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         INIT_DELAYED_WORK(&cp->cp_send_w, rds_send_worker);
 143         INIT_DELAYED_WORK(&cp->cp_recv_w, rds_recv_worker);
 144         INIT_DELAYED_WORK(&cp->cp_conn_w, rds_connect_worker);
 145         INIT_WORK(&cp->cp_down_w, rds_shutdown_worker);
 146         mutex_init(&cp->cp_cm_lock);
 147         cp->cp_flags = 0;
 148 }
 149
 150 /*
 151  * There is only every one 'conn' for a given pair of addresses in the
 152  * system at a time.  They contain messages to be retransmitted and so
 153  * span the lifetime of the actual underlying transport connections.
 154  *
 155  * For now they are not garbage collected once they're created.  They
 156  * are torn down as the module is removed, if ever.
 157  */
 158 static struct rds_connection *__rds_conn_create(struct net *net,
 159                                                 const struct in6_addr *laddr,
 160                                                 const struct in6_addr *faddr,
 161                                                 struct rds_transport *trans,
 162                                                 gfp_t gfp,
 163                                                 int is_outgoing,
 164                                                 int dev_if)
 165 {
 166         struct rds_connection *conn, *parent = NULL;
 167         struct hlist_head *head = rds_conn_bucket(laddr, faddr);
 168         struct rds_transport *loop_trans;
 169         unsigned long flags;
 170         int ret, i;
 171         int npaths = (trans->t_mp_capable ? RDS_MPATH_WORKERS : 1);
 172
 173         rcu_read_lock();
 174         conn = rds_conn_lookup(net, head, laddr, faddr, trans, dev_if);
 175         if (conn &&
 176             conn->c_loopback &&
 177             conn->c_trans != &rds_loop_transport &&
 178             ipv6_addr_equal(laddr, faddr) &&
 179             !is_outgoing) {
 180                 /* This is a looped back IB connection, and we're
 181                  * called by the code handling the incoming connect.
 182                  * We need a second connection object into which we
 183                  * can stick the other QP. */
 184                 parent = conn;
 185                 conn = parent->c_passive;
 186         }
 187         rcu_read_unlock();
 188         if (conn)
 189                 goto out;
 190
 191         conn = kmem_cache_zalloc(rds_conn_slab, gfp);
 192         if (!conn) {
 193                 conn = ERR_PTR(-ENOMEM);
 194                 goto out;
 195         }
 196         conn->c_path = kcalloc(npaths, sizeof(struct rds_conn_path), gfp);
 197         if (!conn->c_path) {
 198                 kmem_cache_free(rds_conn_slab, conn);
 199                 conn = ERR_PTR(-ENOMEM);
 200                 goto out;
 201         }
 202
 203         INIT_HLIST_NODE(&conn->c_hash_node);
 204         conn->c_laddr = *laddr;
 205         conn->c_isv6 = !ipv6_addr_v4mapped(laddr);
 206         conn->c_faddr = *faddr;
 207         conn->c_dev_if = dev_if;
 208
 209 #if IS_ENABLED(CONFIG_IPV6)
 210         /* If the local address is link local, set c_bound_if to be the
 211          * index used for this connection.  Otherwise, set it to 0 as
 212          * the socket is not bound to an interface.  c_bound_if is used
 213          * to look up a socket when a packet is received
 214          */
 215         if (ipv6_addr_type(laddr) & IPV6_ADDR_LINKLOCAL)
 216                 conn->c_bound_if = dev_if;
 217         else
 218 #endif
 219                 conn->c_bound_if = 0;
 220
 221         rds_conn_net_set(conn, net);
 222
 223         ret = rds_cong_get_maps(conn);
 224         if (ret) {
 225                 kfree(conn->c_path);
 226                 kmem_cache_free(rds_conn_slab, conn);
 227                 conn = ERR_PTR(ret);
 228                 goto out;
 229         }
 230
 231         /*
 232          * This is where a connection becomes loopback.  If *any* RDS sockets
 233          * can bind to the destination address then we'd rather the messages
 234          * flow through loopback rather than either transport.
