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         rds_ibdev->odp_capable =
 160                 !!(device->attrs.device_cap_flags &
 161                    IB_DEVICE_ON_DEMAND_PAGING) &&
 162                 !!(device->attrs.odp_caps.per_transport_caps.rc_odp_caps &
 163                    IB_ODP_SUPPORT_WRITE) &&
 164                 !!(device->attrs.odp_caps.per_transport_caps.rc_odp_caps &
 165                    IB_ODP_SUPPORT_READ);
 166
 167         rds_ibdev->fmr_max_remaps = device->attrs.max_map_per_fmr?: 32;
 168         rds_ibdev->max_1m_mrs = device->attrs.max_mr ?
 169                 min_t(unsigned int, (device->attrs.max_mr / 2),
 170                       rds_ib_mr_1m_pool_size) : rds_ib_mr_1m_pool_size;
 171
 172         rds_ibdev->max_8k_mrs = device->attrs.max_mr ?
 173                 min_t(unsigned int, ((device->attrs.max_mr / 2) * RDS_MR_8K_SCALE),
 174                       rds_ib_mr_8k_pool_size) : rds_ib_mr_8k_pool_size;
 175
 176         rds_ibdev->max_initiator_depth = device->attrs.max_qp_init_rd_atom;
 177         rds_ibdev->max_responder_resources = device->attrs.max_qp_rd_atom;
 178
 179         rds_ibdev->vector_load = kcalloc(device->num_comp_vectors,
 180                                          sizeof(int),
 181                                          GFP_KERNEL);
 182         if (!rds_ibdev->vector_load) {
 183                 pr_err("RDS/IB: %s failed to allocate vector memory\n",
 184                         __func__);
 185                 goto put_dev;
 186         }
 187
 188         rds_ibdev->dev = device;
 189         rds_ibdev->pd = ib_alloc_pd(device, 0);
 190         if (IS_ERR(rds_ibdev->pd)) {
 191                 rds_ibdev->pd = NULL;
 192                 goto put_dev;
 193         }
 194         rds_ibdev->rid_hdrs_pool = dma_pool_create(device->name,
 195                                                    device->dma_device,
 196                                                    sizeof(struct rds_header),
 197                                                    L1_CACHE_BYTES, 0);
 198         if (!rds_ibdev->rid_hdrs_pool)
 199                 goto put_dev;
 200
 201         rds_ibdev->mr_1m_pool =
 202                 rds_ib_create_mr_pool(rds_ibdev, RDS_IB_MR_1M_POOL);
 203         if (IS_ERR(rds_ibdev->mr_1m_pool)) {
 204                 rds_ibdev->mr_1m_pool = NULL;
 205                 goto put_dev;
 206         }
 207
 208         rds_ibdev->mr_8k_pool =
 209                 rds_ib_create_mr_pool(rds_ibdev, RDS_IB_MR_8K_POOL);
 210         if (IS_ERR(rds_ibdev->mr_8k_pool)) {
 211                 rds_ibdev->mr_8k_pool = NULL;
 212                 goto put_dev;
 213         }
 214
 215         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",
 216                  device->attrs.max_fmr, rds_ibdev->max_wrs, rds_ibdev->max_sge,
 217                  rds_ibdev->fmr_max_remaps, rds_ibdev->max_1m_mrs,
 218                  rds_ibdev->max_8k_mrs);
 219
 220         pr_info("RDS/IB: %s: %s supported and preferred\n",
 221                 device->name,
 222                 rds_ibdev->use_fastreg ? "FRMR" : "FMR");
 223
 224         down_write(&rds_ib_devices_lock);
 225         list_add_tail_rcu(&rds_ibdev->list, &rds_ib_devices);
 226         up_write(&rds_ib_devices_lock);
 227         refcount_inc(&rds_ibdev->refcount);
 228
 229         ib_set_client_data(device, &rds_ib_client, rds_ibdev);
 230         refcount_inc(&rds_ibdev->refcount);
 231
 232         rds_ib_nodev_connect();
 233
 234 put_dev:
 235         rds_ib_dev_put(rds_ibdev);
 236 }
 237
 238 /*
 239  * New connections use this to find the device to associate with the
 240  * connection.  It's not in the fast path so we're not concerned about the
 241  * performance of the IB call.  (As of this writing, it uses an interrupt
 242  * blocking spinlock to serialize walking a per-device list of all registered
 243  * clients.)
