net/rds/ib.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/in.h>
  35 #include <linux/if.h>
  36 #include <linux/netdevice.h>
  37 #include <linux/inetdevice.h>
  38 #include <linux/if_arp.h>
  39 #include <linux/delay.h>
  40 #include <linux/slab.h>
  41 #include <linux/module.h>
  42 #include <net/addrconf.h>
  43
  44 #include "rds_single_path.h"
  45 #include "rds.h"
  46 #include "ib.h"
  47 #include "ib_mr.h"
  48
  49 static unsigned int rds_ib_mr_1m_pool_size = RDS_MR_1M_POOL_SIZE;
  50 static unsigned int rds_ib_mr_8k_pool_size = RDS_MR_8K_POOL_SIZE;
  51 unsigned int rds_ib_retry_count = RDS_IB_DEFAULT_RETRY_COUNT;
  52 static atomic_t rds_ib_unloading;
  53
  54 module_param(rds_ib_mr_1m_pool_size, int, 0444);
  55 MODULE_PARM_DESC(rds_ib_mr_1m_pool_size, " Max number of 1M mr per HCA");
  56 module_param(rds_ib_mr_8k_pool_size, int, 0444);
  57 MODULE_PARM_DESC(rds_ib_mr_8k_pool_size, " Max number of 8K mr per HCA");
  58 module_param(rds_ib_retry_count, int, 0444);
  59 MODULE_PARM_DESC(rds_ib_retry_count, " Number of hw retries before reporting an error");
  60
  61 /*
  62  * we have a clumsy combination of RCU and a rwsem protecting this list
  63  * because it is used both in the get_mr fast path and while blocking in
  64  * the FMR flushing path.
  65  */
  66 DECLARE_RWSEM(rds_ib_devices_lock);
  67 struct list_head rds_ib_devices;
  68
  69 /* NOTE: if also grabbing ibdev lock, grab this first */
  70 DEFINE_SPINLOCK(ib_nodev_conns_lock);
  71 LIST_HEAD(ib_nodev_conns);
  72
  73 static void rds_ib_nodev_connect(void)
  74 {
  75         struct rds_ib_connection *ic;
  76
  77         spin_lock(&ib_nodev_conns_lock);
  78         list_for_each_entry(ic, &ib_nodev_conns, ib_node)
  79                 rds_conn_connect_if_down(ic->conn);
  80         spin_unlock(&ib_nodev_conns_lock);
  81 }
  82
  83 static void rds_ib_dev_shutdown(struct rds_ib_device *rds_ibdev)
  84 {
  85         struct rds_ib_connection *ic;
  86         unsigned long flags;
  87
  88         spin_lock_irqsave(&rds_ibdev->spinlock, flags);
  89         list_for_each_entry(ic, &rds_ibdev->conn_list, ib_node)
  90                 rds_conn_drop(ic->conn);
  91         spin_unlock_irqrestore(&rds_ibdev->spinlock, flags);
  92 }
  93
  94 /*
  95  * rds_ib_destroy_mr_pool() blocks on a few things and mrs drop references
  96  * from interrupt context so we push freing off into a work struct in krdsd.
