net/rds/connection.c

   1 /*
   2  * Copyright (c) 2006 Oracle.  All rights reserved.
   3  *
   4  * This software is available to you under a choice of one of two
   5  * licenses.  You may choose to be licensed under the terms of the GNU
   6  * General Public License (GPL) Version 2, available from the file
   7  * COPYING in the main directory of this source tree, or the
   8  * OpenIB.org BSD license below:
   9  *
  10  *     Redistribution and use in source and binary forms, with or
  11  *     without modification, are permitted provided that the following
  12  *     conditions are met:
  13  *
  14  *      - Redistributions of source code must retain the above
  15  *        copyright notice, this list of conditions and the following
  16  *        disclaimer.
  17  *
  18  *      - Redistributions in binary form must reproduce the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer in the documentation and/or other materials
  21  *        provided with the distribution.
  22  *
  23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30  * SOFTWARE.
  31  *
  32  */
  33 #include <linux/kernel.h>
  34 #include <linux/list.h>
  35 #include <linux/slab.h>
  36 #include <net/inet_hashtables.h>
  37
  38 #include "rds.h"
  39 #include "loop.h"
  40 #include "rdma.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 static inline int rds_conn_is_sending(struct rds_connection *conn)
  67 {
  68         int ret = 0;
  69
  70         if (!mutex_trylock(&conn->c_send_lock))
  71                 ret = 1;
  72         else
  73                 mutex_unlock(&conn->c_send_lock);
  74
  75         return ret;
  76 }
  77
  78 static struct rds_connection *rds_conn_lookup(struct hlist_head *head,
  79                                               __be32 laddr, __be32 faddr,
  80                                               struct rds_transport *trans)
  81 {
  82         struct rds_connection *conn, *ret = NULL;
  83         struct hlist_node *pos;
  84
  85         hlist_for_each_entry(conn, pos, head, c_hash_node) {
  86                 if (conn->c_faddr == faddr && conn->c_laddr == laddr &&
  87                                 conn->c_trans == trans) {
  88                         ret = conn;
  89                         break;
  90                 }
  91         }
  92         rdsdebug("returning conn %p for %pI4 -> %pI4\n", ret,
  93                  &laddr, &faddr);
  94         return ret;
  95 }
  96
  97 /*
  98  * This is called by transports as they're bringing down a connection.
  99  * It clears partial message state so that the transport can start sending
 100  * and receiving over this connection again in the future.  It is up to
 101  * the transport to have serialized this call with its send and recv.
 102  */
 103 void rds_conn_reset(struct rds_connection *conn)
 104 {
 105         rdsdebug("connection %pI4 to %pI4 reset\n",
 106           &conn->c_laddr, &conn->c_faddr);
 107
 108         rds_stats_inc(s_conn_reset);
 109         rds_send_reset(conn);
 110         conn->c_flags = 0;
 111
 112         /* Do not clear next_rx_seq here, else we cannot distinguish
 113          * retransmitted packets from new packets, and will hand all
 114          * of them to the application. That is not consistent with the
 115          * reliability guarantees of RDS. */
 116 }
 117
 118 /*
 119  * There is only every one 'conn' for a given pair of addresses in the
 120  * system at a time.  They contain messages to be retransmitted and so
 121  * span the lifetime of the actual underlying transport connections.
 122  *
 123  * For now they are not garbage collected once they're created.  They
 124  * are torn down as the module is removed, if ever.
 125  */
 126 static struct rds_connection *__rds_conn_create(__be32 laddr, __be32 faddr,
 127                                        struct rds_transport *trans, gfp_t gfp,
 128                                        int is_outgoing)
 129 {
 130         struct rds_connection *conn, *parent = NULL;
 131         struct hlist_head *head = rds_conn_bucket(laddr, faddr);
 132         unsigned long flags;
 133         int ret;
 134
 135         spin_lock_irqsave(&rds_conn_lock, flags);
 136         conn = rds_conn_lookup(head, laddr, faddr, trans);
 137         if (conn && conn->c_loopback && conn->c_trans != &rds_loop_transport &&
 138             !is_outgoing) {
 139                 /* This is a looped back IB connection, and we're
 140                  * called by the code handling the incoming connect.
