net/rds/ib.c

0001 /*
0002  * Copyright (c) 2006, 2019 Oracle and/or its affiliates. All rights reserved.
0003  *
0004  * This software is available to you under a choice of one of two
0005  * licenses.  You may choose to be licensed under the terms of the GNU
0006  * General Public License (GPL) Version 2, available from the file
0007  * COPYING in the main directory of this source tree, or the
0008  * OpenIB.org BSD license below:
0009  *
0010  *     Redistribution and use in source and binary forms, with or
0011  *     without modification, are permitted provided that the following
0012  *     conditions are met:
0013  *
0014  *      - Redistributions of source code must retain the above
0015  *        copyright notice, this list of conditions and the following
0016  *        disclaimer.
0017  *
0018  *      - Redistributions in binary form must reproduce the above
0019  *        copyright notice, this list of conditions and the following
0020  *        disclaimer in the documentation and/or other materials
0021  *        provided with the distribution.
0022  *
0023  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
0024  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
0025  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
0026  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
0027  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
0028  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
0029  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
0030  * SOFTWARE.
0031  *
0032  */
0033 #include <linux/kernel.h>
0034 #include <linux/in.h>
0035 #include <linux/if.h>
0036 #include <linux/netdevice.h>
0037 #include <linux/inetdevice.h>
0038 #include <linux/if_arp.h>
0039 #include <linux/delay.h>
0040 #include <linux/slab.h>
0041 #include <linux/module.h>
0042 #include <net/addrconf.h>
0043
0044 #include "rds_single_path.h"
0045 #include "rds.h"
0046 #include "ib.h"
0047 #include "ib_mr.h"
0048
0049 static unsigned int rds_ib_mr_1m_pool_size = RDS_MR_1M_POOL_SIZE;
0050 static unsigned int rds_ib_mr_8k_pool_size = RDS_MR_8K_POOL_SIZE;
0051 unsigned int rds_ib_retry_count = RDS_IB_DEFAULT_RETRY_COUNT;
0052 static atomic_t rds_ib_unloading;
0053
0054 module_param(rds_ib_mr_1m_pool_size, int, 0444);
0055 MODULE_PARM_DESC(rds_ib_mr_1m_pool_size, " Max number of 1M mr per HCA");
0056 module_param(rds_ib_mr_8k_pool_size, int, 0444);
0057 MODULE_PARM_DESC(rds_ib_mr_8k_pool_size, " Max number of 8K mr per HCA");
0058 module_param(rds_ib_retry_count, int, 0444);
0059 MODULE_PARM_DESC(rds_ib_retry_count, " Number of hw retries before reporting an error");
0060
0061 /*
0062  * we have a clumsy combination of RCU and a rwsem protecting this list
0063  * because it is used both in the get_mr fast path and while blocking in
0064  * the FMR flushing path.
0065  */
0066 DECLARE_RWSEM(rds_ib_devices_lock);
0067 struct list_head rds_ib_devices;
0068
0069 /* NOTE: if also grabbing ibdev lock, grab this first */
0070 DEFINE_SPINLOCK(ib_nodev_conns_lock);
0071 LIST_HEAD(ib_nodev_conns);
0072
0073 static void rds_ib_nodev_connect(void)
0074 {
0075     struct rds_ib_connection *ic;
0076
0077     spin_lock(&ib_nodev_conns_lock);
0078     list_for_each_entry(ic, &ib_nodev_conns, ib_node)
0079         rds_conn_connect_if_down(ic->conn);
0080     spin_unlock(&ib_nodev_conns_lock);
0081 }
0082
0083 static void rds_ib_dev_shutdown(struct rds_ib_device *rds_ibdev)
0084 {
0085     struct rds_ib_connection *ic;
0086     unsigned long flags;
0087
0088     spin_lock_irqsave(&rds_ibdev->spinlock, flags);
0089     list_for_each_entry(ic, &rds_ibdev->conn_list, ib_node)
0090         rds_conn_path_drop(&ic->conn->c_path[0], true);
0091     spin_unlock_irqrestore(&rds_ibdev->spinlock, flags);
0092 }
0093
0094 /*
0095  * rds_ib_destroy_mr_pool() blocks on a few things and mrs drop references
0096  * from interrupt context so we push freing off into a work struct in krdsd.
