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
  41 #define RDS_CONNECTION_HASH_BITS 12
  42 #define RDS_CONNECTION_HASH_ENTRIES (1 << RDS_CONNECTION_HASH_BITS)
  43 #define RDS_CONNECTION_HASH_MASK (RDS_CONNECTION_HASH_ENTRIES - 1)
  44
  45 /* converting this to RCU is a chore for another day.. */
  46 static DEFINE_SPINLOCK(rds_conn_lock);
  47 static unsigned long rds_conn_count;
  48 static struct hlist_head rds_conn_hash[RDS_CONNECTION_HASH_ENTRIES];
  49 static struct kmem_cache *rds_conn_slab;
  50
  51 static struct hlist_head *rds_conn_bucket(__be32 laddr, __be32 faddr)
  52 {
  53         /* Pass NULL, don't need struct net for hash */
  54         unsigned long hash = inet_ehashfn(NULL,
  55                                           be32_to_cpu(laddr), 0,
  56                                           be32_to_cpu(faddr), 0);
  57         return &rds_conn_hash[hash & RDS_CONNECTION_HASH_MASK];
  58 }
  59
  60 #define rds_conn_info_set(var, test, suffix) do {               \
  61         if (test)                                               \
  62                 var |= RDS_INFO_CONNECTION_FLAG_##suffix;       \
  63 } while (0)
  64
  65 static struct rds_connection *rds_conn_lookup(struct hlist_head *head,
  66                                               __be32 laddr, __be32 faddr,
  67                                               struct rds_transport *trans)
  68 {
  69         struct rds_connection *conn, *ret = NULL;
  70         struct hlist_node *pos;
  71
  72         hlist_for_each_entry(conn, pos, head, c_hash_node) {
  73                 if (conn->c_faddr == faddr && conn->c_laddr == laddr &&
  74                                 conn->c_trans == trans) {
  75                         ret = conn;
  76                         break;
  77                 }
  78         }
  79         rdsdebug("returning conn %p for %pI4 -> %pI4\n", ret,
  80                  &laddr, &faddr);
  81         return ret;
  82 }
  83
  84 /*
  85  * This is called by transports as they're bringing down a connection.
  86  * It clears partial message state so that the transport can start sending
  87  * and receiving over this connection again in the future.  It is up to
  88  * the transport to have serialized this call with its send and recv.
  89  */
  90 void rds_conn_reset(struct rds_connection *conn)
  91 {
  92         rdsdebug("connection %pI4 to %pI4 reset\n",
  93           &conn->c_laddr, &conn->c_faddr);
  94
  95         rds_stats_inc(s_conn_reset);
  96         rds_send_reset(conn);
  97         conn->c_flags = 0;
  98
  99         /* Do not clear next_rx_seq here, else we cannot distinguish
 100          * retransmitted packets from new packets, and will hand all
 101          * of them to the application. That is not consistent with the
 102          * reliability guarantees of RDS. */
 103 }
 104
 105 /*
 106  * There is only every one 'conn' for a given pair of addresses in the
 107  * system at a time.  They contain messages to be retransmitted and so
 108  * span the lifetime of the actual underlying transport connections.
 109  *
 110  * For now they are not garbage collected once they're created.  They
 111  * are torn down as the module is removed, if ever.
 112  */
 113 static struct rds_connection *__rds_conn_create(__be32 laddr, __be32 faddr,
 114                                        struct rds_transport *trans, gfp_t gfp,
 115                                        int is_outgoing)
 116 {
 117         struct rds_connection *conn, *parent = NULL;
 118         struct hlist_head *head = rds_conn_bucket(laddr, faddr);
 119         unsigned long flags;
 120         int ret;
 121
 122         spin_lock_irqsave(&rds_conn_lock, flags);
 123         conn = rds_conn_lookup(head, laddr, faddr, trans);
 124         if (conn && conn->c_loopback && conn->c_trans != &rds_loop_transport &&
 125             !is_outgoing) {
 126                 /* This is a looped back IB connection, and we're
 127                  * called by the code handling the incoming connect.
