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         /* Pass NULL, don't need struct net for hash */
  55         unsigned long hash = inet_ehashfn(NULL,
  56                                           be32_to_cpu(laddr), 0,
  57                                           be32_to_cpu(faddr), 0);
  58         return &rds_conn_hash[hash & RDS_CONNECTION_HASH_MASK];
  59 }
  60
  61 #define rds_conn_info_set(var, test, suffix) do {               \
  62         if (test)                                               \
  63                 var |= RDS_INFO_CONNECTION_FLAG_##suffix;       \
  64 } while (0)
  65
  66 /* rcu read lock must be held or the connection spinlock */
  67 static struct rds_connection *rds_conn_lookup(struct hlist_head *head,
  68                                               __be32 laddr, __be32 faddr,
  69                                               struct rds_transport *trans)
  70 {
  71         struct rds_connection *conn, *ret = NULL;
  72         struct hlist_node *pos;
  73
  74         hlist_for_each_entry_rcu(conn, pos, head, c_hash_node) {
  75                 if (conn->c_faddr == faddr && conn->c_laddr == laddr &&
  76                                 conn->c_trans == trans) {
  77                         ret = conn;
  78                         break;
  79                 }
  80         }
  81         rdsdebug("returning conn %p for %pI4 -> %pI4\n", ret,
  82                  &laddr, &faddr);
  83         return ret;
  84 }
  85
  86 /*
  87  * This is called by transports as they're bringing down a connection.
  88  * It clears partial message state so that the transport can start sending
  89  * and receiving over this connection again in the future.  It is up to
  90  * the transport to have serialized this call with its send and recv.
  91  */
  92 static void rds_conn_reset(struct rds_connection *conn)
  93 {
  94         rdsdebug("connection %pI4 to %pI4 reset\n",
  95           &conn->c_laddr, &conn->c_faddr);
  96
  97         rds_stats_inc(s_conn_reset);
  98         rds_send_reset(conn);
  99         conn->c_flags = 0;
 100
 101         /* Do not clear next_rx_seq here, else we cannot distinguish
 102          * retransmitted packets from new packets, and will hand all
 103          * of them to the application. That is not consistent with the
 104          * reliability guarantees of RDS. */
 105 }
 106
 107 /*
 108  * There is only every one 'conn' for a given pair of addresses in the
 109  * system at a time.  They contain messages to be retransmitted and so
 110  * span the lifetime of the actual underlying transport connections.
 111  *
 112  * For now they are not garbage collected once they're created.  They
 113  * are torn down as the module is removed, if ever.
 114  */
 115 static struct rds_connection *__rds_conn_create(__be32 laddr, __be32 faddr,
 116                                        struct rds_transport *trans, gfp_t gfp,
 117                                        int is_outgoing)
 118 {
 119         struct rds_connection *conn, *parent = NULL;
 120         struct hlist_head *head = rds_conn_bucket(laddr, faddr);
 121         struct rds_transport *loop_trans;
 122         unsigned long flags;
 123         int ret;
 124
 125         rcu_read_lock();
 126         conn = rds_conn_lookup(head, laddr, faddr, trans);
 127         if (conn && conn->c_loopback && conn->c_trans != &rds_loop_transport &&
 128             !is_outgoing) {
 129                 /* This is a looped back IB connection, and we're
 130                  * called by the code handling the incoming connect.
 131                  * We need a second connection object into which we
 132                  * can stick the other QP. */
 133                 parent = conn;
 134                 conn = parent->c_passive;
 135         }
 136         rcu_read_unlock();
 137         if (conn)
 138                 goto out;
 139
 140         conn = kmem_cache_zalloc(rds_conn_slab, gfp);
 141         if (!conn) {
 142                 conn = ERR_PTR(-ENOMEM);
 143                 goto out;
 144         }
 145
 146         INIT_HLIST_NODE(&conn->c_hash_node);
 147         conn->c_laddr = laddr;
 148         conn->c_faddr = faddr;
 149         spin_lock_init(&conn->c_lock);
 150         conn->c_next_tx_seq = 1;
 151
 152         init_waitqueue_head(&conn->c_waitq);
 153         INIT_LIST_HEAD(&conn->c_send_queue);
 154         INIT_LIST_HEAD(&conn->c_retrans);
 155
 156         ret = rds_cong_get_maps(conn);
 157         if (ret) {
 158                 kmem_cache_free(rds_conn_slab, conn);
 159                 conn = ERR_PTR(ret);
 160                 goto out;
 161         }
 162
 163         /*
 164          * This is where a connection becomes loopback.  If *any* RDS sockets
 165          * can bind to the destination address then we'd rather the messages
 166          * flow through loopback rather than either transport.