0097  */
0098 static void rds_ib_dev_free(struct work_struct *work)
0099 {
0100     struct rds_ib_ipaddr *i_ipaddr, *i_next;
0101     struct rds_ib_device *rds_ibdev = container_of(work,
0102                     struct rds_ib_device, free_work);
0103
0104     if (rds_ibdev->mr_8k_pool)
0105         rds_ib_destroy_mr_pool(rds_ibdev->mr_8k_pool);
0106     if (rds_ibdev->mr_1m_pool)
0107         rds_ib_destroy_mr_pool(rds_ibdev->mr_1m_pool);
0108     if (rds_ibdev->pd)
0109         ib_dealloc_pd(rds_ibdev->pd);
0110
0111     list_for_each_entry_safe(i_ipaddr, i_next, &rds_ibdev->ipaddr_list, list) {
0112         list_del(&i_ipaddr->list);
0113         kfree(i_ipaddr);
0114     }
0115
0116     kfree(rds_ibdev->vector_load);
0117
0118     kfree(rds_ibdev);
0119 }
0120
0121 void rds_ib_dev_put(struct rds_ib_device *rds_ibdev)
0122 {
0123     BUG_ON(refcount_read(&rds_ibdev->refcount) == 0);
0124     if (refcount_dec_and_test(&rds_ibdev->refcount))
0125         queue_work(rds_wq, &rds_ibdev->free_work);
0126 }
0127
0128 static int rds_ib_add_one(struct ib_device *device)
0129 {
0130     struct rds_ib_device *rds_ibdev;
0131     int ret;
0132
0133     /* Only handle IB (no iWARP) devices */
0134     if (device->node_type != RDMA_NODE_IB_CA)
0135         return -EOPNOTSUPP;
0136
0137     /* Device must support FRWR */
0138     if (!(device->attrs.device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS))
0139         return -EOPNOTSUPP;
0140
0141     rds_ibdev = kzalloc_node(sizeof(struct rds_ib_device), GFP_KERNEL,
0142                  ibdev_to_node(device));
0143     if (!rds_ibdev)
0144         return -ENOMEM;
0145
0146     spin_lock_init(&rds_ibdev->spinlock);
0147     refcount_set(&rds_ibdev->refcount, 1);
0148     INIT_WORK(&rds_ibdev->free_work, rds_ib_dev_free);
0149
0150     INIT_LIST_HEAD(&rds_ibdev->ipaddr_list);
0151     INIT_LIST_HEAD(&rds_ibdev->conn_list);
0152
0153     rds_ibdev->max_wrs = device->attrs.max_qp_wr;
0154     rds_ibdev->max_sge = min(device->attrs.max_send_sge, RDS_IB_MAX_SGE);
0155
0156     rds_ibdev->odp_capable =
0157         !!(device->attrs.kernel_cap_flags &
0158            IBK_ON_DEMAND_PAGING) &&
0159         !!(device->attrs.odp_caps.per_transport_caps.rc_odp_caps &
0160            IB_ODP_SUPPORT_WRITE) &&
0161         !!(device->attrs.odp_caps.per_transport_caps.rc_odp_caps &
0162            IB_ODP_SUPPORT_READ);
0163
0164     rds_ibdev->max_1m_mrs = device->attrs.max_mr ?
0165         min_t(unsigned int, (device->attrs.max_mr / 2),
0166               rds_ib_mr_1m_pool_size) : rds_ib_mr_1m_pool_size;
0167
0168     rds_ibdev->max_8k_mrs = device->attrs.max_mr ?