 128                  * We need a second connection object into which we
 129                  * can stick the other QP. */
 130                 parent = conn;
 131                 conn = parent->c_passive;
 132         }
 133         spin_unlock_irqrestore(&rds_conn_lock, flags);
 134         if (conn)
 135                 goto out;
 136
 137         conn = kmem_cache_zalloc(rds_conn_slab, gfp);
 138         if (!conn) {
 139                 conn = ERR_PTR(-ENOMEM);
 140                 goto out;
 141         }
 142
 143         INIT_HLIST_NODE(&conn->c_hash_node);
 144         conn->c_laddr = laddr;
 145         conn->c_faddr = faddr;
 146         spin_lock_init(&conn->c_lock);
 147         conn->c_next_tx_seq = 1;
 148
 149         spin_lock_init(&conn->c_send_lock);
 150         atomic_set(&conn->c_send_generation, 1);
 151         INIT_LIST_HEAD(&conn->c_send_queue);
 152         INIT_LIST_HEAD(&conn->c_retrans);
 153
 154         ret = rds_cong_get_maps(conn);
 155         if (ret) {
 156                 kmem_cache_free(rds_conn_slab, conn);
 157                 conn = ERR_PTR(ret);
 158                 goto out;
 159         }
 160
 161         /*
 162          * This is where a connection becomes loopback.  If *any* RDS sockets
 163          * can bind to the destination address then we'd rather the messages
 164          * flow through loopback rather than either transport.
 165          */
 166         if (rds_trans_get_preferred(faddr)) {
 167                 conn->c_loopback = 1;
 168                 if (is_outgoing && trans->t_prefer_loopback) {
 169                         /* "outgoing" connection - and the transport
 170                          * says it wants the connection handled by the
 171                          * loopback transport. This is what TCP does.
 172                          */
 173                         trans = &rds_loop_transport;
 174                 }
 175         }
 176
 177         conn->c_trans = trans;
 178
 179         ret = trans->conn_alloc(conn, gfp);
 180         if (ret) {
 181                 kmem_cache_free(rds_conn_slab, conn);
 182                 conn = ERR_PTR(ret);
 183                 goto out;
 184         }
 185
 186         atomic_set(&conn->c_state, RDS_CONN_DOWN);
 187         conn->c_reconnect_jiffies = 0;
 188         INIT_DELAYED_WORK(&conn->c_send_w, rds_send_worker);
 189         INIT_DELAYED_WORK(&conn->c_recv_w, rds_recv_worker);
 190         INIT_DELAYED_WORK(&conn->c_conn_w, rds_connect_worker);
 191         INIT_WORK(&conn->c_down_w, rds_shutdown_worker);
 192         mutex_init(&conn->c_cm_lock);
 193         conn->c_flags = 0;
 194
 195         rdsdebug("allocated conn %p for %pI4 -> %pI4 over %s %s\n",
 196           conn, &laddr, &faddr,
 197           trans->t_name ? trans->t_name : "[unknown]",
 198           is_outgoing ? "(outgoing)" : "");
 199
 200         /*
 201          * Since we ran without holding the conn lock, someone could
 202          * have created the same conn (either normal or passive) in the
 203          * interim. We check while holding the lock. If we won, we complete
 204          * init and return our conn. If we lost, we rollback and return the
 205          * other one.
