net/rds/ib.c

   1 /*
   2  * Copyright (c) 2006, 2019 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/dmapool.h>
  34 #include <linux/kernel.h>
  35 #include <linux/in.h>
  36 #include <linux/if.h>
  37 #include <linux/netdevice.h>
  38 #include <linux/inetdevice.h>
  39 #include <linux/if_arp.h>
  40 #include <linux/delay.h>
  41 #include <linux/slab.h>
  42 #include <linux/module.h>
  43 #include <net/addrconf.h>
  44
  45 #include "rds_single_path.h"
  46 #include "rds.h"
  47 #include "ib.h"
  48 #include "ib_mr.h"
  49
  50 static unsigned int rds_ib_mr_1m_pool_size = RDS_MR_1M_POOL_SIZE;
  51 static unsigned int rds_ib_mr_8k_pool_size = RDS_MR_8K_POOL_SIZE;
  52 unsigned int rds_ib_retry_count = RDS_IB_DEFAULT_RETRY_COUNT;
  53 static atomic_t rds_ib_unloading;
  54
  55 module_param(rds_ib_mr_1m_pool_size, int, 0444);
  56 MODULE_PARM_DESC(rds_ib_mr_1m_pool_size, " Max number of 1M mr per HCA");
  57 module_param(rds_ib_mr_8k_pool_size, int, 0444);
  58 MODULE_PARM_DESC(rds_ib_mr_8k_pool_size, " Max number of 8K mr per HCA");
  59 module_param(rds_ib_retry_count, int, 0444);
  60 MODULE_PARM_DESC(rds_ib_retry_count, " Number of hw retries before reporting an error");
  61
  62 /*
  63  * we have a clumsy combination of RCU and a rwsem protecting this list
  64  * because it is used both in the get_mr fast path and while blocking in
  65  * the FMR flushing path.
  66  */
  67 DECLARE_RWSEM(rds_ib_devices_lock);
  68 struct list_head rds_ib_devices;
  69
  70 /* NOTE: if also grabbing ibdev lock, grab this first */
  71 DEFINE_SPINLOCK(ib_nodev_conns_lock);
  72 LIST_HEAD(ib_nodev_conns);
  73
  74 static void rds_ib_nodev_connect(void)
  75 {
  76         struct rds_ib_connection *ic;
  77
  78         spin_lock(&ib_nodev_conns_lock);
  79         list_for_each_entry(ic, &ib_nodev_conns, ib_node)
  80                 rds_conn_connect_if_down(ic->conn);
  81         spin_unlock(&ib_nodev_conns_lock);
  82 }
  83
  84 static void rds_ib_dev_shutdown(struct rds_ib_device *rds_ibdev)
  85 {
  86         struct rds_ib_connection *ic;
  87         unsigned long flags;
  88
  89         spin_lock_irqsave(&rds_ibdev->spinlock, flags);
  90         list_for_each_entry(ic, &rds_ibdev->conn_list, ib_node)
  91                 rds_conn_path_drop(&ic->conn->c_path[0], true);
  92         spin_unlock_irqrestore(&rds_ibdev->spinlock, flags);
  93 }
  94
  95 /*
  96  * rds_ib_destroy_mr_pool() blocks on a few things and mrs drop references
  97  * from interrupt context so we push freing off into a work struct in krdsd.
