net/rds/connection.c

   1 /*
   2  * Copyright (c) 2006, 2018 Oracle and/or its affiliates. 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/ipv6.h>
  38 #include <net/inet6_hashtables.h>
  39 #include <net/addrconf.h>
  40
  41 #include "rds.h"
  42 #include "loop.h"
  43
  44 #define RDS_CONNECTION_HASH_BITS 12
  45 #define RDS_CONNECTION_HASH_ENTRIES (1 << RDS_CONNECTION_HASH_BITS)
  46 #define RDS_CONNECTION_HASH_MASK (RDS_CONNECTION_HASH_ENTRIES - 1)
  47
  48 /* converting this to RCU is a chore for another day.. */
  49 static DEFINE_SPINLOCK(rds_conn_lock);
  50 static unsigned long rds_conn_count;
  51 static struct hlist_head rds_conn_hash[RDS_CONNECTION_HASH_ENTRIES];
  52 static struct kmem_cache *rds_conn_slab;
  53
  54 static struct hlist_head *rds_conn_bucket(const struct in6_addr *laddr,
  55                                           const struct in6_addr *faddr)
  56 {
  57         static u32 rds6_hash_secret __read_mostly;
  58         static u32 rds_hash_secret __read_mostly;
  59
  60         u32 lhash, fhash, hash;
  61
  62         net_get_random_once(&rds_hash_secret, sizeof(rds_hash_secret));
  63         net_get_random_once(&rds6_hash_secret, sizeof(rds6_hash_secret));
  64
  65         lhash = (__force u32)laddr->s6_addr32[3];
  66 #if IS_ENABLED(CONFIG_IPV6)
  67         fhash = __ipv6_addr_jhash(faddr, rds6_hash_secret);
  68 #else
  69         fhash = (__force u32)faddr->s6_addr32[3];
  70 #endif
  71         hash = __inet_ehashfn(lhash, 0, fhash, 0, rds_hash_secret);
  72
  73         return &rds_conn_hash[hash & RDS_CONNECTION_HASH_MASK];
  74 }
  75
  76 #define rds_conn_info_set(var, test, suffix) do {               \
  77         if (test)                                               \
  78                 var |= RDS_INFO_CONNECTION_FLAG_##suffix;       \
  79 } while (0)
  80
  81 /* rcu read lock must be held or the connection spinlock */
  82 static struct rds_connection *rds_conn_lookup(struct net *net,
  83                                               struct hlist_head *head,
  84                                               const struct in6_addr *laddr,
  85                                               const struct in6_addr *faddr,
  86                                               struct rds_transport *trans,
  87                                               int dev_if)
  88 {
  89         struct rds_connection *conn, *ret = NULL;
  90
  91         hlist_for_each_entry_rcu(conn, head, c_hash_node) {
  92                 if (ipv6_addr_equal(&conn->c_faddr, faddr) &&
  93                     ipv6_addr_equal(&conn->c_laddr, laddr) &&
  94                     conn->c_trans == trans &&
  95                     net == rds_conn_net(conn) &&
  96                     conn->c_dev_if == dev_if) {
  97                         ret = conn;
  98                         break;
  99                 }
 100         }
 101         rdsdebug("returning conn %p for %pI6c -> %pI6c\n", ret,
 102                  laddr, faddr);
 103         return ret;
 104 }
 105
 106 /*
 107  * This is called by transports as they're bringing down a connection.
 108  * It clears partial message state so that the transport can start sending
 109  * and receiving over this connection again in the future.  It is up to
 110  * the transport to have serialized this call with its send and recv.
 111  */
 112 static void rds_conn_path_reset(struct rds_conn_path *cp)
 113 {
 114         struct rds_connection *conn = cp->cp_conn;
 115
 116         rdsdebug("connection %pI6c to %pI6c reset\n",
 117                  &conn->c_laddr, &conn->c_faddr);
 118
 119         rds_stats_inc(s_conn_reset);
 120         rds_send_path_reset(cp);
 121         cp->cp_flags = 0;
 122
 123         /* Do not clear next_rx_seq here, else we cannot distinguish
 124          * retransmitted packets from new packets, and will hand all
 125          * of them to the application. That is not consistent with the
 126          * reliability guarantees of RDS. */
 127 }
 128
 129 static void __rds_conn_path_init(struct rds_connection *conn,
 130                                  struct rds_conn_path *cp, bool is_outgoing)
 131 {
 132         spin_lock_init(&cp->cp_lock);
 133         cp->cp_next_tx_seq = 1;
 134         init_waitqueue_head(&cp->cp_waitq);
 135         INIT_LIST_HEAD(&cp->cp_send_queue);
 136         INIT_LIST_HEAD(&cp->cp_retrans);
 137
 138         cp->cp_conn = conn;
 139         atomic_set(&cp->cp_state, RDS_CONN_DOWN);
 140         cp->cp_send_gen = 0;
 141         cp->cp_reconnect_jiffies = 0;
 142         cp->cp_conn->c_proposed_version = RDS_PROTOCOL_VERSION;
 143         INIT_DELAYED_WORK(&cp->cp_send_w, rds_send_worker);
 144         INIT_DELAYED_WORK(&cp->cp_recv_w, rds_recv_worker);
 145         INIT_DELAYED_WORK(&cp->cp_conn_w, rds_connect_worker);
 146         INIT_WORK(&cp->cp_down_w, rds_shutdown_worker);
 147         mutex_init(&cp->cp_cm_lock);
 148         cp->cp_flags = 0;
 149 }
 150
 151 /*
 152  * There is only every one 'conn' for a given pair of addresses in the
 153  * system at a time.  They contain messages to be retransmitted and so
 154  * span the lifetime of the actual underlying transport connections.
