net/rds/connection.c

   1 /*
   2  * Copyright (c) 2006 Oracle.  All rights reserved.
   3  *
   4  * This software is available to you under a choice of one of two
   5  * licenses.  You may choose to be licensed under the terms of the GNU
   6  * General Public License (GPL) Version 2, available from the file
   7  * COPYING in the main directory of this source tree, or the
   8  * OpenIB.org BSD license below:
   9  *
  10  *     Redistribution and use in source and binary forms, with or
  11  *     without modification, are permitted provided that the following
  12  *     conditions are met:
  13  *
  14  *      - Redistributions of source code must retain the above
  15  *        copyright notice, this list of conditions and the following
  16  *        disclaimer.
  17  *
  18  *      - Redistributions in binary form must reproduce the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer in the documentation and/or other materials
  21  *        provided with the distribution.
  22  *
  23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30  * SOFTWARE.
  31  *
  32  */
  33 #include <linux/kernel.h>
  34 #include <linux/list.h>
  35 #include <linux/slab.h>
  36 #include <linux/export.h>
  37 #include <net/inet_hashtables.h>
  38
  39 #include "rds.h"
  40 #include "loop.h"
  41
  42 #define RDS_CONNECTION_HASH_BITS 12
  43 #define RDS_CONNECTION_HASH_ENTRIES (1 << RDS_CONNECTION_HASH_BITS)
  44 #define RDS_CONNECTION_HASH_MASK (RDS_CONNECTION_HASH_ENTRIES - 1)
  45
  46 /* converting this to RCU is a chore for another day.. */
  47 static DEFINE_SPINLOCK(rds_conn_lock);
  48 static unsigned long rds_conn_count;
  49 static struct hlist_head rds_conn_hash[RDS_CONNECTION_HASH_ENTRIES];
  50 static struct kmem_cache *rds_conn_slab;
  51
  52 static struct hlist_head *rds_conn_bucket(__be32 laddr, __be32 faddr)
  53 {
  54         static u32 rds_hash_secret __read_mostly;
  55
  56         unsigned long hash;
  57
  58         net_get_random_once(&rds_hash_secret, sizeof(rds_hash_secret));
  59
  60         /* Pass NULL, don't need struct net for hash */
  61         hash = __inet_ehashfn(be32_to_cpu(laddr), 0,
  62                               be32_to_cpu(faddr), 0,
  63                               rds_hash_secret);
  64         return &rds_conn_hash[hash & RDS_CONNECTION_HASH_MASK];
  65 }
  66
  67 #define rds_conn_info_set(var, test, suffix) do {               \
  68         if (test)                                               \
  69                 var |= RDS_INFO_CONNECTION_FLAG_##suffix;       \
  70 } while (0)
  71
  72 /* rcu read lock must be held or the connection spinlock */
  73 static struct rds_connection *rds_conn_lookup(struct net *net,
  74                                               struct hlist_head *head,
  75                                               __be32 laddr, __be32 faddr,
  76                                               struct rds_transport *trans)
  77 {
  78         struct rds_connection *conn, *ret = NULL;
  79
  80         hlist_for_each_entry_rcu(conn, head, c_hash_node) {
  81                 if (conn->c_faddr == faddr && conn->c_laddr == laddr &&
  82                     conn->c_trans == trans && net == rds_conn_net(conn)) {
  83                         ret = conn;
  84                         break;
  85                 }
  86         }
  87         rdsdebug("returning conn %p for %pI4 -> %pI4\n", ret,
  88                  &laddr, &faddr);
  89         return ret;
  90 }
  91
  92 /*
  93  * This is called by transports as they're bringing down a connection.
  94  * It clears partial message state so that the transport can start sending
  95  * and receiving over this connection again in the future.  It is up to
  96  * the transport to have serialized this call with its send and recv.
