net/rds/iw.c

   1 /*
   2  * Copyright (c) 2006 Oracle.  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
  43 #include "rds.h"
  44 #include "iw.h"
  45
  46 unsigned int fastreg_pool_size = RDS_FASTREG_POOL_SIZE;
  47 unsigned int fastreg_message_size = RDS_FASTREG_SIZE + 1; /* +1 allows for unaligned MRs */
  48
  49 module_param(fastreg_pool_size, int, 0444);
  50 MODULE_PARM_DESC(fastreg_pool_size, " Max number of fastreg MRs per device");
  51 module_param(fastreg_message_size, int, 0444);
  52 MODULE_PARM_DESC(fastreg_message_size, " Max size of a RDMA transfer (fastreg MRs)");
  53
  54 struct list_head rds_iw_devices;
  55
  56 /* NOTE: if also grabbing iwdev lock, grab this first */
  57 DEFINE_SPINLOCK(iw_nodev_conns_lock);
  58 LIST_HEAD(iw_nodev_conns);
  59
  60 static void rds_iw_add_one(struct ib_device *device)
  61 {
  62         struct rds_iw_device *rds_iwdev;
  63
  64         /* Only handle iwarp devices */
  65         if (device->node_type != RDMA_NODE_RNIC)
  66                 return;
  67
  68         rds_iwdev = kmalloc(sizeof *rds_iwdev, GFP_KERNEL);
  69         if (!rds_iwdev)
  70                 return;
  71
  72         spin_lock_init(&rds_iwdev->spinlock);
  73
  74         rds_iwdev->dma_local_lkey = !!(device->attrs.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY);
  75         rds_iwdev->max_wrs = device->attrs.max_qp_wr;
  76         rds_iwdev->max_sge = min(device->attrs.max_sge, RDS_IW_MAX_SGE);
  77
  78         rds_iwdev->dev = device;
  79         rds_iwdev->pd = ib_alloc_pd(device);
  80         if (IS_ERR(rds_iwdev->pd))
  81                 goto free_dev;
  82
  83         if (!rds_iwdev->dma_local_lkey) {
  84                 rds_iwdev->mr = ib_get_dma_mr(rds_iwdev->pd,
  85                                         IB_ACCESS_REMOTE_READ |
  86                                         IB_ACCESS_REMOTE_WRITE |
  87                                         IB_ACCESS_LOCAL_WRITE);
  88                 if (IS_ERR(rds_iwdev->mr))
  89                         goto err_pd;
  90         } else
  91                 rds_iwdev->mr = NULL;
  92
  93         rds_iwdev->mr_pool = rds_iw_create_mr_pool(rds_iwdev);
  94         if (IS_ERR(rds_iwdev->mr_pool)) {
  95                 rds_iwdev->mr_pool = NULL;
  96                 goto err_mr;
  97         }
  98
  99         INIT_LIST_HEAD(&rds_iwdev->cm_id_list);
 100         INIT_LIST_HEAD(&rds_iwdev->conn_list);
 101         list_add_tail(&rds_iwdev->list, &rds_iw_devices);
 102
 103         ib_set_client_data(device, &rds_iw_client, rds_iwdev);
 104         return;
 105
 106 err_mr:
 107         if (rds_iwdev->mr)
 108                 ib_dereg_mr(rds_iwdev->mr);
 109 err_pd:
 110         ib_dealloc_pd(rds_iwdev->pd);
 111 free_dev:
 112         kfree(rds_iwdev);
 113 }
 114
 115 static void rds_iw_remove_one(struct ib_device *device, void *client_data)
 116 {
 117         struct rds_iw_device *rds_iwdev = client_data;
 118         struct rds_iw_cm_id *i_cm_id, *next;
 119
 120         if (!rds_iwdev)
 121                 return;
 122
 123         spin_lock_irq(&rds_iwdev->spinlock);
 124         list_for_each_entry_safe(i_cm_id, next, &rds_iwdev->cm_id_list, list) {
 125                 list_del(&i_cm_id->list);
 126                 kfree(i_cm_id);
