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