  97  */
  98 static void rds_conn_path_reset(struct rds_conn_path *cp)
  99 {
 100         struct rds_connection *conn = cp->cp_conn;
 101
 102         rdsdebug("connection %pI4 to %pI4 reset\n",
 103           &conn->c_laddr, &conn->c_faddr);
 104
 105         rds_stats_inc(s_conn_reset);
 106         rds_send_path_reset(cp);
 107         cp->cp_flags = 0;
 108
 109         /* Do not clear next_rx_seq here, else we cannot distinguish
 110          * retransmitted packets from new packets, and will hand all
 111          * of them to the application. That is not consistent with the
 112          * reliability guarantees of RDS. */
 113 }
 114
 115 static void __rds_conn_path_init(struct rds_connection *conn,
 116                                  struct rds_conn_path *cp, bool is_outgoing)
 117 {
 118         spin_lock_init(&cp->cp_lock);
 119         cp->cp_next_tx_seq = 1;
 120         init_waitqueue_head(&cp->cp_waitq);
 121         INIT_LIST_HEAD(&cp->cp_send_queue);
 122         INIT_LIST_HEAD(&cp->cp_retrans);
 123
 124         cp->cp_conn = conn;
 125         atomic_set(&cp->cp_state, RDS_CONN_DOWN);
 126         cp->cp_send_gen = 0;
 127         /* cp_outgoing is per-path. So we can only set it here
 128          * for the single-path transports.
 129          */
 130         if (!conn->c_trans->t_mp_capable)
 131                 cp->cp_outgoing = (is_outgoing ? 1 : 0);
 132         cp->cp_reconnect_jiffies = 0;
 133         INIT_DELAYED_WORK(&cp->cp_send_w, rds_send_worker);
 134         INIT_DELAYED_WORK(&cp->cp_recv_w, rds_recv_worker);
 135         INIT_DELAYED_WORK(&cp->cp_conn_w, rds_connect_worker);
 136         INIT_WORK(&cp->cp_down_w, rds_shutdown_worker);
 137         mutex_init(&cp->cp_cm_lock);
 138         cp->cp_flags = 0;
 139 }
 140
 141 /*
 142  * There is only every one 'conn' for a given pair of addresses in the
 143  * system at a time.  They contain messages to be retransmitted and so
 144  * span the lifetime of the actual underlying transport connections.
 145  *
 146  * For now they are not garbage collected once they're created.  They
 147  * are torn down as the module is removed, if ever.
 148  */
 149 static struct rds_connection *__rds_conn_create(struct net *net,
 150                                                 __be32 laddr, __be32 faddr,
 151                                        struct rds_transport *trans, gfp_t gfp,
 152                                        int is_outgoing)
 153 {
 154         struct rds_connection *conn, *parent = NULL;
 155         struct hlist_head *head = rds_conn_bucket(laddr, faddr);
 156         struct rds_transport *loop_trans;
 157         unsigned long flags;
 158         int ret, i;
 159
 160         rcu_read_lock();
 161         conn = rds_conn_lookup(net, head, laddr, faddr, trans);
 162         if (conn && conn->c_loopback && conn->c_trans != &rds_loop_transport &&
 163             laddr == faddr && !is_outgoing) {
 164                 /* This is a looped back IB connection, and we're
 165                  * called by the code handling the incoming connect.
 166                  * We need a second connection object into which we
 167                  * can stick the other QP. */
 168                 parent = conn;
 169                 conn = parent->c_passive;
 170         }
 171         rcu_read_unlock();
 172         if (conn)
 173                 goto out;
 174
 175         conn = kmem_cache_zalloc(rds_conn_slab, gfp);
 176         if (!conn) {
 177                 conn = ERR_PTR(-ENOMEM);
 178                 goto out;
 179         }
 180
 181         INIT_HLIST_NODE(&conn->c_hash_node);
 182         conn->c_laddr = laddr;
 183         conn->c_faddr = faddr;
 184
 185         rds_conn_net_set(conn, net);
 186
 187         ret = rds_cong_get_maps(conn);
 188         if (ret) {
 189                 kmem_cache_free(rds_conn_slab, conn);
 190                 conn = ERR_PTR(ret);
 191                 goto out;
 192         }
 193
 194         /*
 195          * This is where a connection becomes loopback.  If *any* RDS sockets
 196          * can bind to the destination address then we'd rather the messages
 197          * flow through loopback rather than either transport.