 155  *
 156  * For now they are not garbage collected once they're created.  They
 157  * are torn down as the module is removed, if ever.
 158  */
 159 static struct rds_connection *__rds_conn_create(struct net *net,
 160                                                 const struct in6_addr *laddr,
 161                                                 const struct in6_addr *faddr,
 162                                                 struct rds_transport *trans,
 163                                                 gfp_t gfp,
 164                                                 int is_outgoing,
 165                                                 int dev_if)
 166 {
 167         struct rds_connection *conn, *parent = NULL;
 168         struct hlist_head *head = rds_conn_bucket(laddr, faddr);
 169         struct rds_transport *loop_trans;
 170         unsigned long flags;
 171         int ret, i;
 172         int npaths = (trans->t_mp_capable ? RDS_MPATH_WORKERS : 1);
 173
 174         rcu_read_lock();
 175         conn = rds_conn_lookup(net, head, laddr, faddr, trans, dev_if);
 176         if (conn &&
 177             conn->c_loopback &&
 178             conn->c_trans != &rds_loop_transport &&
 179             ipv6_addr_equal(laddr, faddr) &&
 180             !is_outgoing) {
 181                 /* This is a looped back IB connection, and we're
 182                  * called by the code handling the incoming connect.
 183                  * We need a second connection object into which we
 184                  * can stick the other QP. */
 185                 parent = conn;
 186                 conn = parent->c_passive;
 187         }
 188         rcu_read_unlock();
 189         if (conn)
 190                 goto out;
 191
 192         conn = kmem_cache_zalloc(rds_conn_slab, gfp);
 193         if (!conn) {
 194                 conn = ERR_PTR(-ENOMEM);
 195                 goto out;
 196         }
 197         conn->c_path = kcalloc(npaths, sizeof(struct rds_conn_path), gfp);
 198         if (!conn->c_path) {
 199                 kmem_cache_free(rds_conn_slab, conn);
 200                 conn = ERR_PTR(-ENOMEM);
 201                 goto out;
 202         }
 203
 204         INIT_HLIST_NODE(&conn->c_hash_node);
 205         conn->c_laddr = *laddr;
 206         conn->c_isv6 = !ipv6_addr_v4mapped(laddr);
 207         conn->c_faddr = *faddr;
 208         conn->c_dev_if = dev_if;
 209
 210 #if IS_ENABLED(CONFIG_IPV6)
 211         /* If the local address is link local, set c_bound_if to be the
 212          * index used for this connection.  Otherwise, set it to 0 as
 213          * the socket is not bound to an interface.  c_bound_if is used
 214          * to look up a socket when a packet is received
 215          */
 216         if (ipv6_addr_type(laddr) & IPV6_ADDR_LINKLOCAL)
 217                 conn->c_bound_if = dev_if;
 218         else
 219 #endif
 220                 conn->c_bound_if = 0;
 221
 222         rds_conn_net_set(conn, net);
 223
 224         ret = rds_cong_get_maps(conn);
 225         if (ret) {
 226                 kfree(conn->c_path);
 227                 kmem_cache_free(rds_conn_slab, conn);
 228                 conn = ERR_PTR(ret);
 229                 goto out;
 230         }
 231
 232         /*
 233          * This is where a connection becomes loopback.  If *any* RDS sockets
 234          * can bind to the destination address then we'd rather the messages
 235          * flow through loopback rather than either transport.
 236          */
 237         loop_trans = rds_trans_get_preferred(net, faddr, conn->c_dev_if);
 238         if (loop_trans) {
 239                 rds_trans_put(loop_trans);
 240                 conn->c_loopback = 1;
 241                 if (is_outgoing && trans->t_prefer_loopback) {
 242                         /* "outgoing" connection - and the transport
 243                          * says it wants the connection handled by the
 244                          * loopback transport. This is what TCP does.
