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 <net/inet_hashtables.h>
  37
  38 #include "rds.h"
  39 #include "loop.h"
  40
  41 #define RDS_CONNECTION_HASH_BITS 12
  42 #define RDS_CONNECTION_HASH_ENTRIES (1 << RDS_CONNECTION_HASH_BITS)
  43 #define RDS_CONNECTION_HASH_MASK (RDS_CONNECTION_HASH_ENTRIES - 1)
  44
  45 /* converting this to RCU is a chore for another day.. */
  46 static DEFINE_SPINLOCK(rds_conn_lock);
  47 static unsigned long rds_conn_count;
  48 static struct hlist_head rds_conn_hash[RDS_CONNECTION_HASH_ENTRIES];
  49 static struct kmem_cache *rds_conn_slab;
  50
  51 static struct hlist_head *rds_conn_bucket(__be32 laddr, __be32 faddr)
  52 {
  53         /* Pass NULL, don't need struct net for hash */
  54         unsigned long hash = inet_ehashfn(NULL,
  55                                           be32_to_cpu(laddr), 0,
  56                                           be32_to_cpu(faddr), 0);
  57         return &rds_conn_hash[hash & RDS_CONNECTION_HASH_MASK];
  58 }
  59
  60 #define rds_conn_info_set(var, test, suffix) do {               \
  61         if (test)                                               \
  62                 var |= RDS_INFO_CONNECTION_FLAG_##suffix;       \
  63 } while (0)
  64
  65 /* rcu read lock must be held or the connection spinlock */
  66 static struct rds_connection *rds_conn_lookup(struct hlist_head *head,
  67                                               __be32 laddr, __be32 faddr,
  68                                               struct rds_transport *trans)
  69 {
  70         struct rds_connection *conn, *ret = NULL;
  71         struct hlist_node *pos;
  72
  73         hlist_for_each_entry_rcu(conn, pos, head, c_hash_node) {
  74                 if (conn->c_faddr == faddr && conn->c_laddr == laddr &&
  75                                 conn->c_trans == trans) {
  76                         ret = conn;
  77                         break;
  78                 }
  79         }
  80         rdsdebug("returning conn %p for %pI4 -> %pI4\n", ret,
  81                  &laddr, &faddr);
  82         return ret;
  83 }
  84
  85 /*
  86  * This is called by transports as they're bringing down a connection.
  87  * It clears partial message state so that the transport can start sending
  88  * and receiving over this connection again in the future.  It is up to
  89  * the transport to have serialized this call with its send and recv.
  90  */
  91 void rds_conn_reset(struct rds_connection *conn)
  92 {
  93         rdsdebug("connection %pI4 to %pI4 reset\n",
  94           &conn->c_laddr, &conn->c_faddr);
  95
  96         rds_stats_inc(s_conn_reset);
  97         rds_send_reset(conn);
  98         conn->c_flags = 0;
  99
 100         /* Do not clear next_rx_seq here, else we cannot distinguish
 101          * retransmitted packets from new packets, and will hand all
 102          * of them to the application. That is not consistent with the
 103          * reliability guarantees of RDS. */
 104 }
 105
 106 /*
 107  * There is only every one 'conn' for a given pair of addresses in the
 108  * system at a time.  They contain messages to be retransmitted and so
 109  * span the lifetime of the actual underlying transport connections.
 110  *
 111  * For now they are not garbage collected once they're created.  They
 112  * are torn down as the module is removed, if ever.
 113  */
 114 static struct rds_connection *__rds_conn_create(__be32 laddr, __be32 faddr,
 115                                        struct rds_transport *trans, gfp_t gfp,
 116                                        int is_outgoing)
 117 {
 118         struct rds_connection *conn, *parent = NULL;
 119         struct hlist_head *head = rds_conn_bucket(laddr, faddr);
 120         unsigned long flags;
 121         int ret;
 122
 123         rcu_read_lock();
 124         conn = rds_conn_lookup(head, laddr, faddr, trans);
 125         if (conn && conn->c_loopback && conn->c_trans != &rds_loop_transport &&
 126             !is_outgoing) {
 127                 /* This is a looped back IB connection, and we're
 128                  * called by the code handling the incoming connect.