 198          */
 199         loop_trans = rds_trans_get_preferred(net, faddr);
 200         if (loop_trans) {
 201                 rds_trans_put(loop_trans);
 202                 conn->c_loopback = 1;
 203                 if (is_outgoing && trans->t_prefer_loopback) {
 204                         /* "outgoing" connection - and the transport
 205                          * says it wants the connection handled by the
 206                          * loopback transport. This is what TCP does.
 207                          */
 208                         trans = &rds_loop_transport;
 209                 }
 210         }
 211
 212         conn->c_trans = trans;
 213
 214         init_waitqueue_head(&conn->c_hs_waitq);
 215         for (i = 0; i < RDS_MPATH_WORKERS; i++) {
 216                 __rds_conn_path_init(conn, &conn->c_path[i],
 217                                      is_outgoing);
 218                 conn->c_path[i].cp_index = i;
 219         }
 220         ret = trans->conn_alloc(conn, gfp);
 221         if (ret) {
 222                 kmem_cache_free(rds_conn_slab, conn);
 223                 conn = ERR_PTR(ret);
 224                 goto out;
 225         }
 226
 227         rdsdebug("allocated conn %p for %pI4 -> %pI4 over %s %s\n",
 228           conn, &laddr, &faddr,
 229           trans->t_name ? trans->t_name : "[unknown]",
 230           is_outgoing ? "(outgoing)" : "");
 231
 232         /*
 233          * Since we ran without holding the conn lock, someone could
 234          * have created the same conn (either normal or passive) in the
 235          * interim. We check while holding the lock. If we won, we complete
 236          * init and return our conn. If we lost, we rollback and return the
 237          * other one.
 238          */
 239         spin_lock_irqsave(&rds_conn_lock, flags);
 240         if (parent) {
 241                 /* Creating passive conn */
 242                 if (parent->c_passive) {
 243                         trans->conn_free(conn->c_path[0].cp_transport_data);
 244                         kmem_cache_free(rds_conn_slab, conn);
 245                         conn = parent->c_passive;
 246                 } else {
 247                         parent->c_passive = conn;
 248                         rds_cong_add_conn(conn);
 249                         rds_conn_count++;
 250                 }
 251         } else {
 252                 /* Creating normal conn */
 253                 struct rds_connection *found;
 254
 255                 found = rds_conn_lookup(net, head, laddr, faddr, trans);
 256                 if (found) {
 257                         struct rds_conn_path *cp;
 258                         int i;
 259
 260                         for (i = 0; i < RDS_MPATH_WORKERS; i++) {
 261                                 cp = &conn->c_path[i];
 262                                 /* The ->conn_alloc invocation may have
 263                                  * allocated resource for all paths, so all
 264                                  * of them may have to be freed here.
 265                                  */
 266                                 if (cp->cp_transport_data)
 267                                         trans->conn_free(cp->cp_transport_data);
 268                         }
 269                         kmem_cache_free(rds_conn_slab, conn);
 270                         conn = found;
 271                 } else {
 272                         hlist_add_head_rcu(&conn->c_hash_node, head);
 273                         rds_cong_add_conn(conn);
 274                         rds_conn_count++;
 275                 }
 276         }
 277         spin_unlock_irqrestore(&rds_conn_lock, flags);
 278
 279 out:
 280         return conn;
 281 }
 282
 283 struct rds_connection *rds_conn_create(struct net *net,
 284                                        __be32 laddr, __be32 faddr,
 285                                        struct rds_transport *trans, gfp_t gfp)
 286 {
 287         return __rds_conn_create(net, laddr, faddr, trans, gfp, 0);
 288 }
 289 EXPORT_SYMBOL_GPL(rds_conn_create);
 290
 291 struct rds_connection *rds_conn_create_outgoing(struct net *net,
 292                                                 __be32 laddr, __be32 faddr,
 293                                        struct rds_transport *trans, gfp_t gfp)
 294 {
 295         return __rds_conn_create(net, laddr, faddr, trans, gfp, 1);
 296 }
 297 EXPORT_SYMBOL_GPL(rds_conn_create_outgoing);