  98  */
  99 static void rds_ib_dev_free(struct work_struct *work)
 100 {
 101         struct rds_ib_ipaddr *i_ipaddr, *i_next;
 102         struct rds_ib_device *rds_ibdev = container_of(work,
 103                                         struct rds_ib_device, free_work);
 104
 105         if (rds_ibdev->mr_8k_pool)
 106                 rds_ib_destroy_mr_pool(rds_ibdev->mr_8k_pool);
 107         if (rds_ibdev->mr_1m_pool)
 108                 rds_ib_destroy_mr_pool(rds_ibdev->mr_1m_pool);
 109         if (rds_ibdev->pd)
 110                 ib_dealloc_pd(rds_ibdev->pd);
 111         dma_pool_destroy(rds_ibdev->rid_hdrs_pool);
 112
 113         list_for_each_entry_safe(i_ipaddr, i_next, &rds_ibdev->ipaddr_list, list) {
 114                 list_del(&i_ipaddr->list);
 115                 kfree(i_ipaddr);
 116         }
 117
 118         kfree(rds_ibdev->vector_load);
 119
 120         kfree(rds_ibdev);
 121 }
 122
 123 void rds_ib_dev_put(struct rds_ib_device *rds_ibdev)
 124 {
 125         BUG_ON(refcount_read(&rds_ibdev->refcount) == 0);
 126         if (refcount_dec_and_test(&rds_ibdev->refcount))
 127                 queue_work(rds_wq, &rds_ibdev->free_work);
 128 }
 129
 130 static void rds_ib_add_one(struct ib_device *device)
 131 {
 132         struct rds_ib_device *rds_ibdev;
 133         bool has_fr, has_fmr;
 134
 135         /* Only handle IB (no iWARP) devices */
 136         if (device->node_type != RDMA_NODE_IB_CA)
 137                 return;
 138
 139         rds_ibdev = kzalloc_node(sizeof(struct rds_ib_device), GFP_KERNEL,
 140                                  ibdev_to_node(device));
 141         if (!rds_ibdev)
 142                 return;
 143
 144         spin_lock_init(&rds_ibdev->spinlock);
 145         refcount_set(&rds_ibdev->refcount, 1);
 146         INIT_WORK(&rds_ibdev->free_work, rds_ib_dev_free);
 147
 148         INIT_LIST_HEAD(&rds_ibdev->ipaddr_list);
 149         INIT_LIST_HEAD(&rds_ibdev->conn_list);
 150
 151         rds_ibdev->max_wrs = device->attrs.max_qp_wr;
 152         rds_ibdev->max_sge = min(device->attrs.max_send_sge, RDS_IB_MAX_SGE);
 153
 154         has_fr = (device->attrs.device_cap_flags &
 155                   IB_DEVICE_MEM_MGT_EXTENSIONS);
 156         has_fmr = (device->ops.alloc_fmr && device->ops.dealloc_fmr &&
 157                    device->ops.map_phys_fmr && device->ops.unmap_fmr);
 158         rds_ibdev->use_fastreg = (has_fr && !has_fmr);
 159
 160         rds_ibdev->fmr_max_remaps = device->attrs.max_map_per_fmr?: 32;
 161         rds_ibdev->max_1m_mrs = device->attrs.max_mr ?
 162                 min_t(unsigned int, (device->attrs.max_mr / 2),
 163                       rds_ib_mr_1m_pool_size) : rds_ib_mr_1m_pool_size;
 164
 165         rds_ibdev->max_8k_mrs = device->attrs.max_mr ?
 166                 min_t(unsigned int, ((device->attrs.max_mr / 2) * RDS_MR_8K_SCALE),
 167                       rds_ib_mr_8k_pool_size) : rds_ib_mr_8k_pool_size;
 168
 169         rds_ibdev->max_initiator_depth = device->attrs.max_qp_init_rd_atom;
 170         rds_ibdev->max_responder_resources = device->attrs.max_qp_rd_atom;
 171
 172         rds_ibdev->vector_load = kcalloc(device->num_comp_vectors,
 173                                          sizeof(int),
 174                                          GFP_KERNEL);
 175         if (!rds_ibdev->vector_load) {
 176                 pr_err("RDS/IB: %s failed to allocate vector memory\n",
 177                         __func__);
 178                 goto put_dev;
 179         }
 180
 181         rds_ibdev->dev = device;
 182         rds_ibdev->pd = ib_alloc_pd(device, 0);
 183         if (IS_ERR(rds_ibdev->pd)) {
 184                 rds_ibdev->pd = NULL;
 185                 goto put_dev;
 186         }
 187         rds_ibdev->rid_hdrs_pool = dma_pool_create(device->name,
 188                                                    device->dma_device,
