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                                               u8 tos, 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                     conn->c_tos == tos &&
  96                     net == rds_conn_net(conn) &&
  97                     conn->c_dev_if == dev_if) {
  98                         ret = conn;
  99                         break;
 100                 }
 101         }
 102         rdsdebug("returning conn %p for %pI6c -> %pI6c\n", ret,
 103                  laddr, faddr);
 104         return ret;
 105 }
 106
 107 /*
 108  * This is called by transports as they're bringing down a connection.
 109  * It clears partial message state so that the transport can start sending
 110  * and receiving over this connection again in the future.  It is up to
 111  * the transport to have serialized this call with its send and recv.
 112  */
 113 static void rds_conn_path_reset(struct rds_conn_path *cp)
 114 {
 115         struct rds_connection *conn = cp->cp_conn;
 116
 117         rdsdebug("connection %pI6c to %pI6c reset\n",
 118                  &conn->c_laddr, &conn->c_faddr);
 119
 120         rds_stats_inc(s_conn_reset);
 121         rds_send_path_reset(cp);
 122         cp->cp_flags = 0;
 123
 124         /* Do not clear next_rx_seq here, else we cannot distinguish
 125          * retransmitted packets from new packets, and will hand all
 126          * of them to the application. That is not consistent with the
 127          * reliability guarantees of RDS. */
 128 }
 129
 130 static void __rds_conn_path_init(struct rds_connection *conn,
 131                                  struct rds_conn_path *cp, bool is_outgoing)
 132 {
 133         spin_lock_init(&cp->cp_lock);
 134         cp->cp_next_tx_seq = 1;
 135         init_waitqueue_head(&cp->cp_waitq);
 136         INIT_LIST_HEAD(&cp->cp_send_queue);
 137         INIT_LIST_HEAD(&cp->cp_retrans);
 138
 139         cp->cp_conn = conn;
 140         atomic_set(&cp->cp_state, RDS_CONN_DOWN);
 141         cp->cp_send_gen = 0;
 142         cp->cp_reconnect_jiffies = 0;
 143         cp->cp_conn->c_proposed_version = RDS_PROTOCOL_VERSION;
 144         INIT_DELAYED_WORK(&cp->cp_send_w, rds_send_worker);
 145         INIT_DELAYED_WORK(&cp->cp_recv_w, rds_recv_worker);
 146         INIT_DELAYED_WORK(&cp->cp_conn_w, rds_connect_worker);
 147         INIT_WORK(&cp->cp_down_w, rds_shutdown_worker);
 148         mutex_init(&cp->cp_cm_lock);
 149         cp->cp_flags = 0;
 150 }
 151
 152 /*
 153  * There is only every one 'conn' for a given pair of addresses in the
 154  * system at a time.  They contain messages to be retransmitted and so
 155  * span the lifetime of the actual underlying transport connections.
 156  *
 157  * For now they are not garbage collected once they're created.  They
 158  * are torn down as the module is removed, if ever.
 159  */
 160 static struct rds_connection *__rds_conn_create(struct net *net,
 161                                                 const struct in6_addr *laddr,
 162                                                 const struct in6_addr *faddr,
 163                                                 struct rds_transport *trans,
 164                                                 gfp_t gfp, u8 tos,
 165                                                 int is_outgoing,
 166                                                 int dev_if)
 167 {
 168         struct rds_connection *conn, *parent = NULL;
 169         struct hlist_head *head = rds_conn_bucket(laddr, faddr);
 170         struct rds_transport *loop_trans;
 171         unsigned long flags;
 172         int ret, i;
 173         int npaths = (trans->t_mp_capable ? RDS_MPATH_WORKERS : 1);
 174
 175         rcu_read_lock();
 176         conn = rds_conn_lookup(net, head, laddr, faddr, trans, tos, dev_if);
 177         if (conn &&
 178             conn->c_loopback &&
 179             conn->c_trans != &rds_loop_transport &&
 180             ipv6_addr_equal(laddr, faddr) &&
 181             !is_outgoing) {
 182                 /* This is a looped back IB connection, and we're
 183                  * called by the code handling the incoming connect.