 206          */
 207         spin_lock_irqsave(&rds_conn_lock, flags);
 208         if (parent) {
 209                 /* Creating passive conn */
 210                 if (parent->c_passive) {
 211                         trans->conn_free(conn->c_transport_data);
 212                         kmem_cache_free(rds_conn_slab, conn);
 213                         conn = parent->c_passive;
 214                 } else {
 215                         parent->c_passive = conn;
 216                         rds_cong_add_conn(conn);
 217                         rds_conn_count++;
 218                 }
 219         } else {
 220                 /* Creating normal conn */
 221                 struct rds_connection *found;
 222
 223                 found = rds_conn_lookup(head, laddr, faddr, trans);
 224                 if (found) {
 225                         trans->conn_free(conn->c_transport_data);
 226                         kmem_cache_free(rds_conn_slab, conn);
 227                         conn = found;
 228                 } else {
 229                         hlist_add_head(&conn->c_hash_node, head);
 230                         rds_cong_add_conn(conn);
 231                         rds_conn_count++;
 232                 }
 233         }
 234         spin_unlock_irqrestore(&rds_conn_lock, flags);
 235
 236 out:
 237         return conn;
 238 }
 239
 240 struct rds_connection *rds_conn_create(__be32 laddr, __be32 faddr,
 241                                        struct rds_transport *trans, gfp_t gfp)
 242 {
 243         return __rds_conn_create(laddr, faddr, trans, gfp, 0);
 244 }
 245 EXPORT_SYMBOL_GPL(rds_conn_create);
 246
 247 struct rds_connection *rds_conn_create_outgoing(__be32 laddr, __be32 faddr,
 248                                        struct rds_transport *trans, gfp_t gfp)
 249 {
 250         return __rds_conn_create(laddr, faddr, trans, gfp, 1);
 251 }
 252 EXPORT_SYMBOL_GPL(rds_conn_create_outgoing);
 253
 254 void rds_conn_shutdown(struct rds_connection *conn)
 255 {
 256         /* shut it down unless it's down already */
 257         if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_DOWN)) {
 258                 /*
 259                  * Quiesce the connection mgmt handlers before we start tearing
 260                  * things down. We don't hold the mutex for the entire
 261                  * duration of the shutdown operation, else we may be
 262                  * deadlocking with the CM handler. Instead, the CM event
 263                  * handler is supposed to check for state DISCONNECTING
 264                  */
 265                 mutex_lock(&conn->c_cm_lock);
 266                 if (!rds_conn_transition(conn, RDS_CONN_UP, RDS_CONN_DISCONNECTING)
 267                  && !rds_conn_transition(conn, RDS_CONN_ERROR, RDS_CONN_DISCONNECTING)) {
 268                         rds_conn_error(conn, "shutdown called in state %d\n",
 269                                         atomic_read(&conn->c_state));
 270                         mutex_unlock(&conn->c_cm_lock);
 271                         return;
 272                 }
 273                 mutex_unlock(&conn->c_cm_lock);
 274
 275                 /* verify everybody's out of rds_send_xmit() */
 276                 spin_lock_irq(&conn->c_send_lock);
 277                 spin_unlock_irq(&conn->c_send_lock);
 278
 279                 conn->c_trans->conn_shutdown(conn);
 280                 rds_conn_reset(conn);
 281
 282                 if (!rds_conn_transition(conn, RDS_CONN_DISCONNECTING, RDS_CONN_DOWN)) {
 283                         /* This can happen - eg when we're in the middle of tearing
 284                          * down the connection, and someone unloads the rds module.
 285                          * Quite reproduceable with loopback connections.
 286                          * Mostly harmless.
 287                          */
 288                         rds_conn_error(conn,
 289                                 "%s: failed to transition to state DOWN, "
 290                                 "current state is %d\n",
 291                                 __func__,
 292                                 atomic_read(&conn->c_state));
 293                         return;
 294                 }
 295         }
 296
 297         /* Then reconnect if it's still live.
 298          * The passive side of an IB loopback connection is never added
 299          * to the conn hash, so we never trigger a reconnect on this
 300          * conn - the reconnect is always triggered by the active peer. */
 301         cancel_delayed_work_sync(&conn->c_conn_w);
 302         if (!hlist_unhashed(&conn->c_hash_node))
 303                 rds_queue_reconnect(conn);
 304 }
 305
 306 /*
 307  * Stop and free a connection.
 308  */
 309 void rds_conn_destroy(struct rds_connection *conn)
 310 {
 311         struct rds_message *rm, *rtmp;
 312
 313         rdsdebug("freeing conn %p for %pI4 -> "
 314                  "%pI4\n", conn, &conn->c_laddr,
 315                  &conn->c_faddr);
 316
 317         hlist_del_init(&conn->c_hash_node);
 318
 319         /* wait for the rds thread to shut it down */
 320         atomic_set(&conn->c_state, RDS_CONN_ERROR);
 321         cancel_delayed_work(&conn->c_conn_w);
 322         queue_work(rds_wq, &conn->c_down_w);
 323         flush_workqueue(rds_wq);
 324
 325         /* tear down queued messages */
 326         list_for_each_entry_safe(rm, rtmp,
 327                                  &conn->c_send_queue,
 328                                  m_conn_item) {
 329                 list_del_init(&rm->m_conn_item);
 330                 BUG_ON(!list_empty(&rm->m_sock_item));
 331                 rds_message_put(rm);
 332         }
 333         if (conn->c_xmit_rm)
 334                 rds_message_put(conn->c_xmit_rm);
 335
 336         conn->c_trans->conn_free(conn->c_transport_data);
 337
 338         /*
 339          * The congestion maps aren't freed up here.  They're
 340          * freed by rds_cong_exit() after all the connections
 341          * have been freed.