0169         min_t(unsigned int, ((device->attrs.max_mr / 2) * RDS_MR_8K_SCALE),
0170               rds_ib_mr_8k_pool_size) : rds_ib_mr_8k_pool_size;
0171
0172     rds_ibdev->max_initiator_depth = device->attrs.max_qp_init_rd_atom;
0173     rds_ibdev->max_responder_resources = device->attrs.max_qp_rd_atom;
0174
0175     rds_ibdev->vector_load = kcalloc(device->num_comp_vectors,
0176                      sizeof(int),
0177                      GFP_KERNEL);
0178     if (!rds_ibdev->vector_load) {
0179         pr_err("RDS/IB: %s failed to allocate vector memory\n",
0180             __func__);
0181         ret = -ENOMEM;
0182         goto put_dev;
0183     }
0184
0185     rds_ibdev->dev = device;
0186     rds_ibdev->pd = ib_alloc_pd(device, 0);
0187     if (IS_ERR(rds_ibdev->pd)) {
0188         ret = PTR_ERR(rds_ibdev->pd);
0189         rds_ibdev->pd = NULL;
0190         goto put_dev;
0191     }
0192
0193     rds_ibdev->mr_1m_pool =
0194         rds_ib_create_mr_pool(rds_ibdev, RDS_IB_MR_1M_POOL);
0195     if (IS_ERR(rds_ibdev->mr_1m_pool)) {
0196         ret = PTR_ERR(rds_ibdev->mr_1m_pool);
0197         rds_ibdev->mr_1m_pool = NULL;
0198         goto put_dev;
0199     }
0200
0201     rds_ibdev->mr_8k_pool =
0202         rds_ib_create_mr_pool(rds_ibdev, RDS_IB_MR_8K_POOL);
0203     if (IS_ERR(rds_ibdev->mr_8k_pool)) {
0204         ret = PTR_ERR(rds_ibdev->mr_8k_pool);
0205         rds_ibdev->mr_8k_pool = NULL;
0206         goto put_dev;
0207     }
0208
0209     rdsdebug("RDS/IB: max_mr = %d, max_wrs = %d, max_sge = %d, max_1m_mrs = %d, max_8k_mrs = %d\n",
0210          device->attrs.max_mr, rds_ibdev->max_wrs, rds_ibdev->max_sge,
0211          rds_ibdev->max_1m_mrs, rds_ibdev->max_8k_mrs);
0212
0213     pr_info("RDS/IB: %s: added\n", device->name);
0214
0215     down_write(&rds_ib_devices_lock);
0216     list_add_tail_rcu(&rds_ibdev->list, &rds_ib_devices);
0217     up_write(&rds_ib_devices_lock);
0218     refcount_inc(&rds_ibdev->refcount);
0219
0220     ib_set_client_data(device, &rds_ib_client, rds_ibdev);
0221
0222     rds_ib_nodev_connect();
0223     return 0;
0224
0225 put_dev:
0226     rds_ib_dev_put(rds_ibdev);
0227     return ret;
0228 }
0229
0230 /*
0231  * New connections use this to find the device to associate with the
0232  * connection.  It's not in the fast path so we're not concerned about the
0233  * performance of the IB call.  (As of this writing, it uses an interrupt
0234  * blocking spinlock to serialize walking a per-device list of all registered
0235  * clients.)
0236  *
0237  * RCU is used to handle incoming connections racing with device teardown.
0238  * Rather than use a lock to serialize removal from the client_data and
0239  * getting a new reference, we use an RCU grace period.  The destruction
0240  * path removes the device from client_data and then waits for all RCU
0241  * readers to finish.
0242  *
0243  * A new connection can get NULL from this if its arriving on a
0244  * device that is in the process of being removed.
0245  */
0246 struct rds_ib_device *rds_ib_get_client_data(struct ib_device *device)
0247 {
0248     struct rds_ib_device *rds_ibdev;
0249
0250     rcu_read_lock();
0251     rds_ibdev = ib_get_client_data(device, &rds_ib_client);
0252     if (rds_ibdev)
0253         refcount_inc(&rds_ibdev->refcount);
0254     rcu_read_unlock();
0255     return rds_ibdev;
0256 }
0257
0258 /*
0259  * The IB stack is letting us know that a device is going away.  This can
0260  * happen if the underlying HCA driver is removed or if PCI hotplug is removing
0261  * the pci function, for example.
0262  *
0263  * This can be called at any time and can be racing with any other RDS path.
0264  */
0265 static void rds_ib_remove_one(struct ib_device *device, void *client_data)
0266 {
0267     struct rds_ib_device *rds_ibdev = client_data;
0268
0269     rds_ib_dev_shutdown(rds_ibdev);
0270
0271     /* stop connection attempts from getting a reference to this device. */
0272     ib_set_client_data(device, &rds_ib_client, NULL);
0273
0274     down_write(&rds_ib_devices_lock);
0275     list_del_rcu(&rds_ibdev->list);
0276     up_write(&rds_ib_devices_lock);
0277
0278     /*
0279      * This synchronize rcu is waiting for readers of both the ib
0280      * client data and the devices list to finish before we drop
0281      * both of those references.