 184                  * We need a second connection object into which we
 185                  * can stick the other QP. */
 186                 parent = conn;
 187                 conn = parent->c_passive;
 188         }
 189         rcu_read_unlock();
 190         if (conn)
 191                 goto out;
 192
 193         conn = kmem_cache_zalloc(rds_conn_slab, gfp);
 194         if (!conn) {
 195                 conn = ERR_PTR(-ENOMEM);
 196                 goto out;
 197         }
 198         conn->c_path = kcalloc(npaths, sizeof(struct rds_conn_path), gfp);
 199         if (!conn->c_path) {
 200                 kmem_cache_free(rds_conn_slab, conn);
 201                 conn = ERR_PTR(-ENOMEM);
 202                 goto out;
 203         }
 204
 205         INIT_HLIST_NODE(&conn->c_hash_node);
 206         conn->c_laddr = *laddr;
 207         conn->c_isv6 = !ipv6_addr_v4mapped(laddr);
 208         conn->c_faddr = *faddr;
 209         conn->c_dev_if = dev_if;
 210         conn->c_tos = tos;
 211
 212 #if IS_ENABLED(CONFIG_IPV6)
 213         /* If the local address is link local, set c_bound_if to be the
 214          * index used for this connection.  Otherwise, set it to 0 as
 215          * the socket is not bound to an interface.  c_bound_if is used
 216          * to look up a socket when a packet is received
 217          */
 218         if (ipv6_addr_type(laddr) & IPV6_ADDR_LINKLOCAL)
 219                 conn->c_bound_if = dev_if;
 220         else
 221 #endif
 222                 conn->c_bound_if = 0;
 223
 224         rds_conn_net_set(conn, net);
 225
 226         ret = rds_cong_get_maps(conn);
 227         if (ret) {
 228                 kfree(conn->c_path);
 229                 kmem_cache_free(rds_conn_slab, conn);
 230                 conn = ERR_PTR(ret);
 231                 goto out;
 232         }
 233
 234         /*
 235          * This is where a connection becomes loopback.  If *any* RDS sockets
 236          * can bind to the destination address then we'd rather the messages
 237          * flow through loopback rather than either transport.
 238          */
 239         loop_trans = rds_trans_get_preferred(net, faddr, conn->c_dev_if);
 240         if (loop_trans) {
 241                 rds_trans_put(loop_trans);
 242                 conn->c_loopback = 1;
 243                 if (is_outgoing && trans->t_prefer_loopback) {
 244                         /* "outgoing" connection - and the transport
 245                          * says it wants the connection handled by the
 246                          * loopback transport. This is what TCP does.
 247                          */
 248                         trans = &rds_loop_transport;
 249                 }
 250         }
 251
 252         conn->c_trans = trans;
 253
 254         init_waitqueue_head(&conn->c_hs_waitq);
 255         for (i = 0; i < npaths; i++) {
 256                 __rds_conn_path_init(conn, &conn->c_path[i],
 257                                      is_outgoing);
 258                 conn->c_path[i].cp_index = i;
 259         }
 260         rcu_read_lock();
 261         if (rds_destroy_pending(conn))
 262                 ret = -ENETDOWN;
 263         else
 264                 ret = trans->conn_alloc(conn, GFP_ATOMIC);
 265         if (ret) {
 266                 rcu_read_unlock();
 267                 kfree(conn->c_path);
 268                 kmem_cache_free(rds_conn_slab, conn);
 269                 conn = ERR_PTR(ret);
 270                 goto out;
 271         }
 272
 273         rdsdebug("allocated conn %p for %pI6c -> %pI6c over %s %s\n",
 274                  conn, laddr, faddr,
 275                  strnlen(trans->t_name, sizeof(trans->t_name)) ?
 276                  trans->t_name : "[unknown]", is_outgoing ? "(outgoing)" : "");
 277
 278         /*
 279          * Since we ran without holding the conn lock, someone could
 280          * have created the same conn (either normal or passive) in the
 281          * interim. We check while holding the lock. If we won, we complete
 282          * init and return our conn. If we lost, we rollback and return the
 283          * other one.
 284          */
 285         spin_lock_irqsave(&rds_conn_lock, flags);
 286         if (parent) {
 287                 /* Creating passive conn */
 288                 if (parent->c_passive) {
 289                         trans->conn_free(conn->c_path[0].cp_transport_data);
 290                         kfree(conn->c_path);
 291                         kmem_cache_free(rds_conn_slab, conn);
 292                         conn = parent->c_passive;
 293                 } else {
 294                         parent->c_passive = conn;
 295                         rds_cong_add_conn(conn);
 296                         rds_conn_count++;
 297                 }
 298         } else {
 299                 /* Creating normal conn */
 300                 struct rds_connection *found;
 301
 302                 found = rds_conn_lookup(net, head, laddr, faddr, trans,
 303                                         tos, dev_if);
 304                 if (found) {
 305                         struct rds_conn_path *cp;
 306                         int i;
 307
 308                         for (i = 0; i < npaths; i++) {
 309                                 cp = &conn->c_path[i];
 310                                 /* The ->conn_alloc invocation may have
 311                                  * allocated resource for all paths, so all
 312                                  * of them may have to be freed here.