 129                  * We need a second connection object into which we
 130                  * can stick the other QP. */
 131                 parent = conn;
 132                 conn = parent->c_passive;
 133         }
 134         rcu_read_unlock();
 135         if (conn)
 136                 goto out;
 137
 138         conn = kmem_cache_zalloc(rds_conn_slab, gfp);
 139         if (!conn) {
 140                 conn = ERR_PTR(-ENOMEM);
 141                 goto out;
 142         }
 143
 144         INIT_HLIST_NODE(&conn->c_hash_node);
 145         conn->c_laddr = laddr;
 146         conn->c_faddr = faddr;
 147         spin_lock_init(&conn->c_lock);
 148         conn->c_next_tx_seq = 1;
 149
 150         init_waitqueue_head(&conn->c_waitq);
 151         INIT_LIST_HEAD(&conn->c_send_queue);
 152         INIT_LIST_HEAD(&conn->c_retrans);
 153
 154         ret = rds_cong_get_maps(conn);
 155         if (ret) {
 156                 kmem_cache_free(rds_conn_slab, conn);
 157                 conn = ERR_PTR(ret);
 158                 goto out;
 159         }
 160
 161         /*
 162          * This is where a connection becomes loopback.  If *any* RDS sockets
 163          * can bind to the destination address then we'd rather the messages
 164          * flow through loopback rather than either transport.
 165          */
 166         if (rds_trans_get_preferred(faddr)) {
 167                 conn->c_loopback = 1;
 168                 if (is_outgoing && trans->t_prefer_loopback) {
 169                         /* "outgoing" connection - and the transport
 170                          * says it wants the connection handled by the
 171                          * loopback transport. This is what TCP does.
 172                          */
 173                         trans = &rds_loop_transport;
 174                 }
 175         }
 176
 177         conn->c_trans = trans;
 178
 179         ret = trans->conn_alloc(conn, gfp);
 180         if (ret) {
 181                 kmem_cache_free(rds_conn_slab, conn);
 182                 conn = ERR_PTR(ret);
 183                 goto out;
 184         }
 185
 186         atomic_set(&conn->c_state, RDS_CONN_DOWN);
 187         conn->c_reconnect_jiffies = 0;
 188         INIT_DELAYED_WORK(&conn->c_send_w, rds_send_worker);
 189         INIT_DELAYED_WORK(&conn->c_recv_w, rds_recv_worker);
 190         INIT_DELAYED_WORK(&conn->c_conn_w, rds_connect_worker);
 191         INIT_WORK(&conn->c_down_w, rds_shutdown_worker);
 192         mutex_init(&conn->c_cm_lock);
 193         conn->c_flags = 0;
 194
 195         rdsdebug("allocated conn %p for %pI4 -> %pI4 over %s %s\n",
 196           conn, &laddr, &faddr,
 197           trans->t_name ? trans->t_name : "[unknown]",
 198           is_outgoing ? "(outgoing)" : "");
 199
 200         /*
 201          * Since we ran without holding the conn lock, someone could
 202          * have created the same conn (either normal or passive) in the
 203          * interim. We check while holding the lock. If we won, we complete
 204          * init and return our conn. If we lost, we rollback and return the
 205          * other one.