 298
 299 void rds_conn_shutdown(struct rds_conn_path *cp)
 300 {
 301         struct rds_connection *conn = cp->cp_conn;
 302
 303         /* shut it down unless it's down already */
 304         if (!rds_conn_path_transition(cp, RDS_CONN_DOWN, RDS_CONN_DOWN)) {
 305                 /*
 306                  * Quiesce the connection mgmt handlers before we start tearing
 307                  * things down. We don't hold the mutex for the entire
 308                  * duration of the shutdown operation, else we may be
 309                  * deadlocking with the CM handler. Instead, the CM event
 310                  * handler is supposed to check for state DISCONNECTING
 311                  */
 312                 mutex_lock(&cp->cp_cm_lock);
 313                 if (!rds_conn_path_transition(cp, RDS_CONN_UP,
 314                                               RDS_CONN_DISCONNECTING) &&
 315                     !rds_conn_path_transition(cp, RDS_CONN_ERROR,
 316                                               RDS_CONN_DISCONNECTING)) {
 317                         rds_conn_path_error(cp,
 318                                             "shutdown called in state %d\n",
 319                                             atomic_read(&cp->cp_state));
 320                         mutex_unlock(&cp->cp_cm_lock);
 321                         return;
 322                 }
 323                 mutex_unlock(&cp->cp_cm_lock);
 324
 325                 wait_event(cp->cp_waitq,
 326                            !test_bit(RDS_IN_XMIT, &cp->cp_flags));
 327                 wait_event(cp->cp_waitq,
 328                            !test_bit(RDS_RECV_REFILL, &cp->cp_flags));
 329
 330                 conn->c_trans->conn_path_shutdown(cp);
 331                 rds_conn_path_reset(cp);
 332
 333                 if (!rds_conn_path_transition(cp, RDS_CONN_DISCONNECTING,
 334                                               RDS_CONN_DOWN)) {
 335                         /* This can happen - eg when we're in the middle of tearing
 336                          * down the connection, and someone unloads the rds module.
 337                          * Quite reproduceable with loopback connections.
 338                          * Mostly harmless.
 339                          */
 340                         rds_conn_path_error(cp, "%s: failed to transition "
 341                                             "to state DOWN, current state "
 342                                             "is %d\n", __func__,
 343                                             atomic_read(&cp->cp_state));
 344                         return;
 345                 }
 346         }
 347
 348         /* Then reconnect if it's still live.
 349          * The passive side of an IB loopback connection is never added
 350          * to the conn hash, so we never trigger a reconnect on this
 351          * conn - the reconnect is always triggered by the active peer. */
 352         cancel_delayed_work_sync(&cp->cp_conn_w);
 353         rcu_read_lock();
 354         if (!hlist_unhashed(&conn->c_hash_node)) {
 355                 rcu_read_unlock();
 356                 rds_queue_reconnect(cp);
 357         } else {
 358                 rcu_read_unlock();
 359         }
 360 }
 361
 362 /* destroy a single rds_conn_path. rds_conn_destroy() iterates over
 363  * all paths using rds_conn_path_destroy()
 364  */
 365 static void rds_conn_path_destroy(struct rds_conn_path *cp)
 366 {
 367         struct rds_message *rm, *rtmp;
 368
 369         if (!cp->cp_transport_data)
 370                 return;
 371
 372         rds_conn_path_drop(cp);
 373         flush_work(&cp->cp_down_w);
 374
 375         /* make sure lingering queued work won't try to ref the conn */
 376         cancel_delayed_work_sync(&cp->cp_send_w);
 377         cancel_delayed_work_sync(&cp->cp_recv_w);
 378
 379         /* tear down queued messages */
 380         list_for_each_entry_safe(rm, rtmp,
 381                                  &cp->cp_send_queue,
 382                                  m_conn_item) {
 383                 list_del_init(&rm->m_conn_item);
 384                 BUG_ON(!list_empty(&rm->m_sock_item));
 385                 rds_message_put(rm);
 386         }
 387         if (cp->cp_xmit_rm)
 388                 rds_message_put(cp->cp_xmit_rm);
 389
 390         cp->cp_conn->c_trans->conn_free(cp->cp_transport_data);
 391 }
 392
 393 /*
 394  * Stop and free a connection.
