net/rds/ib.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_single_path.h"
  44 #include "rds.h"
  45 #include "ib.h"
  46 #include "ib_mr.h"
  47
  48 static unsigned int rds_ib_mr_1m_pool_size = RDS_MR_1M_POOL_SIZE;
  49 static unsigned int rds_ib_mr_8k_pool_size = RDS_MR_8K_POOL_SIZE;
  50 unsigned int rds_ib_retry_count = RDS_IB_DEFAULT_RETRY_COUNT;
  51
  52 module_param(rds_ib_mr_1m_pool_size, int, 0444);
  53 MODULE_PARM_DESC(rds_ib_mr_1m_pool_size, " Max number of 1M mr per HCA");
  54 module_param(rds_ib_mr_8k_pool_size, int, 0444);
  55 MODULE_PARM_DESC(rds_ib_mr_8k_pool_size, " Max number of 8K mr per HCA");
  56 module_param(rds_ib_retry_count, int, 0444);
  57 MODULE_PARM_DESC(rds_ib_retry_count, " Number of hw retries before reporting an error");
  58
  59 /*
  60  * we have a clumsy combination of RCU and a rwsem protecting this list
  61  * because it is used both in the get_mr fast path and while blocking in
  62  * the FMR flushing path.
  63  */
  64 DECLARE_RWSEM(rds_ib_devices_lock);
  65 struct list_head rds_ib_devices;
  66
  67 /* NOTE: if also grabbing ibdev lock, grab this first */
  68 DEFINE_SPINLOCK(ib_nodev_conns_lock);
  69 LIST_HEAD(ib_nodev_conns);
  70
  71 static void rds_ib_nodev_connect(void)
  72 {
  73         struct rds_ib_connection *ic;
  74
  75         spin_lock(&ib_nodev_conns_lock);
  76         list_for_each_entry(ic, &ib_nodev_conns, ib_node)
  77                 rds_conn_connect_if_down(ic->conn);
  78         spin_unlock(&ib_nodev_conns_lock);
  79 }
  80
  81 static void rds_ib_dev_shutdown(struct rds_ib_device *rds_ibdev)
  82 {
  83         struct rds_ib_connection *ic;
  84         unsigned long flags;
  85
  86         spin_lock_irqsave(&rds_ibdev->spinlock, flags);
  87         list_for_each_entry(ic, &rds_ibdev->conn_list, ib_node)
  88                 rds_conn_drop(ic->conn);
  89         spin_unlock_irqrestore(&rds_ibdev->spinlock, flags);
  90 }
  91
  92 /*
  93  * rds_ib_destroy_mr_pool() blocks on a few things and mrs drop references
  94  * from interrupt context so we push freing off into a work struct in krdsd.
  95  */
  96 static void rds_ib_dev_free(struct work_struct *work)
  97 {
  98         struct rds_ib_ipaddr *i_ipaddr, *i_next;
  99         struct rds_ib_device *rds_ibdev = container_of(work,
 100                                         struct rds_ib_device, free_work);
 101
 102         if (rds_ibdev->mr_8k_pool)
 103                 rds_ib_destroy_mr_pool(rds_ibdev->mr_8k_pool);
 104         if (rds_ibdev->mr_1m_pool)
 105                 rds_ib_destroy_mr_pool(rds_ibdev->mr_1m_pool);
 106         if (rds_ibdev->pd)
 107                 ib_dealloc_pd(rds_ibdev->pd);
 108
 109         list_for_each_entry_safe(i_ipaddr, i_next, &rds_ibdev->ipaddr_list, list) {
 110                 list_del(&i_ipaddr->list);
 111                 kfree(i_ipaddr);
 112         }
 113
 114         kfree(rds_ibdev->vector_load);
 115
 116         kfree(rds_ibdev);
 117 }
 118
 119 void rds_ib_dev_put(struct rds_ib_device *rds_ibdev)
 120 {
 121         BUG_ON(refcount_read(&rds_ibdev->refcount) == 0);
 122         if (refcount_dec_and_test(&rds_ibdev->refcount))
 123                 queue_work(rds_wq, &rds_ibdev->free_work);
 124 }
 125
 126 static void rds_ib_add_one(struct ib_device *device)
 127 {
 128         struct rds_ib_device *rds_ibdev;
 129
 130         /* Only handle IB (no iWARP) devices */
 131         if (device->node_type != RDMA_NODE_IB_CA)
 132                 return;
 133
 134         rds_ibdev = kzalloc_node(sizeof(struct rds_ib_device), GFP_KERNEL,
 135                                  ibdev_to_node(device));
 136         if (!rds_ibdev)
 137                 return;
 138
 139         spin_lock_init(&rds_ibdev->spinlock);
 140         refcount_set(&rds_ibdev->refcount, 1);
 141         INIT_WORK(&rds_ibdev->free_work, rds_ib_dev_free);
 142
 143         rds_ibdev->max_wrs = device->attrs.max_qp_wr;
 144         rds_ibdev->max_sge = min(device->attrs.max_sge, RDS_IB_MAX_SGE);
 145
 146         rds_ibdev->has_fr = (device->attrs.device_cap_flags &
 147                                   IB_DEVICE_MEM_MGT_EXTENSIONS);
 148         rds_ibdev->has_fmr = (device->alloc_fmr && device->dealloc_fmr &&
 149                             device->map_phys_fmr && device->unmap_fmr);
 150         rds_ibdev->use_fastreg = (rds_ibdev->has_fr && !rds_ibdev->has_fmr);
 151
 152         rds_ibdev->fmr_max_remaps = device->attrs.max_map_per_fmr?: 32;
 153         rds_ibdev->max_1m_mrs = device->attrs.max_mr ?