 189                                                    sizeof(struct rds_header),
 190                                                    L1_CACHE_BYTES, 0);
 191         if (!rds_ibdev->rid_hdrs_pool)
 192                 goto put_dev;
 193
 194         rds_ibdev->mr_1m_pool =
 195                 rds_ib_create_mr_pool(rds_ibdev, RDS_IB_MR_1M_POOL);
 196         if (IS_ERR(rds_ibdev->mr_1m_pool)) {
 197                 rds_ibdev->mr_1m_pool = NULL;
 198                 goto put_dev;
 199         }
 200
 201         rds_ibdev->mr_8k_pool =
 202                 rds_ib_create_mr_pool(rds_ibdev, RDS_IB_MR_8K_POOL);
 203         if (IS_ERR(rds_ibdev->mr_8k_pool)) {
 204                 rds_ibdev->mr_8k_pool = NULL;
 205                 goto put_dev;
 206         }
 207
 208         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",
 209                  device->attrs.max_fmr, rds_ibdev->max_wrs, rds_ibdev->max_sge,
 210                  rds_ibdev->fmr_max_remaps, rds_ibdev->max_1m_mrs,
 211                  rds_ibdev->max_8k_mrs);
 212
 213         pr_info("RDS/IB: %s: %s supported and preferred\n",
 214                 device->name,
 215                 rds_ibdev->use_fastreg ? "FRMR" : "FMR");
 216
 217         down_write(&rds_ib_devices_lock);
 218         list_add_tail_rcu(&rds_ibdev->list, &rds_ib_devices);
 219         up_write(&rds_ib_devices_lock);
 220         refcount_inc(&rds_ibdev->refcount);
 221
 222         ib_set_client_data(device, &rds_ib_client, rds_ibdev);
 223         refcount_inc(&rds_ibdev->refcount);
 224
 225         rds_ib_nodev_connect();
 226
 227 put_dev:
 228         rds_ib_dev_put(rds_ibdev);
 229 }
 230
 231 /*
 232  * New connections use this to find the device to associate with the
 233  * connection.  It's not in the fast path so we're not concerned about the
 234  * performance of the IB call.  (As of this writing, it uses an interrupt
 235  * blocking spinlock to serialize walking a per-device list of all registered
 236  * clients.)
 237  *
 238  * RCU is used to handle incoming connections racing with device teardown.
 239  * Rather than use a lock to serialize removal from the client_data and
 240  * getting a new reference, we use an RCU grace period.  The destruction
 241  * path removes the device from client_data and then waits for all RCU
 242  * readers to finish.
 243  *
 244  * A new connection can get NULL from this if its arriving on a
 245  * device that is in the process of being removed.
 246  */
 247 struct rds_ib_device *rds_ib_get_client_data(struct ib_device *device)
 248 {
 249         struct rds_ib_device *rds_ibdev;
 250
 251         rcu_read_lock();
 252         rds_ibdev = ib_get_client_data(device, &rds_ib_client);
 253         if (rds_ibdev)
 254                 refcount_inc(&rds_ibdev->refcount);
 255         rcu_read_unlock();
 256         return rds_ibdev;
 257 }
 258
 259 /*
 260  * The IB stack is letting us know that a device is going away.  This can
 261  * happen if the underlying HCA driver is removed or if PCI hotplug is removing
 262  * the pci function, for example.
 263  *
 264  * This can be called at any time and can be racing with any other RDS path.
 265  */
 266 static void rds_ib_remove_one(struct ib_device *device, void *client_data)
 267 {
 268         struct rds_ib_device *rds_ibdev = client_data;
 269
 270         if (!rds_ibdev)
 271                 return;
 272
 273         rds_ib_dev_shutdown(rds_ibdev);
 274
 275         /* stop connection attempts from getting a reference to this device. */
 276         ib_set_client_data(device, &rds_ib_client, NULL);
 277
 278         down_write(&rds_ib_devices_lock);
 279         list_del_rcu(&rds_ibdev->list);
 280         up_write(&rds_ib_devices_lock);
 281
 282         /*
 283          * This synchronize rcu is waiting for readers of both the ib
 284          * client data and the devices list to finish before we drop
 285          * both of those references.