 395  *
 396  * This can only be used in very limited circumstances.  It assumes that once
 397  * the conn has been shutdown that no one else is referencing the connection.
 398  * We can only ensure this in the rmmod path in the current code.
 399  */
 400 void rds_conn_destroy(struct rds_connection *conn)
 401 {
 402         unsigned long flags;
 403         int i;
 404         struct rds_conn_path *cp;
 405
 406         rdsdebug("freeing conn %p for %pI4 -> "
 407                  "%pI4\n", conn, &conn->c_laddr,
 408                  &conn->c_faddr);
 409
 410         /* Ensure conn will not be scheduled for reconnect */
 411         spin_lock_irq(&rds_conn_lock);
 412         hlist_del_init_rcu(&conn->c_hash_node);
 413         spin_unlock_irq(&rds_conn_lock);
 414         synchronize_rcu();
 415
 416         /* shut the connection down */
 417         for (i = 0; i < RDS_MPATH_WORKERS; i++) {
 418                 cp = &conn->c_path[i];
 419                 rds_conn_path_destroy(cp);
 420                 BUG_ON(!list_empty(&cp->cp_retrans));
 421         }
 422
 423         /*
 424          * The congestion maps aren't freed up here.  They're
 425          * freed by rds_cong_exit() after all the connections
 426          * have been freed.
 427          */
 428         rds_cong_remove_conn(conn);
 429
 430         kmem_cache_free(rds_conn_slab, conn);
 431
 432         spin_lock_irqsave(&rds_conn_lock, flags);
 433         rds_conn_count--;
 434         spin_unlock_irqrestore(&rds_conn_lock, flags);
 435 }
 436 EXPORT_SYMBOL_GPL(rds_conn_destroy);
 437
 438 static void rds_conn_message_info(struct socket *sock, unsigned int len,
 439                                   struct rds_info_iterator *iter,
 440                                   struct rds_info_lengths *lens,
 441                                   int want_send)
 442 {
 443         struct hlist_head *head;
 444         struct list_head *list;
 445         struct rds_connection *conn;
 446         struct rds_message *rm;
 447         unsigned int total = 0;
 448         unsigned long flags;
 449         size_t i;
 450         int j;
 451
 452         len /= sizeof(struct rds_info_message);
 453
 454         rcu_read_lock();
 455
 456         for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
 457              i++, head++) {
 458                 hlist_for_each_entry_rcu(conn, head, c_hash_node) {
 459                         struct rds_conn_path *cp;
 460
 461                         for (j = 0; j < RDS_MPATH_WORKERS; j++) {
 462                                 cp = &conn->c_path[j];
 463                                 if (want_send)
 464                                         list = &cp->cp_send_queue;
 465                                 else
 466                                         list = &cp->cp_retrans;
 467
 468                                 spin_lock_irqsave(&cp->cp_lock, flags);
 469
 470                                 /* XXX too lazy to maintain counts.. */
 471                                 list_for_each_entry(rm, list, m_conn_item) {
 472                                         total++;
 473                                         if (total <= len)
 474                                                 rds_inc_info_copy(&rm->m_inc,
 475                                                                   iter,
 476                                                                   conn->c_laddr,
 477                                                                   conn->c_faddr,
 478                                                                   0);
 479                                 }
 480
 481                                 spin_unlock_irqrestore(&cp->cp_lock, flags);
 482                                 if (!conn->c_trans->t_mp_capable)
 483                                         break;
 484                         }
 485                 }
 486         }
 487         rcu_read_unlock();
 488
 489         lens->nr = total;
 490         lens->each = sizeof(struct rds_info_message);
 491 }
 492
 493 static void rds_conn_message_info_send(struct socket *sock, unsigned int len,
 494                                        struct rds_info_iterator *iter,