 235          */
 236         loop_trans = rds_trans_get_preferred(net, faddr, conn->c_dev_if);
 237         if (loop_trans) {
 238                 rds_trans_put(loop_trans);
 239                 conn->c_loopback = 1;
 240                 if (trans->t_prefer_loopback) {
 241                         if (likely(is_outgoing)) {
 242                                 /* "outgoing" connection to local address.
 243                                  * Protocol says it wants the connection
 244                                  * handled by the loopback transport.
 245                                  * This is what TCP does.
 246                                  */
 247                                 trans = &rds_loop_transport;
 248                         } else {
 249                                 /* No transport currently in use
 250                                  * should end up here, but if it
 251                                  * does, reset/destroy the connection.
 252                                  */
 253                                 kmem_cache_free(rds_conn_slab, conn);
 254                                 conn = ERR_PTR(-EOPNOTSUPP);
 255                                 goto out;
 256                         }
 257                 }
 258         }
 259
 260         conn->c_trans = trans;
 261
 262         init_waitqueue_head(&conn->c_hs_waitq);
 263         for (i = 0; i < npaths; i++) {
 264                 __rds_conn_path_init(conn, &conn->c_path[i],
 265                                      is_outgoing);
 266                 conn->c_path[i].cp_index = i;
 267         }
 268         rcu_read_lock();
 269         if (rds_destroy_pending(conn))
 270                 ret = -ENETDOWN;
 271         else
 272                 ret = trans->conn_alloc(conn, GFP_ATOMIC);
 273         if (ret) {
 274                 rcu_read_unlock();
 275                 kfree(conn->c_path);
 276                 kmem_cache_free(rds_conn_slab, conn);
 277                 conn = ERR_PTR(ret);
 278                 goto out;
 279         }
 280
 281         rdsdebug("allocated conn %p for %pI6c -> %pI6c over %s %s\n",
 282                  conn, laddr, faddr,
 283                  strnlen(trans->t_name, sizeof(trans->t_name)) ?
 284                  trans->t_name : "[unknown]", is_outgoing ? "(outgoing)" : "");
 285
 286         /*
 287          * Since we ran without holding the conn lock, someone could
 288          * have created the same conn (either normal or passive) in the
 289          * interim. We check while holding the lock. If we won, we complete
 290          * init and return our conn. If we lost, we rollback and return the
 291          * other one.
 292          */
 293         spin_lock_irqsave(&rds_conn_lock, flags);
 294         if (parent) {
 295                 /* Creating passive conn */
 296                 if (parent->c_passive) {
 297                         trans->conn_free(conn->c_path[0].cp_transport_data);
 298                         kfree(conn->c_path);
 299                         kmem_cache_free(rds_conn_slab, conn);
 300                         conn = parent->c_passive;
 301                 } else {
 302                         parent->c_passive = conn;
 303                         rds_cong_add_conn(conn);
 304                         rds_conn_count++;
 305                 }
 306         } else {
 307                 /* Creating normal conn */
 308                 struct rds_connection *found;
 309
 310                 found = rds_conn_lookup(net, head, laddr, faddr, trans,
 311                                         dev_if);
 312                 if (found) {
 313                         struct rds_conn_path *cp;
 314                         int i;
 315
 316                         for (i = 0; i < npaths; i++) {
 317                                 cp = &conn->c_path[i];
 318                                 /* The ->conn_alloc invocation may have
 319                                  * allocated resource for all paths, so all
 320                                  * of them may have to be freed here.