 286          */
 287         synchronize_rcu();
 288         rds_ib_dev_put(rds_ibdev);
 289         rds_ib_dev_put(rds_ibdev);
 290 }
 291
 292 struct ib_client rds_ib_client = {
 293         .name   = "rds_ib",
 294         .add    = rds_ib_add_one,
 295         .remove = rds_ib_remove_one
 296 };
 297
 298 static int rds_ib_conn_info_visitor(struct rds_connection *conn,
 299                                     void *buffer)
 300 {
 301         struct rds_info_rdma_connection *iinfo = buffer;
 302         struct rds_ib_connection *ic = conn->c_transport_data;
 303
 304         /* We will only ever look at IB transports */
 305         if (conn->c_trans != &rds_ib_transport)
 306                 return 0;
 307         if (conn->c_isv6)
 308                 return 0;
 309
 310         iinfo->src_addr = conn->c_laddr.s6_addr32[3];
 311         iinfo->dst_addr = conn->c_faddr.s6_addr32[3];
 312         if (ic) {
 313                 iinfo->tos = conn->c_tos;
 314                 iinfo->sl = ic->i_sl;
 315         }
 316
 317         memset(&iinfo->src_gid, 0, sizeof(iinfo->src_gid));
 318         memset(&iinfo->dst_gid, 0, sizeof(iinfo->dst_gid));
 319         if (rds_conn_state(conn) == RDS_CONN_UP) {
 320                 struct rds_ib_device *rds_ibdev;
 321
 322                 rdma_read_gids(ic->i_cm_id, (union ib_gid *)&iinfo->src_gid,
 323                                (union ib_gid *)&iinfo->dst_gid);
 324
 325                 rds_ibdev = ic->rds_ibdev;
 326                 iinfo->max_send_wr = ic->i_send_ring.w_nr;
 327                 iinfo->max_recv_wr = ic->i_recv_ring.w_nr;
 328                 iinfo->max_send_sge = rds_ibdev->max_sge;
 329                 rds_ib_get_mr_info(rds_ibdev, iinfo);
 330                 iinfo->cache_allocs = atomic_read(&ic->i_cache_allocs);
 331         }
 332         return 1;
 333 }
 334
 335 #if IS_ENABLED(CONFIG_IPV6)
 336 /* IPv6 version of rds_ib_conn_info_visitor(). */
 337 static int rds6_ib_conn_info_visitor(struct rds_connection *conn,
 338                                      void *buffer)
 339 {
 340         struct rds6_info_rdma_connection *iinfo6 = buffer;
 341         struct rds_ib_connection *ic = conn->c_transport_data;
 342
 343         /* We will only ever look at IB transports */
 344         if (conn->c_trans != &rds_ib_transport)
 345                 return 0;
 346
 347         iinfo6->src_addr = conn->c_laddr;
 348         iinfo6->dst_addr = conn->c_faddr;
 349         if (ic) {
 350                 iinfo6->tos = conn->c_tos;
 351                 iinfo6->sl = ic->i_sl;
 352         }
 353
 354         memset(&iinfo6->src_gid, 0, sizeof(iinfo6->src_gid));
 355         memset(&iinfo6->dst_gid, 0, sizeof(iinfo6->dst_gid));
 356
 357         if (rds_conn_state(conn) == RDS_CONN_UP) {
 358                 struct rds_ib_device *rds_ibdev;
 359
 360                 rdma_read_gids(ic->i_cm_id, (union ib_gid *)&iinfo6->src_gid,
 361                                (union ib_gid *)&iinfo6->dst_gid);
 362                 rds_ibdev = ic->rds_ibdev;
 363                 iinfo6->max_send_wr = ic->i_send_ring.w_nr;
 364                 iinfo6->max_recv_wr = ic->i_recv_ring.w_nr;
 365                 iinfo6->max_send_sge = rds_ibdev->max_sge;
 366                 rds6_ib_get_mr_info(rds_ibdev, iinfo6);
 367                 iinfo6->cache_allocs = atomic_read(&ic->i_cache_allocs);
 368         }
 369         return 1;
 370 }
 371 #endif
 372
 373 static void rds_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 rds_info_rdma_connection) + 7) / 8];
 378
 379         rds_for_each_conn_info(sock, len, iter, lens,
 380                                 rds_ib_conn_info_visitor,
 381                                 buffer,
 382                                 sizeof(struct rds_info_rdma_connection));
 383 }
 384
 385 #if IS_ENABLED(CONFIG_IPV6)
 386 /* IPv6 version of rds_ib_ic_info(). */
 387 static void rds6_ib_ic_info(struct socket *sock, unsigned int len,
 388                             struct rds_info_iterator *iter,
 389                             struct rds_info_lengths *lens)
 390 {
 391         u64 buffer[(sizeof(struct rds6_info_rdma_connection) + 7) / 8];
 392
 393         rds_for_each_conn_info(sock, len, iter, lens,
 394                                rds6_ib_conn_info_visitor,
 395                                buffer,
 396                                sizeof(struct rds6_info_rdma_connection));
 397 }
 398 #endif
 399
 400 /*
 401  * Early RDS/IB was built to only bind to an address if there is an IPoIB
 402  * device with that address set.