 313                                  */
 314                                 if (cp->cp_transport_data)
 315                                         trans->conn_free(cp->cp_transport_data);
 316                         }
 317                         kfree(conn->c_path);
 318                         kmem_cache_free(rds_conn_slab, conn);
 319                         conn = found;
 320                 } else {
 321                         conn->c_my_gen_num = rds_gen_num;
 322                         conn->c_peer_gen_num = 0;
 323                         hlist_add_head_rcu(&conn->c_hash_node, head);
 324                         rds_cong_add_conn(conn);
 325                         rds_conn_count++;
 326                 }
 327         }
 328         spin_unlock_irqrestore(&rds_conn_lock, flags);
 329         rcu_read_unlock();
 330
 331 out:
 332         return conn;
 333 }
 334
 335 struct rds_connection *rds_conn_create(struct net *net,
 336                                        const struct in6_addr *laddr,
 337                                        const struct in6_addr *faddr,
 338                                        struct rds_transport *trans, u8 tos,
 339                                        gfp_t gfp, int dev_if)
 340 {
 341         return __rds_conn_create(net, laddr, faddr, trans, gfp, tos, 0, dev_if);
 342 }
 343 EXPORT_SYMBOL_GPL(rds_conn_create);
 344
 345 struct rds_connection *rds_conn_create_outgoing(struct net *net,
 346                                                 const struct in6_addr *laddr,
 347                                                 const struct in6_addr *faddr,
 348                                                 struct rds_transport *trans,
 349                                                 u8 tos, gfp_t gfp, int dev_if)
 350 {
 351         return __rds_conn_create(net, laddr, faddr, trans, gfp, tos, 1, dev_if);
 352 }
 353 EXPORT_SYMBOL_GPL(rds_conn_create_outgoing);
 354
 355 void rds_conn_shutdown(struct rds_conn_path *cp)
 356 {
 357         struct rds_connection *conn = cp->cp_conn;
 358
 359         /* shut it down unless it's down already */
 360         if (!rds_conn_path_transition(cp, RDS_CONN_DOWN, RDS_CONN_DOWN)) {
 361                 /*
 362                  * Quiesce the connection mgmt handlers before we start tearing
 363                  * things down. We don't hold the mutex for the entire
 364                  * duration of the shutdown operation, else we may be
 365                  * deadlocking with the CM handler. Instead, the CM event
 366                  * handler is supposed to check for state DISCONNECTING
 367                  */
 368                 mutex_lock(&cp->cp_cm_lock);
 369                 if (!rds_conn_path_transition(cp, RDS_CONN_UP,
 370                                               RDS_CONN_DISCONNECTING) &&
 371                     !rds_conn_path_transition(cp, RDS_CONN_ERROR,
 372                                               RDS_CONN_DISCONNECTING)) {
 373                         rds_conn_path_error(cp,
 374                                             "shutdown called in state %d\n",
 375                                             atomic_read(&cp->cp_state));
 376                         mutex_unlock(&cp->cp_cm_lock);
 377                         return;
 378                 }
 379                 mutex_unlock(&cp->cp_cm_lock);
 380
 381                 wait_event(cp->cp_waitq,
 382                            !test_bit(RDS_IN_XMIT, &cp->cp_flags));
 383                 wait_event(cp->cp_waitq,
 384                            !test_bit(RDS_RECV_REFILL, &cp->cp_flags));
 385
 386                 conn->c_trans->conn_path_shutdown(cp);
 387                 rds_conn_path_reset(cp);
 388
 389                 if (!rds_conn_path_transition(cp, RDS_CONN_DISCONNECTING,
 390                                               RDS_CONN_DOWN) &&
 391                     !rds_conn_path_transition(cp, RDS_CONN_ERROR,
 392                                               RDS_CONN_DOWN)) {
 393                         /* This can happen - eg when we're in the middle of tearing
 394                          * down the connection, and someone unloads the rds module.