 167          */
 168         loop_trans = rds_trans_get_preferred(faddr);
 169         if (loop_trans) {
 170                 rds_trans_put(loop_trans);
 171                 conn->c_loopback = 1;
 172                 if (is_outgoing && trans->t_prefer_loopback) {
 173                         /* "outgoing" connection - and the transport
 174                          * says it wants the connection handled by the
 175                          * loopback transport. This is what TCP does.
 176                          */
 177                         trans = &rds_loop_transport;
 178                 }
 179         }
 180
 181         if (trans == NULL) {
 182                 kmem_cache_free(rds_conn_slab, conn);
 183                 conn = ERR_PTR(-ENODEV);
 184                 goto out;
 185         }
 186
 187         conn->c_trans = trans;
 188
 189         ret = trans->conn_alloc(conn, gfp);
 190         if (ret) {
 191                 kmem_cache_free(rds_conn_slab, conn);
 192                 conn = ERR_PTR(ret);
 193                 goto out;
 194         }
 195
 196         atomic_set(&conn->c_state, RDS_CONN_DOWN);
 197         conn->c_reconnect_jiffies = 0;
 198         INIT_DELAYED_WORK(&conn->c_send_w, rds_send_worker);
 199         INIT_DELAYED_WORK(&conn->c_recv_w, rds_recv_worker);
 200         INIT_DELAYED_WORK(&conn->c_conn_w, rds_connect_worker);
 201         INIT_WORK(&conn->c_down_w, rds_shutdown_worker);
 202         mutex_init(&conn->c_cm_lock);
 203         conn->c_flags = 0;
 204
 205         rdsdebug("allocated conn %p for %pI4 -> %pI4 over %s %s\n",
 206           conn, &laddr, &faddr,
 207           trans->t_name ? trans->t_name : "[unknown]",
 208           is_outgoing ? "(outgoing)" : "");
 209
 210         /*
 211          * Since we ran without holding the conn lock, someone could
 212          * have created the same conn (either normal or passive) in the
 213          * interim. We check while holding the lock. If we won, we complete
 214          * init and return our conn. If we lost, we rollback and return the
 215          * other one.
 216          */
 217         spin_lock_irqsave(&rds_conn_lock, flags);
 218         if (parent) {
 219                 /* Creating passive conn */
 220                 if (parent->c_passive) {
 221                         trans->conn_free(conn->c_transport_data);
 222                         kmem_cache_free(rds_conn_slab, conn);
 223                         conn = parent->c_passive;
 224                 } else {
 225                         parent->c_passive = conn;
 226                         rds_cong_add_conn(conn);
 227                         rds_conn_count++;
 228                 }
 229         } else {
 230                 /* Creating normal conn */
 231                 struct rds_connection *found;
 232
 233                 found = rds_conn_lookup(head, laddr, faddr, trans);
 234                 if (found) {
 235                         trans->conn_free(conn->c_transport_data);
 236                         kmem_cache_free(rds_conn_slab, conn);
 237                         conn = found;
 238                 } else {
 239                         hlist_add_head_rcu(&conn->c_hash_node, head);
 240                         rds_cong_add_conn(conn);
 241                         rds_conn_count++;
 242                 }
 243         }
 244         spin_unlock_irqrestore(&rds_conn_lock, flags);
 245
 246 out:
 247         return conn;
 248 }
 249
 250 struct rds_connection *rds_conn_create(__be32 laddr, __be32 faddr,
 251                                        struct rds_transport *trans, gfp_t gfp)
 252 {
 253         return __rds_conn_create(laddr, faddr, trans, gfp, 0);
 254 }
 255 EXPORT_SYMBOL_GPL(rds_conn_create);
 256
 257 struct rds_connection *rds_conn_create_outgoing(__be32 laddr, __be32 faddr,
 258                                        struct rds_transport *trans, gfp_t gfp)
 259 {
 260         return __rds_conn_create(laddr, faddr, trans, gfp, 1);
 261 }
 262 EXPORT_SYMBOL_GPL(rds_conn_create_outgoing);
 263
 264 void rds_conn_shutdown(struct rds_connection *conn)
 265 {
 266         /* shut it down unless it's down already */
 267         if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_DOWN)) {
 268                 /*
 269                  * Quiesce the connection mgmt handlers before we start tearing