 141                  * We need a second connection object into which we
 142                  * can stick the other QP. */
 143                 parent = conn;
 144                 conn = parent->c_passive;
 145         }
 146         spin_unlock_irqrestore(&rds_conn_lock, flags);
 147         if (conn)
 148                 goto out;
 149
 150         conn = kmem_cache_zalloc(rds_conn_slab, gfp);
 151         if (!conn) {
 152                 conn = ERR_PTR(-ENOMEM);
 153                 goto out;
 154         }
 155
 156         INIT_HLIST_NODE(&conn->c_hash_node);
 157         conn->c_laddr = laddr;
 158         conn->c_faddr = faddr;
 159         spin_lock_init(&conn->c_lock);
 160         conn->c_next_tx_seq = 1;
 161
 162         mutex_init(&conn->c_send_lock);
 163         INIT_LIST_HEAD(&conn->c_send_queue);
 164         INIT_LIST_HEAD(&conn->c_retrans);
 165
 166         ret = rds_cong_get_maps(conn);
 167         if (ret) {
 168                 kmem_cache_free(rds_conn_slab, conn);
 169                 conn = ERR_PTR(ret);
 170                 goto out;
 171         }
 172
 173         /*
 174          * This is where a connection becomes loopback.  If *any* RDS sockets
 175          * can bind to the destination address then we'd rather the messages
 176          * flow through loopback rather than either transport.
 177          */
 178         if (rds_trans_get_preferred(faddr)) {
 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         conn->c_trans = trans;
 190
 191         ret = trans->conn_alloc(conn, gfp);
 192         if (ret) {
 193                 kmem_cache_free(rds_conn_slab, conn);
 194                 conn = ERR_PTR(ret);
 195                 goto out;
 196         }
 197
 198         atomic_set(&conn->c_state, RDS_CONN_DOWN);
 199         conn->c_reconnect_jiffies = 0;
 200         INIT_DELAYED_WORK(&conn->c_send_w, rds_send_worker);
 201         INIT_DELAYED_WORK(&conn->c_recv_w, rds_recv_worker);
 202         INIT_DELAYED_WORK(&conn->c_conn_w, rds_connect_worker);
 203         INIT_WORK(&conn->c_down_w, rds_shutdown_worker);
 204         mutex_init(&conn->c_cm_lock);
 205         conn->c_flags = 0;
 206
 207         rdsdebug("allocated conn %p for %pI4 -> %pI4 over %s %s\n",
 208           conn, &laddr, &faddr,
 209           trans->t_name ? trans->t_name : "[unknown]",
 210           is_outgoing ? "(outgoing)" : "");
 211
 212         /*
 213          * Since we ran without holding the conn lock, someone could
 214          * have created the same conn (either normal or passive) in the
 215          * interim. We check while holding the lock. If we won, we complete
 216          * init and return our conn. If we lost, we rollback and return the
 217          * other one.
