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_reset(struct rds_connection *conn)
  99 {
 100         rdsdebug("connection %pI4 to %pI4 reset\n",
 101           &conn->c_laddr, &conn->c_faddr);
 102
 103         rds_stats_inc(s_conn_reset);
 104         rds_send_reset(conn);
 105         conn->c_flags = 0;
 106
 107         /* Do not clear next_rx_seq here, else we cannot distinguish
 108          * retransmitted packets from new packets, and will hand all
 109          * of them to the application. That is not consistent with the
 110          * reliability guarantees of RDS. */
 111 }
 112
 113 /*
 114  * There is only every one 'conn' for a given pair of addresses in the
 115  * system at a time.  They contain messages to be retransmitted and so
 116  * span the lifetime of the actual underlying transport connections.
 117  *
 118  * For now they are not garbage collected once they're created.  They
 119  * are torn down as the module is removed, if ever.
 120  */
 121 static struct rds_connection *__rds_conn_create(struct net *net,
 122                                                 __be32 laddr, __be32 faddr,
 123                                        struct rds_transport *trans, gfp_t gfp,
 124                                        int is_outgoing)
 125 {
 126         struct rds_connection *conn, *parent = NULL;
 127         struct hlist_head *head = rds_conn_bucket(laddr, faddr);
 128         struct rds_transport *loop_trans;
 129         unsigned long flags;
 130         int ret;
 131
 132         rcu_read_lock();
 133         conn = rds_conn_lookup(net, head, laddr, faddr, trans);
 134         if (conn && conn->c_loopback && conn->c_trans != &rds_loop_transport &&
 135             laddr == faddr && !is_outgoing) {
 136                 /* This is a looped back IB connection, and we're
 137                  * called by the code handling the incoming connect.
 138                  * We need a second connection object into which we
 139                  * can stick the other QP. */
 140                 parent = conn;
 141                 conn = parent->c_passive;
 142         }
 143         rcu_read_unlock();
 144         if (conn)
 145                 goto out;
 146
 147         conn = kmem_cache_zalloc(rds_conn_slab, gfp);
 148         if (!conn) {
 149                 conn = ERR_PTR(-ENOMEM);
 150                 goto out;
 151         }
 152
 153         INIT_HLIST_NODE(&conn->c_hash_node);
 154         conn->c_laddr = laddr;
 155         conn->c_faddr = faddr;
 156         spin_lock_init(&conn->c_lock);
 157         conn->c_next_tx_seq = 1;
 158         rds_conn_net_set(conn, net);
 159
 160         init_waitqueue_head(&conn->c_waitq);
 161         INIT_LIST_HEAD(&conn->c_send_queue);
 162         INIT_LIST_HEAD(&conn->c_retrans);
 163
 164         ret = rds_cong_get_maps(conn);
 165         if (ret) {
 166                 kmem_cache_free(rds_conn_slab, conn);
 167                 conn = ERR_PTR(ret);
 168                 goto out;
 169         }
 170
 171         /*
 172          * This is where a connection becomes loopback.  If *any* RDS sockets
 173          * can bind to the destination address then we'd rather the messages
 174          * flow through loopback rather than either transport.
 175          */
 176         loop_trans = rds_trans_get_preferred(net, faddr);
 177         if (loop_trans) {
 178                 rds_trans_put(loop_trans);
 179                 conn->c_loopback = 1;
 180                 if (is_outgoing && trans->t_prefer_loopback) {
 181                         /* "outgoing" connection - and the transport
 182                          * says it wants the connection handled by the
 183                          * loopback transport. This is what TCP does.