 321                                  */
 322                                 if (cp->cp_transport_data)
 323                                         trans->conn_free(cp->cp_transport_data);
 324                         }
 325                         kfree(conn->c_path);
 326                         kmem_cache_free(rds_conn_slab, conn);
 327                         conn = found;
 328                 } else {
 329                         conn->c_my_gen_num = rds_gen_num;
 330                         conn->c_peer_gen_num = 0;
 331                         hlist_add_head_rcu(&conn->c_hash_node, head);
 332                         rds_cong_add_conn(conn);
 333                         rds_conn_count++;
 334                 }
 335         }
 336         spin_unlock_irqrestore(&rds_conn_lock, flags);
 337         rcu_read_unlock();
 338
 339 out:
 340         return conn;
 341 }
 342
 343 struct rds_connection *rds_conn_create(struct net *net,
 344                                        const struct in6_addr *laddr,
 345                                        const struct in6_addr *faddr,
 346                                        struct rds_transport *trans, gfp_t gfp,
 347                                        int dev_if)
 348 {
 349         return __rds_conn_create(net, laddr, faddr, trans, gfp, 0, dev_if);
 350 }
 351 EXPORT_SYMBOL_GPL(rds_conn_create);
 352
 353 struct rds_connection *rds_conn_create_outgoing(struct net *net,
 354                                                 const struct in6_addr *laddr,
 355                                                 const struct in6_addr *faddr,
 356                                                 struct rds_transport *trans,
 357                                                 gfp_t gfp, int dev_if)
 358 {
 359         return __rds_conn_create(net, laddr, faddr, trans, gfp, 1, dev_if);
 360 }
 361 EXPORT_SYMBOL_GPL(rds_conn_create_outgoing);
 362
 363 void rds_conn_shutdown(struct rds_conn_path *cp)
 364 {
 365         struct rds_connection *conn = cp->cp_conn;
 366
 367         /* shut it down unless it's down already */
 368         if (!rds_conn_path_transition(cp, RDS_CONN_DOWN, RDS_CONN_DOWN)) {
 369                 /*
 370                  * Quiesce the connection mgmt handlers before we start tearing
 371                  * things down. We don't hold the mutex for the entire
 372                  * duration of the shutdown operation, else we may be
 373                  * deadlocking with the CM handler. Instead, the CM event
 374                  * handler is supposed to check for state DISCONNECTING
 375                  */
 376                 mutex_lock(&cp->cp_cm_lock);
 377                 if (!rds_conn_path_transition(cp, RDS_CONN_UP,
 378                                               RDS_CONN_DISCONNECTING) &&
 379                     !rds_conn_path_transition(cp, RDS_CONN_ERROR,
 380                                               RDS_CONN_DISCONNECTING)) {
 381                         rds_conn_path_error(cp,
 382                                             "shutdown called in state %d\n",
 383                                             atomic_read(&cp->cp_state));
 384                         mutex_unlock(&cp->cp_cm_lock);
 385                         return;
 386                 }
 387                 mutex_unlock(&cp->cp_cm_lock);
 388
 389                 wait_event(cp->cp_waitq,
 390                            !test_bit(RDS_IN_XMIT, &cp->cp_flags));
 391                 wait_event(cp->cp_waitq,
 392                            !test_bit(RDS_RECV_REFILL, &cp->cp_flags));
 393
 394                 conn->c_trans->conn_path_shutdown(cp);
 395                 rds_conn_path_reset(cp);
 396
 397                 if (!rds_conn_path_transition(cp, RDS_CONN_DISCONNECTING,
 398                                               RDS_CONN_DOWN) &&
 399                     !rds_conn_path_transition(cp, RDS_CONN_ERROR,
 400                                               RDS_CONN_DOWN)) {
 401                         /* This can happen - eg when we're in the middle of tearing
 402                          * down the connection, and someone unloads the rds module.
 403                          * Quite reproducible with loopback connections.
 404                          * Mostly harmless.
 405                          *
 406                          * Note that this also happens with rds-tcp because
 407                          * we could have triggered rds_conn_path_drop in irq
 408                          * mode from rds_tcp_state change on the receipt of
 409                          * a FIN, thus we need to recheck for RDS_CONN_ERROR
 410                          * here.
 411                          */
 412                         rds_conn_path_error(cp, "%s: failed to transition "
 413                                             "to state DOWN, current state "
 414                                             "is %d\n", __func__,
 415                                             atomic_read(&cp->cp_state));
 416                         return;
 417                 }
 418         }
 419
 420         /* Then reconnect if it's still live.