 218          */
 219         spin_lock_irqsave(&rds_conn_lock, flags);
 220         if (parent) {
 221                 /* Creating passive conn */
 222                 if (parent->c_passive) {
 223                         trans->conn_free(conn->c_transport_data);
 224                         kmem_cache_free(rds_conn_slab, conn);
 225                         conn = parent->c_passive;
 226                 } else {
 227                         parent->c_passive = conn;
 228                         rds_cong_add_conn(conn);
 229                         rds_conn_count++;
 230                 }
 231         } else {
 232                 /* Creating normal conn */
 233                 struct rds_connection *found;
 234
 235                 found = rds_conn_lookup(head, laddr, faddr, trans);
 236                 if (found) {
 237                         trans->conn_free(conn->c_transport_data);
 238                         kmem_cache_free(rds_conn_slab, conn);
 239                         conn = found;
 240                 } else {
 241                         hlist_add_head(&conn->c_hash_node, head);
 242                         rds_cong_add_conn(conn);
 243                         rds_conn_count++;
 244                 }
 245         }
 246         spin_unlock_irqrestore(&rds_conn_lock, flags);
 247
 248 out:
 249         return conn;
 250 }
 251
 252 struct rds_connection *rds_conn_create(__be32 laddr, __be32 faddr,
 253                                        struct rds_transport *trans, gfp_t gfp)
 254 {
 255         return __rds_conn_create(laddr, faddr, trans, gfp, 0);
 256 }
 257 EXPORT_SYMBOL_GPL(rds_conn_create);
 258
 259 struct rds_connection *rds_conn_create_outgoing(__be32 laddr, __be32 faddr,
 260                                        struct rds_transport *trans, gfp_t gfp)
 261 {
 262         return __rds_conn_create(laddr, faddr, trans, gfp, 1);
 263 }
 264 EXPORT_SYMBOL_GPL(rds_conn_create_outgoing);
 265
 266 void rds_conn_shutdown(struct rds_connection *conn)
 267 {
 268         /* shut it down unless it's down already */
 269         if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_DOWN)) {
 270                 /*
 271                  * Quiesce the connection mgmt handlers before we start tearing
 272                  * things down. We don't hold the mutex for the entire
 273                  * duration of the shutdown operation, else we may be
 274                  * deadlocking with the CM handler. Instead, the CM event
 275                  * handler is supposed to check for state DISCONNECTING
 276                  */
 277                 mutex_lock(&conn->c_cm_lock);
 278                 if (!rds_conn_transition(conn, RDS_CONN_UP, RDS_CONN_DISCONNECTING)
 279                  && !rds_conn_transition(conn, RDS_CONN_ERROR, RDS_CONN_DISCONNECTING)) {
 280                         rds_conn_error(conn, "shutdown called in state %d\n",
 281                                         atomic_read(&conn->c_state));
 282                         mutex_unlock(&conn->c_cm_lock);
 283                         return;
 284                 }
 285                 mutex_unlock(&conn->c_cm_lock);
 286
 287                 mutex_lock(&conn->c_send_lock);
 288                 conn->c_trans->conn_shutdown(conn);
 289                 rds_conn_reset(conn);
 290                 mutex_unlock(&conn->c_send_lock);
 291
 292                 if (!rds_conn_transition(conn, RDS_CONN_DISCONNECTING, RDS_CONN_DOWN)) {
 293                         /* This can happen - eg when we're in the middle of tearing
 294                          * down the connection, and someone unloads the rds module.
 295                          * Quite reproduceable with loopback connections.
 296                          * Mostly harmless.
 297                          */
 298                         rds_conn_error(conn,
 299                                 "%s: failed to transition to state DOWN, "
 300                                 "current state is %d\n",
 301                                 __func__,
 302                                 atomic_read(&conn->c_state));
 303                         return;
 304                 }
 305         }
 306
 307         /* Then reconnect if it's still live.
 308          * The passive side of an IB loopback connection is never added
 309          * to the conn hash, so we never trigger a reconnect on this
 310          * conn - the reconnect is always triggered by the active peer. */
 311         cancel_delayed_work_sync(&conn->c_conn_w);
 312         if (!hlist_unhashed(&conn->c_hash_node))
 313                 rds_queue_reconnect(conn);
 314 }
 315
 316 /*
 317  * Stop and free a connection.
 318  */
 319 void rds_conn_destroy(struct rds_connection *conn)
 320 {
 321         struct rds_message *rm, *rtmp;
 322
 323         rdsdebug("freeing conn %p for %pI4 -> "
 324                  "%pI4\n", conn, &conn->c_laddr,
 325                  &conn->c_faddr);
 326
 327         hlist_del_init(&conn->c_hash_node);
 328
 329         /* wait for the rds thread to shut it down */
 330         atomic_set(&conn->c_state, RDS_CONN_ERROR);
 331         cancel_delayed_work(&conn->c_conn_w);
 332         queue_work(rds_wq, &conn->c_down_w);
 333         flush_workqueue(rds_wq);
 334
 335         /* tear down queued messages */
 336         list_for_each_entry_safe(rm, rtmp,
 337                                  &conn->c_send_queue,
 338                                  m_conn_item) {
 339                 list_del_init(&rm->m_conn_item);
 340                 BUG_ON(!list_empty(&rm->m_sock_item));
 341                 rds_message_put(rm);
 342         }
 343         if (conn->c_xmit_rm)
 344                 rds_message_put(conn->c_xmit_rm);
 345
 346         conn->c_trans->conn_free(conn->c_transport_data);
 347
 348         /*
 349          * The congestion maps aren't freed up here.  They're
 350          * freed by rds_cong_exit() after all the connections
 351          * have been freed.