 395                          * Quite reproducible with loopback connections.
 396                          * Mostly harmless.
 397                          *
 398                          * Note that this also happens with rds-tcp because
 399                          * we could have triggered rds_conn_path_drop in irq
 400                          * mode from rds_tcp_state change on the receipt of
 401                          * a FIN, thus we need to recheck for RDS_CONN_ERROR
 402                          * here.
 403                          */
 404                         rds_conn_path_error(cp, "%s: failed to transition "
 405                                             "to state DOWN, current state "
 406                                             "is %d\n", __func__,
 407                                             atomic_read(&cp->cp_state));
 408                         return;
 409                 }
 410         }
 411
 412         /* Then reconnect if it's still live.
 413          * The passive side of an IB loopback connection is never added
 414          * to the conn hash, so we never trigger a reconnect on this
 415          * conn - the reconnect is always triggered by the active peer. */
 416         cancel_delayed_work_sync(&cp->cp_conn_w);
 417         rcu_read_lock();
 418         if (!hlist_unhashed(&conn->c_hash_node)) {
 419                 rcu_read_unlock();
 420                 rds_queue_reconnect(cp);
 421         } else {
 422                 rcu_read_unlock();
 423         }
 424 }
 425
 426 /* destroy a single rds_conn_path. rds_conn_destroy() iterates over
 427  * all paths using rds_conn_path_destroy()
 428  */
 429 static void rds_conn_path_destroy(struct rds_conn_path *cp)
 430 {
 431         struct rds_message *rm, *rtmp;
 432
 433         if (!cp->cp_transport_data)
 434                 return;
 435
 436         /* make sure lingering queued work won't try to ref the conn */
 437         cancel_delayed_work_sync(&cp->cp_send_w);
 438         cancel_delayed_work_sync(&cp->cp_recv_w);
 439
 440         rds_conn_path_drop(cp, true);
 441         flush_work(&cp->cp_down_w);
 442
 443         /* tear down queued messages */
 444         list_for_each_entry_safe(rm, rtmp,
 445                                  &cp->cp_send_queue,
 446                                  m_conn_item) {
 447                 list_del_init(&rm->m_conn_item);
 448                 BUG_ON(!list_empty(&rm->m_sock_item));
 449                 rds_message_put(rm);
 450         }
 451         if (cp->cp_xmit_rm)
 452                 rds_message_put(cp->cp_xmit_rm);
 453
 454         WARN_ON(delayed_work_pending(&cp->cp_send_w));
 455         WARN_ON(delayed_work_pending(&cp->cp_recv_w));
 456         WARN_ON(delayed_work_pending(&cp->cp_conn_w));
 457         WARN_ON(work_pending(&cp->cp_down_w));
 458
 459         cp->cp_conn->c_trans->conn_free(cp->cp_transport_data);
 460 }
 461
 462 /*
 463  * Stop and free a connection.
 464  *
 465  * This can only be used in very limited circumstances.  It assumes that once
 466  * the conn has been shutdown that no one else is referencing the connection.
 467  * We can only ensure this in the rmmod path in the current code.
 468  */
 469 void rds_conn_destroy(struct rds_connection *conn)
 470 {
 471         unsigned long flags;
 472         int i;
 473         struct rds_conn_path *cp;
 474         int npaths = (conn->c_trans->t_mp_capable ? RDS_MPATH_WORKERS : 1);
 475
 476         rdsdebug("freeing conn %p for %pI4 -> "
 477                  "%pI4\n", conn, &conn->c_laddr,
 478                  &conn->c_faddr);
 479
 480         /* Ensure conn will not be scheduled for reconnect */
 481         spin_lock_irq(&rds_conn_lock);
 482         hlist_del_init_rcu(&conn->c_hash_node);
 483         spin_unlock_irq(&rds_conn_lock);
 484         synchronize_rcu();
 485
 486         /* shut the connection down */
 487         for (i = 0; i < npaths; i++) {
 488                 cp = &conn->c_path[i];
 489                 rds_conn_path_destroy(cp);
 490                 BUG_ON(!list_empty(&cp->cp_retrans));
 491         }
 492
 493         /*
 494          * The congestion maps aren't freed up here.  They're
 495          * freed by rds_cong_exit() after all the connections
 496          * have been freed.