 184                          */
 185                         trans = &rds_loop_transport;
 186                 }
 187         }
 188
 189         if (trans == NULL) {
 190                 kmem_cache_free(rds_conn_slab, conn);
 191                 conn = ERR_PTR(-ENODEV);
 192                 goto out;
 193         }
 194
 195         conn->c_trans = trans;
 196
 197         ret = trans->conn_alloc(conn, gfp);
 198         if (ret) {
 199                 kmem_cache_free(rds_conn_slab, conn);
 200                 conn = ERR_PTR(ret);
 201                 goto out;
 202         }
 203
 204         atomic_set(&conn->c_state, RDS_CONN_DOWN);
 205         conn->c_send_gen = 0;
 206         conn->c_outgoing = (is_outgoing ? 1 : 0);
 207         conn->c_reconnect_jiffies = 0;
 208         INIT_DELAYED_WORK(&conn->c_send_w, rds_send_worker);
 209         INIT_DELAYED_WORK(&conn->c_recv_w, rds_recv_worker);
 210         INIT_DELAYED_WORK(&conn->c_conn_w, rds_connect_worker);
 211         INIT_WORK(&conn->c_down_w, rds_shutdown_worker);
 212         mutex_init(&conn->c_cm_lock);
 213         conn->c_flags = 0;
 214
 215         rdsdebug("allocated conn %p for %pI4 -> %pI4 over %s %s\n",
 216           conn, &laddr, &faddr,
 217           trans->t_name ? trans->t_name : "[unknown]",
 218           is_outgoing ? "(outgoing)" : "");
 219
 220         /*
 221          * Since we ran without holding the conn lock, someone could
 222          * have created the same conn (either normal or passive) in the
 223          * interim. We check while holding the lock. If we won, we complete
 224          * init and return our conn. If we lost, we rollback and return the
 225          * other one.
 226          */
 227         spin_lock_irqsave(&rds_conn_lock, flags);
 228         if (parent) {
 229                 /* Creating passive conn */
 230                 if (parent->c_passive) {
 231                         trans->conn_free(conn->c_transport_data);
 232                         kmem_cache_free(rds_conn_slab, conn);
 233                         conn = parent->c_passive;
 234                 } else {
 235                         parent->c_passive = conn;
 236                         rds_cong_add_conn(conn);
 237                         rds_conn_count++;
 238                 }
 239         } else {
 240                 /* Creating normal conn */
 241                 struct rds_connection *found;
 242
 243                 found = rds_conn_lookup(net, head, laddr, faddr, trans);
 244                 if (found) {
 245                         trans->conn_free(conn->c_transport_data);
 246                         kmem_cache_free(rds_conn_slab, conn);
 247                         conn = found;
 248                 } else {
 249                         hlist_add_head_rcu(&conn->c_hash_node, head);
 250                         rds_cong_add_conn(conn);
 251                         rds_conn_count++;
 252                 }
 253         }
 254         spin_unlock_irqrestore(&rds_conn_lock, flags);
 255
 256 out:
 257         return conn;
 258 }
 259
 260 struct rds_connection *rds_conn_create(struct net *net,
 261                                        __be32 laddr, __be32 faddr,
 262                                        struct rds_transport *trans, gfp_t gfp)
 263 {
 264         return __rds_conn_create(net, laddr, faddr, trans, gfp, 0);
 265 }
 266 EXPORT_SYMBOL_GPL(rds_conn_create);
 267
 268 struct rds_connection *rds_conn_create_outgoing(struct net *net,
 269                                                 __be32 laddr, __be32 faddr,
 270                                        struct rds_transport *trans, gfp_t gfp)
 271 {
 272         return __rds_conn_create(net, laddr, faddr, trans, gfp, 1);
 273 }
 274 EXPORT_SYMBOL_GPL(rds_conn_create_outgoing);
 275
 276 void rds_conn_shutdown(struct rds_connection *conn)
 277 {
 278         /* shut it down unless it's down already */
 279         if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_DOWN)) {
 280                 /*
 281                  * Quiesce the connection mgmt handlers before we start tearing
 282                  * things down. We don't hold the mutex for the entire
 283                  * duration of the shutdown operation, else we may be
 284                  * deadlocking with the CM handler. Instead, the CM event
 285                  * handler is supposed to check for state DISCONNECTING
 286                  */
 287                 mutex_lock(&conn->c_cm_lock);
 288                 if (!rds_conn_transition(conn, RDS_CONN_UP, RDS_CONN_DISCONNECTING)
 289                  && !rds_conn_transition(conn, RDS_CONN_ERROR, RDS_CONN_DISCONNECTING)) {
 290                         rds_conn_error(conn, "shutdown called in state %d\n",
 291                                         atomic_read(&conn->c_state));
 292                         mutex_unlock(&conn->c_cm_lock);
 293                         return;
 294                 }
 295                 mutex_unlock(&conn->c_cm_lock);
 296
 297                 wait_event(conn->c_waitq,
 298                            !test_bit(RDS_IN_XMIT, &conn->c_flags));
 299                 wait_event(conn->c_waitq,
 300                            !test_bit(RDS_RECV_REFILL, &conn->c_flags));
 301
 302                 conn->c_trans->conn_shutdown(conn);
 303                 rds_conn_reset(conn);
 304
 305                 if (!rds_conn_transition(conn, RDS_CONN_DISCONNECTING, RDS_CONN_DOWN)) {
 306                         /* This can happen - eg when we're in the middle of tearing
 307                          * down the connection, and someone unloads the rds module.