  97  */
  98 static void rds_ib_dev_free(struct work_struct *work)
  99 {
 100         struct rds_ib_ipaddr *i_ipaddr, *i_next;
 101         struct rds_ib_device *rds_ibdev = container_of(work,
 102                                         struct rds_ib_device, free_work);
 103
 104         if (rds_ibdev->mr_8k_pool)
 105                 rds_ib_destroy_mr_pool(rds_ibdev->mr_8k_pool);
 106         if (rds_ibdev->mr_1m_pool)
 107                 rds_ib_destroy_mr_pool(rds_ibdev->mr_1m_pool);
 108         if (rds_ibdev->pd)
 109                 ib_dealloc_pd(rds_ibdev->pd);
 110
 111         list_for_each_entry_safe(i_ipaddr, i_next, &rds_ibdev->ipaddr_list, list) {
 112                 list_del(&i_ipaddr->list);
 113                 kfree(i_ipaddr);
 114         }
 115
 116         kfree(rds_ibdev->vector_load);
 117
 118         kfree(rds_ibdev);
 119 }
 120
 121 void rds_ib_dev_put(struct rds_ib_device *rds_ibdev)
 122 {
 123         BUG_ON(refcount_read(&rds_ibdev->refcount) == 0);
 124         if (refcount_dec_and_test(&rds_ibdev->refcount))
 125                 queue_work(rds_wq, &rds_ibdev->free_work);
 126 }
 127
 128 static void rds_ib_add_one(struct ib_device *device)
 129 {
 130         struct rds_ib_device *rds_ibdev;
 131         bool has_fr, has_fmr;
 132
 133         /* Only handle IB (no iWARP) devices */
 134         if (device->node_type != RDMA_NODE_IB_CA)
 135                 return;
 136
 137         rds_ibdev = kzalloc_node(sizeof(struct rds_ib_device), GFP_KERNEL,
 138                                  ibdev_to_node(device));
 139         if (!rds_ibdev)
 140                 return;
 141
 142         spin_lock_init(&rds_ibdev->spinlock);
 143         refcount_set(&rds_ibdev->refcount, 1);
 144         INIT_WORK(&rds_ibdev->free_work, rds_ib_dev_free);
 145
 146         rds_ibdev->max_wrs = device->attrs.max_qp_wr;
 147         rds_ibdev->max_sge = min(device->attrs.max_send_sge, RDS_IB_MAX_SGE);
 148
 149         has_fr = (device->attrs.device_cap_flags &
 150                   IB_DEVICE_MEM_MGT_EXTENSIONS);
 151         has_fmr = (device->ops.alloc_fmr && device->ops.dealloc_fmr &&
 152                    device->ops.map_phys_fmr && device->ops.unmap_fmr);
 153         rds_ibdev->use_fastreg = (has_fr && !has_fmr);
 154
 155         rds_ibdev->fmr_max_remaps = device->attrs.max_map_per_fmr?: 32;
 156         rds_ibdev->max_1m_mrs = device->attrs.max_mr ?
 157                 min_t(unsigned int, (device->attrs.max_mr / 2),
 158                       rds_ib_mr_1m_pool_size) : rds_ib_mr_1m_pool_size;
 159
 160         rds_ibdev->max_8k_mrs = device->attrs.max_mr ?
 161                 min_t(unsigned int, ((device->attrs.max_mr / 2) * RDS_MR_8K_SCALE),
 162                       rds_ib_mr_8k_pool_size) : rds_ib_mr_8k_pool_size;
 163
 164         rds_ibdev->max_initiator_depth = device->attrs.max_qp_init_rd_atom;
 165         rds_ibdev->max_responder_resources = device->attrs.max_qp_rd_atom;
 166
 167         rds_ibdev->vector_load = kcalloc(device->num_comp_vectors,
 168                                          sizeof(int),
 169                                          GFP_KERNEL);
 170         if (!rds_ibdev->vector_load) {
 171                 pr_err("RDS/IB: %s failed to allocate vector memory\n",
 172                         __func__);
 173                 goto put_dev;
 174         }
 175
 176         rds_ibdev->dev = device;
 177         rds_ibdev->pd = ib_alloc_pd(device, 0);
 178         if (IS_ERR(rds_ibdev->pd)) {
 179                 rds_ibdev->pd = NULL;
 180                 goto put_dev;
 181         }
 182
 183         rds_ibdev->mr_1m_pool =
 184                 rds_ib_create_mr_pool(rds_ibdev, RDS_IB_MR_1M_POOL);
 185         if (IS_ERR(rds_ibdev->mr_1m_pool)) {