 421          * The passive side of an IB loopback connection is never added
 422          * to the conn hash, so we never trigger a reconnect on this
 423          * conn - the reconnect is always triggered by the active peer. */
 424         cancel_delayed_work_sync(&cp->cp_conn_w);
 425         rcu_read_lock();
 426         if (!hlist_unhashed(&conn->c_hash_node)) {
 427                 rcu_read_unlock();
 428                 rds_queue_reconnect(cp);
 429         } else {
 430                 rcu_read_unlock();
 431         }
 432 }
 433
 434 /* destroy a single rds_conn_path. rds_conn_destroy() iterates over
 435  * all paths using rds_conn_path_destroy()
 436  */
 437 static void rds_conn_path_destroy(struct rds_conn_path *cp)
 438 {
 439         struct rds_message *rm, *rtmp;
 440
 441         if (!cp->cp_transport_data)
 442                 return;
 443
 444         /* make sure lingering queued work won't try to ref the conn */
 445         cancel_delayed_work_sync(&cp->cp_send_w);
 446         cancel_delayed_work_sync(&cp->cp_recv_w);
 447
 448         rds_conn_path_drop(cp, true);
 449         flush_work(&cp->cp_down_w);
 450
 451         /* tear down queued messages */
 452         list_for_each_entry_safe(rm, rtmp,
 453                                  &cp->cp_send_queue,
 454                                  m_conn_item) {
 455                 list_del_init(&rm->m_conn_item);
 456                 BUG_ON(!list_empty(&rm->m_sock_item));
 457                 rds_message_put(rm);
 458         }
 459         if (cp->cp_xmit_rm)
 460                 rds_message_put(cp->cp_xmit_rm);
 461
 462         WARN_ON(delayed_work_pending(&cp->cp_send_w));
 463         WARN_ON(delayed_work_pending(&cp->cp_recv_w));
 464         WARN_ON(delayed_work_pending(&cp->cp_conn_w));
 465         WARN_ON(work_pending(&cp->cp_down_w));
 466
 467         cp->cp_conn->c_trans->conn_free(cp->cp_transport_data);
 468 }
 469
 470 /*
 471  * Stop and free a connection.
 472  *
 473  * This can only be used in very limited circumstances.  It assumes that once
 474  * the conn has been shutdown that no one else is referencing the connection.
 475  * We can only ensure this in the rmmod path in the current code.
 476  */
 477 void rds_conn_destroy(struct rds_connection *conn)
 478 {
 479         unsigned long flags;
 480         int i;
 481         struct rds_conn_path *cp;
 482         int npaths = (conn->c_trans->t_mp_capable ? RDS_MPATH_WORKERS : 1);
 483
 484         rdsdebug("freeing conn %p for %pI4 -> "
 485                  "%pI4\n", conn, &conn->c_laddr,
 486                  &conn->c_faddr);
 487
 488         /* Ensure conn will not be scheduled for reconnect */
 489         spin_lock_irq(&rds_conn_lock);
 490         hlist_del_init_rcu(&conn->c_hash_node);
 491         spin_unlock_irq(&rds_conn_lock);
 492         synchronize_rcu();
 493
 494         /* shut the connection down */
 495         for (i = 0; i < npaths; i++) {
 496                 cp = &conn->c_path[i];
 497                 rds_conn_path_destroy(cp);
 498                 BUG_ON(!list_empty(&cp->cp_retrans));
 499         }
 500
 501         /*
 502          * The congestion maps aren't freed up here.  They're
 503          * freed by rds_cong_exit() after all the connections
 504          * have been freed.