 308                          * Quite reproduceable with loopback connections.
 309                          * Mostly harmless.
 310                          */
 311                         rds_conn_error(conn,
 312                                 "%s: failed to transition to state DOWN, "
 313                                 "current state is %d\n",
 314                                 __func__,
 315                                 atomic_read(&conn->c_state));
 316                         return;
 317                 }
 318         }
 319
 320         /* Then reconnect if it's still live.
 321          * The passive side of an IB loopback connection is never added
 322          * to the conn hash, so we never trigger a reconnect on this
 323          * conn - the reconnect is always triggered by the active peer. */
 324         cancel_delayed_work_sync(&conn->c_conn_w);
 325         rcu_read_lock();
 326         if (!hlist_unhashed(&conn->c_hash_node)) {
 327                 rcu_read_unlock();
 328                 if (conn->c_trans->t_type != RDS_TRANS_TCP ||
 329                     conn->c_outgoing == 1)
 330                         rds_queue_reconnect(conn);
 331         } else {
 332                 rcu_read_unlock();
 333         }
 334 }
 335
 336 /*
 337  * Stop and free a connection.
 338  *
 339  * This can only be used in very limited circumstances.  It assumes that once
 340  * the conn has been shutdown that no one else is referencing the connection.
 341  * We can only ensure this in the rmmod path in the current code.
 342  */
 343 void rds_conn_destroy(struct rds_connection *conn)
 344 {
 345         struct rds_message *rm, *rtmp;
 346         unsigned long flags;
 347
 348         rdsdebug("freeing conn %p for %pI4 -> "
 349                  "%pI4\n", conn, &conn->c_laddr,
 350                  &conn->c_faddr);
 351
 352         /* Ensure conn will not be scheduled for reconnect */
 353         spin_lock_irq(&rds_conn_lock);
 354         hlist_del_init_rcu(&conn->c_hash_node);
 355         spin_unlock_irq(&rds_conn_lock);
 356         synchronize_rcu();
 357
 358         /* shut the connection down */
 359         rds_conn_drop(conn);
 360         flush_work(&conn->c_down_w);
 361
 362         /* make sure lingering queued work won't try to ref the conn */
 363         cancel_delayed_work_sync(&conn->c_send_w);
 364         cancel_delayed_work_sync(&conn->c_recv_w);
 365
 366         /* tear down queued messages */
 367         list_for_each_entry_safe(rm, rtmp,
 368                                  &conn->c_send_queue,
 369                                  m_conn_item) {
 370                 list_del_init(&rm->m_conn_item);
 371                 BUG_ON(!list_empty(&rm->m_sock_item));
 372                 rds_message_put(rm);
 373         }
 374         if (conn->c_xmit_rm)
 375                 rds_message_put(conn->c_xmit_rm);
 376
 377         conn->c_trans->conn_free(conn->c_transport_data);
 378
 379         /*
 380          * The congestion maps aren't freed up here.  They're
 381          * freed by rds_cong_exit() after all the connections
 382          * have been freed.
