net/rxrpc/conn_client.c

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