 505          */
 506         rds_cong_remove_conn(conn);
 507
 508         kfree(conn->c_path);
 509         kmem_cache_free(rds_conn_slab, conn);
 510
 511         spin_lock_irqsave(&rds_conn_lock, flags);
 512         rds_conn_count--;
 513         spin_unlock_irqrestore(&rds_conn_lock, flags);
 514 }
 515 EXPORT_SYMBOL_GPL(rds_conn_destroy);
 516
 517 static void __rds_inc_msg_cp(struct rds_incoming *inc,
 518                              struct rds_info_iterator *iter,
 519                              void *saddr, void *daddr, int flip, bool isv6)
 520 {
 521 #if IS_ENABLED(CONFIG_IPV6)
 522         if (isv6)
 523                 rds6_inc_info_copy(inc, iter, saddr, daddr, flip);
 524         else
 525 #endif
 526                 rds_inc_info_copy(inc, iter, *(__be32 *)saddr,
 527                                   *(__be32 *)daddr, flip);
 528 }
 529
 530 static void rds_conn_message_info_cmn(struct socket *sock, unsigned int len,
 531                                       struct rds_info_iterator *iter,
 532                                       struct rds_info_lengths *lens,
 533                                       int want_send, bool isv6)
 534 {
 535         struct hlist_head *head;
 536         struct list_head *list;
 537         struct rds_connection *conn;
 538         struct rds_message *rm;
 539         unsigned int total = 0;
 540         unsigned long flags;
 541         size_t i;
 542         int j;
 543
 544         if (isv6)
 545                 len /= sizeof(struct rds6_info_message);
 546         else
 547                 len /= sizeof(struct rds_info_message);
 548
 549         rcu_read_lock();
 550
 551         for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
 552              i++, head++) {
 553                 hlist_for_each_entry_rcu(conn, head, c_hash_node) {
 554                         struct rds_conn_path *cp;
 555                         int npaths;
 556
 557                         if (!isv6 && conn->c_isv6)
 558                                 continue;
 559
 560                         npaths = (conn->c_trans->t_mp_capable ?
 561                                  RDS_MPATH_WORKERS : 1);
 562
 563                         for (j = 0; j < npaths; j++) {
 564                                 cp = &conn->c_path[j];
 565                                 if (want_send)
 566                                         list = &cp->cp_send_queue;
 567                                 else
 568                                         list = &cp->cp_retrans;
 569
 570                                 spin_lock_irqsave(&cp->cp_lock, flags);
 571
 572                                 /* XXX too lazy to maintain counts.. */
 573                                 list_for_each_entry(rm, list, m_conn_item) {
 574                                         total++;
 575                                         if (total <= len)
 576                                                 __rds_inc_msg_cp(&rm->m_inc,
 577                                                                  iter,
 578                                                                  &conn->c_laddr,
 579                                                                  &conn->c_faddr,
 580                                                                  0, isv6);
 581                                 }
 582
 583                                 spin_unlock_irqrestore(&cp->cp_lock, flags);
 584                         }
 585                 }
 586         }
 587         rcu_read_unlock();
 588
 589         lens->nr = total;
 590         if (isv6)
 591                 lens->each = sizeof(struct rds6_info_message);
 592         else
 593                 lens->each = sizeof(struct rds_info_message);
 594 }
 595
 596 static void rds_conn_message_info(struct socket *sock, unsigned int len,
 597                                   struct rds_info_iterator *iter,
 598                                   struct rds_info_lengths *lens,
 599                                   int want_send)
 600 {
 601         rds_conn_message_info_cmn(sock, len, iter, lens, want_send, false);
 602 }
 603
 604 #if IS_ENABLED(CONFIG_IPV6)
 605 static void rds6_conn_message_info(struct socket *sock, unsigned int len,
 606                                    struct rds_info_iterator *iter,
 607                                    struct rds_info_lengths *lens,
 608                                    int want_send)
 609 {
 610         rds_conn_message_info_cmn(sock, len, iter, lens, want_send, true);
 611 }
 612 #endif
 613
 614 static void rds_conn_message_info_send(struct socket *sock, unsigned int len,
 615                                        struct rds_info_iterator *iter,