 342          */
 343         rds_cong_remove_conn(conn);
 344
 345         BUG_ON(!list_empty(&conn->c_retrans));
 346         kmem_cache_free(rds_conn_slab, conn);
 347
 348         rds_conn_count--;
 349 }
 350 EXPORT_SYMBOL_GPL(rds_conn_destroy);
 351
 352 static void rds_conn_message_info(struct socket *sock, unsigned int len,
 353                                   struct rds_info_iterator *iter,
 354                                   struct rds_info_lengths *lens,
 355                                   int want_send)
 356 {
 357         struct hlist_head *head;
 358         struct hlist_node *pos;
 359         struct list_head *list;
 360         struct rds_connection *conn;
 361         struct rds_message *rm;
 362         unsigned long flags;
 363         unsigned int total = 0;
 364         size_t i;
 365
 366         len /= sizeof(struct rds_info_message);
 367
 368         spin_lock_irqsave(&rds_conn_lock, flags);
 369
 370         for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
 371              i++, head++) {
 372                 hlist_for_each_entry(conn, pos, head, c_hash_node) {
 373                         if (want_send)
 374                                 list = &conn->c_send_queue;
 375                         else
 376                                 list = &conn->c_retrans;
 377
 378                         spin_lock(&conn->c_lock);
 379
 380                         /* XXX too lazy to maintain counts.. */
 381                         list_for_each_entry(rm, list, m_conn_item) {
 382                                 total++;
 383                                 if (total <= len)
 384                                         rds_inc_info_copy(&rm->m_inc, iter,
 385                                                           conn->c_laddr,
 386                                                           conn->c_faddr, 0);
 387                         }
 388
 389                         spin_unlock(&conn->c_lock);
 390                 }
 391         }
 392
 393         spin_unlock_irqrestore(&rds_conn_lock, flags);
 394
 395         lens->nr = total;
 396         lens->each = sizeof(struct rds_info_message);
 397 }
 398
 399 static void rds_conn_message_info_send(struct socket *sock, unsigned int len,
 400                                        struct rds_info_iterator *iter,
 401                                        struct rds_info_lengths *lens)
 402 {
 403         rds_conn_message_info(sock, len, iter, lens, 1);
 404 }
 405
 406 static void rds_conn_message_info_retrans(struct socket *sock,
 407                                           unsigned int len,
 408                                           struct rds_info_iterator *iter,
 409                                           struct rds_info_lengths *lens)
 410 {
 411         rds_conn_message_info(sock, len, iter, lens, 0);
 412 }
 413
 414 void rds_for_each_conn_info(struct socket *sock, unsigned int len,
 415                           struct rds_info_iterator *iter,
 416                           struct rds_info_lengths *lens,
 417                           int (*visitor)(struct rds_connection *, void *),
 418                           size_t item_len)
 419 {
 420         uint64_t buffer[(item_len + 7) / 8];
 421         struct hlist_head *head;
 422         struct hlist_node *pos;
 423         struct hlist_node *tmp;
 424         struct rds_connection *conn;
 425         unsigned long flags;
 426         size_t i;
 427
 428         spin_lock_irqsave(&rds_conn_lock, flags);
 429
 430         lens->nr = 0;
 431         lens->each = item_len;
 432
 433         for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
 434              i++, head++) {
 435                 hlist_for_each_entry_safe(conn, pos, tmp, head, c_hash_node) {
 436
 437                         /* XXX no c_lock usage.. */
 438                         if (!visitor(conn, buffer))
 439                                 continue;