 245                          */
 246                         trans = &rds_loop_transport;
 247                 }
 248         }
 249
 250         conn->c_trans = trans;
 251
 252         init_waitqueue_head(&conn->c_hs_waitq);
 253         for (i = 0; i < npaths; i++) {
 254                 __rds_conn_path_init(conn, &conn->c_path[i],
 255                                      is_outgoing);
 256                 conn->c_path[i].cp_index = i;
 257         }
 258         rcu_read_lock();
 259         if (rds_destroy_pending(conn))
 260                 ret = -ENETDOWN;
 261         else
 262                 ret = trans->conn_alloc(conn, GFP_ATOMIC);
 263         if (ret) {
 264                 rcu_read_unlock();
 265                 kfree(conn->c_path);
 266                 kmem_cache_free(rds_conn_slab, conn);
 267                 conn = ERR_PTR(ret);
 268                 goto out;
 269         }
 270
 271         rdsdebug("allocated conn %p for %pI6c -> %pI6c over %s %s\n",
 272                  conn, laddr, faddr,
 273                  strnlen(trans->t_name, sizeof(trans->t_name)) ?
 274                  trans->t_name : "[unknown]", is_outgoing ? "(outgoing)" : "");
 275
 276         /*
 277          * Since we ran without holding the conn lock, someone could
 278          * have created the same conn (either normal or passive) in the
 279          * interim. We check while holding the lock. If we won, we complete
 280          * init and return our conn. If we lost, we rollback and return the
 281          * other one.
 282          */
 283         spin_lock_irqsave(&rds_conn_lock, flags);
 284         if (parent) {
 285                 /* Creating passive conn */
 286                 if (parent->c_passive) {
 287                         trans->conn_free(conn->c_path[0].cp_transport_data);
 288                         kfree(conn->c_path);
 289                         kmem_cache_free(rds_conn_slab, conn);
 290                         conn = parent->c_passive;
 291                 } else {
 292                         parent->c_passive = conn;
 293                         rds_cong_add_conn(conn);
 294                         rds_conn_count++;
 295                 }
 296         } else {
 297                 /* Creating normal conn */
 298                 struct rds_connection *found;
 299
 300                 found = rds_conn_lookup(net, head, laddr, faddr, trans,
 301                                         dev_if);
 302                 if (found) {
 303                         struct rds_conn_path *cp;
 304                         int i;
 305
 306                         for (i = 0; i < npaths; i++) {
 307                                 cp = &conn->c_path[i];
 308                                 /* The ->conn_alloc invocation may have
 309                                  * allocated resource for all paths, so all
 310                                  * of them may have to be freed here.
 311                                  */
 312                                 if (cp->cp_transport_data)
 313                                         trans->conn_free(cp->cp_transport_data);
 314                         }
 315                         kfree(conn->c_path);
 316                         kmem_cache_free(rds_conn_slab, conn);
 317                         conn = found;
 318                 } else {
 319                         conn->c_my_gen_num = rds_gen_num;
 320                         conn->c_peer_gen_num = 0;
 321                         hlist_add_head_rcu(&conn->c_hash_node, head);
 322                         rds_cong_add_conn(conn);
 323                         rds_conn_count++;
 324                 }
 325         }
 326         spin_unlock_irqrestore(&rds_conn_lock, flags);
 327         rcu_read_unlock();
 328
 329 out:
 330         return conn;
 331 }
 332
 333 struct rds_connection *rds_conn_create(struct net *net,
 334                                        const struct in6_addr *laddr,
 335                                        const struct in6_addr *faddr,
 336                                        struct rds_transport *trans, gfp_t gfp,
 337                                        int dev_if)
 338 {
 339         return __rds_conn_create(net, laddr, faddr, trans, gfp, 0, dev_if);
 340 }
 341 EXPORT_SYMBOL_GPL(rds_conn_create);
 342
 343 struct rds_connection *rds_conn_create_outgoing(struct net *net,
 344                                                 const struct in6_addr *laddr,
 345                                                 const struct in6_addr *faddr,
 346                                                 struct rds_transport *trans,
 347                                                 gfp_t gfp, int dev_if)
 348 {
 349         return __rds_conn_create(net, laddr, faddr, trans, gfp, 1, dev_if);
 350 }
 351 EXPORT_SYMBOL_GPL(rds_conn_create_outgoing);
 352
 353 void rds_conn_shutdown(struct rds_conn_path *cp)
 354 {
 355         struct rds_connection *conn = cp->cp_conn;
 356
 357         /* shut it down unless it's down already */