 616                                        struct rds_info_lengths *lens)
 617 {
 618         rds_conn_message_info(sock, len, iter, lens, 1);
 619 }
 620
 621 #if IS_ENABLED(CONFIG_IPV6)
 622 static void rds6_conn_message_info_send(struct socket *sock, unsigned int len,
 623                                         struct rds_info_iterator *iter,
 624                                         struct rds_info_lengths *lens)
 625 {
 626         rds6_conn_message_info(sock, len, iter, lens, 1);
 627 }
 628 #endif
 629
 630 static void rds_conn_message_info_retrans(struct socket *sock,
 631                                           unsigned int len,
 632                                           struct rds_info_iterator *iter,
 633                                           struct rds_info_lengths *lens)
 634 {
 635         rds_conn_message_info(sock, len, iter, lens, 0);
 636 }
 637
 638 #if IS_ENABLED(CONFIG_IPV6)
 639 static void rds6_conn_message_info_retrans(struct socket *sock,
 640                                            unsigned int len,
 641                                            struct rds_info_iterator *iter,
 642                                            struct rds_info_lengths *lens)
 643 {
 644         rds6_conn_message_info(sock, len, iter, lens, 0);
 645 }
 646 #endif
 647
 648 void rds_for_each_conn_info(struct socket *sock, unsigned int len,
 649                           struct rds_info_iterator *iter,
 650                           struct rds_info_lengths *lens,
 651                           int (*visitor)(struct rds_connection *, void *),
 652                           u64 *buffer,
 653                           size_t item_len)
 654 {
 655         struct hlist_head *head;
 656         struct rds_connection *conn;
 657         size_t i;
 658
 659         rcu_read_lock();
 660
 661         lens->nr = 0;
 662         lens->each = item_len;
 663
 664         for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
 665              i++, head++) {
 666                 hlist_for_each_entry_rcu(conn, head, c_hash_node) {
 667
 668                         /* XXX no c_lock usage.. */
 669                         if (!visitor(conn, buffer))
 670                                 continue;
 671
 672                         /* We copy as much as we can fit in the buffer,
 673                          * but we count all items so that the caller
 674                          * can resize the buffer. */
 675                         if (len >= item_len) {
 676                                 rds_info_copy(iter, buffer, item_len);
 677                                 len -= item_len;
 678                         }
 679                         lens->nr++;
 680                 }
 681         }
 682         rcu_read_unlock();
 683 }
 684 EXPORT_SYMBOL_GPL(rds_for_each_conn_info);
 685
 686 static void rds_walk_conn_path_info(struct socket *sock, unsigned int len,
 687                                     struct rds_info_iterator *iter,
 688                                     struct rds_info_lengths *lens,
 689                                     int (*visitor)(struct rds_conn_path *, void *),
 690                                     u64 *buffer,
 691                                     size_t item_len)
 692 {
 693         struct hlist_head *head;
 694         struct rds_connection *conn;
 695         size_t i;
 696
 697         rcu_read_lock();
 698
 699         lens->nr = 0;
 700         lens->each = item_len;
 701
 702         for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
 703              i++, head++) {
 704                 hlist_for_each_entry_rcu(conn, head, c_hash_node) {
 705                         struct rds_conn_path *cp;
 706
 707                         /* XXX We only copy the information from the first
 708                          * path for now.  The problem is that if there are
 709                          * more than one underlying paths, we cannot report
 710                          * information of all of them using the existing
 711                          * API.  For example, there is only one next_tx_seq,
 712                          * which path's next_tx_seq should we report?  It is
 713                          * a bug in the design of MPRDS.
 714                          */
 715                         cp = conn->c_path;
 716
 717                         /* XXX no cp_lock usage.. */
 718                         if (!visitor(cp, buffer))
 719                                 continue;
 720
 721                         /* We copy as much as we can fit in the buffer,
 722                          * but we count all items so that the caller
 723                          * can resize the buffer.