 244  *
 245  * RCU is used to handle incoming connections racing with device teardown.
 246  * Rather than use a lock to serialize removal from the client_data and
 247  * getting a new reference, we use an RCU grace period.  The destruction
 248  * path removes the device from client_data and then waits for all RCU
 249  * readers to finish.
 250  *
 251  * A new connection can get NULL from this if its arriving on a
 252  * device that is in the process of being removed.
 253  */
 254 struct rds_ib_device *rds_ib_get_client_data(struct ib_device *device)
 255 {
 256         struct rds_ib_device *rds_ibdev;
 257
 258         rcu_read_lock();
 259         rds_ibdev = ib_get_client_data(device, &rds_ib_client);
 260         if (rds_ibdev)
 261                 refcount_inc(&rds_ibdev->refcount);
 262         rcu_read_unlock();
 263         return rds_ibdev;
 264 }
 265
 266 /*
 267  * The IB stack is letting us know that a device is going away.  This can
 268  * happen if the underlying HCA driver is removed or if PCI hotplug is removing
 269  * the pci function, for example.
 270  *
 271  * This can be called at any time and can be racing with any other RDS path.
 272  */
 273 static void rds_ib_remove_one(struct ib_device *device, void *client_data)
 274 {
 275         struct rds_ib_device *rds_ibdev = client_data;
 276
 277         if (!rds_ibdev)
 278                 return;
 279
 280         rds_ib_dev_shutdown(rds_ibdev);
 281
 282         /* stop connection attempts from getting a reference to this device. */
 283         ib_set_client_data(device, &rds_ib_client, NULL);
 284
 285         down_write(&rds_ib_devices_lock);
 286         list_del_rcu(&rds_ibdev->list);
 287         up_write(&rds_ib_devices_lock);
 288
 289         /*
 290          * This synchronize rcu is waiting for readers of both the ib
 291          * client data and the devices list to finish before we drop
 292          * both of those references.
 293          */
 294         synchronize_rcu();
 295         rds_ib_dev_put(rds_ibdev);
 296         rds_ib_dev_put(rds_ibdev);
 297 }
 298
 299 struct ib_client rds_ib_client = {
 300         .name   = "rds_ib",
 301         .add    = rds_ib_add_one,
 302         .remove = rds_ib_remove_one
 303 };
 304
 305 static int rds_ib_conn_info_visitor(struct rds_connection *conn,
 306                                     void *buffer)
 307 {
 308         struct rds_info_rdma_connection *iinfo = buffer;
 309         struct rds_ib_connection *ic = conn->c_transport_data;
 310
 311         /* We will only ever look at IB transports */
 312         if (conn->c_trans != &rds_ib_transport)
 313                 return 0;
 314         if (conn->c_isv6)
 315                 return 0;
 316
 317         iinfo->src_addr = conn->c_laddr.s6_addr32[3];
 318         iinfo->dst_addr = conn->c_faddr.s6_addr32[3];
 319         if (ic) {
 320                 iinfo->tos = conn->c_tos;
 321                 iinfo->sl = ic->i_sl;
 322         }
 323
 324         memset(&iinfo->src_gid, 0, sizeof(iinfo->src_gid));
 325         memset(&iinfo->dst_gid, 0, sizeof(iinfo->dst_gid));
 326         if (rds_conn_state(conn) == RDS_CONN_UP) {
 327                 struct rds_ib_device *rds_ibdev;
 328
 329                 rdma_read_gids(ic->i_cm_id, (union ib_gid *)&iinfo->src_gid,
 330                                (union ib_gid *)&iinfo->dst_gid);
 331
 332                 rds_ibdev = ic->rds_ibdev;
 333                 iinfo->max_send_wr = ic->i_send_ring.w_nr;
 334                 iinfo->max_recv_wr = ic->i_recv_ring.w_nr;