 358         if (!rds_conn_path_transition(cp, RDS_CONN_DOWN, RDS_CONN_DOWN)) {
 359                 /*
 360                  * Quiesce the connection mgmt handlers before we start tearing
 361                  * things down. We don't hold the mutex for the entire
 362                  * duration of the shutdown operation, else we may be
 363                  * deadlocking with the CM handler. Instead, the CM event
 364                  * handler is supposed to check for state DISCONNECTING
 365                  */
 366                 mutex_lock(&cp->cp_cm_lock);
 367                 if (!rds_conn_path_transition(cp, RDS_CONN_UP,
 368                                               RDS_CONN_DISCONNECTING) &&
 369                     !rds_conn_path_transition(cp, RDS_CONN_ERROR,
 370                                               RDS_CONN_DISCONNECTING)) {
 371                         rds_conn_path_error(cp,
 372                                             "shutdown called in state %d\n",
 373                                             atomic_read(&cp->cp_state));
 374                         mutex_unlock(&cp->cp_cm_lock);
 375                         return;
 376                 }
 377                 mutex_unlock(&cp->cp_cm_lock);
 378
 379                 wait_event(cp->cp_waitq,
 380                            !test_bit(RDS_IN_XMIT, &cp->cp_flags));
 381                 wait_event(cp->cp_waitq,
 382                            !test_bit(RDS_RECV_REFILL, &cp->cp_flags));
 383
 384                 conn->c_trans->conn_path_shutdown(cp);
 385                 rds_conn_path_reset(cp);
 386
 387                 if (!rds_conn_path_transition(cp, RDS_CONN_DISCONNECTING,
 388                                               RDS_CONN_DOWN) &&
 389                     !rds_conn_path_transition(cp, RDS_CONN_ERROR,
 390                                               RDS_CONN_DOWN)) {
 391                         /* This can happen - eg when we're in the middle of tearing
 392                          * down the connection, and someone unloads the rds module.
 393                          * Quite reproducible with loopback connections.
 394                          * Mostly harmless.
 395                          *
 396                          * Note that this also happens with rds-tcp because
 397                          * we could have triggered rds_conn_path_drop in irq
 398                          * mode from rds_tcp_state change on the receipt of
 399                          * a FIN, thus we need to recheck for RDS_CONN_ERROR
 400                          * here.
 401                          */
 402                         rds_conn_path_error(cp, "%s: failed to transition "
 403                                             "to state DOWN, current state "
 404                                             "is %d\n", __func__,
 405                                             atomic_read(&cp->cp_state));
 406                         return;
 407                 }
 408         }
 409
 410         /* Then reconnect if it's still live.
 411          * The passive side of an IB loopback connection is never added
 412          * to the conn hash, so we never trigger a reconnect on this
 413          * conn - the reconnect is always triggered by the active peer. */
 414         cancel_delayed_work_sync(&cp->cp_conn_w);
 415         rcu_read_lock();
 416         if (!hlist_unhashed(&conn->c_hash_node)) {
 417                 rcu_read_unlock();
 418                 rds_queue_reconnect(cp);
 419         } else {
 420                 rcu_read_unlock();
 421         }
 422 }
 423
 424 /* destroy a single rds_conn_path. rds_conn_destroy() iterates over
 425  * all paths using rds_conn_path_destroy()
 426  */
 427 static void rds_conn_path_destroy(struct rds_conn_path *cp)
 428 {
 429         struct rds_message *rm, *rtmp;
 430
 431         if (!cp->cp_transport_data)
 432                 return;
 433
 434         /* make sure lingering queued work won't try to ref the conn */
 435         cancel_delayed_work_sync(&cp->cp_send_w);
 436         cancel_delayed_work_sync(&cp->cp_recv_w);
 437
 438         rds_conn_path_drop(cp, true);
 439         flush_work(&cp->cp_down_w);
 440
 441         /* tear down queued messages */
 442         list_for_each_entry_safe(rm, rtmp,
 443                                  &cp->cp_send_queue,
 444                                  m_conn_item) {
 445                 list_del_init(&rm->m_conn_item);
 446                 BUG_ON(!list_empty(&rm->m_sock_item));
 447                 rds_message_put(rm);
 448         }
 449         if (cp->cp_xmit_rm)
 450                 rds_message_put(cp->cp_xmit_rm);
 451
 452         WARN_ON(delayed_work_pending(&cp->cp_send_w));
 453         WARN_ON(delayed_work_pending(&cp->cp_recv_w));
 454         WARN_ON(delayed_work_pending(&cp->cp_conn_w));
 455         WARN_ON(work_pending(&cp->cp_down_w));
 456
 457         cp->cp_conn->c_trans->conn_free(cp->cp_transport_data);
 458 }
 459
 460 /*
 461  * Stop and free a connection.