 127         }
 128         spin_unlock_irq(&rds_iwdev->spinlock);
 129
 130         rds_iw_destroy_conns(rds_iwdev);
 131
 132         if (rds_iwdev->mr_pool)
 133                 rds_iw_destroy_mr_pool(rds_iwdev->mr_pool);
 134
 135         if (rds_iwdev->mr)
 136                 ib_dereg_mr(rds_iwdev->mr);
 137
 138         ib_dealloc_pd(rds_iwdev->pd);
 139
 140         list_del(&rds_iwdev->list);
 141         kfree(rds_iwdev);
 142 }
 143
 144 struct ib_client rds_iw_client = {
 145         .name   = "rds_iw",
 146         .add    = rds_iw_add_one,
 147         .remove = rds_iw_remove_one
 148 };
 149
 150 static int rds_iw_conn_info_visitor(struct rds_connection *conn,
 151                                     void *buffer)
 152 {
 153         struct rds_info_rdma_connection *iinfo = buffer;
 154         struct rds_iw_connection *ic;
 155
 156         /* We will only ever look at IB transports */
 157         if (conn->c_trans != &rds_iw_transport)
 158                 return 0;
 159
 160         iinfo->src_addr = conn->c_laddr;
 161         iinfo->dst_addr = conn->c_faddr;
 162
 163         memset(&iinfo->src_gid, 0, sizeof(iinfo->src_gid));
 164         memset(&iinfo->dst_gid, 0, sizeof(iinfo->dst_gid));
 165         if (rds_conn_state(conn) == RDS_CONN_UP) {
 166                 struct rds_iw_device *rds_iwdev;
 167                 struct rdma_dev_addr *dev_addr;
 168
 169                 ic = conn->c_transport_data;
 170                 dev_addr = &ic->i_cm_id->route.addr.dev_addr;
 171
 172                 rdma_addr_get_sgid(dev_addr, (union ib_gid *) &iinfo->src_gid);
 173                 rdma_addr_get_dgid(dev_addr, (union ib_gid *) &iinfo->dst_gid);
 174
 175                 rds_iwdev = ib_get_client_data(ic->i_cm_id->device, &rds_iw_client);
 176                 iinfo->max_send_wr = ic->i_send_ring.w_nr;
 177                 iinfo->max_recv_wr = ic->i_recv_ring.w_nr;
 178                 iinfo->max_send_sge = rds_iwdev->max_sge;
 179                 rds_iw_get_mr_info(rds_iwdev, iinfo);
 180         }
 181         return 1;
 182 }
 183
 184 static void rds_iw_ic_info(struct socket *sock, unsigned int len,
 185                            struct rds_info_iterator *iter,
 186                            struct rds_info_lengths *lens)
 187 {
 188         rds_for_each_conn_info(sock, len, iter, lens,
 189                                 rds_iw_conn_info_visitor,
 190                                 sizeof(struct rds_info_rdma_connection));
 191 }
 192
 193
 194 /*
 195  * Early RDS/IB was built to only bind to an address if there is an IPoIB
 196  * device with that address set.
 197  *
 198  * If it were me, I'd advocate for something more flexible.  Sending and
 199  * receiving should be device-agnostic.  Transports would try and maintain
 200  * connections between peers who have messages queued.  Userspace would be
 201  * allowed to influence which paths have priority.  We could call userspace
 202  * asserting this policy "routing".
 203  */
 204 static int rds_iw_laddr_check(struct net *net, __be32 addr)
 205 {
 206         int ret;
 207         struct rdma_cm_id *cm_id;
 208         struct sockaddr_in sin;
 209
 210         /* Create a CMA ID and try to bind it. This catches both
 211          * IB and iWARP capable NICs.