 186                 rds_ibdev->mr_1m_pool = NULL;
 187                 goto put_dev;
 188         }
 189
 190         rds_ibdev->mr_8k_pool =
 191                 rds_ib_create_mr_pool(rds_ibdev, RDS_IB_MR_8K_POOL);
 192         if (IS_ERR(rds_ibdev->mr_8k_pool)) {
 193                 rds_ibdev->mr_8k_pool = NULL;
 194                 goto put_dev;
 195         }
 196
 197         rdsdebug("RDS/IB: max_mr = %d, max_wrs = %d, max_sge = %d, fmr_max_remaps = %d, max_1m_mrs = %d, max_8k_mrs = %d\n",
 198                  device->attrs.max_fmr, rds_ibdev->max_wrs, rds_ibdev->max_sge,
 199                  rds_ibdev->fmr_max_remaps, rds_ibdev->max_1m_mrs,
 200                  rds_ibdev->max_8k_mrs);
 201
 202         pr_info("RDS/IB: %s: %s supported and preferred\n",
 203                 device->name,
 204                 rds_ibdev->use_fastreg ? "FRMR" : "FMR");
 205
 206         INIT_LIST_HEAD(&rds_ibdev->ipaddr_list);
 207         INIT_LIST_HEAD(&rds_ibdev->conn_list);
 208
 209         down_write(&rds_ib_devices_lock);
 210         list_add_tail_rcu(&rds_ibdev->list, &rds_ib_devices);
 211         up_write(&rds_ib_devices_lock);
 212         refcount_inc(&rds_ibdev->refcount);
 213
 214         ib_set_client_data(device, &rds_ib_client, rds_ibdev);
 215         refcount_inc(&rds_ibdev->refcount);
 216
 217         rds_ib_nodev_connect();
 218
 219 put_dev:
 220         rds_ib_dev_put(rds_ibdev);
 221 }
 222
 223 /*
 224  * New connections use this to find the device to associate with the
 225  * connection.  It's not in the fast path so we're not concerned about the
 226  * performance of the IB call.  (As of this writing, it uses an interrupt
 227  * blocking spinlock to serialize walking a per-device list of all registered
 228  * clients.)
 229  *
 230  * RCU is used to handle incoming connections racing with device teardown.
 231  * Rather than use a lock to serialize removal from the client_data and
 232  * getting a new reference, we use an RCU grace period.  The destruction
 233  * path removes the device from client_data and then waits for all RCU
 234  * readers to finish.
 235  *
 236  * A new connection can get NULL from this if its arriving on a
 237  * device that is in the process of being removed.
 238  */
 239 struct rds_ib_device *rds_ib_get_client_data(struct ib_device *device)
 240 {
 241         struct rds_ib_device *rds_ibdev;
 242
 243         rcu_read_lock();
 244         rds_ibdev = ib_get_client_data(device, &rds_ib_client);
 245         if (rds_ibdev)
 246                 refcount_inc(&rds_ibdev->refcount);
 247         rcu_read_unlock();
 248         return rds_ibdev;
 249 }
 250
 251 /*
 252  * The IB stack is letting us know that a device is going away.  This can
 253  * happen if the underlying HCA driver is removed or if PCI hotplug is removing
 254  * the pci function, for example.
 255  *
 256  * This can be called at any time and can be racing with any other RDS path.
 257  */
 258 static void rds_ib_remove_one(struct ib_device *device, void *client_data)
 259 {
 260         struct rds_ib_device *rds_ibdev = client_data;
 261
 262         if (!rds_ibdev)
 263                 return;
 264
 265         rds_ib_dev_shutdown(rds_ibdev);
 266
 267         /* stop connection attempts from getting a reference to this device. */
 268         ib_set_client_data(device, &rds_ib_client, NULL);
 269
 270         down_write(&rds_ib_devices_lock);
 271         list_del_rcu(&rds_ibdev->list);
 272         up_write(&rds_ib_devices_lock);
 273
 274         /*
 275          * This synchronize rcu is waiting for readers of both the ib
 276          * client data and the devices list to finish before we drop
 277          * both of those references.