 335                 iinfo->max_send_sge = rds_ibdev->max_sge;
 336                 rds_ib_get_mr_info(rds_ibdev, iinfo);
 337                 iinfo->cache_allocs = atomic_read(&ic->i_cache_allocs);
 338         }
 339         return 1;
 340 }
 341
 342 #if IS_ENABLED(CONFIG_IPV6)
 343 /* IPv6 version of rds_ib_conn_info_visitor(). */
 344 static int rds6_ib_conn_info_visitor(struct rds_connection *conn,
 345                                      void *buffer)
 346 {
 347         struct rds6_info_rdma_connection *iinfo6 = buffer;
 348         struct rds_ib_connection *ic = conn->c_transport_data;
 349
 350         /* We will only ever look at IB transports */
 351         if (conn->c_trans != &rds_ib_transport)
 352                 return 0;
 353
 354         iinfo6->src_addr = conn->c_laddr;
 355         iinfo6->dst_addr = conn->c_faddr;
 356         if (ic) {
 357                 iinfo6->tos = conn->c_tos;
 358                 iinfo6->sl = ic->i_sl;
 359         }
 360
 361         memset(&iinfo6->src_gid, 0, sizeof(iinfo6->src_gid));
 362         memset(&iinfo6->dst_gid, 0, sizeof(iinfo6->dst_gid));
 363
 364         if (rds_conn_state(conn) == RDS_CONN_UP) {
 365                 struct rds_ib_device *rds_ibdev;
 366
 367                 rdma_read_gids(ic->i_cm_id, (union ib_gid *)&iinfo6->src_gid,
 368                                (union ib_gid *)&iinfo6->dst_gid);
 369                 rds_ibdev = ic->rds_ibdev;
 370                 iinfo6->max_send_wr = ic->i_send_ring.w_nr;
 371                 iinfo6->max_recv_wr = ic->i_recv_ring.w_nr;
 372                 iinfo6->max_send_sge = rds_ibdev->max_sge;
 373                 rds6_ib_get_mr_info(rds_ibdev, iinfo6);
 374                 iinfo6->cache_allocs = atomic_read(&ic->i_cache_allocs);
 375         }
 376         return 1;
 377 }
 378 #endif
 379
 380 static void rds_ib_ic_info(struct socket *sock, unsigned int len,
 381                            struct rds_info_iterator *iter,
 382                            struct rds_info_lengths *lens)
 383 {
 384         u64 buffer[(sizeof(struct rds_info_rdma_connection) + 7) / 8];
 385
 386         rds_for_each_conn_info(sock, len, iter, lens,
 387                                 rds_ib_conn_info_visitor,
 388                                 buffer,
 389                                 sizeof(struct rds_info_rdma_connection));
 390 }
 391
 392 #if IS_ENABLED(CONFIG_IPV6)
 393 /* IPv6 version of rds_ib_ic_info(). */
 394 static void rds6_ib_ic_info(struct socket *sock, unsigned int len,
 395                             struct rds_info_iterator *iter,
 396                             struct rds_info_lengths *lens)
 397 {
 398         u64 buffer[(sizeof(struct rds6_info_rdma_connection) + 7) / 8];
 399
 400         rds_for_each_conn_info(sock, len, iter, lens,
 401                                rds6_ib_conn_info_visitor,
 402                                buffer,
 403                                sizeof(struct rds6_info_rdma_connection));
 404 }
 405 #endif
 406
 407 /*
 408  * Early RDS/IB was built to only bind to an address if there is an IPoIB
 409  * device with that address set.
 410  *
 411  * If it were me, I'd advocate for something more flexible.  Sending and
 412  * receiving should be device-agnostic.  Transports would try and maintain
 413  * connections between peers who have messages queued.  Userspace would be
 414  * allowed to influence which paths have priority.  We could call userspace
 415  * asserting this policy "routing".