 497          */
 498         rds_cong_remove_conn(conn);
 499
 500         kfree(conn->c_path);
 501         kmem_cache_free(rds_conn_slab, conn);
 502
 503         spin_lock_irqsave(&rds_conn_lock, flags);
 504         rds_conn_count--;
 505         spin_unlock_irqrestore(&rds_conn_lock, flags);
 506 }
 507 EXPORT_SYMBOL_GPL(rds_conn_destroy);
 508
 509 static void __rds_inc_msg_cp(struct rds_incoming *inc,
 510                              struct rds_info_iterator *iter,
 511                              void *saddr, void *daddr, int flip, bool isv6)
 512 {
 513 #if IS_ENABLED(CONFIG_IPV6)
 514         if (isv6)
 515                 rds6_inc_info_copy(inc, iter, saddr, daddr, flip);
 516         else
 517 #endif
 518                 rds_inc_info_copy(inc, iter, *(__be32 *)saddr,
 519                                   *(__be32 *)daddr, flip);
 520 }
 521
 522 static void rds_conn_message_info_cmn(struct socket *sock, unsigned int len,
 523                                       struct rds_info_iterator *iter,
 524                                       struct rds_info_lengths *lens,
 525                                       int want_send, bool isv6)
 526 {
 527         struct hlist_head *head;
 528         struct list_head *list;
 529         struct rds_connection *conn;
 530         struct rds_message *rm;
 531         unsigned int total = 0;
 532         unsigned long flags;
 533         size_t i;
 534         int j;
 535
 536         if (isv6)
 537                 len /= sizeof(struct rds6_info_message);
 538         else
 539                 len /= sizeof(struct rds_info_message);
 540
 541         rcu_read_lock();
 542
 543         for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
 544              i++, head++) {
 545                 hlist_for_each_entry_rcu(conn, head, c_hash_node) {
 546                         struct rds_conn_path *cp;
 547                         int npaths;
 548
 549                         if (!isv6 && conn->c_isv6)
 550                                 continue;
 551
 552                         npaths = (conn->c_trans->t_mp_capable ?
 553                                  RDS_MPATH_WORKERS : 1);
 554
 555                         for (j = 0; j < npaths; j++) {
 556                                 cp = &conn->c_path[j];
 557                                 if (want_send)
 558                                         list = &cp->cp_send_queue;
 559                                 else
 560                                         list = &cp->cp_retrans;
 561
 562                                 spin_lock_irqsave(&cp->cp_lock, flags);
 563
 564                                 /* XXX too lazy to maintain counts.. */
 565                                 list_for_each_entry(rm, list, m_conn_item) {
 566                                         total++;
 567                                         if (total <= len)
 568                                                 __rds_inc_msg_cp(&rm->m_inc,
 569                                                                  iter,
 570                                                                  &conn->c_laddr,
 571                                                                  &conn->c_faddr,
 572                                                                  0, isv6);
 573                                 }
 574
 575                                 spin_unlock_irqrestore(&cp->cp_lock, flags);
 576                         }
 577                 }
 578         }
 579         rcu_read_unlock();
 580
 581         lens->nr = total;
 582         if (isv6)
 583                 lens->each = sizeof(struct rds6_info_message);
 584         else
 585                 lens->each = sizeof(struct rds_info_message);
 586 }
 587
 588 static void rds_conn_message_info(struct socket *sock, unsigned int len,
 589                                   struct rds_info_iterator *iter,
 590                                   struct rds_info_lengths *lens,
 591                                   int want_send)
 592 {
 593         rds_conn_message_info_cmn(sock, len, iter, lens, want_send, false);
 594 }
 595
 596 #if IS_ENABLED(CONFIG_IPV6)
 597 static void rds6_conn_message_info(struct socket *sock, unsigned int len,
 598                                    struct rds_info_iterator *iter,
 599                                    struct rds_info_lengths *lens,
 600                                    int want_send)
 601 {
 602         rds_conn_message_info_cmn(sock, len, iter, lens, want_send, true);
 603 }
 604 #endif
 605
 606 static void rds_conn_message_info_send(struct socket *sock, unsigned int len,
 607                                        struct rds_info_iterator *iter,