 278          */
 279         synchronize_rcu();
 280         rds_ib_dev_put(rds_ibdev);
 281         rds_ib_dev_put(rds_ibdev);
 282 }
 283
 284 struct ib_client rds_ib_client = {
 285         .name   = "rds_ib",
 286         .add    = rds_ib_add_one,
 287         .remove = rds_ib_remove_one
 288 };
 289
 290 static int rds_ib_conn_info_visitor(struct rds_connection *conn,
 291                                     void *buffer)
 292 {
 293         struct rds_info_rdma_connection *iinfo = buffer;
 294         struct rds_ib_connection *ic;
 295
 296         /* We will only ever look at IB transports */
 297         if (conn->c_trans != &rds_ib_transport)
 298                 return 0;
 299         if (conn->c_isv6)
 300                 return 0;
 301
 302         iinfo->src_addr = conn->c_laddr.s6_addr32[3];
 303         iinfo->dst_addr = conn->c_faddr.s6_addr32[3];
 304         iinfo->tos = conn->c_tos;
 305
 306         memset(&iinfo->src_gid, 0, sizeof(iinfo->src_gid));
 307         memset(&iinfo->dst_gid, 0, sizeof(iinfo->dst_gid));
 308         if (rds_conn_state(conn) == RDS_CONN_UP) {
 309                 struct rds_ib_device *rds_ibdev;
 310
 311                 ic = conn->c_transport_data;
 312
 313                 rdma_read_gids(ic->i_cm_id, (union ib_gid *)&iinfo->src_gid,
 314                                (union ib_gid *)&iinfo->dst_gid);
 315
 316                 rds_ibdev = ic->rds_ibdev;
 317                 iinfo->max_send_wr = ic->i_send_ring.w_nr;
 318                 iinfo->max_recv_wr = ic->i_recv_ring.w_nr;
 319                 iinfo->max_send_sge = rds_ibdev->max_sge;
 320                 rds_ib_get_mr_info(rds_ibdev, iinfo);
 321         }
 322         return 1;
 323 }
 324
 325 #if IS_ENABLED(CONFIG_IPV6)
 326 /* IPv6 version of rds_ib_conn_info_visitor(). */
 327 static int rds6_ib_conn_info_visitor(struct rds_connection *conn,
 328                                      void *buffer)
 329 {
 330         struct rds6_info_rdma_connection *iinfo6 = buffer;
 331         struct rds_ib_connection *ic;
 332
 333         /* We will only ever look at IB transports */
 334         if (conn->c_trans != &rds_ib_transport)
 335                 return 0;
 336
 337         iinfo6->src_addr = conn->c_laddr;
 338         iinfo6->dst_addr = conn->c_faddr;
 339
 340         memset(&iinfo6->src_gid, 0, sizeof(iinfo6->src_gid));
 341         memset(&iinfo6->dst_gid, 0, sizeof(iinfo6->dst_gid));
 342
 343         if (rds_conn_state(conn) == RDS_CONN_UP) {
 344                 struct rds_ib_device *rds_ibdev;
 345
 346                 ic = conn->c_transport_data;
 347                 rdma_read_gids(ic->i_cm_id, (union ib_gid *)&iinfo6->src_gid,
 348                                (union ib_gid *)&iinfo6->dst_gid);
 349                 rds_ibdev = ic->rds_ibdev;
 350                 iinfo6->max_send_wr = ic->i_send_ring.w_nr;
 351                 iinfo6->max_recv_wr = ic->i_recv_ring.w_nr;
 352                 iinfo6->max_send_sge = rds_ibdev->max_sge;
 353                 rds6_ib_get_mr_info(rds_ibdev, iinfo6);
 354         }
 355         return 1;
 356 }
 357 #endif
 358
 359 static void rds_ib_ic_info(struct socket *sock, unsigned int len,
 360                            struct rds_info_iterator *iter,
 361                            struct rds_info_lengths *lens)
 362 {
 363         u64 buffer[(sizeof(struct rds_info_rdma_connection) + 7) / 8];
 364
 365         rds_for_each_conn_info(sock, len, iter, lens,
 366                                 rds_ib_conn_info_visitor,
 367                                 buffer,
 368                                 sizeof(struct rds_info_rdma_connection));
 369 }
 370
 371 #if IS_ENABLED(CONFIG_IPV6)
 372 /* IPv6 version of rds_ib_ic_info(). */
 373 static void rds6_ib_ic_info(struct socket *sock, unsigned int len,
 374                             struct rds_info_iterator *iter,
 375                             struct rds_info_lengths *lens)
 376 {
 377         u64 buffer[(sizeof(struct rds6_info_rdma_connection) + 7) / 8];
 378
 379         rds_for_each_conn_info(sock, len, iter, lens,
 380                                rds6_ib_conn_info_visitor,
 381                                buffer,
 382                                sizeof(struct rds6_info_rdma_connection));
 383 }
 384 #endif
 385
 386 /*
 387  * Early RDS/IB was built to only bind to an address if there is an IPoIB
 388  * device with that address set.
