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 hlist_head *head,
  74                                               __be32 laddr, __be32 faddr,
  75                                               struct rds_transport *trans)
  76 {
  77         struct rds_connection *conn, *ret = NULL;
  78
  79         hlist_for_each_entry_rcu(conn, head, c_hash_node) {
  80                 if (conn->c_faddr == faddr && conn->c_laddr == laddr &&
  81                                 conn->c_trans == trans) {
  82                         ret = conn;
  83                         break;
  84                 }
  85         }
  86         rdsdebug("returning conn %p for %pI4 -> %pI4\n", ret,
  87                  &laddr, &faddr);
  88         return ret;
  89 }
  90
  91 /*
  92  * This is called by transports as they're bringing down a connection.
  93  * It clears partial message state so that the transport can start sending
  94  * and receiving over this connection again in the future.  It is up to
  95  * the transport to have serialized this call with its send and recv.
  96  */
  97 static void rds_conn_reset(struct rds_connection *conn)
  98 {
  99         rdsdebug("connection %pI4 to %pI4 reset\n",
 100           &conn->c_laddr, &conn->c_faddr);
 101
 102         rds_stats_inc(s_conn_reset);
 103         rds_send_reset(conn);
 104         conn->c_flags = 0;
 105
 106         /* Do not clear next_rx_seq here, else we cannot distinguish
 107          * retransmitted packets from new packets, and will hand all
 108          * of them to the application. That is not consistent with the
 109          * reliability guarantees of RDS. */
 110 }
 111
 112 /*
 113  * There is only every one 'conn' for a given pair of addresses in the
 114  * system at a time.  They contain messages to be retransmitted and so
 115  * span the lifetime of the actual underlying transport connections.
 116  *
 117  * For now they are not garbage collected once they're created.  They
 118  * are torn down as the module is removed, if ever.
 119  */
 120 static struct rds_connection *__rds_conn_create(__be32 laddr, __be32 faddr,
 121                                        struct rds_transport *trans, gfp_t gfp,
 122                                        int is_outgoing)
 123 {
 124         struct rds_connection *conn, *parent = NULL;
 125         struct hlist_head *head = rds_conn_bucket(laddr, faddr);
 126         struct rds_transport *loop_trans;
 127         unsigned long flags;
 128         int ret;
 129
 130         rcu_read_lock();
 131         conn = rds_conn_lookup(head, laddr, faddr, trans);
 132         if (conn && conn->c_loopback && conn->c_trans != &rds_loop_transport &&
 133             laddr == faddr && !is_outgoing) {
 134                 /* This is a looped back IB connection, and we're
 135                  * called by the code handling the incoming connect.
 136                  * We need a second connection object into which we
 137                  * can stick the other QP. */
 138                 parent = conn;
 139                 conn = parent->c_passive;
 140         }
 141         rcu_read_unlock();
 142         if (conn)
 143                 goto out;
 144
 145         conn = kmem_cache_zalloc(rds_conn_slab, gfp);
 146         if (!conn) {
 147                 conn = ERR_PTR(-ENOMEM);
 148                 goto out;
 149         }
 150
 151         INIT_HLIST_NODE(&conn->c_hash_node);
 152         conn->c_laddr = laddr;
 153         conn->c_faddr = faddr;
 154         spin_lock_init(&conn->c_lock);
 155         conn->c_next_tx_seq = 1;
 156
 157         init_waitqueue_head(&conn->c_waitq);
 158         INIT_LIST_HEAD(&conn->c_send_queue);
 159         INIT_LIST_HEAD(&conn->c_retrans);
 160
 161         ret = rds_cong_get_maps(conn);
 162         if (ret) {
 163                 kmem_cache_free(rds_conn_slab, conn);
 164                 conn = ERR_PTR(ret);
 165                 goto out;
 166         }
 167
 168         /*
 169          * This is where a connection becomes loopback.  If *any* RDS sockets
 170          * can bind to the destination address then we'd rather the messages
 171          * flow through loopback rather than either transport.