 270                  * things down. We don't hold the mutex for the entire
 271                  * duration of the shutdown operation, else we may be
 272                  * deadlocking with the CM handler. Instead, the CM event
 273                  * handler is supposed to check for state DISCONNECTING
 274                  */
 275                 mutex_lock(&conn->c_cm_lock);
 276                 if (!rds_conn_transition(conn, RDS_CONN_UP, RDS_CONN_DISCONNECTING)
 277                  && !rds_conn_transition(conn, RDS_CONN_ERROR, RDS_CONN_DISCONNECTING)) {
 278                         rds_conn_error(conn, "shutdown called in state %d\n",
 279                                         atomic_read(&conn->c_state));
 280                         mutex_unlock(&conn->c_cm_lock);
 281                         return;
 282                 }
 283                 mutex_unlock(&conn->c_cm_lock);
 284
 285                 wait_event(conn->c_waitq,
 286                            !test_bit(RDS_IN_XMIT, &conn->c_flags));
 287
 288                 conn->c_trans->conn_shutdown(conn);
 289                 rds_conn_reset(conn);
 290
 291                 if (!rds_conn_transition(conn, RDS_CONN_DISCONNECTING, RDS_CONN_DOWN)) {
 292                         /* This can happen - eg when we're in the middle of tearing
 293                          * down the connection, and someone unloads the rds module.
 294                          * Quite reproduceable with loopback connections.
 295                          * Mostly harmless.
 296                          */
 297                         rds_conn_error(conn,
 298                                 "%s: failed to transition to state DOWN, "
 299                                 "current state is %d\n",
 300                                 __func__,
 301                                 atomic_read(&conn->c_state));
 302                         return;
 303                 }
 304         }
 305
 306         /* Then reconnect if it's still live.
 307          * The passive side of an IB loopback connection is never added
 308          * to the conn hash, so we never trigger a reconnect on this
 309          * conn - the reconnect is always triggered by the active peer. */
 310         cancel_delayed_work_sync(&conn->c_conn_w);
 311         rcu_read_lock();
 312         if (!hlist_unhashed(&conn->c_hash_node)) {
 313                 rcu_read_unlock();
 314                 rds_queue_reconnect(conn);
 315         } else {
 316                 rcu_read_unlock();
 317         }
 318 }
 319
 320 /*
 321  * Stop and free a connection.
 322  *
 323  * This can only be used in very limited circumstances.  It assumes that once
 324  * the conn has been shutdown that no one else is referencing the connection.
 325  * We can only ensure this in the rmmod path in the current code.
 326  */
 327 void rds_conn_destroy(struct rds_connection *conn)
 328 {
 329         struct rds_message *rm, *rtmp;
 330         unsigned long flags;
 331
 332         rdsdebug("freeing conn %p for %pI4 -> "
 333                  "%pI4\n", conn, &conn->c_laddr,
 334                  &conn->c_faddr);
 335
 336         /* Ensure conn will not be scheduled for reconnect */
 337         spin_lock_irq(&rds_conn_lock);
 338         hlist_del_init_rcu(&conn->c_hash_node);
 339         spin_unlock_irq(&rds_conn_lock);
 340         synchronize_rcu();
 341
 342         /* shut the connection down */
 343         rds_conn_drop(conn);
 344         flush_work(&conn->c_down_w);
 345
 346         /* make sure lingering queued work won't try to ref the conn */
 347         cancel_delayed_work_sync(&conn->c_send_w);
 348         cancel_delayed_work_sync(&conn->c_recv_w);
 349
 350         /* tear down queued messages */
 351         list_for_each_entry_safe(rm, rtmp,
 352                                  &conn->c_send_queue,
 353                                  m_conn_item) {
 354                 list_del_init(&rm->m_conn_item);
 355                 BUG_ON(!list_empty(&rm->m_sock_item));
 356                 rds_message_put(rm);
 357         }
 358         if (conn->c_xmit_rm)
 359                 rds_message_put(conn->c_xmit_rm);
 360
 361         conn->c_trans->conn_free(conn->c_transport_data);
 362
 363         /*
 364          * The congestion maps aren't freed up here.  They're
 365          * freed by rds_cong_exit() after all the connections
 366          * have been freed.