 608                                        struct rds_info_lengths *lens)
 609 {
 610         rds_conn_message_info(sock, len, iter, lens, 1);
 611 }
 612
 613 #if IS_ENABLED(CONFIG_IPV6)
 614 static void rds6_conn_message_info_send(struct socket *sock, unsigned int len,
 615                                         struct rds_info_iterator *iter,
 616                                         struct rds_info_lengths *lens)
 617 {
 618         rds6_conn_message_info(sock, len, iter, lens, 1);
 619 }
 620 #endif
 621
 622 static void rds_conn_message_info_retrans(struct socket *sock,
 623                                           unsigned int len,
 624                                           struct rds_info_iterator *iter,
 625                                           struct rds_info_lengths *lens)
 626 {
 627         rds_conn_message_info(sock, len, iter, lens, 0);
 628 }
 629
 630 #if IS_ENABLED(CONFIG_IPV6)
 631 static void rds6_conn_message_info_retrans(struct socket *sock,
 632                                            unsigned int len,
 633                                            struct rds_info_iterator *iter,
 634                                            struct rds_info_lengths *lens)
 635 {
 636         rds6_conn_message_info(sock, len, iter, lens, 0);
 637 }
 638 #endif
 639
 640 void rds_for_each_conn_info(struct socket *sock, unsigned int len,
 641                           struct rds_info_iterator *iter,
 642                           struct rds_info_lengths *lens,
 643                           int (*visitor)(struct rds_connection *, void *),
 644                           u64 *buffer,
 645                           size_t item_len)
 646 {
 647         struct hlist_head *head;
 648         struct rds_connection *conn;
 649         size_t i;
 650
 651         rcu_read_lock();
 652
 653         lens->nr = 0;
 654         lens->each = item_len;
 655
 656         for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
 657              i++, head++) {
 658                 hlist_for_each_entry_rcu(conn, head, c_hash_node) {
 659
 660                         /* XXX no c_lock usage.. */
 661                         if (!visitor(conn, buffer))
 662                                 continue;
 663
 664                         /* We copy as much as we can fit in the buffer,
 665                          * but we count all items so that the caller
 666                          * can resize the buffer. */
 667                         if (len >= item_len) {
 668                                 rds_info_copy(iter, buffer, item_len);
 669                                 len -= item_len;
 670                         }
 671                         lens->nr++;
 672                 }
 673         }
 674         rcu_read_unlock();
 675 }
 676 EXPORT_SYMBOL_GPL(rds_for_each_conn_info);
 677
 678 static void rds_walk_conn_path_info(struct socket *sock, unsigned int len,
 679                                     struct rds_info_iterator *iter,
 680                                     struct rds_info_lengths *lens,
 681                                     int (*visitor)(struct rds_conn_path *, void *),
 682                                     u64 *buffer,
 683                                     size_t item_len)
 684 {
 685         struct hlist_head *head;
 686         struct rds_connection *conn;
 687         size_t i;
 688
 689         rcu_read_lock();
 690
 691         lens->nr = 0;
 692         lens->each = item_len;
 693
 694         for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
 695              i++, head++) {
 696                 hlist_for_each_entry_rcu(conn, head, c_hash_node) {
 697                         struct rds_conn_path *cp;
 698
 699                         /* XXX We only copy the information from the first
 700                          * path for now.  The problem is that if there are
 701                          * more than one underlying paths, we cannot report
 702                          * information of all of them using the existing
 703                          * API.  For example, there is only one next_tx_seq,
 704                          * which path's next_tx_seq should we report?  It is
 705                          * a bug in the design of MPRDS.
 706                          */
 707                         cp = conn->c_path;
 708
 709                         /* XXX no cp_lock usage.. */
 710                         if (!visitor(cp, buffer))
 711                                 continue;
 712
 713                         /* We copy as much as we can fit in the buffer,
 714                          * but we count all items so that the caller
 715                          * can resize the buffer.