 352          */
 353         rds_cong_remove_conn(conn);
 354
 355         BUG_ON(!list_empty(&conn->c_retrans));
 356         kmem_cache_free(rds_conn_slab, conn);
 357
 358         rds_conn_count--;
 359 }
 360 EXPORT_SYMBOL_GPL(rds_conn_destroy);
 361
 362 static void rds_conn_message_info(struct socket *sock, unsigned int len,
 363                                   struct rds_info_iterator *iter,
 364                                   struct rds_info_lengths *lens,
 365                                   int want_send)
 366 {
 367         struct hlist_head *head;
 368         struct hlist_node *pos;
 369         struct list_head *list;
 370         struct rds_connection *conn;
 371         struct rds_message *rm;
 372         unsigned long flags;
 373         unsigned int total = 0;
 374         size_t i;
 375
 376         len /= sizeof(struct rds_info_message);
 377
 378         spin_lock_irqsave(&rds_conn_lock, flags);
 379
 380         for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
 381              i++, head++) {
 382                 hlist_for_each_entry(conn, pos, head, c_hash_node) {
 383                         if (want_send)
 384                                 list = &conn->c_send_queue;
 385                         else
 386                                 list = &conn->c_retrans;
 387
 388                         spin_lock(&conn->c_lock);
 389
 390                         /* XXX too lazy to maintain counts.. */
 391                         list_for_each_entry(rm, list, m_conn_item) {
 392                                 total++;
 393                                 if (total <= len)
 394                                         rds_inc_info_copy(&rm->m_inc, iter,
 395                                                           conn->c_laddr,
 396                                                           conn->c_faddr, 0);
 397                         }
 398
 399                         spin_unlock(&conn->c_lock);
 400                 }
 401         }
 402
 403         spin_unlock_irqrestore(&rds_conn_lock, flags);
 404
 405         lens->nr = total;
 406         lens->each = sizeof(struct rds_info_message);
 407 }
 408
 409 static void rds_conn_message_info_send(struct socket *sock, unsigned int len,
 410                                        struct rds_info_iterator *iter,
 411                                        struct rds_info_lengths *lens)
 412 {
 413         rds_conn_message_info(sock, len, iter, lens, 1);
 414 }
 415
 416 static void rds_conn_message_info_retrans(struct socket *sock,
 417                                           unsigned int len,
 418                                           struct rds_info_iterator *iter,
 419                                           struct rds_info_lengths *lens)
 420 {
 421         rds_conn_message_info(sock, len, iter, lens, 0);
 422 }
 423
 424 void rds_for_each_conn_info(struct socket *sock, unsigned int len,
 425                           struct rds_info_iterator *iter,
 426                           struct rds_info_lengths *lens,
 427                           int (*visitor)(struct rds_connection *, void *),
 428                           size_t item_len)
 429 {
 430         uint64_t buffer[(item_len + 7) / 8];
 431         struct hlist_head *head;
 432         struct hlist_node *pos;
 433         struct hlist_node *tmp;
 434         struct rds_connection *conn;
 435         unsigned long flags;
 436         size_t i;
 437
 438         spin_lock_irqsave(&rds_conn_lock, flags);
 439
 440         lens->nr = 0;
 441         lens->each = item_len;
 442
 443         for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
 444              i++, head++) {
 445                 hlist_for_each_entry_safe(conn, pos, tmp, head, c_hash_node) {
 446
 447                         /* XXX no c_lock usage.. */
 448                         if (!visitor(conn, buffer))
 449                                 continue;