 206          */
 207         spin_lock_irqsave(&rds_conn_lock, flags);
 208         if (parent) {
 209                 /* Creating passive conn */
 210                 if (parent->c_passive) {
 211                         trans->conn_free(conn->c_transport_data);
 212                         kmem_cache_free(rds_conn_slab, conn);
 213                         conn = parent->c_passive;
 214                 } else {
 215                         parent->c_passive = conn;
 216                         rds_cong_add_conn(conn);
 217                         rds_conn_count++;
 218                 }
 219         } else {
 220                 /* Creating normal conn */
 221                 struct rds_connection *found;
 222
 223                 found = rds_conn_lookup(head, laddr, faddr, trans);
 224                 if (found) {
 225                         trans->conn_free(conn->c_transport_data);
 226                         kmem_cache_free(rds_conn_slab, conn);
 227                         conn = found;
 228                 } else {
 229                         hlist_add_head_rcu(&conn->c_hash_node, head);
 230                         rds_cong_add_conn(conn);
 231                         rds_conn_count++;
 232                 }
 233         }
 234         spin_unlock_irqrestore(&rds_conn_lock, flags);
 235
 236 out:
 237         return conn;
 238 }
 239
 240 struct rds_connection *rds_conn_create(__be32 laddr, __be32 faddr,
 241                                        struct rds_transport *trans, gfp_t gfp)
 242 {
 243         return __rds_conn_create(laddr, faddr, trans, gfp, 0);
 244 }
 245 EXPORT_SYMBOL_GPL(rds_conn_create);
 246
 247 struct rds_connection *rds_conn_create_outgoing(__be32 laddr, __be32 faddr,
 248                                        struct rds_transport *trans, gfp_t gfp)
 249 {
 250         return __rds_conn_create(laddr, faddr, trans, gfp, 1);
 251 }
 252 EXPORT_SYMBOL_GPL(rds_conn_create_outgoing);
 253
 254 void rds_conn_shutdown(struct rds_connection *conn)
 255 {
 256         /* shut it down unless it's down already */
 257         if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_DOWN)) {
 258                 /*
 259                  * Quiesce the connection mgmt handlers before we start tearing
 260                  * things down. We don't hold the mutex for the entire
 261                  * duration of the shutdown operation, else we may be
 262                  * deadlocking with the CM handler. Instead, the CM event
 263                  * handler is supposed to check for state DISCONNECTING
 264                  */
 265                 mutex_lock(&conn->c_cm_lock);
 266                 if (!rds_conn_transition(conn, RDS_CONN_UP, RDS_CONN_DISCONNECTING)
 267                  && !rds_conn_transition(conn, RDS_CONN_ERROR, RDS_CONN_DISCONNECTING)) {
 268                         rds_conn_error(conn, "shutdown called in state %d\n",
 269                                         atomic_read(&conn->c_state));
 270                         mutex_unlock(&conn->c_cm_lock);
 271                         return;
 272                 }
 273                 mutex_unlock(&conn->c_cm_lock);
 274
 275                 wait_event(conn->c_waitq,
 276                            !test_bit(RDS_IN_XMIT, &conn->c_flags));
 277
 278                 conn->c_trans->conn_shutdown(conn);
 279                 rds_conn_reset(conn);
 280
 281                 if (!rds_conn_transition(conn, RDS_CONN_DISCONNECTING, RDS_CONN_DOWN)) {
 282                         /* This can happen - eg when we're in the middle of tearing
 283                          * down the connection, and someone unloads the rds module.
 284                          * Quite reproduceable with loopback connections.
 285                          * Mostly harmless.
 286                          */
 287                         rds_conn_error(conn,
 288                                 "%s: failed to transition to state DOWN, "
 289                                 "current state is %d\n",
 290                                 __func__,
 291                                 atomic_read(&conn->c_state));
 292                         return;
 293                 }
 294         }
 295
 296         /* Then reconnect if it's still live.
 297          * The passive side of an IB loopback connection is never added
 298          * to the conn hash, so we never trigger a reconnect on this
 299          * conn - the reconnect is always triggered by the active peer. */
 300         cancel_delayed_work_sync(&conn->c_conn_w);
 301         rcu_read_lock();
 302         if (!hlist_unhashed(&conn->c_hash_node)) {
 303                 rcu_read_unlock();
 304                 rds_queue_reconnect(conn);
 305         } else {
 306                 rcu_read_unlock();
 307         }
 308 }
 309
 310 /*
 311  * Stop and free a connection.
 312  */
 313 void rds_conn_destroy(struct rds_connection *conn)
 314 {
 315         struct rds_message *rm, *rtmp;
 316
 317         rdsdebug("freeing conn %p for %pI4 -> "
 318                  "%pI4\n", conn, &conn->c_laddr,
 319                  &conn->c_faddr);
 320
 321         /* Ensure conn will not be scheduled for reconnect */
 322         spin_lock_irq(&rds_conn_lock);
 323         hlist_del_init_rcu(&conn->c_hash_node);
 324         spin_unlock_irq(&rds_conn_lock);
 325
 326         synchronize_rcu();
 327
 328         rds_conn_shutdown(conn);
 329
 330         /* tear down queued messages */
 331         list_for_each_entry_safe(rm, rtmp,
 332                                  &conn->c_send_queue,
 333                                  m_conn_item) {
 334                 list_del_init(&rm->m_conn_item);
 335                 BUG_ON(!list_empty(&rm->m_sock_item));
 336                 rds_message_put(rm);
 337         }
 338         if (conn->c_xmit_rm)
 339                 rds_message_put(conn->c_xmit_rm);
 340
 341         conn->c_trans->conn_free(conn->c_transport_data);
 342
 343         /*
 344          * The congestion maps aren't freed up here.  They're
 345          * freed by rds_cong_exit() after all the connections
 346          * have been freed.