 172          */
 173         loop_trans = rds_trans_get_preferred(faddr);
 174         if (loop_trans) {
 175                 rds_trans_put(loop_trans);
 176                 conn->c_loopback = 1;
 177                 if (is_outgoing && trans->t_prefer_loopback) {
 178                         /* "outgoing" connection - and the transport
 179                          * says it wants the connection handled by the
 180                          * loopback transport. This is what TCP does.
 181                          */
 182                         trans = &rds_loop_transport;
 183                 }
 184         }
 185
 186         conn->c_trans = trans;
 187
 188         ret = trans->conn_alloc(conn, gfp);
 189         if (ret) {
 190                 kmem_cache_free(rds_conn_slab, conn);
 191                 conn = ERR_PTR(ret);
 192                 goto out;
 193         }
 194
 195         atomic_set(&conn->c_state, RDS_CONN_DOWN);
 196         conn->c_send_gen = 0;
 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 list_head *list;
 386         struct rds_connection *conn;
 387         struct rds_message *rm;
 388         unsigned int total = 0;
 389         unsigned long flags;
 390         size_t i;
 391
 392         len /= sizeof(struct rds_info_message);
 393
 394         rcu_read_lock();
 395
 396         for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
 397              i++, head++) {
 398                 hlist_for_each_entry_rcu(conn, head, c_hash_node) {
 399                         if (want_send)
 400                                 list = &conn->c_send_queue;
 401                         else
 402                                 list = &conn->c_retrans;
 403
 404                         spin_lock_irqsave(&conn->c_lock, flags);
 405
 406                         /* XXX too lazy to maintain counts.. */
 407                         list_for_each_entry(rm, list, m_conn_item) {
 408                                 total++;
 409                                 if (total <= len)
 410                                         rds_inc_info_copy(&rm->m_inc, iter,
 411                                                           conn->c_laddr,
 412                                                           conn->c_faddr, 0);
 413                         }
 414
 415                         spin_unlock_irqrestore(&conn->c_lock, flags);
 416                 }
 417         }
 418         rcu_read_unlock();
 419
 420         lens->nr = total;
 421         lens->each = sizeof(struct rds_info_message);
 422 }
 423
 424 static void rds_conn_message_info_send(struct socket *sock, unsigned int len,
 425                                        struct rds_info_iterator *iter,
 426                                        struct rds_info_lengths *lens)
 427 {
 428         rds_conn_message_info(sock, len, iter, lens, 1);
 429 }
 430
 431 static void rds_conn_message_info_retrans(struct socket *sock,
 432                                           unsigned int len,
 433                                           struct rds_info_iterator *iter,
 434                                           struct rds_info_lengths *lens)
 435 {
 436         rds_conn_message_info(sock, len, iter, lens, 0);
 437 }
 438
 439 void rds_for_each_conn_info(struct socket *sock, unsigned int len,
 440                           struct rds_info_iterator *iter,
 441                           struct rds_info_lengths *lens,
 442                           int (*visitor)(struct rds_connection *, void *),
 443                           size_t item_len)
 444 {
 445         uint64_t buffer[(item_len + 7) / 8];
 446         struct hlist_head *head;
 447         struct rds_connection *conn;
 448         size_t i;
 449
 450         rcu_read_lock();
 451
 452         lens->nr = 0;
 453         lens->each = item_len;
 454
 455         for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
 456              i++, head++) {
 457                 hlist_for_each_entry_rcu(conn, head, c_hash_node) {
 458
 459                         /* XXX no c_lock usage.. */
 460                         if (!visitor(conn, buffer))
 461                                 continue;
 462
 463                         /* We copy as much as we can fit in the buffer,