 462  *
 463  * This can only be used in very limited circumstances.  It assumes that once
 464  * the conn has been shutdown that no one else is referencing the connection.
 465  * We can only ensure this in the rmmod path in the current code.
 466  */
 467 void rds_conn_destroy(struct rds_connection *conn)
 468 {
 469         unsigned long flags;
 470         int i;
 471         struct rds_conn_path *cp;
 472         int npaths = (conn->c_trans->t_mp_capable ? RDS_MPATH_WORKERS : 1);
 473
 474         rdsdebug("freeing conn %p for %pI4 -> "
 475                  "%pI4\n", conn, &conn->c_laddr,
 476                  &conn->c_faddr);
 477
 478         /* Ensure conn will not be scheduled for reconnect */
 479         spin_lock_irq(&rds_conn_lock);
 480         hlist_del_init_rcu(&conn->c_hash_node);
 481         spin_unlock_irq(&rds_conn_lock);
 482         synchronize_rcu();
 483
 484         /* shut the connection down */
 485         for (i = 0; i < npaths; i++) {
 486                 cp = &conn->c_path[i];
 487                 rds_conn_path_destroy(cp);
 488                 BUG_ON(!list_empty(&cp->cp_retrans));
 489         }
 490
 491         /*
 492          * The congestion maps aren't freed up here.  They're
 493          * freed by rds_cong_exit() after all the connections
 494          * have been freed.
 495          */
 496         rds_cong_remove_conn(conn);
 497
 498         kfree(conn->c_path);
 499         kmem_cache_free(rds_conn_slab, conn);
 500
 501         spin_lock_irqsave(&rds_conn_lock, flags);
 502         rds_conn_count--;
 503         spin_unlock_irqrestore(&rds_conn_lock, flags);
 504 }
 505 EXPORT_SYMBOL_GPL(rds_conn_destroy);
 506
 507 static void __rds_inc_msg_cp(struct rds_incoming *inc,
 508                              struct rds_info_iterator *iter,
 509                              void *saddr, void *daddr, int flip, bool isv6)
 510 {
 511 #if IS_ENABLED(CONFIG_IPV6)
 512         if (isv6)
 513                 rds6_inc_info_copy(inc, iter, saddr, daddr, flip);
 514         else
 515 #endif
 516                 rds_inc_info_copy(inc, iter, *(__be32 *)saddr,
 517                                   *(__be32 *)daddr, flip);
 518 }
 519
 520 static void rds_conn_message_info_cmn(struct socket *sock, unsigned int len,
 521                                       struct rds_info_iterator *iter,
 522                                       struct rds_info_lengths *lens,
 523                                       int want_send, bool isv6)
 524 {
 525         struct hlist_head *head;
 526         struct list_head *list;
 527         struct rds_connection *conn;
 528         struct rds_message *rm;
 529         unsigned int total = 0;
 530         unsigned long flags;
 531         size_t i;
 532         int j;
 533
 534         if (isv6)
 535                 len /= sizeof(struct rds6_info_message);
 536         else
 537                 len /= sizeof(struct rds_info_message);
 538
 539         rcu_read_lock();
 540
 541         for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
 542              i++, head++) {
 543                 hlist_for_each_entry_rcu(conn, head, c_hash_node) {
 544                         struct rds_conn_path *cp;
 545                         int npaths;
 546
 547                         if (!isv6 && conn->c_isv6)
 548                                 continue;
 549
 550                         npaths = (conn->c_trans->t_mp_capable ?