 450
 451                         /* We copy as much as we can fit in the buffer,
 452                          * but we count all items so that the caller
 453                          * can resize the buffer. */
 454                         if (len >= item_len) {
 455                                 rds_info_copy(iter, buffer, item_len);
 456                                 len -= item_len;
 457                         }
 458                         lens->nr++;
 459                 }
 460         }
 461
 462         spin_unlock_irqrestore(&rds_conn_lock, flags);
 463 }
 464 EXPORT_SYMBOL_GPL(rds_for_each_conn_info);
 465
 466 static int rds_conn_info_visitor(struct rds_connection *conn,
 467                                   void *buffer)
 468 {
 469         struct rds_info_connection *cinfo = buffer;
 470
 471         cinfo->next_tx_seq = conn->c_next_tx_seq;
 472         cinfo->next_rx_seq = conn->c_next_rx_seq;
 473         cinfo->laddr = conn->c_laddr;
 474         cinfo->faddr = conn->c_faddr;
 475         strncpy(cinfo->transport, conn->c_trans->t_name,
 476                 sizeof(cinfo->transport));
 477         cinfo->flags = 0;
 478
 479         rds_conn_info_set(cinfo->flags,
 480                           rds_conn_is_sending(conn), SENDING);
 481         /* XXX Future: return the state rather than these funky bits */
 482         rds_conn_info_set(cinfo->flags,
 483                           atomic_read(&conn->c_state) == RDS_CONN_CONNECTING,
 484                           CONNECTING);
 485         rds_conn_info_set(cinfo->flags,
 486                           atomic_read(&conn->c_state) == RDS_CONN_UP,
 487                           CONNECTED);
 488         return 1;
 489 }
 490
 491 static void rds_conn_info(struct socket *sock, unsigned int len,
 492                           struct rds_info_iterator *iter,
 493                           struct rds_info_lengths *lens)
 494 {
 495         rds_for_each_conn_info(sock, len, iter, lens,
 496                                 rds_conn_info_visitor,
 497                                 sizeof(struct rds_info_connection));
 498 }
 499
 500 int __init rds_conn_init(void)
 501 {
 502         rds_conn_slab = kmem_cache_create("rds_connection",
 503                                           sizeof(struct rds_connection),
 504                                           0, 0, NULL);
 505         if (!rds_conn_slab)
 506                 return -ENOMEM;
 507
 508         rds_info_register_func(RDS_INFO_CONNECTIONS, rds_conn_info);
 509         rds_info_register_func(RDS_INFO_SEND_MESSAGES,
 510                                rds_conn_message_info_send);
 511         rds_info_register_func(RDS_INFO_RETRANS_MESSAGES,
 512                                rds_conn_message_info_retrans);
 513
 514         return 0;
 515 }
 516
 517 void rds_conn_exit(void)
 518 {
 519         rds_loop_exit();
 520
 521         WARN_ON(!hlist_empty(rds_conn_hash));
 522
 523         kmem_cache_destroy(rds_conn_slab);
 524
 525         rds_info_deregister_func(RDS_INFO_CONNECTIONS, rds_conn_info);
 526         rds_info_deregister_func(RDS_INFO_SEND_MESSAGES,
 527                                  rds_conn_message_info_send);
 528         rds_info_deregister_func(RDS_INFO_RETRANS_MESSAGES,
 529                                  rds_conn_message_info_retrans);
 530 }
 531
 532 /*
 533  * Force a disconnect
 534  */
 535 void rds_conn_drop(struct rds_connection *conn)
 536 {
 537         atomic_set(&conn->c_state, RDS_CONN_ERROR);
 538         queue_work(rds_wq, &conn->c_down_w);
 539 }
 540 EXPORT_SYMBOL_GPL(rds_conn_drop);
 541
 542 /*
 543  * An error occurred on the connection
 544  */
 545 void
 546 __rds_conn_error(struct rds_connection *conn, const char *fmt, ...)
 547 {
 548         va_list ap;
 549
 550         va_start(ap, fmt);
 551         vprintk(fmt, ap);
 552         va_end(ap);
 553
 554         rds_conn_drop(conn);
 555 }