linux/net/rds/connection.c
<<
>>
Prefs
   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.. */
  47static DEFINE_SPINLOCK(rds_conn_lock);
  48static unsigned long rds_conn_count;
  49static struct hlist_head rds_conn_hash[RDS_CONNECTION_HASH_ENTRIES];
  50static struct kmem_cache *rds_conn_slab;
  51
  52static 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 */
  67static 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 */
  92static 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 */
 115static 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        conn->c_trans = trans;
 182
 183        ret = trans->conn_alloc(conn, gfp);
 184        if (ret) {
 185                kmem_cache_free(rds_conn_slab, conn);
 186                conn = ERR_PTR(ret);
 187                goto out;
 188        }
 189
 190        atomic_set(&conn->c_state, RDS_CONN_DOWN);
 191        conn->c_reconnect_jiffies = 0;
 192        INIT_DELAYED_WORK(&conn->c_send_w, rds_send_worker);
 193        INIT_DELAYED_WORK(&conn->c_recv_w, rds_recv_worker);
 194        INIT_DELAYED_WORK(&conn->c_conn_w, rds_connect_worker);
 195        INIT_WORK(&conn->c_down_w, rds_shutdown_worker);
 196        mutex_init(&conn->c_cm_lock);
 197        conn->c_flags = 0;
 198
 199        rdsdebug("allocated conn %p for %pI4 -> %pI4 over %s %s\n",
 200          conn, &laddr, &faddr,
 201          trans->t_name ? trans->t_name : "[unknown]",
 202          is_outgoing ? "(outgoing)" : "");
 203
 204        /*
 205         * Since we ran without holding the conn lock, someone could
 206         * have created the same conn (either normal or passive) in the
 207         * interim. We check while holding the lock. If we won, we complete
 208         * init and return our conn. If we lost, we rollback and return the
 209         * other one.
 210         */
 211        spin_lock_irqsave(&rds_conn_lock, flags);
 212        if (parent) {
 213                /* Creating passive conn */
 214                if (parent->c_passive) {
 215                        trans->conn_free(conn->c_transport_data);
 216                        kmem_cache_free(rds_conn_slab, conn);
 217                        conn = parent->c_passive;
 218                } else {
 219                        parent->c_passive = conn;
 220                        rds_cong_add_conn(conn);
 221                        rds_conn_count++;
 222                }
 223        } else {
 224                /* Creating normal conn */
 225                struct rds_connection *found;
 226
 227                found = rds_conn_lookup(head, laddr, faddr, trans);
 228                if (found) {
 229                        trans->conn_free(conn->c_transport_data);
 230                        kmem_cache_free(rds_conn_slab, conn);
 231                        conn = found;
 232                } else {
 233                        hlist_add_head_rcu(&conn->c_hash_node, head);
 234                        rds_cong_add_conn(conn);
 235                        rds_conn_count++;
 236                }
 237        }
 238        spin_unlock_irqrestore(&rds_conn_lock, flags);
 239
 240out:
 241        return conn;
 242}
 243
 244struct rds_connection *rds_conn_create(__be32 laddr, __be32 faddr,
 245                                       struct rds_transport *trans, gfp_t gfp)
 246{
 247        return __rds_conn_create(laddr, faddr, trans, gfp, 0);
 248}
 249EXPORT_SYMBOL_GPL(rds_conn_create);
 250
 251struct rds_connection *rds_conn_create_outgoing(__be32 laddr, __be32 faddr,
 252                                       struct rds_transport *trans, gfp_t gfp)
 253{
 254        return __rds_conn_create(laddr, faddr, trans, gfp, 1);
 255}
 256EXPORT_SYMBOL_GPL(rds_conn_create_outgoing);
 257
 258void rds_conn_shutdown(struct rds_connection *conn)
 259{
 260        /* shut it down unless it's down already */
 261        if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_DOWN)) {
 262                /*
 263                 * Quiesce the connection mgmt handlers before we start tearing
 264                 * things down. We don't hold the mutex for the entire
 265                 * duration of the shutdown operation, else we may be