 464                          * but we count all items so that the caller
 465                          * can resize the buffer. */
 466                         if (len >= item_len) {
 467                                 rds_info_copy(iter, buffer, item_len);
 468                                 len -= item_len;
 469                         }
 470                         lens->nr++;
 471                 }
 472         }
 473         rcu_read_unlock();
 474 }
 475 EXPORT_SYMBOL_GPL(rds_for_each_conn_info);
 476
 477 static int rds_conn_info_visitor(struct rds_connection *conn,
 478                                   void *buffer)
 479 {
 480         struct rds_info_connection *cinfo = buffer;
 481
 482         cinfo->next_tx_seq = conn->c_next_tx_seq;
 483         cinfo->next_rx_seq = conn->c_next_rx_seq;
 484         cinfo->laddr = conn->c_laddr;
 485         cinfo->faddr = conn->c_faddr;
 486         strncpy(cinfo->transport, conn->c_trans->t_name,
 487                 sizeof(cinfo->transport));
 488         cinfo->flags = 0;
 489
 490         rds_conn_info_set(cinfo->flags, test_bit(RDS_IN_XMIT, &conn->c_flags),
 491                           SENDING);
 492         /* XXX Future: return the state rather than these funky bits */
 493         rds_conn_info_set(cinfo->flags,
 494                           atomic_read(&conn->c_state) == RDS_CONN_CONNECTING,
 495                           CONNECTING);
 496         rds_conn_info_set(cinfo->flags,
 497                           atomic_read(&conn->c_state) == RDS_CONN_UP,
 498                           CONNECTED);
 499         return 1;
 500 }
 501
 502 static void rds_conn_info(struct socket *sock, unsigned int len,
 503                           struct rds_info_iterator *iter,
 504                           struct rds_info_lengths *lens)
 505 {
 506         rds_for_each_conn_info(sock, len, iter, lens,
 507                                 rds_conn_info_visitor,
 508                                 sizeof(struct rds_info_connection));
 509 }
 510
 511 int rds_conn_init(void)
 512 {
 513         rds_conn_slab = kmem_cache_create("rds_connection",
 514                                           sizeof(struct rds_connection),
 515                                           0, 0, NULL);
 516         if (!rds_conn_slab)
 517                 return -ENOMEM;
 518
 519         rds_info_register_func(RDS_INFO_CONNECTIONS, rds_conn_info);
 520         rds_info_register_func(RDS_INFO_SEND_MESSAGES,
 521                                rds_conn_message_info_send);
 522         rds_info_register_func(RDS_INFO_RETRANS_MESSAGES,
 523                                rds_conn_message_info_retrans);
 524
 525         return 0;
 526 }
 527
 528 void rds_conn_exit(void)
 529 {
 530         rds_loop_exit();
 531
 532         WARN_ON(!hlist_empty(rds_conn_hash));
 533
 534         kmem_cache_destroy(rds_conn_slab);
 535
 536         rds_info_deregister_func(RDS_INFO_CONNECTIONS, rds_conn_info);
 537         rds_info_deregister_func(RDS_INFO_SEND_MESSAGES,
 538                                  rds_conn_message_info_send);
 539         rds_info_deregister_func(RDS_INFO_RETRANS_MESSAGES,
 540                                  rds_conn_message_info_retrans);
 541 }
 542
 543 /*
 544  * Force a disconnect
 545  */
 546 void rds_conn_drop(struct rds_connection *conn)
 547 {
 548         atomic_set(&conn->c_state, RDS_CONN_ERROR);
 549         queue_work(rds_wq, &conn->c_down_w);
 550 }
 551 EXPORT_SYMBOL_GPL(rds_conn_drop);
 552
 553 /*
 554  * If the connection is down, trigger a connect. We may have scheduled a
 555  * delayed reconnect however - in this case we should not interfere.
 556  */
 557 void rds_conn_connect_if_down(struct rds_connection *conn)
 558 {
 559         if (rds_conn_state(conn) == RDS_CONN_DOWN &&
 560             !test_and_set_bit(RDS_RECONNECT_PENDING, &conn->c_flags))
 561                 queue_delayed_work(rds_wq, &conn->c_conn_w, 0);
 562 }
 563 EXPORT_SYMBOL_GPL(rds_conn_connect_if_down);
 564
 565 /*
 566  * An error occurred on the connection
 567  */
 568 void
 569 __rds_conn_error(struct rds_connection *conn, const char *fmt, ...)
 570 {
 571         va_list ap;
 572
 573         va_start(ap, fmt);
 574         vprintk(fmt, ap);
 575         va_end(ap);
 576
 577         rds_conn_drop(conn);
 578 }