 551                                  RDS_MPATH_WORKERS : 1);
 552
 553                         for (j = 0; j < npaths; j++) {
 554                                 cp = &conn->c_path[j];
 555                                 if (want_send)
 556                                         list = &cp->cp_send_queue;
 557                                 else
 558                                         list = &cp->cp_retrans;
 559
 560                                 spin_lock_irqsave(&cp->cp_lock, flags);
 561
 562                                 /* XXX too lazy to maintain counts.. */
 563                                 list_for_each_entry(rm, list, m_conn_item) {
 564                                         total++;
 565                                         if (total <= len)
 566                                                 __rds_inc_msg_cp(&rm->m_inc,
 567                                                                  iter,
 568                                                                  &conn->c_laddr,
 569                                                                  &conn->c_faddr,
 570                                                                  0, isv6);
 571                                 }
 572
 573                                 spin_unlock_irqrestore(&cp->cp_lock, flags);
 574                         }
 575                 }
 576         }
 577         rcu_read_unlock();
 578
 579         lens->nr = total;
 580         if (isv6)
 581                 lens->each = sizeof(struct rds6_info_message);
 582         else
 583                 lens->each = sizeof(struct rds_info_message);
 584 }
 585
 586 static void rds_conn_message_info(struct socket *sock, unsigned int len,
 587                                   struct rds_info_iterator *iter,
 588                                   struct rds_info_lengths *lens,
 589                                   int want_send)
 590 {
 591         rds_conn_message_info_cmn(sock, len, iter, lens, want_send, false);
 592 }
 593
 594 #if IS_ENABLED(CONFIG_IPV6)
 595 static void rds6_conn_message_info(struct socket *sock, unsigned int len,
 596                                    struct rds_info_iterator *iter,
 597                                    struct rds_info_lengths *lens,
 598                                    int want_send)
 599 {
 600         rds_conn_message_info_cmn(sock, len, iter, lens, want_send, true);
 601 }
 602 #endif
 603
 604 static void rds_conn_message_info_send(struct socket *sock, unsigned int len,
 605                                        struct rds_info_iterator *iter,
 606                                        struct rds_info_lengths *lens)
 607 {
 608         rds_conn_message_info(sock, len, iter, lens, 1);
 609 }
 610
 611 #if IS_ENABLED(CONFIG_IPV6)
 612 static void rds6_conn_message_info_send(struct socket *sock, unsigned int len,
 613                                         struct rds_info_iterator *iter,
 614                                         struct rds_info_lengths *lens)
 615 {
 616         rds6_conn_message_info(sock, len, iter, lens, 1);
 617 }
 618 #endif
 619
 620 static void rds_conn_message_info_retrans(struct socket *sock,
 621                                           unsigned int len,
 622                                           struct rds_info_iterator *iter,
 623                                           struct rds_info_lengths *lens)
 624 {
 625         rds_conn_message_info(sock, len, iter, lens, 0);
 626 }
 627
 628 #if IS_ENABLED(CONFIG_IPV6)
 629 static void rds6_conn_message_info_retrans(struct socket *sock,
 630                                            unsigned int len,
 631                                            struct rds_info_iterator *iter,
 632                                            struct rds_info_lengths *lens)
 633 {
 634         rds6_conn_message_info(sock, len, iter, lens, 0);
 635 }
 636 #endif
 637
 638 void rds_for_each_conn_info(struct socket *sock, unsigned int len,
 639                           struct rds_info_iterator *iter,
 640                           struct rds_info_lengths *lens,
 641                           int (*visitor)(struct rds_connection *, void *),
 642                           u64 *buffer,
 643                           size_t item_len)
 644 {
 645         struct hlist_head *head;
 646         struct rds_connection *conn;
 647         size_t i;
 648
 649         rcu_read_lock();
 650
 651         lens->nr = 0;
 652         lens->each = item_len;
 653
 654         for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
 655              i++, head++) {
 656                 hlist_for_each_entry_rcu(conn, head, c_hash_node) {
 657
 658                         /* XXX no c_lock usage.. */
 659                         if (!visitor(conn, buffer))
 660                                 continue;
 661
 662                         /* We copy as much as we can fit in the buffer,
 663                          * but we count all items so that the caller
 664                          * can resize the buffer. */
 665                         if (len >= item_len) {
 666                                 rds_info_copy(iter, buffer, item_len);
 667                                 len -= item_len;
 668                         }
 669                         lens->nr++;
 670                 }
 671         }
 672         rcu_read_unlock();
 673 }
 674 EXPORT_SYMBOL_GPL(rds_for_each_conn_info);
 675
 676 static void rds_walk_conn_path_info(struct socket *sock, unsigned int len,
 677                                     struct rds_info_iterator *iter,
 678                                     struct rds_info_lengths *lens,
 679                                     int (*visitor)(struct rds_conn_path *, void *),
 680                                     u64 *buffer,
 681                                     size_t item_len)
 682 {
 683         struct hlist_head *head;
 684         struct rds_connection *conn;
 685         size_t i;
 686
 687         rcu_read_lock();
 688
 689         lens->nr = 0;
 690         lens->each = item_len;
 691
 692         for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
 693              i++, head++) {
 694                 hlist_for_each_entry_rcu(conn, head, c_hash_node) {
 695                         struct rds_conn_path *cp;
 696
 697                         /* XXX We only copy the information from the first
 698                          * path for now.  The problem is that if there are
 699                          * more than one underlying paths, we cannot report
 700                          * information of all of them using the existing
 701                          * API.  For example, there is only one next_tx_seq,
 702                          * which path's next_tx_seq should we report?  It is
 703                          * a bug in the design of MPRDS.