 266                 * deadlocking with the CM handler. Instead, the CM event
 267                 * handler is supposed to check for state DISCONNECTING
 268                 */
 269                mutex_lock(&conn->c_cm_lock);
 270                if (!rds_conn_transition(conn, RDS_CONN_UP, RDS_CONN_DISCONNECTING)
 271                 && !rds_conn_transition(conn, RDS_CONN_ERROR, RDS_CONN_DISCONNECTING)) {
 272                        rds_conn_error(conn, "shutdown called in state %d\n",
 273                                        atomic_read(&conn->c_state));
 274                        mutex_unlock(&conn->c_cm_lock);
 275                        return;
 276                }
 277                mutex_unlock(&conn->c_cm_lock);
 278
 279                wait_event(conn->c_waitq,
 280                           !test_bit(RDS_IN_XMIT, &conn->c_flags));
 281
 282                conn->c_trans->conn_shutdown(conn);
 283                rds_conn_reset(conn);
 284
 285                if (!rds_conn_transition(conn, RDS_CONN_DISCONNECTING, RDS_CONN_DOWN)) {
 286                        /* This can happen - eg when we're in the middle of tearing
 287                         * down the connection, and someone unloads the rds module.
 288                         * Quite reproduceable with loopback connections.
 289                         * Mostly harmless.
 290                         */
 291                        rds_conn_error(conn,
 292                                "%s: failed to transition to state DOWN, "
 293                                "current state is %d\n",
 294                                __func__,
 295                                atomic_read(&conn->c_state));
 296                        return;
 297                }
 298        }
 299
 300        /* Then reconnect if it's still live.
 301         * The passive side of an IB loopback connection is never added
 302         * to the conn hash, so we never trigger a reconnect on this
 303         * conn - the reconnect is always triggered by the active peer. */
 304        cancel_delayed_work_sync(&conn->c_conn_w);
 305        rcu_read_lock();
 306        if (!hlist_unhashed(&conn->c_hash_node)) {
 307                rcu_read_unlock();
 308                rds_queue_reconnect(conn);
 309        } else {
 310                rcu_read_unlock();
 311        }
 312}
 313
 314/*
 315 * Stop and free a connection.
 316 *
 317 * This can only be used in very limited circumstances.  It assumes that once
 318 * the conn has been shutdown that no one else is referencing the connection.
 319 * We can only ensure this in the rmmod path in the current code.
 320 */
 321void rds_conn_destroy(struct rds_connection *conn)
 322{
 323        struct rds_message *rm, *rtmp;
 324        unsigned long flags;
 325
 326        rdsdebug("freeing conn %p for %pI4 -> "
 327                 "%pI4\n", conn, &conn->c_laddr,
 328                 &conn->c_faddr);
 329
 330        /* Ensure conn will not be scheduled for reconnect */
 331        spin_lock_irq(&rds_conn_lock);
 332        hlist_del_init_rcu(&conn->c_hash_node);
 333        spin_unlock_irq(&rds_conn_lock);
 334        synchronize_rcu();
 335
 336        /* shut the connection down */
 337        rds_conn_drop(conn);
 338        flush_work(&conn->c_down_w);
 339
 340        /* make sure lingering queued work won't try to ref the conn */
 341        cancel_delayed_work_sync(&conn->c_send_w);
 342        cancel_delayed_work_sync(&conn->c_recv_w);
 343
 344        /* tear down queued messages */
 345        list_for_each_entry_safe(rm, rtmp,
 346                                 &conn->c_send_queue,
 347                                 m_conn_item) {
 348                list_del_init(&rm->m_conn_item);
 349                BUG_ON(!list_empty(&rm->m_sock_item));
 350                rds_message_put(rm);
 351        }
 352        if (conn->c_xmit_rm)
 353                rds_message_put(conn->c_xmit_rm);
 354
 355        conn->c_trans->conn_free(conn->c_transport_data);
 356
 357        /*
 358         * The congestion maps aren't freed up here.  They're
 359         * freed by rds_cong_exit() after all the connections
 360         * have been freed.
