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 <net/inet_hashtables.h>
  36
  37 #include "rds.h"
  38 #include "loop.h"
  39 #include "rdma.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 inline int rds_conn_is_sending(struct rds_connection *conn)
  66 {
  67         int ret = 0;
  68
  69         if (!mutex_trylock(&conn->c_send_lock))
  70                 ret = 1;
  71         else
  72                 mutex_unlock(&conn->c_send_lock);
  73
  74         return ret;
  75 }
  76
  77 static struct rds_connection *rds_conn_lookup(struct hlist_head *head,
  78                                               __be32 laddr, __be32 faddr,
  79                                               struct rds_transport *trans)
  80 {
  81         struct rds_connection *conn, *ret = NULL;
  82         struct hlist_node *pos;
  83
  84         hlist_for_each_entry(conn, pos, head, c_hash_node) {
  85                 if (conn->c_faddr == faddr && conn->c_laddr == laddr &&
  86                                 conn->c_trans == trans) {
  87                         ret = conn;
  88                         break;
  89                 }
  90         }
  91         rdsdebug("returning conn %p for %pI4 -> %pI4\n", ret,
  92                  &laddr, &faddr);
  93         return ret;
  94 }
  95
  96 /*
  97  * This is called by transports as they're bringing down a connection.
  98  * It clears partial message state so that the transport can start sending
  99  * and receiving over this connection again in the future.  It is up to
 100  * the transport to have serialized this call with its send and recv.
 101  */
 102 void rds_conn_reset(struct rds_connection *conn)
 103 {
 104         rdsdebug("connection %pI4 to %pI4 reset\n",
 105           &conn->c_laddr, &conn->c_faddr);
 106
 107         rds_stats_inc(s_conn_reset);
 108         rds_send_reset(conn);
 109         conn->c_flags = 0;
 110
 111         /* Do not clear next_rx_seq here, else we cannot distinguish
 112          * retransmitted packets from new packets, and will hand all
 113          * of them to the application. That is not consistent with the
 114          * reliability guarantees of RDS. */
 115 }
 116
 117 /*
 118  * There is only every one 'conn' for a given pair of addresses in the
 119  * system at a time.  They contain messages to be retransmitted and so
 120  * span the lifetime of the actual underlying transport connections.
 121  *
 122  * For now they are not garbage collected once they're created.  They
 123  * are torn down as the module is removed, if ever.
 124  */
 125 static struct rds_connection *__rds_conn_create(__be32 laddr, __be32 faddr,
 126                                        struct rds_transport *trans, gfp_t gfp,
 127                                        int is_outgoing)
 128 {
 129         struct rds_connection *conn, *parent = NULL;
 130         struct hlist_head *head = rds_conn_bucket(laddr, faddr);
 131         unsigned long flags;
 132         int ret;
 133
 134         spin_lock_irqsave(&rds_conn_lock, flags);
 135         conn = rds_conn_lookup(head, laddr, faddr, trans);
 136         if (conn
 137          && conn->c_loopback
 138          && conn->c_trans != &rds_loop_transport
 139          && !is_outgoing) {
 140                 /* This is a looped back IB connection, and we're
 141                  * called by the code handling the incoming connect.
 142                  * We need a second connection object into which we
 143                  * can stick the other QP. */
 144                 parent = conn;
 145                 conn = parent->c_passive;
 146         }
 147         spin_unlock_irqrestore(&rds_conn_lock, flags);
 148         if (conn)
 149                 goto out;
 150
 151         conn = kmem_cache_zalloc(rds_conn_slab, gfp);
 152         if (conn == NULL) {
 153                 conn = ERR_PTR(-ENOMEM);
 154                 goto out;
 155         }
 156
 157         INIT_HLIST_NODE(&conn->c_hash_node);
 158         conn->c_laddr = laddr;
 159         conn->c_faddr = faddr;
 160         spin_lock_init(&conn->c_lock);
 161         conn->c_next_tx_seq = 1;
 162
 163         mutex_init(&conn->c_send_lock);
 164         INIT_LIST_HEAD(&conn->c_send_queue);
 165         INIT_LIST_HEAD(&conn->c_retrans);
 166
 167         ret = rds_cong_get_maps(conn);
 168         if (ret) {
 169                 kmem_cache_free(rds_conn_slab, conn);
 170                 conn = ERR_PTR(ret);
 171                 goto out;
 172         }
 173
 174         /*
 175          * This is where a connection becomes loopback.  If *any* RDS sockets
 176          * can bind to the destination address then we'd rather the messages
 177          * flow through loopback rather than either transport.