 704                          */
 705                         cp = conn->c_path;
 706
 707                         /* XXX no cp_lock usage.. */
 708                         if (!visitor(cp, buffer))
 709                                 continue;
 710
 711                         /* We copy as much as we can fit in the buffer,
 712                          * but we count all items so that the caller
 713                          * can resize the buffer.
 714                          */
 715                         if (len >= item_len) {
 716                                 rds_info_copy(iter, buffer, item_len);
 717                                 len -= item_len;
 718                         }
 719                         lens->nr++;
 720                 }
 721         }
 722         rcu_read_unlock();
 723 }
 724
 725 static int rds_conn_info_visitor(struct rds_conn_path *cp, void *buffer)
 726 {
 727         struct rds_info_connection *cinfo = buffer;
 728         struct rds_connection *conn = cp->cp_conn;
 729
 730         if (conn->c_isv6)
 731                 return 0;
 732
 733         cinfo->next_tx_seq = cp->cp_next_tx_seq;
 734         cinfo->next_rx_seq = cp->cp_next_rx_seq;
 735         cinfo->laddr = conn->c_laddr.s6_addr32[3];
 736         cinfo->faddr = conn->c_faddr.s6_addr32[3];
 737         strncpy(cinfo->transport, conn->c_trans->t_name,
 738                 sizeof(cinfo->transport));
 739         cinfo->flags = 0;
 740
 741         rds_conn_info_set(cinfo->flags, test_bit(RDS_IN_XMIT, &cp->cp_flags),
 742                           SENDING);
 743         /* XXX Future: return the state rather than these funky bits */
 744         rds_conn_info_set(cinfo->flags,
 745                           atomic_read(&cp->cp_state) == RDS_CONN_CONNECTING,
 746                           CONNECTING);
 747         rds_conn_info_set(cinfo->flags,
 748                           atomic_read(&cp->cp_state) == RDS_CONN_UP,
 749                           CONNECTED);
 750         return 1;
 751 }
 752
 753 #if IS_ENABLED(CONFIG_IPV6)
 754 static int rds6_conn_info_visitor(struct rds_conn_path *cp, void *buffer)
 755 {
 756         struct rds6_info_connection *cinfo6 = buffer;
 757         struct rds_connection *conn = cp->cp_conn;
 758
 759         cinfo6->next_tx_seq = cp->cp_next_tx_seq;
 760         cinfo6->next_rx_seq = cp->cp_next_rx_seq;
 761         cinfo6->laddr = conn->c_laddr;
 762         cinfo6->faddr = conn->c_faddr;
 763         strncpy(cinfo6->transport, conn->c_trans->t_name,
 764                 sizeof(cinfo6->transport));
 765         cinfo6->flags = 0;
 766
 767         rds_conn_info_set(cinfo6->flags, test_bit(RDS_IN_XMIT, &cp->cp_flags),
 768                           SENDING);
 769         /* XXX Future: return the state rather than these funky bits */
 770         rds_conn_info_set(cinfo6->flags,
 771                           atomic_read(&cp->cp_state) == RDS_CONN_CONNECTING,
 772                           CONNECTING);
 773         rds_conn_info_set(cinfo6->flags,
 774                           atomic_read(&cp->cp_state) == RDS_CONN_UP,
 775                           CONNECTED);
 776         /* Just return 1 as there is no error case. This is a helper function
 777          * for rds_walk_conn_path_info() and it wants a return value.