 361         */
 362        rds_cong_remove_conn(conn);
 363
 364        BUG_ON(!list_empty(&conn->c_retrans));
 365        kmem_cache_free(rds_conn_slab, conn);
 366
 367        spin_lock_irqsave(&rds_conn_lock, flags);
 368        rds_conn_count--;
 369        spin_unlock_irqrestore(&rds_conn_lock, flags);
 370}
 371EXPORT_SYMBOL_GPL(rds_conn_destroy);
 372
 373static void rds_conn_message_info(struct socket *sock, unsigned int len,
 374                                  struct rds_info_iterator *iter,
 375                                  struct rds_info_lengths *lens,
 376                                  int want_send)
 377{
 378        struct hlist_head *head;
 379        struct hlist_node *pos;
 380        struct list_head *list;
 381        struct rds_connection *conn;
 382        struct rds_message *rm;
 383        unsigned int total = 0;
 384        unsigned long flags;
 385        size_t i;
 386
 387        len /= sizeof(struct rds_info_message);
 388
 389        rcu_read_lock();
 390
 391        for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
 392             i++, head++) {
 393                hlist_for_each_entry_rcu(conn, pos, head, c_hash_node) {
 394                        if (want_send)
 395                                list = &conn->c_send_queue;
 396                        else
 397                                list = &conn->c_retrans;
 398
 399                        spin_lock_irqsave(&conn->c_lock, flags);
 400
 401                        /* XXX too lazy to maintain counts.. */
 402                        list_for_each_entry(rm, list, m_conn_item) {
 403                                total++;
 404                                if (total <= len)
 405                                        rds_inc_info_copy(&rm->m_inc, iter,
 406                                                          conn->c_laddr,
 407                                                          conn->c_faddr, 0);
 408                        }
 409
 410                        spin_unlock_irqrestore(&conn->c_lock, flags);
 411                }
 412        }
 413        rcu_read_unlock();
 414
 415        lens->nr = total;
 416        lens->each = sizeof(struct rds_info_message);
 417}
 418
 419static void rds_conn_message_info_send(struct socket *sock, unsigned int len,
 420                                       struct rds_info_iterator *iter,
 421                                       struct rds_info_lengths *lens)
 422{
 423        rds_conn_message_info(sock, len, iter, lens, 1);
 424}
 425
 426static void rds_conn_message_info_retrans(struct socket *sock,
 427                                          unsigned int len,
 428                                          struct rds_info_iterator *iter,
 429                                          struct rds_info_lengths *lens)
 430{
 431        rds_conn_message_info(sock, len, iter, lens, 0);
 432}
 433
 434void rds_for_each_conn_info(struct socket *sock, unsigned int len,
 435                          struct rds_info_iterator *iter,
 436                          struct rds_info_lengths *lens,
 437                          int (*visitor)(struct rds_connection *, void *),
 438                          size_t item_len)
 439{
 440        uint64_t buffer[(item_len + 7) / 8];
 441        struct hlist_head *head;
 442        struct hlist_node *pos;
 443        struct rds_connection *conn;
 444        size_t i;
 445
 446        rcu_read_lock();
 447
 448        lens->nr = 0;
 449        lens->each = item_len;
 450
 451        for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
 452             i++, head++) {
 453                hlist_for_each_entry_rcu(conn, pos, head, c_hash_node) {
 454
 455                        /* XXX no c_lock usage.. */
 456                        if (!visitor(conn, buffer))
 457                                continue;
 458
 459                        /* We copy as much as we can fit in the buffer,
 460                         * but we count all items so that the caller
 461                         * can resize the buffer. */
 462                        if (len >= item_len) {