 403  *
 404  * If it were me, I'd advocate for something more flexible.  Sending and
 405  * receiving should be device-agnostic.  Transports would try and maintain
 406  * connections between peers who have messages queued.  Userspace would be
 407  * allowed to influence which paths have priority.  We could call userspace
 408  * asserting this policy "routing".
 409  */
 410 static int rds_ib_laddr_check(struct net *net, const struct in6_addr *addr,
 411                               __u32 scope_id)
 412 {
 413         int ret;
 414         struct rdma_cm_id *cm_id;
 415 #if IS_ENABLED(CONFIG_IPV6)
 416         struct sockaddr_in6 sin6;
 417 #endif
 418         struct sockaddr_in sin;
 419         struct sockaddr *sa;
 420         bool isv4;
 421
 422         isv4 = ipv6_addr_v4mapped(addr);
 423         /* Create a CMA ID and try to bind it. This catches both
 424          * IB and iWARP capable NICs.
 425          */
 426         cm_id = rdma_create_id(&init_net, rds_rdma_cm_event_handler,
 427                                NULL, RDMA_PS_TCP, IB_QPT_RC);
 428         if (IS_ERR(cm_id))
 429                 return PTR_ERR(cm_id);
 430
 431         if (isv4) {
 432                 memset(&sin, 0, sizeof(sin));
 433                 sin.sin_family = AF_INET;
 434                 sin.sin_addr.s_addr = addr->s6_addr32[3];
 435                 sa = (struct sockaddr *)&sin;
 436         } else {
 437 #if IS_ENABLED(CONFIG_IPV6)
 438                 memset(&sin6, 0, sizeof(sin6));
 439                 sin6.sin6_family = AF_INET6;
 440                 sin6.sin6_addr = *addr;
 441                 sin6.sin6_scope_id = scope_id;
 442                 sa = (struct sockaddr *)&sin6;
 443
 444                 /* XXX Do a special IPv6 link local address check here.  The
 445                  * reason is that rdma_bind_addr() always succeeds with IPv6
 446                  * link local address regardless it is indeed configured in a
 447                  * system.
 448                  */
 449                 if (ipv6_addr_type(addr) & IPV6_ADDR_LINKLOCAL) {
 450                         struct net_device *dev;
 451
 452                         if (scope_id == 0) {
 453                                 ret = -EADDRNOTAVAIL;
 454                                 goto out;
 455                         }
 456
 457                         /* Use init_net for now as RDS is not network
 458                          * name space aware.
 459                          */
 460                         dev = dev_get_by_index(&init_net, scope_id);
 461                         if (!dev) {
 462                                 ret = -EADDRNOTAVAIL;
 463                                 goto out;
 464                         }
 465                         if (!ipv6_chk_addr(&init_net, addr, dev, 1)) {
 466                                 dev_put(dev);
 467                                 ret = -EADDRNOTAVAIL;
 468                                 goto out;
 469                         }
 470                         dev_put(dev);
 471                 }
 472 #else
 473                 ret = -EADDRNOTAVAIL;
 474                 goto out;
 475 #endif
 476         }
 477
 478         /* rdma_bind_addr will only succeed for IB & iWARP devices */
 479         ret = rdma_bind_addr(cm_id, sa);
 480         /* due to this, we will claim to support iWARP devices unless we
 481            check node_type. */
 482         if (ret || !cm_id->device ||
 483             cm_id->device->node_type != RDMA_NODE_IB_CA)
 484                 ret = -EADDRNOTAVAIL;
 485
 486         rdsdebug("addr %pI6c%%%u ret %d node type %d\n",
 487                  addr, scope_id, ret,
 488                  cm_id->device ? cm_id->device->node_type : -1);
 489
 490 out:
 491         rdma_destroy_id(cm_id);
 492
 493         return ret;
 494 }
 495
 496 static void rds_ib_unregister_client(void)
 497 {
 498         ib_unregister_client(&rds_ib_client);
 499         /* wait for rds_ib_dev_free() to complete */
 500         flush_workqueue(rds_wq);
 501 }
 502
 503 static void rds_ib_set_unloading(void)
 504 {
 505         atomic_set(&rds_ib_unloading, 1);
 506 }
 507
 508 static bool rds_ib_is_unloading(struct rds_connection *conn)
 509 {
 510         struct rds_conn_path *cp = &conn->c_path[0];
 511
 512         return (test_bit(RDS_DESTROY_PENDING, &cp->cp_flags) ||
 513                 atomic_read(&rds_ib_unloading) != 0);
 514 }
 515
 516 void rds_ib_exit(void)
 517 {
 518         rds_ib_set_unloading();
 519         synchronize_rcu();
 520         rds_info_deregister_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info);
 521 #if IS_ENABLED(CONFIG_IPV6)
 522         rds_info_deregister_func(RDS6_INFO_IB_CONNECTIONS, rds6_ib_ic_info);
 523 #endif
 524         rds_ib_unregister_client();
 525         rds_ib_destroy_nodev_conns();
 526         rds_ib_sysctl_exit();
 527         rds_ib_recv_exit();
 528         rds_trans_unregister(&rds_ib_transport);
 529         rds_ib_mr_exit();
 530 }
 531
 532 static u8 rds_ib_get_tos_map(u8 tos)
 533 {
 534         /* 1:1 user to transport map for RDMA transport.