 440
 441                         /* We copy as much as we can fit in the buffer,
 442                          * but we count all items so that the caller
 443                          * can resize the buffer. */
 444                         if (len >= item_len) {
 445                                 rds_info_copy(iter, buffer, item_len);
 446                                 len -= item_len;
 447                         }
 448                         lens->nr++;
 449                 }
 450         }
 451
 452         spin_unlock_irqrestore(&rds_conn_lock, flags);
 453 }
 454 EXPORT_SYMBOL_GPL(rds_for_each_conn_info);
 455
 456 static int rds_conn_info_visitor(struct rds_connection *conn,
 457                                   void *buffer)
 458 {
 459         struct rds_info_connection *cinfo = buffer;
 460
 461         cinfo->next_tx_seq = conn->c_next_tx_seq;
 462         cinfo->next_rx_seq = conn->c_next_rx_seq;
 463         cinfo->laddr = conn->c_laddr;
 464         cinfo->faddr = conn->c_faddr;
 465         strncpy(cinfo->transport, conn->c_trans->t_name,
 466                 sizeof(cinfo->transport));
 467         cinfo->flags = 0;
 468
 469         rds_conn_info_set(cinfo->flags,
 470                           spin_is_locked(&conn->c_send_lock), SENDING);
 471         /* XXX Future: return the state rather than these funky bits */
 472         rds_conn_info_set(cinfo->flags,
 473                           atomic_read(&conn->c_state) == RDS_CONN_CONNECTING,
 474                           CONNECTING);
 475         rds_conn_info_set(cinfo->flags,
 476                           atomic_read(&conn->c_state) == RDS_CONN_UP,
 477                           CONNECTED);
 478         return 1;
 479 }
 480
 481 static void rds_conn_info(struct socket *sock, unsigned int len,
 482                           struct rds_info_iterator *iter,
 483                           struct rds_info_lengths *lens)
 484 {
 485         rds_for_each_conn_info(sock, len, iter, lens,
 486                                 rds_conn_info_visitor,
 487                                 sizeof(struct rds_info_connection));
 488 }
 489
 490 int __init rds_conn_init(void)
 491 {
 492         rds_conn_slab = kmem_cache_create("rds_connection",
 493                                           sizeof(struct rds_connection),
 494                                           0, 0, NULL);
 495         if (!rds_conn_slab)
 496                 return -ENOMEM;
 497
 498         rds_info_register_func(RDS_INFO_CONNECTIONS, rds_conn_info);
 499         rds_info_register_func(RDS_INFO_SEND_MESSAGES,
 500                                rds_conn_message_info_send);
 501         rds_info_register_func(RDS_INFO_RETRANS_MESSAGES,
 502                                rds_conn_message_info_retrans);
 503
 504         return 0;
 505 }
 506
 507 void rds_conn_exit(void)
 508 {
 509         rds_loop_exit();
 510
 511         WARN_ON(!hlist_empty(rds_conn_hash));
 512
 513         kmem_cache_destroy(rds_conn_slab);
 514
 515         rds_info_deregister_func(RDS_INFO_CONNECTIONS, rds_conn_info);
 516         rds_info_deregister_func(RDS_INFO_SEND_MESSAGES,
 517                                  rds_conn_message_info_send);
 518         rds_info_deregister_func(RDS_INFO_RETRANS_MESSAGES,
 519                                  rds_conn_message_info_retrans);
 520 }
 521
 522 /*
 523  * Force a disconnect
 524  */
 525 void rds_conn_drop(struct rds_connection *conn)
 526 {
 527         atomic_set(&conn->c_state, RDS_CONN_ERROR);
 528         queue_work(rds_wq, &conn->c_down_w);
 529 }
 530 EXPORT_SYMBOL_GPL(rds_conn_drop);
 531
 532 /*
 533  * An error occurred on the connection
 534  */
 535 void
 536 __rds_conn_error(struct rds_connection *conn, const char *fmt, ...)
 537 {
 538         va_list ap;
 539
 540         va_start(ap, fmt);
 541         vprintk(fmt, ap);
 542         va_end(ap);
 543
 544         rds_conn_drop(conn);
 545 }