 383          */
 384         rds_cong_remove_conn(conn);
 385
 386         BUG_ON(!list_empty(&conn->c_retrans));
 387         kmem_cache_free(rds_conn_slab, conn);
 388
 389         spin_lock_irqsave(&rds_conn_lock, flags);
 390         rds_conn_count--;
 391         spin_unlock_irqrestore(&rds_conn_lock, flags);
 392 }
 393 EXPORT_SYMBOL_GPL(rds_conn_destroy);
 394
 395 static void rds_conn_message_info(struct socket *sock, unsigned int len,
 396                                   struct rds_info_iterator *iter,
 397                                   struct rds_info_lengths *lens,
 398                                   int want_send)
 399 {
 400         struct hlist_head *head;
 401         struct list_head *list;
 402         struct rds_connection *conn;
 403         struct rds_message *rm;
 404         unsigned int total = 0;
 405         unsigned long flags;
 406         size_t i;
 407
 408         len /= sizeof(struct rds_info_message);
 409
 410         rcu_read_lock();
 411
 412         for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
 413              i++, head++) {
 414                 hlist_for_each_entry_rcu(conn, head, c_hash_node) {
 415                         if (want_send)
 416                                 list = &conn->c_send_queue;
 417                         else
 418                                 list = &conn->c_retrans;
 419
 420                         spin_lock_irqsave(&conn->c_lock, flags);
 421
 422                         /* XXX too lazy to maintain counts.. */
 423                         list_for_each_entry(rm, list, m_conn_item) {
 424                                 total++;
 425                                 if (total <= len)
 426                                         rds_inc_info_copy(&rm->m_inc, iter,
 427                                                           conn->c_laddr,
 428                                                           conn->c_faddr, 0);
 429                         }
 430
 431                         spin_unlock_irqrestore(&conn->c_lock, flags);
 432                 }
 433         }
 434         rcu_read_unlock();
 435
 436         lens->nr = total;
 437         lens->each = sizeof(struct rds_info_message);
 438 }
 439
 440 static void rds_conn_message_info_send(struct socket *sock, unsigned int len,
 441                                        struct rds_info_iterator *iter,
 442                                        struct rds_info_lengths *lens)
 443 {
 444         rds_conn_message_info(sock, len, iter, lens, 1);
 445 }
 446
 447 static void rds_conn_message_info_retrans(struct socket *sock,
 448                                           unsigned int len,
 449                                           struct rds_info_iterator *iter,
 450                                           struct rds_info_lengths *lens)
 451 {
 452         rds_conn_message_info(sock, len, iter, lens, 0);
 453 }
 454
 455 void rds_for_each_conn_info(struct socket *sock, unsigned int len,
 456                           struct rds_info_iterator *iter,
 457                           struct rds_info_lengths *lens,
 458                           int (*visitor)(struct rds_connection *, void *),
 459                           size_t item_len)
 460 {
 461         uint64_t buffer[(item_len + 7) / 8];
 462         struct hlist_head *head;
 463         struct rds_connection *conn;
 464         size_t i;
 465
 466         rcu_read_lock();
 467
 468         lens->nr = 0;
 469         lens->each = item_len;
 470
 471         for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
 472              i++, head++) {
 473                 hlist_for_each_entry_rcu(conn, head, c_hash_node) {
 474
 475                         /* XXX no c_lock usage.. */
 476                         if (!visitor(conn, buffer))
 477                                 continue;
 478
 479                         /* We copy as much as we can fit in the buffer,
 480                          * but we count all items so that the caller
 481                          * can resize the buffer. */
 482                         if (len >= item_len) {