 778          */
 779         return 1;
 780 }
 781 #endif
 782
 783 static void rds_conn_info(struct socket *sock, unsigned int len,
 784                           struct rds_info_iterator *iter,
 785                           struct rds_info_lengths *lens)
 786 {
 787         u64 buffer[(sizeof(struct rds_info_connection) + 7) / 8];
 788
 789         rds_walk_conn_path_info(sock, len, iter, lens,
 790                                 rds_conn_info_visitor,
 791                                 buffer,
 792                                 sizeof(struct rds_info_connection));
 793 }
 794
 795 #if IS_ENABLED(CONFIG_IPV6)
 796 static void rds6_conn_info(struct socket *sock, unsigned int len,
 797                            struct rds_info_iterator *iter,
 798                            struct rds_info_lengths *lens)
 799 {
 800         u64 buffer[(sizeof(struct rds6_info_connection) + 7) / 8];
 801
 802         rds_walk_conn_path_info(sock, len, iter, lens,
 803                                 rds6_conn_info_visitor,
 804                                 buffer,
 805                                 sizeof(struct rds6_info_connection));
 806 }
 807 #endif
 808
 809 int rds_conn_init(void)
 810 {
 811         int ret;
 812
 813         ret = rds_loop_net_init(); /* register pernet callback */
 814         if (ret)
 815                 return ret;
 816
 817         rds_conn_slab = kmem_cache_create("rds_connection",
 818                                           sizeof(struct rds_connection),
 819                                           0, 0, NULL);
 820         if (!rds_conn_slab) {
 821                 rds_loop_net_exit();
 822                 return -ENOMEM;
 823         }
 824
 825         rds_info_register_func(RDS_INFO_CONNECTIONS, rds_conn_info);
 826         rds_info_register_func(RDS_INFO_SEND_MESSAGES,
 827                                rds_conn_message_info_send);
 828         rds_info_register_func(RDS_INFO_RETRANS_MESSAGES,
 829                                rds_conn_message_info_retrans);
 830 #if IS_ENABLED(CONFIG_IPV6)
 831         rds_info_register_func(RDS6_INFO_CONNECTIONS, rds6_conn_info);
 832         rds_info_register_func(RDS6_INFO_SEND_MESSAGES,
 833                                rds6_conn_message_info_send);
 834         rds_info_register_func(RDS6_INFO_RETRANS_MESSAGES,
 835                                rds6_conn_message_info_retrans);
 836 #endif
 837         return 0;
 838 }
 839
 840 void rds_conn_exit(void)
 841 {
 842         rds_loop_net_exit(); /* unregister pernet callback */
 843         rds_loop_exit();
 844
 845         WARN_ON(!hlist_empty(rds_conn_hash));
 846
 847         kmem_cache_destroy(rds_conn_slab);
 848
 849         rds_info_deregister_func(RDS_INFO_CONNECTIONS, rds_conn_info);
 850         rds_info_deregister_func(RDS_INFO_SEND_MESSAGES,
 851                                  rds_conn_message_info_send);
 852         rds_info_deregister_func(RDS_INFO_RETRANS_MESSAGES,
 853                                  rds_conn_message_info_retrans);
 854 #if IS_ENABLED(CONFIG_IPV6)
 855         rds_info_deregister_func(RDS6_INFO_CONNECTIONS, rds6_conn_info);
 856         rds_info_deregister_func(RDS6_INFO_SEND_MESSAGES,
 857                                  rds6_conn_message_info_send);
 858         rds_info_deregister_func(RDS6_INFO_RETRANS_MESSAGES,
 859                                  rds6_conn_message_info_retrans);
 860 #endif
 861 }
 862
 863 /*
 864  * Force a disconnect
 865  */
 866 void rds_conn_path_drop(struct rds_conn_path *cp, bool destroy)
 867 {
 868         atomic_set(&cp->cp_state, RDS_CONN_ERROR);
 869
 870         rcu_read_lock();
 871         if (!destroy && rds_destroy_pending(cp->cp_conn)) {
 872                 rcu_read_unlock();
 873                 return;
 874         }
 875         queue_work(rds_wq, &cp->cp_down_w);
 876         rcu_read_unlock();
 877 }
 878 EXPORT_SYMBOL_GPL(rds_conn_path_drop);
 879
 880 void rds_conn_drop(struct rds_connection *conn)
 881 {
 882         WARN_ON(conn->c_trans->t_mp_capable);
 883         rds_conn_path_drop(&conn->c_path[0], false);
 884 }
 885 EXPORT_SYMBOL_GPL(rds_conn_drop);
 886
 887 /*
 888  * If the connection is down, trigger a connect. We may have scheduled a
 889  * delayed reconnect however - in this case we should not interfere.
 890  */
 891 void rds_conn_path_connect_if_down(struct rds_conn_path *cp)
 892 {
 893         rcu_read_lock();
 894         if (rds_destroy_pending(cp->cp_conn)) {
 895                 rcu_read_unlock();
 896                 return;
 897         }
 898         if (rds_conn_path_state(cp) == RDS_CONN_DOWN &&
 899             !test_and_set_bit(RDS_RECONNECT_PENDING, &cp->cp_flags))
 900                 queue_delayed_work(rds_wq, &cp->cp_conn_w, 0);
 901         rcu_read_unlock();
 902 }
 903 EXPORT_SYMBOL_GPL(rds_conn_path_connect_if_down);
 904
 905 void rds_conn_connect_if_down(struct rds_connection *conn)
 906 {
 907         WARN_ON(conn->c_trans->t_mp_capable);
 908         rds_conn_path_connect_if_down(&conn->c_path[0]);
 909 }
 910 EXPORT_SYMBOL_GPL(rds_conn_connect_if_down);
 911
 912 void
 913 __rds_conn_path_error(struct rds_conn_path *cp, const char *fmt, ...)
 914 {
 915         va_list ap;
 916
 917         va_start(ap, fmt);
 918         vprintk(fmt, ap);
 919         va_end(ap);
 920
 921         rds_conn_path_drop(cp, false);
 922 }