 535          * In future, if custom map is desired, hook can export
 536          * user configurable map.
 537          */
 538         return tos;
 539 }
 540
 541 struct rds_transport rds_ib_transport = {
 542         .laddr_check            = rds_ib_laddr_check,
 543         .xmit_path_complete     = rds_ib_xmit_path_complete,
 544         .xmit                   = rds_ib_xmit,
 545         .xmit_rdma              = rds_ib_xmit_rdma,
 546         .xmit_atomic            = rds_ib_xmit_atomic,
 547         .recv_path              = rds_ib_recv_path,
 548         .conn_alloc             = rds_ib_conn_alloc,
 549         .conn_free              = rds_ib_conn_free,
 550         .conn_path_connect      = rds_ib_conn_path_connect,
 551         .conn_path_shutdown     = rds_ib_conn_path_shutdown,
 552         .inc_copy_to_user       = rds_ib_inc_copy_to_user,
 553         .inc_free               = rds_ib_inc_free,
 554         .cm_initiate_connect    = rds_ib_cm_initiate_connect,
 555         .cm_handle_connect      = rds_ib_cm_handle_connect,
 556         .cm_connect_complete    = rds_ib_cm_connect_complete,
 557         .stats_info_copy        = rds_ib_stats_info_copy,
 558         .exit                   = rds_ib_exit,
 559         .get_mr                 = rds_ib_get_mr,
 560         .sync_mr                = rds_ib_sync_mr,
 561         .free_mr                = rds_ib_free_mr,
 562         .flush_mrs              = rds_ib_flush_mrs,
 563         .get_tos_map            = rds_ib_get_tos_map,
 564         .t_owner                = THIS_MODULE,
 565         .t_name                 = "infiniband",
 566         .t_unloading            = rds_ib_is_unloading,
 567         .t_type                 = RDS_TRANS_IB
 568 };
 569
 570 int rds_ib_init(void)
 571 {
 572         int ret;
 573
 574         INIT_LIST_HEAD(&rds_ib_devices);
 575
 576         ret = rds_ib_mr_init();
 577         if (ret)
 578                 goto out;
 579
 580         ret = ib_register_client(&rds_ib_client);
 581         if (ret)
 582                 goto out_mr_exit;
 583
 584         ret = rds_ib_sysctl_init();
 585         if (ret)
 586                 goto out_ibreg;
 587
 588         ret = rds_ib_recv_init();
 589         if (ret)
 590                 goto out_sysctl;
 591
 592         rds_trans_register(&rds_ib_transport);
 593
 594         rds_info_register_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info);
 595 #if IS_ENABLED(CONFIG_IPV6)
 596         rds_info_register_func(RDS6_INFO_IB_CONNECTIONS, rds6_ib_ic_info);
 597 #endif
 598
 599         goto out;
 600
 601 out_sysctl:
 602         rds_ib_sysctl_exit();
 603 out_ibreg:
 604         rds_ib_unregister_client();
 605 out_mr_exit:
 606         rds_ib_mr_exit();
 607 out:
 608         return ret;
 609 }
 610
 611 MODULE_LICENSE("GPL");