0282      */
0283     synchronize_rcu();
0284     rds_ib_dev_put(rds_ibdev);
0285     rds_ib_dev_put(rds_ibdev);
0286 }
0287
0288 struct ib_client rds_ib_client = {
0289     .name   = "rds_ib",
0290     .add    = rds_ib_add_one,
0291     .remove = rds_ib_remove_one
0292 };
0293
0294 static int rds_ib_conn_info_visitor(struct rds_connection *conn,
0295                     void *buffer)
0296 {
0297     struct rds_info_rdma_connection *iinfo = buffer;
0298     struct rds_ib_connection *ic = conn->c_transport_data;
0299
0300     /* We will only ever look at IB transports */
0301     if (conn->c_trans != &rds_ib_transport)
0302         return 0;
0303     if (conn->c_isv6)
0304         return 0;
0305
0306     iinfo->src_addr = conn->c_laddr.s6_addr32[3];
0307     iinfo->dst_addr = conn->c_faddr.s6_addr32[3];
0308     if (ic) {
0309         iinfo->tos = conn->c_tos;
0310         iinfo->sl = ic->i_sl;
0311     }
0312
0313     memset(&iinfo->src_gid, 0, sizeof(iinfo->src_gid));
0314     memset(&iinfo->dst_gid, 0, sizeof(iinfo->dst_gid));
0315     if (rds_conn_state(conn) == RDS_CONN_UP) {
0316         struct rds_ib_device *rds_ibdev;
0317
0318         rdma_read_gids(ic->i_cm_id, (union ib_gid *)&iinfo->src_gid,
0319                    (union ib_gid *)&iinfo->dst_gid);
0320
0321         rds_ibdev = ic->rds_ibdev;
0322         iinfo->max_send_wr = ic->i_send_ring.w_nr;
0323         iinfo->max_recv_wr = ic->i_recv_ring.w_nr;
0324         iinfo->max_send_sge = rds_ibdev->max_sge;
0325         rds_ib_get_mr_info(rds_ibdev, iinfo);
0326         iinfo->cache_allocs = atomic_read(&ic->i_cache_allocs);
0327     }
0328     return 1;
0329 }
0330
0331 #if IS_ENABLED(CONFIG_IPV6)
0332 /* IPv6 version of rds_ib_conn_info_visitor(). */
0333 static int rds6_ib_conn_info_visitor(struct rds_connection *conn,
0334                      void *buffer)
0335 {
0336     struct rds6_info_rdma_connection *iinfo6 = buffer;
0337     struct rds_ib_connection *ic = conn->c_transport_data;
0338
0339     /* We will only ever look at IB transports */
0340     if (conn->c_trans != &rds_ib_transport)
0341         return 0;
0342
0343     iinfo6->src_addr = conn->c_laddr;
0344     iinfo6->dst_addr = conn->c_faddr;
0345     if (ic) {
0346         iinfo6->tos = conn->c_tos;
0347         iinfo6->sl = ic->i_sl;
0348     }
0349
0350     memset(&iinfo6->src_gid, 0, sizeof(iinfo6->src_gid));
0351     memset(&iinfo6->dst_gid, 0, sizeof(iinfo6->dst_gid));
0352
0353     if (rds_conn_state(conn) == RDS_CONN_UP) {
0354         struct rds_ib_device *rds_ibdev;
0355
0356         rdma_read_gids(ic->i_cm_id, (union ib_gid *)&iinfo6->src_gid,
0357                    (union ib_gid *)&iinfo6->dst_gid);
0358         rds_ibdev = ic->rds_ibdev;
0359         iinfo6->max_send_wr = ic->i_send_ring.w_nr;
0360         iinfo6->max_recv_wr = ic->i_recv_ring.w_nr;
0361         iinfo6->max_send_sge = rds_ibdev->max_sge;
0362         rds6_ib_get_mr_info(rds_ibdev, iinfo6);
0363         iinfo6->cache_allocs = atomic_read(&ic->i_cache_allocs);
0364     }
0365     return 1;
0366 }
0367 #endif
0368
0369 static void rds_ib_ic_info(struct socket *sock, unsigned int len,
0370                struct rds_info_iterator *iter,
0371                struct rds_info_lengths *lens)
0372 {
0373     u64 buffer[(sizeof(struct rds_info_rdma_connection) + 7) / 8];
0374
0375     rds_for_each_conn_info(sock, len, iter, lens,
0376                 rds_ib_conn_info_visitor,
0377                 buffer,
0378                 sizeof(struct rds_info_rdma_connection));
0379 }
0380
0381 #if IS_ENABLED(CONFIG_IPV6)
0382 /* IPv6 version of rds_ib_ic_info(). */
0383 static void rds6_ib_ic_info(struct socket *sock, unsigned int len,
0384                 struct rds_info_iterator *iter,
0385                 struct rds_info_lengths *lens)
0386 {
0387     u64 buffer[(sizeof(struct rds6_info_rdma_connection) + 7) / 8];
0388
0389     rds_for_each_conn_info(sock, len, iter, lens,
0390                    rds6_ib_conn_info_visitor,
0391                    buffer,
0392                    sizeof(struct rds6_info_rdma_connection));
0393 }
0394 #endif
0395
0396 /*
0397  * Early RDS/IB was built to only bind to an address if there is an IPoIB
0398  * device with that address set.