 178          */
 179         if (rds_trans_get_preferred(faddr)) {
 180                 conn->c_loopback = 1;
 181                 if (is_outgoing && trans->t_prefer_loopback) {
 182                         /* "outgoing" connection - and the transport
 183                          * says it wants the connection handled by the
 184                          * loopback transport. This is what TCP does.
 185                          */
 186                         trans = &rds_loop_transport;
 187                 }
 188         }
 189
 190         conn->c_trans = trans;
 191
 192         ret = trans->conn_alloc(conn, gfp);
 193         if (ret) {
 194                 kmem_cache_free(rds_conn_slab, conn);
 195                 conn = ERR_PTR(ret);
 196                 goto out;
 197         }
 198
 199         atomic_set(&conn->c_state, RDS_CONN_DOWN);
 200         conn->c_reconnect_jiffies = 0;
 201         INIT_DELAYED_WORK(&conn->c_send_w, rds_send_worker);
 202         INIT_DELAYED_WORK(&conn->c_recv_w, rds_recv_worker);
 203         INIT_DELAYED_WORK(&conn->c_conn_w, rds_connect_worker);
 204         INIT_WORK(&conn->c_down_w, rds_shutdown_worker);
 205         mutex_init(&conn->c_cm_lock);
 206         conn->c_flags = 0;
 207
 208         rdsdebug("allocated conn %p for %pI4 -> %pI4 over %s %s\n",
 209           conn, &laddr, &faddr,
 210           trans->t_name ? trans->t_name : "[unknown]",
 211           is_outgoing ? "(outgoing)" : "");
 212
 213         /*
 214          * Since we ran without holding the conn lock, someone could
 215          * have created the same conn (either normal or passive) in the
 216          * interim. We check while holding the lock. If we won, we complete
 217          * init and return our conn. If we lost, we rollback and return the
 218          * other one.
 219          */
 220         spin_lock_irqsave(&rds_conn_lock, flags);
 221         if (parent) {
 222                 /* Creating passive conn */
 223                 if (parent->c_passive) {
 224                         trans->conn_free(conn->c_transport_data);
 225                         kmem_cache_free(rds_conn_slab, conn);
 226                         conn = parent->c_passive;
 227                 } else {
 228                         parent->c_passive = conn;
 229                         rds_cong_add_conn(conn);
 230                         rds_conn_count++;
 231                 }
 232         } else {
 233                 /* Creating normal conn */
 234                 struct rds_connection *found;
 235
 236                 found = rds_conn_lookup(head, laddr, faddr, trans);
 237                 if (found) {
 238                         trans->conn_free(conn->c_transport_data);
 239                         kmem_cache_free(rds_conn_slab, conn);
 240                         conn = found;
 241                 } else {
 242                         hlist_add_head(&conn->c_hash_node, head);
 243                         rds_cong_add_conn(conn);
 244                         rds_conn_count++;
 245                 }
 246         }
 247         spin_unlock_irqrestore(&rds_conn_lock, flags);
 248
 249 out:
 250         return conn;
 251 }
 252
 253 struct rds_connection *rds_conn_create(__be32 laddr, __be32 faddr,
 254                                        struct rds_transport *trans, gfp_t gfp)
 255 {
 256         return __rds_conn_create(laddr, faddr, trans, gfp, 0);
 257 }
 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
 265 void rds_conn_destroy(struct rds_connection *conn)
 266 {
 267         struct rds_message *rm, *rtmp;
 268
 269         rdsdebug("freeing conn %p for %pI4 -> "
 270                  "%pI4\n", conn, &conn->c_laddr,
 271                  &conn->c_faddr);
 272
 273         hlist_del_init(&conn->c_hash_node);
 274
 275         /* wait for the rds thread to shut it down */
 276         atomic_set(&conn->c_state, RDS_CONN_ERROR);
 277         cancel_delayed_work(&conn->c_conn_w);
 278         queue_work(rds_wq, &conn->c_down_w);
 279         flush_workqueue(rds_wq);
 280
 281         /* tear down queued messages */
 282         list_for_each_entry_safe(rm, rtmp,
 283                                  &conn->c_send_queue,
 284                                  m_conn_item) {
 285                 list_del_init(&rm->m_conn_item);
 286                 BUG_ON(!list_empty(&rm->m_sock_item));
 287                 rds_message_put(rm);
 288         }
 289         if (conn->c_xmit_rm)
 290                 rds_message_put(conn->c_xmit_rm);
 291
 292         conn->c_trans->conn_free(conn->c_transport_data);
 293
 294         /*
 295          * The congestion maps aren't freed up here.  They're
 296          * freed by rds_cong_exit() after all the connections
 297          * have been freed.