 389  *
 390  * If it were me, I'd advocate for something more flexible.  Sending and
 391  * receiving should be device-agnostic.  Transports would try and maintain
 392  * connections between peers who have messages queued.  Userspace would be
 393  * allowed to influence which paths have priority.  We could call userspace
 394  * asserting this policy "routing".
 395  */
 396 static int rds_ib_laddr_check(struct net *net, const struct in6_addr *addr,
 397                               __u32 scope_id)
 398 {
 399         int ret;
 400         struct rdma_cm_id *cm_id;
 401 #if IS_ENABLED(CONFIG_IPV6)
 402         struct sockaddr_in6 sin6;
 403 #endif
 404         struct sockaddr_in sin;
 405         struct sockaddr *sa;
 406         bool isv4;
 407
 408         isv4 = ipv6_addr_v4mapped(addr);
 409         /* Create a CMA ID and try to bind it. This catches both
 410          * IB and iWARP capable NICs.
 411          */
 412         cm_id = rdma_create_id(&init_net, rds_rdma_cm_event_handler,
 413                                NULL, RDMA_PS_TCP, IB_QPT_RC);
 414         if (IS_ERR(cm_id))
 415                 return PTR_ERR(cm_id);
 416
 417         if (isv4) {
 418                 memset(&sin, 0, sizeof(sin));
 419                 sin.sin_family = AF_INET;
 420                 sin.sin_addr.s_addr = addr->s6_addr32[3];
 421                 sa = (struct sockaddr *)&sin;
 422         } else {
 423 #if IS_ENABLED(CONFIG_IPV6)
 424                 memset(&sin6, 0, sizeof(sin6));
 425                 sin6.sin6_family = AF_INET6;
 426                 sin6.sin6_addr = *addr;
 427                 sin6.sin6_scope_id = scope_id;
 428                 sa = (struct sockaddr *)&sin6;
 429
 430                 /* XXX Do a special IPv6 link local address check here.  The
 431                  * reason is that rdma_bind_addr() always succeeds with IPv6
 432                  * link local address regardless it is indeed configured in a
 433                  * system.
 434                  */
 435                 if (ipv6_addr_type(addr) & IPV6_ADDR_LINKLOCAL) {
 436                         struct net_device *dev;
 437
 438                         if (scope_id == 0) {
 439                                 ret = -EADDRNOTAVAIL;
 440                                 goto out;
 441                         }
 442
 443                         /* Use init_net for now as RDS is not network
 444                          * name space aware.
 445                          */
 446                         dev = dev_get_by_index(&init_net, scope_id);
 447                         if (!dev) {
 448                                 ret = -EADDRNOTAVAIL;
 449                                 goto out;
 450                         }
 451                         if (!ipv6_chk_addr(&init_net, addr, dev, 1)) {
 452                                 dev_put(dev);
 453                                 ret = -EADDRNOTAVAIL;
 454                                 goto out;
 455                         }
 456                         dev_put(dev);
 457                 }
 458 #else
 459                 ret = -EADDRNOTAVAIL;
 460                 goto out;
 461 #endif
 462         }
 463
 464         /* rdma_bind_addr will only succeed for IB & iWARP devices */
 465         ret = rdma_bind_addr(cm_id, sa);
 466         /* due to this, we will claim to support iWARP devices unless we
 467            check node_type. */
 468         if (ret || !cm_id->device ||
 469             cm_id->device->node_type != RDMA_NODE_IB_CA)
 470                 ret = -EADDRNOTAVAIL;
 471
 472         rdsdebug("addr %pI6c%%%u ret %d node type %d\n",
 473                  addr, scope_id, ret,
 474                  cm_id->device ? cm_id->device->node_type : -1);
 475
 476 out:
 477         rdma_destroy_id(cm_id);
 478
 479         return ret;
 480 }
 481
 482 static void rds_ib_unregister_client(void)
 483 {
 484         ib_unregister_client(&rds_ib_client);
 485         /* wait for rds_ib_dev_free() to complete */
 486         flush_workqueue(rds_wq);
 487 }
 488
 489 static void rds_ib_set_unloading(void)
 490 {
 491         atomic_set(&rds_ib_unloading, 1);
 492 }
 493
 494 static bool rds_ib_is_unloading(struct rds_connection *conn)
 495 {
 496         struct rds_conn_path *cp = &conn->c_path[0];
 497
 498         return (test_bit(RDS_DESTROY_PENDING, &cp->cp_flags) ||
 499                 atomic_read(&rds_ib_unloading) != 0);
 500 }
 501
 502 void rds_ib_exit(void)
 503 {
 504         rds_ib_set_unloading();
 505         synchronize_rcu();
 506         rds_info_deregister_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info);
 507 #if IS_ENABLED(CONFIG_IPV6)
 508         rds_info_deregister_func(RDS6_INFO_IB_CONNECTIONS, rds6_ib_ic_info);
 509 #endif
 510         rds_ib_unregister_client();
 511         rds_ib_destroy_nodev_conns();
 512         rds_ib_sysctl_exit();
 513         rds_ib_recv_exit();
 514         rds_trans_unregister(&rds_ib_transport);
 515         rds_ib_mr_exit();
 516 }
 517
 518 static u8 rds_ib_get_tos_map(u8 tos)
 519 {
 520         /* 1:1 user to transport map for RDMA transport.