 495                                        struct rds_info_lengths *lens)
 496 {
 497         rds_conn_message_info(sock, len, iter, lens, 1);
 498 }
 499
 500 static void rds_conn_message_info_retrans(struct socket *sock,
 501                                           unsigned int len,
 502                                           struct rds_info_iterator *iter,
 503                                           struct rds_info_lengths *lens)
 504 {
 505         rds_conn_message_info(sock, len, iter, lens, 0);
 506 }
 507
 508 void rds_for_each_conn_info(struct socket *sock, unsigned int len,
 509                           struct rds_info_iterator *iter,
 510                           struct rds_info_lengths *lens,
 511                           int (*visitor)(struct rds_connection *, void *),
 512                           size_t item_len)
 513 {
 514         uint64_t buffer[(item_len + 7) / 8];
 515         struct hlist_head *head;
 516         struct rds_connection *conn;
 517         size_t i;
 518
 519         rcu_read_lock();
 520
 521         lens->nr = 0;
 522         lens->each = item_len;
 523
 524         for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
 525              i++, head++) {
 526                 hlist_for_each_entry_rcu(conn, head, c_hash_node) {
 527
 528                         /* XXX no c_lock usage.. */
 529                         if (!visitor(conn, buffer))
 530                                 continue;
 531
 532                         /* We copy as much as we can fit in the buffer,
 533                          * but we count all items so that the caller
 534                          * can resize the buffer. */
 535                         if (len >= item_len) {
 536                                 rds_info_copy(iter, buffer, item_len);
 537                                 len -= item_len;
 538                         }
 539                         lens->nr++;
 540                 }
 541         }
 542         rcu_read_unlock();
 543 }
 544 EXPORT_SYMBOL_GPL(rds_for_each_conn_info);
 545
 546 void rds_walk_conn_path_info(struct socket *sock, unsigned int len,
 547                              struct rds_info_iterator *iter,
 548                              struct rds_info_lengths *lens,
 549                              int (*visitor)(struct rds_conn_path *, void *),
 550                              size_t item_len)
 551 {
 552         u64  buffer[(item_len + 7) / 8];
 553         struct hlist_head *head;
 554         struct rds_connection *conn;
 555         size_t i;
 556         int j;
 557
 558         rcu_read_lock();
 559
 560         lens->nr = 0;
 561         lens->each = item_len;
 562
 563         for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
 564              i++, head++) {
 565                 hlist_for_each_entry_rcu(conn, head, c_hash_node) {
 566                         struct rds_conn_path *cp;
 567
 568                         for (j = 0; j < RDS_MPATH_WORKERS; j++) {
 569                                 cp = &conn->c_path[j];
 570
 571                                 /* XXX no cp_lock usage.. */
 572                                 if (!visitor(cp, buffer))
 573                                         continue;
 574                                 if (!conn->c_trans->t_mp_capable)
 575                                         break;
 576                         }
 577
 578                         /* We copy as much as we can fit in the buffer,
 579                          * but we count all items so that the caller
 580                          * can resize the buffer.
 581                          */
 582                         if (len >= item_len) {
 583                                 rds_info_copy(iter, buffer, item_len);
 584                                 len -= item_len;
 585                         }
 586                         lens->nr++;
 587                 }
 588         }
 589         rcu_read_unlock();
 590 }
 591
 592 static int rds_conn_info_visitor(struct rds_conn_path *cp, void *buffer)
 593 {
 594         struct rds_info_connection *cinfo = buffer;
 595
 596         cinfo->next_tx_seq = cp->cp_next_tx_seq;
 597         cinfo->next_rx_seq = cp->cp_next_rx_seq;