0399  *
0400  * If it were me, I'd advocate for something more flexible.  Sending and
0401  * receiving should be device-agnostic.  Transports would try and maintain
0402  * connections between peers who have messages queued.  Userspace would be
0403  * allowed to influence which paths have priority.  We could call userspace
0404  * asserting this policy "routing".
0405  */
0406 static int rds_ib_laddr_check(struct net *net, const struct in6_addr *addr,
0407                   __u32 scope_id)
0408 {
0409     int ret;
0410     struct rdma_cm_id *cm_id;
0411 #if IS_ENABLED(CONFIG_IPV6)
0412     struct sockaddr_in6 sin6;
0413 #endif
0414     struct sockaddr_in sin;
0415     struct sockaddr *sa;
0416     bool isv4;
0417
0418     isv4 = ipv6_addr_v4mapped(addr);
0419     /* Create a CMA ID and try to bind it. This catches both
0420      * IB and iWARP capable NICs.
0421      */
0422     cm_id = rdma_create_id(&init_net, rds_rdma_cm_event_handler,
0423                    NULL, RDMA_PS_TCP, IB_QPT_RC);
0424     if (IS_ERR(cm_id))
0425         return PTR_ERR(cm_id);
0426
0427     if (isv4) {
0428         memset(&sin, 0, sizeof(sin));
0429         sin.sin_family = AF_INET;
0430         sin.sin_addr.s_addr = addr->s6_addr32[3];
0431         sa = (struct sockaddr *)&sin;
0432     } else {
0433 #if IS_ENABLED(CONFIG_IPV6)
0434         memset(&sin6, 0, sizeof(sin6));
0435         sin6.sin6_family = AF_INET6;
0436         sin6.sin6_addr = *addr;
0437         sin6.sin6_scope_id = scope_id;
0438         sa = (struct sockaddr *)&sin6;
0439
0440         /* XXX Do a special IPv6 link local address check here.  The
0441          * reason is that rdma_bind_addr() always succeeds with IPv6
0442          * link local address regardless it is indeed configured in a
0443          * system.
0444          */
0445         if (ipv6_addr_type(addr) & IPV6_ADDR_LINKLOCAL) {
0446             struct net_device *dev;
0447
0448             if (scope_id == 0) {
0449                 ret = -EADDRNOTAVAIL;
0450                 goto out;
0451             }
0452
0453             /* Use init_net for now as RDS is not network
0454              * name space aware.
0455              */
0456             dev = dev_get_by_index(&init_net, scope_id);
0457             if (!dev) {
0458                 ret = -EADDRNOTAVAIL;
0459                 goto out;
0460             }
0461             if (!ipv6_chk_addr(&init_net, addr, dev, 1)) {
0462                 dev_put(dev);
0463                 ret = -EADDRNOTAVAIL;
0464                 goto out;
0465             }
0466             dev_put(dev);
0467         }
0468 #else
0469         ret = -EADDRNOTAVAIL;
0470         goto out;
0471 #endif
0472     }
0473
0474     /* rdma_bind_addr will only succeed for IB & iWARP devices */
0475     ret = rdma_bind_addr(cm_id, sa);
0476     /* due to this, we will claim to support iWARP devices unless we
0477        check node_type. */
0478     if (ret || !cm_id->device ||
0479         cm_id->device->node_type != RDMA_NODE_IB_CA)
0480         ret = -EADDRNOTAVAIL;
0481
0482     rdsdebug("addr %pI6c%%%u ret %d node type %d\n",
0483          addr, scope_id, ret,
0484          cm_id->device ? cm_id->device->node_type : -1);
0485
0486 out:
0487     rdma_destroy_id(cm_id);
0488
0489     return ret;
0490 }
0491
0492 static void rds_ib_unregister_client(void)
0493 {
0494     ib_unregister_client(&rds_ib_client);
0495     /* wait for rds_ib_dev_free() to complete */
0496     flush_workqueue(rds_wq);