 367          */
 368         rds_cong_remove_conn(conn);
 369
 370         BUG_ON(!list_empty(&conn->c_retrans));
 371         kmem_cache_free(rds_conn_slab, conn);
 372
 373         spin_lock_irqsave(&rds_conn_lock, flags);
 374         rds_conn_count--;
 375         spin_unlock_irqrestore(&rds_conn_lock, flags);
 376 }
 377 EXPORT_SYMBOL_GPL(rds_conn_destroy);
 378
 379 static void rds_conn_message_info(struct socket *sock, unsigned int len,
 380                                   struct rds_info_iterator *iter,
 381                                   struct rds_info_lengths *lens,
 382                                   int want_send)
 383 {
 384         struct hlist_head *head;
 385         struct hlist_node *pos;
 386         struct list_head *list;
 387         struct rds_connection *conn;
 388         struct rds_message *rm;
 389         unsigned int total = 0;
 390         unsigned long flags;
 391         size_t i;
 392
 393         len /= sizeof(struct rds_info_message);
 394
 395         rcu_read_lock();
 396
 397         for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
 398              i++, head++) {
 399                 hlist_for_each_entry_rcu(conn, pos, head, c_hash_node) {
 400                         if (want_send)
 401                                 list = &conn->c_send_queue;
 402                         else
 403                                 list = &conn->c_retrans;
 404
 405                         spin_lock_irqsave(&conn->c_lock, flags);
 406
 407                         /* XXX too lazy to maintain counts.. */
 408                         list_for_each_entry(rm, list, m_conn_item) {
 409                                 total++;
 410                                 if (total <= len)
 411                                         rds_inc_info_copy(&rm->m_inc, iter,
 412                                                           conn->c_laddr,
 413                                                           conn->c_faddr, 0);
 414                         }
 415
 416                         spin_unlock_irqrestore(&conn->c_lock, flags);
 417                 }
 418         }
 419         rcu_read_unlock();
 420
 421         lens->nr = total;
 422         lens->each = sizeof(struct rds_info_message);
 423 }
 424
 425 static void rds_conn_message_info_send(struct socket *sock, unsigned int len,
 426                                        struct rds_info_iterator *iter,
 427                                        struct rds_info_lengths *lens)
 428 {
 429         rds_conn_message_info(sock, len, iter, lens, 1);
 430 }
 431
 432 static void rds_conn_message_info_retrans(struct socket *sock,
 433                                           unsigned int len,
 434                                           struct rds_info_iterator *iter,
 435                                           struct rds_info_lengths *lens)
 436 {
 437         rds_conn_message_info(sock, len, iter, lens, 0);
 438 }
 439
 440 void rds_for_each_conn_info(struct socket *sock, unsigned int len,
 441                           struct rds_info_iterator *iter,
 442                           struct rds_info_lengths *lens,
 443                           int (*visitor)(struct rds_connection *, void *),
 444                           size_t item_len)
 445 {
 446         uint64_t buffer[(item_len + 7) / 8];
 447         struct hlist_head *head;
 448         struct hlist_node *pos;
 449         struct rds_connection *conn;
 450         size_t i;
 451
 452         rcu_read_lock();
 453
 454         lens->nr = 0;
 455         lens->each = item_len;
 456
 457         for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
 458              i++, head++) {
 459                 hlist_for_each_entry_rcu(conn, pos, head, c_hash_node) {
 460
 461                         /* XXX no c_lock usage.. */
 462                         if (!visitor(conn, buffer))
 463                                 continue;
 464
 465                         /* We copy as much as we can fit in the buffer,
 466                          * but we count all items so that the caller
 467                          * can resize the buffer. */