 716                          */
 717                         if (len >= item_len) {
 718                                 rds_info_copy(iter, buffer, item_len);
 719                                 len -= item_len;
 720                         }
 721                         lens->nr++;
 722                 }
 723         }
 724         rcu_read_unlock();
 725 }
 726
 727 static int rds_conn_info_visitor(struct rds_conn_path *cp, void *buffer)
 728 {
 729         struct rds_info_connection *cinfo = buffer;
 730         struct rds_connection *conn = cp->cp_conn;
 731
 732         if (conn->c_isv6)
 733                 return 0;
 734
 735         cinfo->next_tx_seq = cp->cp_next_tx_seq;
 736         cinfo->next_rx_seq = cp->cp_next_rx_seq;
 737         cinfo->laddr = conn->c_laddr.s6_addr32[3];
 738         cinfo->faddr = conn->c_faddr.s6_addr32[3];
 739         cinfo->tos = conn->c_tos;
 740         strncpy(cinfo->transport, conn->c_trans->t_name,
 741                 sizeof(cinfo->transport));
 742         cinfo->flags = 0;
 743
 744         rds_conn_info_set(cinfo->flags, test_bit(RDS_IN_XMIT, &cp->cp_flags),
 745                           SENDING);
 746         /* XXX Future: return the state rather than these funky bits */
 747         rds_conn_info_set(cinfo->flags,
 748                           atomic_read(&cp->cp_state) == RDS_CONN_CONNECTING,
 749                           CONNECTING);
 750         rds_conn_info_set(cinfo->flags,
 751                           atomic_read(&cp->cp_state) == RDS_CONN_UP,
 752                           CONNECTED);
 753         return 1;
 754 }
 755
 756 #if IS_ENABLED(CONFIG_IPV6)
 757 static int rds6_conn_info_visitor(struct rds_conn_path *cp, void *buffer)
 758 {
 759         struct rds6_info_connection *cinfo6 = buffer;
 760         struct rds_connection *conn = cp->cp_conn;
 761
 762         cinfo6->next_tx_seq = cp->cp_next_tx_seq;
 763         cinfo6->next_rx_seq = cp->cp_next_rx_seq;
 764         cinfo6->laddr = conn->c_laddr;
 765         cinfo6->faddr = conn->c_faddr;
 766         strncpy(cinfo6->transport, conn->c_trans->t_name,
 767                 sizeof(cinfo6->transport));
 768         cinfo6->flags = 0;
 769
 770         rds_conn_info_set(cinfo6->flags, test_bit(RDS_IN_XMIT, &cp->cp_flags),
 771                           SENDING);
 772         /* XXX Future: return the state rather than these funky bits */
 773         rds_conn_info_set(cinfo6->flags,
 774                           atomic_read(&cp->cp_state) == RDS_CONN_CONNECTING,
 775                           CONNECTING);
 776         rds_conn_info_set(cinfo6->flags,
 777                           atomic_read(&cp->cp_state) == RDS_CONN_UP,
 778                           CONNECTED);
 779         /* Just return 1 as there is no error case. This is a helper function
 780          * for rds_walk_conn_path_info() and it wants a return value.
 781          */
 782         return 1;
 783 }
 784 #endif
 785
 786 static void rds_conn_info(struct socket *sock, unsigned int len,
 787                           struct rds_info_iterator *iter,
 788                           struct rds_info_lengths *lens)
 789 {
 790         u64 buffer[(sizeof(struct rds_info_connection) + 7) / 8];
 791
 792         rds_walk_conn_path_info(sock, len, iter, lens,
 793                                 rds_conn_info_visitor,
 794                                 buffer,
 795                                 sizeof(struct rds_info_connection));
 796 }
 797
 798 #if IS_ENABLED(CONFIG_IPV6)
 799 static void rds6_conn_info(struct socket *sock, unsigned int len,
 800                            struct rds_info_iterator *iter,
 801                            struct rds_info_lengths *lens)
 802 {
 803         u64 buffer[(sizeof(struct rds6_info_connection) + 7) / 8];
 804
 805         rds_walk_conn_path_info(sock, len, iter, lens,
 806                                 rds6_conn_info_visitor,
 807                                 buffer,
 808                                 sizeof(struct rds6_info_connection));
 809 }
 810 #endif
 811
 812 int rds_conn_init(void)
 813 {
 814         int ret;
 815
 816         ret = rds_loop_net_init(); /* register pernet callback */
 817         if (ret)
 818                 return ret;
 819