 598         cinfo->laddr = cp->cp_conn->c_laddr;
 599         cinfo->faddr = cp->cp_conn->c_faddr;
 600         strncpy(cinfo->transport, cp->cp_conn->c_trans->t_name,
 601                 sizeof(cinfo->transport));
 602         cinfo->flags = 0;
 603
 604         rds_conn_info_set(cinfo->flags, test_bit(RDS_IN_XMIT, &cp->cp_flags),
 605                           SENDING);
 606         /* XXX Future: return the state rather than these funky bits */
 607         rds_conn_info_set(cinfo->flags,
 608                           atomic_read(&cp->cp_state) == RDS_CONN_CONNECTING,
 609                           CONNECTING);
 610         rds_conn_info_set(cinfo->flags,
 611                           atomic_read(&cp->cp_state) == RDS_CONN_UP,
 612                           CONNECTED);
 613         return 1;
 614 }
 615
 616 static void rds_conn_info(struct socket *sock, unsigned int len,
 617                           struct rds_info_iterator *iter,
 618                           struct rds_info_lengths *lens)
 619 {
 620         rds_walk_conn_path_info(sock, len, iter, lens,
 621                                 rds_conn_info_visitor,
 622                                 sizeof(struct rds_info_connection));
 623 }
 624
 625 int rds_conn_init(void)
 626 {
 627         rds_conn_slab = kmem_cache_create("rds_connection",
 628                                           sizeof(struct rds_connection),
 629                                           0, 0, NULL);
 630         if (!rds_conn_slab)
 631                 return -ENOMEM;
 632
 633         rds_info_register_func(RDS_INFO_CONNECTIONS, rds_conn_info);
 634         rds_info_register_func(RDS_INFO_SEND_MESSAGES,
 635                                rds_conn_message_info_send);
 636         rds_info_register_func(RDS_INFO_RETRANS_MESSAGES,
 637                                rds_conn_message_info_retrans);
 638
 639         return 0;
 640 }
 641
 642 void rds_conn_exit(void)
 643 {
 644         rds_loop_exit();
 645
 646         WARN_ON(!hlist_empty(rds_conn_hash));
 647
 648         kmem_cache_destroy(rds_conn_slab);
 649
 650         rds_info_deregister_func(RDS_INFO_CONNECTIONS, rds_conn_info);
 651         rds_info_deregister_func(RDS_INFO_SEND_MESSAGES,
 652                                  rds_conn_message_info_send);
 653         rds_info_deregister_func(RDS_INFO_RETRANS_MESSAGES,
 654                                  rds_conn_message_info_retrans);
 655 }
 656
 657 /*
 658  * Force a disconnect
 659  */
 660 void rds_conn_path_drop(struct rds_conn_path *cp)
 661 {
 662         atomic_set(&cp->cp_state, RDS_CONN_ERROR);
 663         queue_work(rds_wq, &cp->cp_down_w);
 664 }
 665 EXPORT_SYMBOL_GPL(rds_conn_path_drop);
 666
 667 void rds_conn_drop(struct rds_connection *conn)
 668 {
 669         WARN_ON(conn->c_trans->t_mp_capable);
 670         rds_conn_path_drop(&conn->c_path[0]);
 671 }
 672 EXPORT_SYMBOL_GPL(rds_conn_drop);
 673
 674 /*
 675  * If the connection is down, trigger a connect. We may have scheduled a
 676  * delayed reconnect however - in this case we should not interfere.
 677  */
 678 void rds_conn_path_connect_if_down(struct rds_conn_path *cp)
 679 {
 680         if (rds_conn_path_state(cp) == RDS_CONN_DOWN &&
 681             !test_and_set_bit(RDS_RECONNECT_PENDING, &cp->cp_flags))
 682                 queue_delayed_work(rds_wq, &cp->cp_conn_w, 0);
 683 }
 684
 685 void rds_conn_connect_if_down(struct rds_connection *conn)
 686 {
 687         WARN_ON(conn->c_trans->t_mp_capable);
 688         rds_conn_path_connect_if_down(&conn->c_path[0]);
 689 }
 690 EXPORT_SYMBOL_GPL(rds_conn_connect_if_down);
 691
 692 /*
 693  * An error occurred on the connection
 694  */
 695 void
 696 __rds_conn_error(struct rds_connection *conn, const char *fmt, ...)
 697 {
 698         va_list ap;
 699
 700         va_start(ap, fmt);
 701         vprintk(fmt, ap);
 702         va_end(ap);
 703
 704         rds_conn_drop(conn);
 705 }
 706
 707 void
 708 __rds_conn_path_error(struct rds_conn_path *cp, const char *fmt, ...)
 709 {
 710         va_list ap;
 711
 712         va_start(ap, fmt);
 713         vprintk(fmt, ap);
 714         va_end(ap);
 715
 716         rds_conn_path_drop(cp);
 717 }