 298          */
 299         rds_cong_remove_conn(conn);
 300
 301         BUG_ON(!list_empty(&conn->c_retrans));
 302         kmem_cache_free(rds_conn_slab, conn);
 303
 304         rds_conn_count--;
 305 }
 306
 307 static void rds_conn_message_info(struct socket *sock, unsigned int len,
 308                                   struct rds_info_iterator *iter,
 309                                   struct rds_info_lengths *lens,
 310                                   int want_send)
 311 {
 312         struct hlist_head *head;
 313         struct hlist_node *pos;
 314         struct list_head *list;
 315         struct rds_connection *conn;
 316         struct rds_message *rm;
 317         unsigned long flags;
 318         unsigned int total = 0;
 319         size_t i;
 320
 321         len /= sizeof(struct rds_info_message);
 322
 323         spin_lock_irqsave(&rds_conn_lock, flags);
 324
 325         for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
 326              i++, head++) {
 327                 hlist_for_each_entry(conn, pos, head, c_hash_node) {
 328                         if (want_send)
 329                                 list = &conn->c_send_queue;
 330                         else
 331                                 list = &conn->c_retrans;
 332
 333                         spin_lock(&conn->c_lock);
 334
 335                         /* XXX too lazy to maintain counts.. */
 336                         list_for_each_entry(rm, list, m_conn_item) {
 337                                 total++;
 338                                 if (total <= len)
 339                                         rds_inc_info_copy(&rm->m_inc, iter,
 340                                                           conn->c_laddr,
 341                                                           conn->c_faddr, 0);
 342                         }
 343
 344                         spin_unlock(&conn->c_lock);
 345                 }
 346         }
 347
 348         spin_unlock_irqrestore(&rds_conn_lock, flags);
 349
 350         lens->nr = total;
 351         lens->each = sizeof(struct rds_info_message);
 352 }
 353
 354 static void rds_conn_message_info_send(struct socket *sock, unsigned int len,
 355                                        struct rds_info_iterator *iter,
 356                                        struct rds_info_lengths *lens)
 357 {
 358         rds_conn_message_info(sock, len, iter, lens, 1);
 359 }
 360
 361 static void rds_conn_message_info_retrans(struct socket *sock,
 362                                           unsigned int len,
 363                                           struct rds_info_iterator *iter,
 364                                           struct rds_info_lengths *lens)
 365 {
 366         rds_conn_message_info(sock, len, iter, lens, 0);
 367 }
 368
 369 void rds_for_each_conn_info(struct socket *sock, unsigned int len,
 370                           struct rds_info_iterator *iter,
 371                           struct rds_info_lengths *lens,
 372                           int (*visitor)(struct rds_connection *, void *),
 373                           size_t item_len)
 374 {
 375         uint64_t buffer[(item_len + 7) / 8];
 376         struct hlist_head *head;
 377         struct hlist_node *pos;
 378         struct hlist_node *tmp;
 379         struct rds_connection *conn;
 380         unsigned long flags;
 381         size_t i;
 382
 383         spin_lock_irqsave(&rds_conn_lock, flags);
 384
 385         lens->nr = 0;
 386         lens->each = item_len;
 387
 388         for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
 389              i++, head++) {
 390                 hlist_for_each_entry_safe(conn, pos, tmp, head, c_hash_node) {
 391
 392                         /* XXX no c_lock usage.. */
 393                         if (!visitor(conn, buffer))