 212          */
 213         cm_id = rdma_create_id(&init_net, NULL, NULL, RDMA_PS_TCP, IB_QPT_RC);
 214         if (IS_ERR(cm_id))
 215                 return PTR_ERR(cm_id);
 216
 217         memset(&sin, 0, sizeof(sin));
 218         sin.sin_family = AF_INET;
 219         sin.sin_addr.s_addr = addr;
 220
 221         /* rdma_bind_addr will only succeed for IB & iWARP devices */
 222         ret = rdma_bind_addr(cm_id, (struct sockaddr *)&sin);
 223         /* due to this, we will claim to support IB devices unless we
 224            check node_type. */
 225         if (ret || !cm_id->device ||
 226             cm_id->device->node_type != RDMA_NODE_RNIC)
 227                 ret = -EADDRNOTAVAIL;
 228
 229         rdsdebug("addr %pI4 ret %d node type %d\n",
 230                 &addr, ret,
 231                 cm_id->device ? cm_id->device->node_type : -1);
 232
 233         rdma_destroy_id(cm_id);
 234
 235         return ret;
 236 }
 237
 238 void rds_iw_exit(void)
 239 {
 240         rds_info_deregister_func(RDS_INFO_IWARP_CONNECTIONS, rds_iw_ic_info);
 241         rds_iw_destroy_nodev_conns();
 242         ib_unregister_client(&rds_iw_client);
 243         rds_iw_sysctl_exit();
 244         rds_iw_recv_exit();
 245         rds_trans_unregister(&rds_iw_transport);
 246 }
 247
 248 struct rds_transport rds_iw_transport = {
 249         .laddr_check            = rds_iw_laddr_check,
 250         .xmit_complete          = rds_iw_xmit_complete,
 251         .xmit                   = rds_iw_xmit,
 252         .xmit_rdma              = rds_iw_xmit_rdma,
 253         .recv                   = rds_iw_recv,
 254         .conn_alloc             = rds_iw_conn_alloc,
 255         .conn_free              = rds_iw_conn_free,
 256         .conn_connect           = rds_iw_conn_connect,
 257         .conn_shutdown          = rds_iw_conn_shutdown,
 258         .inc_copy_to_user       = rds_iw_inc_copy_to_user,
 259         .inc_free               = rds_iw_inc_free,
 260         .cm_initiate_connect    = rds_iw_cm_initiate_connect,
 261         .cm_handle_connect      = rds_iw_cm_handle_connect,
 262         .cm_connect_complete    = rds_iw_cm_connect_complete,
 263         .stats_info_copy        = rds_iw_stats_info_copy,
 264         .exit                   = rds_iw_exit,
 265         .get_mr                 = rds_iw_get_mr,
 266         .sync_mr                = rds_iw_sync_mr,
 267         .free_mr                = rds_iw_free_mr,
 268         .flush_mrs              = rds_iw_flush_mrs,
 269         .t_owner                = THIS_MODULE,
 270         .t_name                 = "iwarp",
 271         .t_type                 = RDS_TRANS_IWARP,
 272         .t_prefer_loopback      = 1,
 273 };
 274
 275 int rds_iw_init(void)
 276 {
 277         int ret;
 278
 279         INIT_LIST_HEAD(&rds_iw_devices);
 280
 281         ret = ib_register_client(&rds_iw_client);
 282         if (ret)
 283                 goto out;
 284
 285         ret = rds_iw_sysctl_init();
 286         if (ret)
 287                 goto out_ibreg;
 288
 289         ret = rds_iw_recv_init();
 290         if (ret)
 291                 goto out_sysctl;
 292
 293         ret = rds_trans_register(&rds_iw_transport);
 294         if (ret)
 295                 goto out_recv;
 296
 297         rds_info_register_func(RDS_INFO_IWARP_CONNECTIONS, rds_iw_ic_info);
 298
 299         goto out;
 300
 301 out_recv:
 302         rds_iw_recv_exit();
 303 out_sysctl:
 304         rds_iw_sysctl_exit();
 305 out_ibreg:
 306         ib_unregister_client(&rds_iw_client);
 307 out:
 308         return ret;
 309 }
 310
 311 MODULE_LICENSE("GPL");
 312