 347          */
 348         rds_cong_remove_conn(conn);
 349
 350         BUG_ON(!list_empty(&conn->c_retrans));
 351         kmem_cache_free(rds_conn_slab, conn);
 352
 353         rds_conn_count--;
 354 }
 355 EXPORT_SYMBOL_GPL(rds_conn_destroy);
 356
 357 static void rds_conn_message_info(struct socket *sock, unsigned int len,
 358                                   struct rds_info_iterator *iter,
 359                                   struct rds_info_lengths *lens,
 360                                   int want_send)
 361 {
 362         struct hlist_head *head;
 363         struct hlist_node *pos;
 364         struct list_head *list;
 365         struct rds_connection *conn;
 366         struct rds_message *rm;
 367         unsigned int total = 0;
 368         unsigned long flags;
 369         size_t i;
 370
 371         len /= sizeof(struct rds_info_message);
 372
 373         rcu_read_lock();
 374
 375         for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
 376              i++, head++) {
 377                 hlist_for_each_entry_rcu(conn, pos, head, c_hash_node) {
 378                         if (want_send)
 379                                 list = &conn->c_send_queue;
 380                         else
 381                                 list = &conn->c_retrans;
 382
 383                         spin_lock_irqsave(&conn->c_lock, flags);
 384
 385                         /* XXX too lazy to maintain counts.. */
 386                         list_for_each_entry(rm, list, m_conn_item) {
 387                                 total++;
 388                                 if (total <= len)
 389                                         rds_inc_info_copy(&rm->m_inc, iter,
 390                                                           conn->c_laddr,
 391                                                           conn->c_faddr, 0);
 392                         }
 393
 394                         spin_unlock_irqrestore(&conn->c_lock, flags);
 395                 }
 396         }
 397         rcu_read_unlock();
 398
 399         lens->nr = total;
 400         lens->each = sizeof(struct rds_info_message);
 401 }
 402
 403 static void rds_conn_message_info_send(struct socket *sock, unsigned int len,
 404                                        struct rds_info_iterator *iter,
 405                                        struct rds_info_lengths *lens)
 406 {
 407         rds_conn_message_info(sock, len, iter, lens, 1);
 408 }
 409
 410 static void rds_conn_message_info_retrans(struct socket *sock,
 411                                           unsigned int len,
 412                                           struct rds_info_iterator *iter,
 413                                           struct rds_info_lengths *lens)
 414 {
 415         rds_conn_message_info(sock, len, iter, lens, 0);
 416 }
 417
 418 void rds_for_each_conn_info(struct socket *sock, unsigned int len,
 419                           struct rds_info_iterator *iter,
 420                           struct rds_info_lengths *lens,
 421                           int (*visitor)(struct rds_connection *, void *),
 422                           size_t item_len)
 423 {
 424         uint64_t buffer[(item_len + 7) / 8];
 425         struct hlist_head *head;
 426         struct hlist_node *pos;
 427         struct rds_connection *conn;
 428         size_t i;
 429
 430         rcu_read_lock();
 431
 432         lens->nr = 0;
 433         lens->each = item_len;
 434
 435         for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
 436              i++, head++) {
 437                 hlist_for_each_entry_rcu(conn, pos, head, c_hash_node) {
 438
 439                         /* XXX no c_lock usage.. */
 440                         if (!visitor(conn, buffer))
 441                                 continue;
 442
 443                         /* We copy as much as we can fit in the buffer,
 444                          * but we count all items so that the caller
 445                          * can resize the buffer. */
 446                         if (len >= item_len) {
