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