 416  */
 417 static int rds_ib_laddr_check(struct net *net, const struct in6_addr *addr,
 418                               __u32 scope_id)
 419 {
 420         int ret;
 421         struct rdma_cm_id *cm_id;
 422 #if IS_ENABLED(CONFIG_IPV6)
 423         struct sockaddr_in6 sin6;
 424 #endif
 425         struct sockaddr_in sin;
 426         struct sockaddr *sa;
 427         bool isv4;
 428
 429         isv4 = ipv6_addr_v4mapped(addr);
 430         /* Create a CMA ID and try to bind it. This catches both
 431          * IB and iWARP capable NICs.
 432          */
 433         cm_id = rdma_create_id(&init_net, rds_rdma_cm_event_handler,
 434                                NULL, RDMA_PS_TCP, IB_QPT_RC);
 435         if (IS_ERR(cm_id))
 436                 return PTR_ERR(cm_id);
 437
 438         if (isv4) {
 439                 memset(&sin, 0, sizeof(sin));
 440                 sin.sin_family = AF_INET;
 441                 sin.sin_addr.s_addr = addr->s6_addr32[3];
 442                 sa = (struct sockaddr *)&sin;
 443         } else {
 444 #if IS_ENABLED(CONFIG_IPV6)
 445                 memset(&sin6, 0, sizeof(sin6));
 446                 sin6.sin6_family = AF_INET6;
 447                 sin6.sin6_addr = *addr;
 448                 sin6.sin6_scope_id = scope_id;
 449                 sa = (struct sockaddr *)&sin6;
 450
 451                 /* XXX Do a special IPv6 link local address check here.  The
 452                  * reason is that rdma_bind_addr() always succeeds with IPv6
 453                  * link local address regardless it is indeed configured in a
 454                  * system.
 455                  */
 456                 if (ipv6_addr_type(addr) & IPV6_ADDR_LINKLOCAL) {
 457                         struct net_device *dev;
 458
 459                         if (scope_id == 0) {
 460                                 ret = -EADDRNOTAVAIL;
 461                                 goto out;
 462                         }
 463
 464                         /* Use init_net for now as RDS is not network
 465                          * name space aware.
 466                          */
 467                         dev = dev_get_by_index(&init_net, scope_id);
 468                         if (!dev) {
 469                                 ret = -EADDRNOTAVAIL;
 470                                 goto out;
 471                         }
 472                         if (!ipv6_chk_addr(&init_net, addr, dev, 1)) {
 473                                 dev_put(dev);
 474                                 ret = -EADDRNOTAVAIL;
 475                                 goto out;
 476                         }
 477                         dev_put(dev);
 478                 }
 479 #else
 480                 ret = -EADDRNOTAVAIL;
 481                 goto out;
 482 #endif
 483         }
 484
 485         /* rdma_bind_addr will only succeed for IB & iWARP devices */
 486         ret = rdma_bind_addr(cm_id, sa);
 487         /* due to this, we will claim to support iWARP devices unless we
 488            check node_type. */
 489         if (ret || !cm_id->device ||
 490             cm_id->device->node_type != RDMA_NODE_IB_CA)
 491                 ret = -EADDRNOTAVAIL;
 492
 493         rdsdebug("addr %pI6c%%%u ret %d node type %d\n",
 494                  addr, scope_id, ret,
 495                  cm_id->device ? cm_id->device->node_type : -1);
 496
 497 out:
 498         rdma_destroy_id(cm_id);
 499
 500         return ret;
 501 }
 502
 503 static void rds_ib_unregister_client(void)
 504 {
 505         ib_unregister_client(&rds_ib_client);
 506         /* wait for rds_ib_dev_free() to complete */
 507         flush_workqueue(rds_wq);
 508 }
 509
 510 static void rds_ib_set_unloading(void)
 511 {
 512         atomic_set(&rds_ib_unloading, 1);
 513 }
 514
 515 static bool rds_ib_is_unloading(struct rds_connection *conn)
 516 {
 517         struct rds_conn_path *cp = &conn->c_path[0];
 518
 519         return (test_bit(RDS_DESTROY_PENDING, &cp->cp_flags) ||
 520                 atomic_read(&rds_ib_unloading) != 0);
 521 }
 522
 523 void rds_ib_exit(void)
 524 {
 525         rds_ib_set_unloading();
 526         synchronize_rcu();
 527         rds_info_deregister_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info);
 528 #if IS_ENABLED(CONFIG_IPV6)
 529         rds_info_deregister_func(RDS6_INFO_IB_CONNECTIONS, rds6_ib_ic_info);
 530 #endif
 531         rds_ib_unregister_client();
 532         rds_ib_destroy_nodev_conns();
 533         rds_ib_sysctl_exit();
 534         rds_ib_recv_exit();
 535         rds_trans_unregister(&rds_ib_transport);
 536         rds_ib_mr_exit();
 537 }
 538
 539 static u8 rds_ib_get_tos_map(u8 tos)
 540 {
 541         /* 1:1 user to transport map for RDMA transport.