 820         rds_conn_slab = kmem_cache_create("rds_connection",
 821                                           sizeof(struct rds_connection),
 822                                           0, 0, NULL);
 823         if (!rds_conn_slab) {
 824                 rds_loop_net_exit();
 825                 return -ENOMEM;
 826         }
 827
 828         rds_info_register_func(RDS_INFO_CONNECTIONS, rds_conn_info);
 829         rds_info_register_func(RDS_INFO_SEND_MESSAGES,
 830                                rds_conn_message_info_send);
 831         rds_info_register_func(RDS_INFO_RETRANS_MESSAGES,
 832                                rds_conn_message_info_retrans);
 833 #if IS_ENABLED(CONFIG_IPV6)
 834         rds_info_register_func(RDS6_INFO_CONNECTIONS, rds6_conn_info);
 835         rds_info_register_func(RDS6_INFO_SEND_MESSAGES,
 836                                rds6_conn_message_info_send);
 837         rds_info_register_func(RDS6_INFO_RETRANS_MESSAGES,
 838                                rds6_conn_message_info_retrans);
 839 #endif
 840         return 0;
 841 }
 842
 843 void rds_conn_exit(void)
 844 {
 845         rds_loop_net_exit(); /* unregister pernet callback */
 846         rds_loop_exit();
 847
 848         WARN_ON(!hlist_empty(rds_conn_hash));
 849
 850         kmem_cache_destroy(rds_conn_slab);
 851
 852         rds_info_deregister_func(RDS_INFO_CONNECTIONS, rds_conn_info);
 853         rds_info_deregister_func(RDS_INFO_SEND_MESSAGES,
 854                                  rds_conn_message_info_send);
 855         rds_info_deregister_func(RDS_INFO_RETRANS_MESSAGES,
 856                                  rds_conn_message_info_retrans);
 857 #if IS_ENABLED(CONFIG_IPV6)
 858         rds_info_deregister_func(RDS6_INFO_CONNECTIONS, rds6_conn_info);
 859         rds_info_deregister_func(RDS6_INFO_SEND_MESSAGES,
 860                                  rds6_conn_message_info_send);
 861         rds_info_deregister_func(RDS6_INFO_RETRANS_MESSAGES,
 862                                  rds6_conn_message_info_retrans);
 863 #endif
 864 }
 865
 866 /*
 867  * Force a disconnect
 868  */
 869 void rds_conn_path_drop(struct rds_conn_path *cp, bool destroy)
 870 {
 871         atomic_set(&cp->cp_state, RDS_CONN_ERROR);
 872
 873         rcu_read_lock();
 874         if (!destroy && rds_destroy_pending(cp->cp_conn)) {
 875                 rcu_read_unlock();
 876                 return;
 877         }
 878         queue_work(rds_wq, &cp->cp_down_w);
 879         rcu_read_unlock();
 880 }
 881 EXPORT_SYMBOL_GPL(rds_conn_path_drop);
 882
 883 void rds_conn_drop(struct rds_connection *conn)
 884 {
 885         WARN_ON(conn->c_trans->t_mp_capable);
 886         rds_conn_path_drop(&conn->c_path[0], false);
 887 }
 888 EXPORT_SYMBOL_GPL(rds_conn_drop);
 889
 890 /*
 891  * If the connection is down, trigger a connect. We may have scheduled a
 892  * delayed reconnect however - in this case we should not interfere.
 893  */
 894 void rds_conn_path_connect_if_down(struct rds_conn_path *cp)
 895 {
 896         rcu_read_lock();
 897         if (rds_destroy_pending(cp->cp_conn)) {
 898                 rcu_read_unlock();
 899                 return;
 900         }
 901         if (rds_conn_path_state(cp) == RDS_CONN_DOWN &&
 902             !test_and_set_bit(RDS_RECONNECT_PENDING, &cp->cp_flags))
 903                 queue_delayed_work(rds_wq, &cp->cp_conn_w, 0);
 904         rcu_read_unlock();
 905 }
 906 EXPORT_SYMBOL_GPL(rds_conn_path_connect_if_down);
 907
 908 /* Check connectivity of all paths
 909  */
 910 void rds_check_all_paths(struct rds_connection *conn)
 911 {
 912         int i = 0;
 913
 914         do {
 915                 rds_conn_path_connect_if_down(&conn->c_path[i]);
 916         } while (++i < conn->c_npaths);
 917 }
 918
 919 void rds_conn_connect_if_down(struct rds_connection *conn)
 920 {
 921         WARN_ON(conn->c_trans->t_mp_capable);
 922         rds_conn_path_connect_if_down(&conn->c_path[0]);
 923 }
 924 EXPORT_SYMBOL_GPL(rds_conn_connect_if_down);
 925
 926 void
 927 __rds_conn_path_error(struct rds_conn_path *cp, const char *fmt, ...)
 928 {
 929         va_list ap;
 930
 931         va_start(ap, fmt);
 932         vprintk(fmt, ap);
 933         va_end(ap);
 934
 935         rds_conn_path_drop(cp, false);
 936 }