0497 }
0498
0499 static void rds_ib_set_unloading(void)
0500 {
0501     atomic_set(&rds_ib_unloading, 1);
0502 }
0503
0504 static bool rds_ib_is_unloading(struct rds_connection *conn)
0505 {
0506     struct rds_conn_path *cp = &conn->c_path[0];
0507
0508     return (test_bit(RDS_DESTROY_PENDING, &cp->cp_flags) ||
0509         atomic_read(&rds_ib_unloading) != 0);
0510 }
0511
0512 void rds_ib_exit(void)
0513 {
0514     rds_ib_set_unloading();
0515     synchronize_rcu();
0516     rds_info_deregister_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info);
0517 #if IS_ENABLED(CONFIG_IPV6)
0518     rds_info_deregister_func(RDS6_INFO_IB_CONNECTIONS, rds6_ib_ic_info);
0519 #endif
0520     rds_ib_unregister_client();
0521     rds_ib_destroy_nodev_conns();
0522     rds_ib_sysctl_exit();
0523     rds_ib_recv_exit();
0524     rds_trans_unregister(&rds_ib_transport);
0525     rds_ib_mr_exit();
0526 }
0527
0528 static u8 rds_ib_get_tos_map(u8 tos)
0529 {
0530     /* 1:1 user to transport map for RDMA transport.
0531      * In future, if custom map is desired, hook can export
0532      * user configurable map.
0533      */
0534     return tos;
0535 }
0536
0537 struct rds_transport rds_ib_transport = {
0538     .laddr_check        = rds_ib_laddr_check,
0539     .xmit_path_complete = rds_ib_xmit_path_complete,
0540     .xmit           = rds_ib_xmit,
0541     .xmit_rdma      = rds_ib_xmit_rdma,
0542     .xmit_atomic        = rds_ib_xmit_atomic,
0543     .recv_path      = rds_ib_recv_path,
0544     .conn_alloc     = rds_ib_conn_alloc,
0545     .conn_free      = rds_ib_conn_free,
0546     .conn_path_connect  = rds_ib_conn_path_connect,
0547     .conn_path_shutdown = rds_ib_conn_path_shutdown,
0548     .inc_copy_to_user   = rds_ib_inc_copy_to_user,
0549     .inc_free       = rds_ib_inc_free,
0550     .cm_initiate_connect    = rds_ib_cm_initiate_connect,
0551     .cm_handle_connect  = rds_ib_cm_handle_connect,
0552     .cm_connect_complete    = rds_ib_cm_connect_complete,
0553     .stats_info_copy    = rds_ib_stats_info_copy,
0554     .exit           = rds_ib_exit,
0555     .get_mr         = rds_ib_get_mr,
0556     .sync_mr        = rds_ib_sync_mr,
0557     .free_mr        = rds_ib_free_mr,
0558     .flush_mrs      = rds_ib_flush_mrs,
0559     .get_tos_map        = rds_ib_get_tos_map,
0560     .t_owner        = THIS_MODULE,
0561     .t_name         = "infiniband",
0562     .t_unloading        = rds_ib_is_unloading,
0563     .t_type         = RDS_TRANS_IB
0564 };
0565
0566 int rds_ib_init(void)
0567 {
0568     int ret;
0569
0570     INIT_LIST_HEAD(&rds_ib_devices);
0571
0572     ret = rds_ib_mr_init();
0573     if (ret)
0574         goto out;
0575
0576     ret = ib_register_client(&rds_ib_client);
0577     if (ret)
0578         goto out_mr_exit;
0579
0580     ret = rds_ib_sysctl_init();
0581     if (ret)
0582         goto out_ibreg;
0583
0584     ret = rds_ib_recv_init();
0585     if (ret)
0586         goto out_sysctl;
0587
0588     rds_trans_register(&rds_ib_transport);
0589
0590     rds_info_register_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info);
0591 #if IS_ENABLED(CONFIG_IPV6)
0592     rds_info_register_func(RDS6_INFO_IB_CONNECTIONS, rds6_ib_ic_info);
0593 #endif
0594
0595     goto out;
0596
0597 out_sysctl:
0598     rds_ib_sysctl_exit();
0599 out_ibreg:
0600     rds_ib_unregister_client();
0601 out_mr_exit:
0602     rds_ib_mr_exit();
0603 out:
0604     return ret;
0605 }
0606
0607 MODULE_LICENSE("GPL");