 447                                 rds_info_copy(iter, buffer, item_len);
 448                                 len -= item_len;
 449                         }
 450                         lens->nr++;
 451                 }
 452         }
 453         rcu_read_unlock();
 454 }
 455 EXPORT_SYMBOL_GPL(rds_for_each_conn_info);
 456
 457 static int rds_conn_info_visitor(struct rds_connection *conn,
 458                                   void *buffer)
 459 {
 460         struct rds_info_connection *cinfo = buffer;
 461
 462         cinfo->next_tx_seq = conn->c_next_tx_seq;
 463         cinfo->next_rx_seq = conn->c_next_rx_seq;
 464         cinfo->laddr = conn->c_laddr;
 465         cinfo->faddr = conn->c_faddr;
 466         strncpy(cinfo->transport, conn->c_trans->t_name,
 467                 sizeof(cinfo->transport));
 468         cinfo->flags = 0;
 469
 470         rds_conn_info_set(cinfo->flags, test_bit(RDS_IN_XMIT, &conn->c_flags),
 471                           SENDING);
 472         /* XXX Future: return the state rather than these funky bits */
 473         rds_conn_info_set(cinfo->flags,
 474                           atomic_read(&conn->c_state) == RDS_CONN_CONNECTING,
 475                           CONNECTING);
 476         rds_conn_info_set(cinfo->flags,
 477                           atomic_read(&conn->c_state) == RDS_CONN_UP,
 478                           CONNECTED);
 479         return 1;
 480 }
 481
 482 static void rds_conn_info(struct socket *sock, unsigned int len,
 483                           struct rds_info_iterator *iter,
 484                           struct rds_info_lengths *lens)
 485 {
 486         rds_for_each_conn_info(sock, len, iter, lens,
 487                                 rds_conn_info_visitor,
 488                                 sizeof(struct rds_info_connection));
 489 }
 490
 491 int __init rds_conn_init(void)
 492 {
 493         rds_conn_slab = kmem_cache_create("rds_connection",
 494                                           sizeof(struct rds_connection),
 495                                           0, 0, NULL);
 496         if (!rds_conn_slab)
 497                 return -ENOMEM;
 498
 499         rds_info_register_func(RDS_INFO_CONNECTIONS, rds_conn_info);
 500         rds_info_register_func(RDS_INFO_SEND_MESSAGES,
 501                                rds_conn_message_info_send);
 502         rds_info_register_func(RDS_INFO_RETRANS_MESSAGES,
 503                                rds_conn_message_info_retrans);
 504
 505         return 0;
 506 }
 507
 508 void rds_conn_exit(void)
 509 {
 510         rds_loop_exit();
 511
 512         WARN_ON(!hlist_empty(rds_conn_hash));
 513
 514         kmem_cache_destroy(rds_conn_slab);
 515
 516         rds_info_deregister_func(RDS_INFO_CONNECTIONS, rds_conn_info);
 517         rds_info_deregister_func(RDS_INFO_SEND_MESSAGES,
 518                                  rds_conn_message_info_send);
 519         rds_info_deregister_func(RDS_INFO_RETRANS_MESSAGES,
 520                                  rds_conn_message_info_retrans);
 521 }
 522
 523 /*
 524  * Force a disconnect
 525  */
 526 void rds_conn_drop(struct rds_connection *conn)
 527 {
 528         atomic_set(&conn->c_state, RDS_CONN_ERROR);
 529         queue_work(rds_wq, &conn->c_down_w);
 530 }
 531 EXPORT_SYMBOL_GPL(rds_conn_drop);
 532
 533 /*
 534  * If the connection is down, trigger a connect. We may have scheduled a
 535  * delayed reconnect however - in this case we should not interfere.
 536  */
 537 void rds_conn_connect_if_down(struct rds_connection *conn)
 538 {
 539         if (rds_conn_state(conn) == RDS_CONN_DOWN &&
 540             !test_and_set_bit(RDS_RECONNECT_PENDING, &conn->c_flags))
 541                 queue_delayed_work(rds_wq, &conn->c_conn_w, 0);
 542 }
 543 EXPORT_SYMBOL_GPL(rds_conn_connect_if_down);
 544
 545 /*
 546  * An error occurred on the connection
 547  */
 548 void
 549 __rds_conn_error(struct rds_connection *conn, const char *fmt, ...)
 550 {
 551         va_list ap;
 552
 553         va_start(ap, fmt);
 554         vprintk(fmt, ap);
 555         va_end(ap);
 556
 557         rds_conn_drop(conn);
 558 }