 521          * In future, if custom map is desired, hook can export
 522          * user configurable map.
 523          */
 524         return tos;
 525 }
 526
 527 struct rds_transport rds_ib_transport = {
 528         .laddr_check            = rds_ib_laddr_check,
 529         .xmit_path_complete     = rds_ib_xmit_path_complete,
 530         .xmit                   = rds_ib_xmit,
 531         .xmit_rdma              = rds_ib_xmit_rdma,
 532         .xmit_atomic            = rds_ib_xmit_atomic,
 533         .recv_path              = rds_ib_recv_path,
 534         .conn_alloc             = rds_ib_conn_alloc,
 535         .conn_free              = rds_ib_conn_free,
 536         .conn_path_connect      = rds_ib_conn_path_connect,
 537         .conn_path_shutdown     = rds_ib_conn_path_shutdown,
 538         .inc_copy_to_user       = rds_ib_inc_copy_to_user,
 539         .inc_free               = rds_ib_inc_free,
 540         .cm_initiate_connect    = rds_ib_cm_initiate_connect,
 541         .cm_handle_connect      = rds_ib_cm_handle_connect,
 542         .cm_connect_complete    = rds_ib_cm_connect_complete,
 543         .stats_info_copy        = rds_ib_stats_info_copy,
 544         .exit                   = rds_ib_exit,
 545         .get_mr                 = rds_ib_get_mr,
 546         .sync_mr                = rds_ib_sync_mr,
 547         .free_mr                = rds_ib_free_mr,
 548         .flush_mrs              = rds_ib_flush_mrs,
 549         .get_tos_map            = rds_ib_get_tos_map,
 550         .t_owner                = THIS_MODULE,
 551         .t_name                 = "infiniband",
 552         .t_unloading            = rds_ib_is_unloading,
 553         .t_type                 = RDS_TRANS_IB
 554 };
 555
 556 int rds_ib_init(void)
 557 {
 558         int ret;
 559
 560         INIT_LIST_HEAD(&rds_ib_devices);
 561
 562         ret = rds_ib_mr_init();
 563         if (ret)
 564                 goto out;
 565
 566         ret = ib_register_client(&rds_ib_client);
 567         if (ret)
 568                 goto out_mr_exit;
 569
 570         ret = rds_ib_sysctl_init();
 571         if (ret)
 572                 goto out_ibreg;
 573
 574         ret = rds_ib_recv_init();
 575         if (ret)
 576                 goto out_sysctl;
 577
 578         rds_trans_register(&rds_ib_transport);
 579
 580         rds_info_register_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info);
 581 #if IS_ENABLED(CONFIG_IPV6)
 582         rds_info_register_func(RDS6_INFO_IB_CONNECTIONS, rds6_ib_ic_info);
 583 #endif
 584
 585         goto out;
 586
 587 out_sysctl:
 588         rds_ib_sysctl_exit();
 589 out_ibreg:
 590         rds_ib_unregister_client();
 591 out_mr_exit:
 592         rds_ib_mr_exit();
 593 out:
 594         return ret;
 595 }
 596
 597 MODULE_LICENSE("GPL");