 463                                rds_info_copy(iter, buffer, item_len);
 464                                len -= item_len;
 465                        }
 466                        lens->nr++;
 467                }
 468        }
 469        rcu_read_unlock();
 470}
 471EXPORT_SYMBOL_GPL(rds_for_each_conn_info);
 472
 473static int rds_conn_info_visitor(struct rds_connection *conn,
 474                                  void *buffer)
 475{
 476        struct rds_info_connection *cinfo = buffer;
 477
 478        cinfo->next_tx_seq = conn->c_next_tx_seq;
 479        cinfo->next_rx_seq = conn->c_next_rx_seq;
 480        cinfo->laddr = conn->c_laddr;
 481        cinfo->faddr = conn->c_faddr;
 482        strncpy(cinfo->transport, conn->c_trans->t_name,
 483                sizeof(cinfo->transport));
 484        cinfo->flags = 0;
 485
 486        rds_conn_info_set(cinfo->flags, test_bit(RDS_IN_XMIT, &conn->c_flags),
 487                          SENDING);
 488        /* XXX Future: return the state rather than these funky bits */
 489        rds_conn_info_set(cinfo->flags,
 490                          atomic_read(&conn->c_state) == RDS_CONN_CONNECTING,
 491                          CONNECTING);
 492        rds_conn_info_set(cinfo->flags,
 493                          atomic_read(&conn->c_state) == RDS_CONN_UP,
 494                          CONNECTED);
 495        return 1;
 496}
 497
 498static void rds_conn_info(struct socket *sock, unsigned int len,
 499                          struct rds_info_iterator *iter,
 500                          struct rds_info_lengths *lens)
 501{
 502        rds_for_each_conn_info(sock, len, iter, lens,
 503                                rds_conn_info_visitor,
 504                                sizeof(struct rds_info_connection));
 505}
 506
 507int rds_conn_init(void)
 508{
 509        rds_conn_slab = kmem_cache_create("rds_connection",
 510                                          sizeof(struct rds_connection),
 511                                          0, 0, NULL);
 512        if (!rds_conn_slab)
 513                return -ENOMEM;
 514
 515        rds_info_register_func(RDS_INFO_CONNECTIONS, rds_conn_info);
 516        rds_info_register_func(RDS_INFO_SEND_MESSAGES,
 517                               rds_conn_message_info_send);
 518        rds_info_register_func(RDS_INFO_RETRANS_MESSAGES,
 519                               rds_conn_message_info_retrans);
 520
 521        return 0;
 522}
 523
 524void rds_conn_exit(void)
 525{
 526        rds_loop_exit();
 527
 528        WARN_ON(!hlist_empty(rds_conn_hash));
 529
 530        kmem_cache_destroy(rds_conn_slab);
 531
 532        rds_info_deregister_func(RDS_INFO_CONNECTIONS, rds_conn_info);
 533        rds_info_deregister_func(RDS_INFO_SEND_MESSAGES,
 534                                 rds_conn_message_info_send);
 535        rds_info_deregister_func(RDS_INFO_RETRANS_MESSAGES,
 536                                 rds_conn_message_info_retrans);
 537}
 538
 539/*
 540 * Force a disconnect
 541 */
 542void rds_conn_drop(struct rds_connection *conn)
 543{
 544        atomic_set(&conn->c_state, RDS_CONN_ERROR);
 545        queue_work(rds_wq, &conn->c_down_w);
 546}
 547EXPORT_SYMBOL_GPL(rds_conn_drop);
 548
 549/*
 550 * If the connection is down, trigger a connect. We may have scheduled a
 551 * delayed reconnect however - in this case we should not interfere.
 552 */
 553void rds_conn_connect_if_down(struct rds_connection *conn)
 554{
 555        if (rds_conn_state(conn) == RDS_CONN_DOWN &&
 556            !test_and_set_bit(RDS_RECONNECT_PENDING, &conn->c_flags))
 557                queue_delayed_work(rds_wq, &conn->c_conn_w, 0);
 558}
 559EXPORT_SYMBOL_GPL(rds_conn_connect_if_down);
 560
 561/*
 562 * An error occurred on the connection
 563 */
 564void
 565__rds_conn_error(struct rds_connection *conn, const char *fmt, ...)
 566{
 567        va_list ap;
 568
 569        va_start(ap, fmt);
 570        vprintk(fmt, ap);
 571        va_end(ap);
 572
 573        rds_conn_drop(conn);
 574}
 575
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.