 468                         if (len >= item_len) {
 469                                 rds_info_copy(iter, buffer, item_len);
 470                                 len -= item_len;
 471                         }
 472                         lens->nr++;
 473                 }
 474         }
 475         rcu_read_unlock();
 476 }
 477 EXPORT_SYMBOL_GPL(rds_for_each_conn_info);
 478
 479 static int rds_conn_info_visitor(struct rds_connection *conn,
 480                                   void *buffer)
 481 {
 482         struct rds_info_connection *cinfo = buffer;
 483
 484         cinfo->next_tx_seq = conn->c_next_tx_seq;
 485         cinfo->next_rx_seq = conn->c_next_rx_seq;
 486         cinfo->laddr = conn->c_laddr;
 487         cinfo->faddr = conn->c_faddr;
 488         strncpy(cinfo->transport, conn->c_trans->t_name,
 489                 sizeof(cinfo->transport));
 490         cinfo->flags = 0;
 491
 492         rds_conn_info_set(cinfo->flags, test_bit(RDS_IN_XMIT, &conn->c_flags),
 493                           SENDING);
 494         /* XXX Future: return the state rather than these funky bits */
 495         rds_conn_info_set(cinfo->flags,
 496                           atomic_read(&conn->c_state) == RDS_CONN_CONNECTING,
 497                           CONNECTING);
 498         rds_conn_info_set(cinfo->flags,
 499                           atomic_read(&conn->c_state) == RDS_CONN_UP,
 500                           CONNECTED);
 501         return 1;
 502 }
 503
 504 static void rds_conn_info(struct socket *sock, unsigned int len,
 505                           struct rds_info_iterator *iter,
 506                           struct rds_info_lengths *lens)
 507 {
 508         rds_for_each_conn_info(sock, len, iter, lens,
 509                                 rds_conn_info_visitor,
 510                                 sizeof(struct rds_info_connection));
 511 }
 512
 513 int rds_conn_init(void)
 514 {
 515         rds_conn_slab = kmem_cache_create("rds_connection",
 516                                           sizeof(struct rds_connection),
 517                                           0, 0, NULL);
 518         if (!rds_conn_slab)
 519                 return -ENOMEM;
 520
 521         rds_info_register_func(RDS_INFO_CONNECTIONS, rds_conn_info);
 522         rds_info_register_func(RDS_INFO_SEND_MESSAGES,
 523                                rds_conn_message_info_send);
 524         rds_info_register_func(RDS_INFO_RETRANS_MESSAGES,
 525                                rds_conn_message_info_retrans);
 526
 527         return 0;
 528 }
 529
 530 void rds_conn_exit(void)
 531 {
 532         rds_loop_exit();
 533
 534         WARN_ON(!hlist_empty(rds_conn_hash));
 535
 536         kmem_cache_destroy(rds_conn_slab);
 537
 538         rds_info_deregister_func(RDS_INFO_CONNECTIONS, rds_conn_info);
 539         rds_info_deregister_func(RDS_INFO_SEND_MESSAGES,
 540                                  rds_conn_message_info_send);
 541         rds_info_deregister_func(RDS_INFO_RETRANS_MESSAGES,
 542                                  rds_conn_message_info_retrans);
 543 }
 544
 545 /*
 546  * Force a disconnect
 547  */
 548 void rds_conn_drop(struct rds_connection *conn)
 549 {
 550         atomic_set(&conn->c_state, RDS_CONN_ERROR);
 551         queue_work(rds_wq, &conn->c_down_w);
 552 }
 553 EXPORT_SYMBOL_GPL(rds_conn_drop);
 554
 555 /*
 556  * If the connection is down, trigger a connect. We may have scheduled a
 557  * delayed reconnect however - in this case we should not interfere.
 558  */
 559 void rds_conn_connect_if_down(struct rds_connection *conn)
 560 {
 561         if (rds_conn_state(conn) == RDS_CONN_DOWN &&
 562             !test_and_set_bit(RDS_RECONNECT_PENDING, &conn->c_flags))
 563                 queue_delayed_work(rds_wq, &conn->c_conn_w, 0);
 564 }
 565 EXPORT_SYMBOL_GPL(rds_conn_connect_if_down);
 566
 567 /*
 568  * An error occurred on the connection
 569  */
 570 void
 571 __rds_conn_error(struct rds_connection *conn, const char *fmt, ...)
 572 {
 573         va_list ap;
 574
 575         va_start(ap, fmt);
 576         vprintk(fmt, ap);
 577         va_end(ap);
 578
 579         rds_conn_drop(conn);
 580 }