 724                          */
 725                         if (len >= item_len) {
 726                                 rds_info_copy(iter, buffer, item_len);
 727                                 len -= item_len;
 728                         }
 729                         lens->nr++;
 730                 }
 731         }
 732         rcu_read_unlock();
 733 }
 734
 735 static int rds_conn_info_visitor(struct rds_conn_path *cp, void *buffer)
 736 {
 737         struct rds_info_connection *cinfo = buffer;
 738         struct rds_connection *conn = cp->cp_conn;
 739
 740         if (conn->c_isv6)
 741                 return 0;
 742
 743         cinfo->next_tx_seq = cp->cp_next_tx_seq;
 744         cinfo->next_rx_seq = cp->cp_next_rx_seq;
 745         cinfo->laddr = conn->c_laddr.s6_addr32[3];
 746         cinfo->faddr = conn->c_faddr.s6_addr32[3];
 747         strncpy(cinfo->transport, conn->c_trans->t_name,
 748                 sizeof(cinfo->transport));
 749         cinfo->flags = 0;
 750
 751         rds_conn_info_set(cinfo->flags, test_bit(RDS_IN_XMIT, &cp->cp_flags),
 752                           SENDING);
 753         /* XXX Future: return the state rather than these funky bits */
 754         rds_conn_info_set(cinfo->flags,
 755                           atomic_read(&cp->cp_state) == RDS_CONN_CONNECTING,
 756                           CONNECTING);
 757         rds_conn_info_set(cinfo->flags,
 758                           atomic_read(&cp->cp_state) == RDS_CONN_UP,
 759                           CONNECTED);
 760         return 1;
 761 }
 762
 763 #if IS_ENABLED(CONFIG_IPV6)
 764 static int rds6_conn_info_visitor(struct rds_conn_path *cp, void *buffer)
 765 {
 766         struct rds6_info_connection *cinfo6 = buffer;
 767         struct rds_connection *conn = cp->cp_conn;
 768
 769         cinfo6->next_tx_seq = cp->cp_next_tx_seq;
 770         cinfo6->next_rx_seq = cp->cp_next_rx_seq;
 771         cinfo6->laddr = conn->c_laddr;
 772         cinfo6->faddr = conn->c_faddr;
 773         strncpy(cinfo6->transport, conn->c_trans->t_name,
 774                 sizeof(cinfo6->transport));
 775         cinfo6->flags = 0;
 776
 777         rds_conn_info_set(cinfo6->flags, test_bit(RDS_IN_XMIT, &cp->cp_flags),
 778                           SENDING);
 779         /* XXX Future: return the state rather than these funky bits */
 780         rds_conn_info_set(cinfo6->flags,
 781                           atomic_read(&cp->cp_state) == RDS_CONN_CONNECTING,
 782                           CONNECTING);
 783         rds_conn_info_set(cinfo6->flags,
 784                           atomic_read(&cp->cp_state) == RDS_CONN_UP,
 785                           CONNECTED);
 786         /* Just return 1 as there is no error case. This is a helper function
 787          * for rds_walk_conn_path_info() and it wants a return value.
 788          */
 789         return 1;
 790 }
 791 #endif
 792
 793 static void rds_conn_info(struct socket *sock, unsigned int len,
 794                           struct rds_info_iterator *iter,
 795                           struct rds_info_lengths *lens)
 796 {
 797         u64 buffer[(sizeof(struct rds_info_connection) + 7) / 8];
 798
 799         rds_walk_conn_path_info(sock, len, iter, lens,
 800                                 rds_conn_info_visitor,
 801                                 buffer,
 802                                 sizeof(struct rds_info_connection));
 803 }
 804
 805 #if IS_ENABLED(CONFIG_IPV6)
 806 static void rds6_conn_info(struct socket *sock, unsigned int len,
 807                            struct rds_info_iterator *iter,
 808                            struct rds_info_lengths *lens)
 809 {
 810         u64 buffer[(sizeof(struct rds6_info_connection) + 7) / 8];
 811
 812         rds_walk_conn_path_info(sock, len, iter, lens,
 813                                 rds6_conn_info_visitor,
 814                                 buffer,
 815                                 sizeof(struct rds6_info_connection));
 816 }
 817 #endif
 818
 819 int rds_conn_init(void)
 820 {
 821         int ret;
 822