 542          * In future, if custom map is desired, hook can export
 543          * user configurable map.
 544          */
 545         return tos;
 546 }
 547
 548 struct rds_transport rds_ib_transport = {
 549         .laddr_check            = rds_ib_laddr_check,
 550         .xmit_path_complete     = rds_ib_xmit_path_complete,
 551         .xmit                   = rds_ib_xmit,
 552         .xmit_rdma              = rds_ib_xmit_rdma,
 553         .xmit_atomic            = rds_ib_xmit_atomic,
 554         .recv_path              = rds_ib_recv_path,
 555         .conn_alloc             = rds_ib_conn_alloc,
 556         .conn_free              = rds_ib_conn_free,
 557         .conn_path_connect      = rds_ib_conn_path_connect,
 558         .conn_path_shutdown     = rds_ib_conn_path_shutdown,
 559         .inc_copy_to_user       = rds_ib_inc_copy_to_user,
 560         .inc_free               = rds_ib_inc_free,
 561         .cm_initiate_connect    = rds_ib_cm_initiate_connect,
 562         .cm_handle_connect      = rds_ib_cm_handle_connect,
 563         .cm_connect_complete    = rds_ib_cm_connect_complete,
 564         .stats_info_copy        = rds_ib_stats_info_copy,
 565         .exit                   = rds_ib_exit,
 566         .get_mr                 = rds_ib_get_mr,
 567         .sync_mr                = rds_ib_sync_mr,
 568         .free_mr                = rds_ib_free_mr,
 569         .flush_mrs              = rds_ib_flush_mrs,
 570         .get_tos_map            = rds_ib_get_tos_map,
 571         .t_owner                = THIS_MODULE,
 572         .t_name                 = "infiniband",
 573         .t_unloading            = rds_ib_is_unloading,
 574         .t_type                 = RDS_TRANS_IB
 575 };
 576
 577 int rds_ib_init(void)
 578 {
 579         int ret;
 580
 581         INIT_LIST_HEAD(&rds_ib_devices);
 582
 583         ret = rds_ib_mr_init();
 584         if (ret)
 585                 goto out;
 586
 587         ret = ib_register_client(&rds_ib_client);
 588         if (ret)
 589                 goto out_mr_exit;
 590
 591         ret = rds_ib_sysctl_init();
 592         if (ret)
 593                 goto out_ibreg;
 594
 595         ret = rds_ib_recv_init();
 596         if (ret)
 597                 goto out_sysctl;
 598
 599         rds_trans_register(&rds_ib_transport);
 600
 601         rds_info_register_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info);
 602 #if IS_ENABLED(CONFIG_IPV6)
 603         rds_info_register_func(RDS6_INFO_IB_CONNECTIONS, rds6_ib_ic_info);
 604 #endif
 605
 606         goto out;
 607
 608 out_sysctl:
 609         rds_ib_sysctl_exit();
 610 out_ibreg:
 611         rds_ib_unregister_client();
 612 out_mr_exit:
 613         rds_ib_mr_exit();
 614 out:
 615         return ret;
 616 }
 617
 618 MODULE_LICENSE("GPL");