (!rxnet->kill_all_client_conns) {
1031                 /* If the number of connections is over the reap limit, we
1032                  * expedite discard by reducing the expiry timeout.  We must,
1033                  * however, have at least a short grace period to be able to do
1034                  * final-ACK or ABORT retransmission.
1035                  */
1036                 expiry = rxrpc_conn_idle_client_expiry;
1037                 if (nr_conns > rxrpc_reap_client_connections)
1038                         expiry = rxrpc_conn_idle_client_fast_expiry;
1039                 if (conn->local->service_closed)
1040                         expiry = rxrpc_closed_conn_expiry * HZ;
1041
1042                 conn_expires_at = conn->idle_timestamp + expiry;
1043
1044                 now = READ_ONCE(jiffies);
1045                 if (time_after(conn_expires_at, now))
1046                         goto not_yet_expired;
1047         }
1048
1049         atomic_dec(&conn->active);
1050         trace_rxrpc_client(conn, -1, rxrpc_client_discard);
1051         list_del_init(&conn->cache_link);
1052
1053         spin_unlock(&rxnet->client_conn_cache_lock);
1054
1055         rxrpc_unbundle_conn(conn);
1056         /* Drop the ->cache_link ref */
1057         rxrpc_put_connection(conn, rxrpc_conn_put_discard_idle);
1058
1059         nr_conns--;
1060         goto next;
1061
1062 not_yet_expired:
1063         /* The connection at the front of the queue hasn't yet expired, so
1064          * schedule the work item for that point if we discarded something.
1065          *
1066          * We don't worry if the work item is already scheduled - it can look
1067          * after rescheduling itself at a later time.  We could cancel it, but
1068          * then things get messier.
1069          */
1070         _debug("not yet");
1071         if (!rxnet->kill_all_client_conns)
1072                 timer_reduce(&rxnet->client_conn_reap_timer, conn_expires_at);
1073
1074 out:
1075         spin_unlock(&rxnet->client_conn_cache_lock);
1076         mutex_unlock(&rxnet->client_conn_discard_lock);
1077         _leave("");
1078 }
1079
1080 /*
1081  * Preemptively destroy all the client connection records rather than waiting
1082  * for them to time out
1083  */
1084 void rxrpc_destroy_all_client_connections(struct rxrpc_net *rxnet)
1085 {
1086         _enter("");
1087
1088         spin_lock(&rxnet->client_conn_cache_lock);
1089         rxnet->kill_all_client_conns = true;
1090         spin_unlock(&rxnet->client_conn_cache_lock);
1091
1092         del_timer_sync(&rxnet->client_conn_reap_timer);
1093
1094         if (!rxrpc_queue_work(&rxnet->client_conn_reaper))
1095                 _debug("destroy: queue failed");
1096
1097         _leave("");
1098 }
1099
1100 /*
1101  * Clean up the client connections on a local endpoint.
1102  */
1103 void rxrpc_clean_up_local_conns(struct rxrpc_local *local)
1104 {
1105         struct rxrpc_connection *conn, *tmp;
1106         struct rxrpc_net *rxnet = local->rxnet;
1107         LIST_HEAD(graveyard);
1108
1109         _enter("");
1110
1111         spin_lock(&rxnet->client_conn_cache_lock);
1112
1113         list_for_each_entry_safe(conn, tmp, &rxnet->idle_client_conns,
1114                                  cache_link) {
1115                 if (conn->local == local) {
1116                         atomic_dec(&conn->active);
1117                         trace_rxrpc_client(conn, -1, rxrpc_client_discard);
1118                         list_move(&conn->cache_link, &graveyard);
1119                 }
1120         }
1121
1122         spin_unlock(&rxnet->client_conn_cache_lock);
1123
1124         while (!list_empty(&graveyard)) {
1125                 conn = list_entry(graveyard.next,
1126                                   struct rxrpc_connection, cache_link);
1127                 list_del_init(&conn->cache_link);
1128                 rxrpc_unbundle_conn(conn);
1129                 rxrpc_put_connection(conn, rxrpc_conn_put_local_dead);
1130         }
1131
1132         _leave(" [culled]");
1133 }