 483                                 rds_info_copy(iter, buffer, item_len);
 484                                 len -= item_len;
 485                         }
 486                         lens->nr++;
 487                 }
 488         }
 489         rcu_read_unlock();
 490 }
 491 EXPORT_SYMBOL_GPL(rds_for_each_conn_info);
 492
 493 static int rds_conn_info_visitor(struct rds_connection *conn,
 494                                   void *buffer)
 495 {
 496         struct rds_info_connection *cinfo = buffer;
 497
 498         cinfo->next_tx_seq = conn->c_next_tx_seq;
 499         cinfo->next_rx_seq = conn->c_next_rx_seq;
 500         cinfo->laddr = conn->c_laddr;
 501         cinfo->faddr = conn->c_faddr;
 502         strncpy(cinfo->transport, conn->c_trans->t_name,
 503                 sizeof(cinfo->transport));
 504         cinfo->flags = 0;
 505
 506         rds_conn_info_set(cinfo->flags, test_bit(RDS_IN_XMIT, &conn->c_flags),
 507                           SENDING);
 508         /* XXX Future: return the state rather than these funky bits */
 509         rds_conn_info_set(cinfo->flags,
 510                           atomic_read(&conn->c_state) == RDS_CONN_CONNECTING,
 511                           CONNECTING);
 512         rds_conn_info_set(cinfo->flags,
 513                           atomic_read(&conn->c_state) == RDS_CONN_UP,
 514                           CONNECTED);
 515         return 1;
 516 }
 517
 518 static void rds_conn_info(struct socket *sock, unsigned int len,
 519                           struct rds_info_iterator *iter,
 520                           struct rds_info_lengths *lens)
 521 {
 522         rds_for_each_conn_info(sock, len, iter, lens,
 523                                 rds_conn_info_visitor,
 524                                 sizeof(struct rds_info_connection));
 525 }
 526
 527 int rds_conn_init(void)
 528 {
 529         rds_conn_slab = kmem_cache_create("rds_connection",
 530                                           sizeof(struct rds_connection),
 531                                           0, 0, NULL);
 532         if (!rds_conn_slab)
 533                 return -ENOMEM;
 534
 535         rds_info_register_func(RDS_INFO_CONNECTIONS, rds_conn_info);
 536         rds_info_register_func(RDS_INFO_SEND_MESSAGES,
 537                                rds_conn_message_info_send);
 538         rds_info_register_func(RDS_INFO_RETRANS_MESSAGES,
 539                                rds_conn_message_info_retrans);
 540
 541         return 0;
 542 }
 543
 544 void rds_conn_exit(void)
 545 {
 546         rds_loop_exit();
 547
 548         WARN_ON(!hlist_empty(rds_conn_hash));
 549
 550         kmem_cache_destroy(rds_conn_slab);
 551
 552         rds_info_deregister_func(RDS_INFO_CONNECTIONS, rds_conn_info);
 553         rds_info_deregister_func(RDS_INFO_SEND_MESSAGES,
 554                                  rds_conn_message_info_send);
 555         rds_info_deregister_func(RDS_INFO_RETRANS_MESSAGES,
 556                                  rds_conn_message_info_retrans);
 557 }
 558
 559 /*
 560  * Force a disconnect
 561  */
 562 void rds_conn_drop(struct rds_connection *conn)
 563 {
 564         atomic_set(&conn->c_state, RDS_CONN_ERROR);
 565         queue_work(rds_wq, &conn->c_down_w);
 566 }
 567 EXPORT_SYMBOL_GPL(rds_conn_drop);
 568
 569 /*
 570  * If the connection is down, trigger a connect. We may have scheduled a
 571  * delayed reconnect however - in this case we should not interfere.
 572  */
 573 void rds_conn_connect_if_down(struct rds_connection *conn)
 574 {
 575         if (rds_conn_state(conn) == RDS_CONN_DOWN &&
 576             !test_and_set_bit(RDS_RECONNECT_PENDING, &conn->c_flags))
 577                 queue_delayed_work(rds_wq, &conn->c_conn_w, 0);
 578 }
 579 EXPORT_SYMBOL_GPL(rds_conn_connect_if_down);
 580
 581 /*
 582  * An error occurred on the connection
 583  */
 584 void
 585 __rds_conn_error(struct rds_connection *conn, const char *fmt, ...)
 586 {
 587         va_list ap;
 588
 589         va_start(ap, fmt);
 590         vprintk(fmt, ap);
 591         va_end(ap);
 592
 593         rds_conn_drop(conn);
 594 }