 823         ret = rds_loop_net_init(); /* register pernet callback */
 824         if (ret)
 825                 return ret;
 826
 827         rds_conn_slab = kmem_cache_create("rds_connection",
 828                                           sizeof(struct rds_connection),
 829                                           0, 0, NULL);
 830         if (!rds_conn_slab) {
 831                 rds_loop_net_exit();
 832                 return -ENOMEM;
 833         }
 834
 835         rds_info_register_func(RDS_INFO_CONNECTIONS, rds_conn_info);
 836         rds_info_register_func(RDS_INFO_SEND_MESSAGES,
 837                                rds_conn_message_info_send);
 838         rds_info_register_func(RDS_INFO_RETRANS_MESSAGES,
 839                                rds_conn_message_info_retrans);
 840 #if IS_ENABLED(CONFIG_IPV6)
 841         rds_info_register_func(RDS6_INFO_CONNECTIONS, rds6_conn_info);
 842         rds_info_register_func(RDS6_INFO_SEND_MESSAGES,
 843                                rds6_conn_message_info_send);
 844         rds_info_register_func(RDS6_INFO_RETRANS_MESSAGES,
 845                                rds6_conn_message_info_retrans);
 846 #endif
 847         return 0;
 848 }
 849
 850 void rds_conn_exit(void)
 851 {
 852         rds_loop_net_exit(); /* unregister pernet callback */
 853         rds_loop_exit();
 854
 855         WARN_ON(!hlist_empty(rds_conn_hash));
 856
 857         kmem_cache_destroy(rds_conn_slab);
 858
 859         rds_info_deregister_func(RDS_INFO_CONNECTIONS, rds_conn_info);
 860         rds_info_deregister_func(RDS_INFO_SEND_MESSAGES,
 861                                  rds_conn_message_info_send);
 862         rds_info_deregister_func(RDS_INFO_RETRANS_MESSAGES,
 863                                  rds_conn_message_info_retrans);
 864 #if IS_ENABLED(CONFIG_IPV6)
 865         rds_info_deregister_func(RDS6_INFO_CONNECTIONS, rds6_conn_info);
 866         rds_info_deregister_func(RDS6_INFO_SEND_MESSAGES,
 867                                  rds6_conn_message_info_send);
 868         rds_info_deregister_func(RDS6_INFO_RETRANS_MESSAGES,
 869                                  rds6_conn_message_info_retrans);
 870 #endif
 871 }
 872
 873 /*
 874  * Force a disconnect
 875  */
 876 void rds_conn_path_drop(struct rds_conn_path *cp, bool destroy)
 877 {
 878         atomic_set(&cp->cp_state, RDS_CONN_ERROR);
 879
 880         rcu_read_lock();
 881         if (!destroy && rds_destroy_pending(cp->cp_conn)) {
 882                 rcu_read_unlock();
 883                 return;
 884         }
 885         queue_work(rds_wq, &cp->cp_down_w);
 886         rcu_read_unlock();
 887 }
 888 EXPORT_SYMBOL_GPL(rds_conn_path_drop);
 889
 890 void rds_conn_drop(struct rds_connection *conn)
 891 {
 892         WARN_ON(conn->c_trans->t_mp_capable);
 893         rds_conn_path_drop(&conn->c_path[0], false);
 894 }
 895 EXPORT_SYMBOL_GPL(rds_conn_drop);
 896
 897 /*
 898  * If the connection is down, trigger a connect. We may have scheduled a
 899  * delayed reconnect however - in this case we should not interfere.
 900  */
 901 void rds_conn_path_connect_if_down(struct rds_conn_path *cp)
 902 {
 903         rcu_read_lock();
 904         if (rds_destroy_pending(cp->cp_conn)) {
 905                 rcu_read_unlock();
 906                 return;
 907         }
 908         if (rds_conn_path_state(cp) == RDS_CONN_DOWN &&
 909             !test_and_set_bit(RDS_RECONNECT_PENDING, &cp->cp_flags))
 910                 queue_delayed_work(rds_wq, &cp->cp_conn_w, 0);
 911         rcu_read_unlock();
 912 }
 913 EXPORT_SYMBOL_GPL(rds_conn_path_connect_if_down);
 914
 915 void rds_conn_connect_if_down(struct rds_connection *conn)
 916 {
 917         WARN_ON(conn->c_trans->t_mp_capable);
 918         rds_conn_path_connect_if_down(&conn->c_path[0]);
 919 }
 920 EXPORT_SYMBOL_GPL(rds_conn_connect_if_down);
 921
 922 void
 923 __rds_conn_path_error(struct rds_conn_path *cp, const char *fmt, ...)
 924 {
 925         va_list ap;
 926
 927         va_start(ap, fmt);
 928         vprintk(fmt, ap);
 929         va_end(ap);
 930
 931         rds_conn_path_drop(cp, false);
 932 }