 394                                 continue;
 395
 396                         /* We copy as much as we can fit in the buffer,
 397                          * but we count all items so that the caller
 398                          * can resize the buffer. */
 399                         if (len >= item_len) {
 400                                 rds_info_copy(iter, buffer, item_len);
 401                                 len -= item_len;
 402                         }
 403                         lens->nr++;
 404                 }
 405         }
 406
 407         spin_unlock_irqrestore(&rds_conn_lock, flags);
 408 }
 409
 410 static int rds_conn_info_visitor(struct rds_connection *conn,
 411                                   void *buffer)
 412 {
 413         struct rds_info_connection *cinfo = buffer;
 414
 415         cinfo->next_tx_seq = conn->c_next_tx_seq;
 416         cinfo->next_rx_seq = conn->c_next_rx_seq;
 417         cinfo->laddr = conn->c_laddr;
 418         cinfo->faddr = conn->c_faddr;
 419         strncpy(cinfo->transport, conn->c_trans->t_name,
 420                 sizeof(cinfo->transport));
 421         cinfo->flags = 0;
 422
 423         rds_conn_info_set(cinfo->flags,
 424                           rds_conn_is_sending(conn), SENDING);
 425         /* XXX Future: return the state rather than these funky bits */
 426         rds_conn_info_set(cinfo->flags,
 427                           atomic_read(&conn->c_state) == RDS_CONN_CONNECTING,
 428                           CONNECTING);
 429         rds_conn_info_set(cinfo->flags,
 430                           atomic_read(&conn->c_state) == RDS_CONN_UP,
 431                           CONNECTED);
 432         return 1;
 433 }
 434
 435 static void rds_conn_info(struct socket *sock, unsigned int len,
 436                           struct rds_info_iterator *iter,
 437                           struct rds_info_lengths *lens)
 438 {
 439         rds_for_each_conn_info(sock, len, iter, lens,
 440                                 rds_conn_info_visitor,
 441                                 sizeof(struct rds_info_connection));
 442 }
 443
 444 int __init rds_conn_init(void)
 445 {
 446         rds_conn_slab = kmem_cache_create("rds_connection",
 447                                           sizeof(struct rds_connection),
 448                                           0, 0, NULL);
 449         if (rds_conn_slab == NULL)
 450                 return -ENOMEM;
 451
 452         rds_info_register_func(RDS_INFO_CONNECTIONS, rds_conn_info);
 453         rds_info_register_func(RDS_INFO_SEND_MESSAGES,
 454                                rds_conn_message_info_send);
 455         rds_info_register_func(RDS_INFO_RETRANS_MESSAGES,
 456                                rds_conn_message_info_retrans);
 457
 458         return 0;
 459 }
 460
 461 void rds_conn_exit(void)
 462 {
 463         rds_loop_exit();
 464
 465         WARN_ON(!hlist_empty(rds_conn_hash));
 466
 467         kmem_cache_destroy(rds_conn_slab);
 468
 469         rds_info_deregister_func(RDS_INFO_CONNECTIONS, rds_conn_info);
 470         rds_info_deregister_func(RDS_INFO_SEND_MESSAGES,
 471                                  rds_conn_message_info_send);
 472         rds_info_deregister_func(RDS_INFO_RETRANS_MESSAGES,
 473                                  rds_conn_message_info_retrans);
 474 }
 475
 476 /*
 477  * Force a disconnect
 478  */
 479 void rds_conn_drop(struct rds_connection *conn)
 480 {
 481         atomic_set(&conn->c_state, RDS_CONN_ERROR);
 482         queue_work(rds_wq, &conn->c_down_w);
 483 }
 484
 485 /*
 486  * An error occurred on the connection
 487  */
 488 void
 489 __rds_conn_error(struct rds_connection *conn, const char *fmt, ...)
 490 {
 491         va_list ap;
 492
 493         va_start(ap, fmt);
 494         vprintk(fmt, ap);
 495         va_end(ap);
 496
 497         rds_conn_drop(conn);
 498 }