 154                 min_t(unsigned int, (device->attrs.max_mr / 2),
 155                       rds_ib_mr_1m_pool_size) : rds_ib_mr_1m_pool_size;
 156
 157         rds_ibdev->max_8k_mrs = device->attrs.max_mr ?
 158                 min_t(unsigned int, ((device->attrs.max_mr / 2) * RDS_MR_8K_SCALE),
 159                       rds_ib_mr_8k_pool_size) : rds_ib_mr_8k_pool_size;
 160
 161         rds_ibdev->max_initiator_depth = device->attrs.max_qp_init_rd_atom;
 162         rds_ibdev->max_responder_resources = device->attrs.max_qp_rd_atom;
 163
 164         rds_ibdev->vector_load = kzalloc(sizeof(int) * device->num_comp_vectors,
 165                                          GFP_KERNEL);
 166         if (!rds_ibdev->vector_load) {
 167                 pr_err("RDS/IB: %s failed to allocate vector memory\n",
 168                         __func__);
 169                 goto put_dev;
 170         }
 171
 172         rds_ibdev->dev = device;
 173         rds_ibdev->pd = ib_alloc_pd(device, 0);
 174         if (IS_ERR(rds_ibdev->pd)) {
 175                 rds_ibdev->pd = NULL;
 176                 goto put_dev;
 177         }
 178
 179         rds_ibdev->mr_1m_pool =
 180                 rds_ib_create_mr_pool(rds_ibdev, RDS_IB_MR_1M_POOL);
 181         if (IS_ERR(rds_ibdev->mr_1m_pool)) {
 182                 rds_ibdev->mr_1m_pool = NULL;
 183                 goto put_dev;
 184         }
 185
 186         rds_ibdev->mr_8k_pool =
 187                 rds_ib_create_mr_pool(rds_ibdev, RDS_IB_MR_8K_POOL);
 188         if (IS_ERR(rds_ibdev->mr_8k_pool)) {
 189                 rds_ibdev->mr_8k_pool = NULL;
 190                 goto put_dev;
 191         }
 192
 193         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",
 194                  device->attrs.max_fmr, rds_ibdev->max_wrs, rds_ibdev->max_sge,
 195                  rds_ibdev->fmr_max_remaps, rds_ibdev->max_1m_mrs,
 196                  rds_ibdev->max_8k_mrs);
 197
 198         pr_info("RDS/IB: %s: %s supported and preferred\n",
 199                 device->name,
 200                 rds_ibdev->use_fastreg ? "FRMR" : "FMR");
 201
 202         INIT_LIST_HEAD(&rds_ibdev->ipaddr_list);
 203         INIT_LIST_HEAD(&rds_ibdev->conn_list);
 204
 205         down_write(&rds_ib_devices_lock);
 206         list_add_tail_rcu(&rds_ibdev->list, &rds_ib_devices);
 207         up_write(&rds_ib_devices_lock);
 208         refcount_inc(&rds_ibdev->refcount);
 209
 210         ib_set_client_data(device, &rds_ib_client, rds_ibdev);
 211         refcount_inc(&rds_ibdev->refcount);
 212
 213         rds_ib_nodev_connect();
 214
 215 put_dev:
 216         rds_ib_dev_put(rds_ibdev);
 217 }
 218
 219 /*
 220  * New connections use this to find the device to associate with the
 221  * connection.  It's not in the fast path so we're not concerned about the
 222  * performance of the IB call.  (As of this writing, it uses an interrupt
 223  * blocking spinlock to serialize walking a per-device list of all registered
 224  * clients.)
 225  *
 226  * RCU is used to handle incoming connections racing with device teardown.
 227  * Rather than use a lock to serialize removal from the client_data and
 228  * getting a new reference, we use an RCU grace period.  The destruction
 229  * path removes the device from client_data and then waits for all RCU
 230  * readers to finish.
 231  *
 232  * A new connection can get NULL from this if its arriving on a
 233  * device that is in the process of being removed.
 234  */
 235 struct rds_ib_device *rds_ib_get_client_data(struct ib_device *device)
 236 {
 237         struct rds_ib_device *rds_ibdev;
 238
 239         rcu_read_lock();
 240         rds_ibdev = ib_get_client_data(device, &rds_ib_client);
 241         if (rds_ibdev)
 242                 refcount_inc(&rds_ibdev->refcount);
 243         rcu_read_unlock();
 244         return rds_ibdev;
 245 }
 246
 247 /*
 248  * The IB stack is letting us know that a device is going away.  This can
 249  * happen if the underlying HCA driver is removed or if PCI hotplug is removing
 250  * the pci function, for example.
 251  *
 252  * This can be called at any time and can be racing with any other RDS path.
 253  */
 254 static void rds_ib_remove_one(struct ib_device *device, void *client_data)
 255 {
 256         struct rds_ib_device *rds_ibdev = client_data;
 257
 258         if (!rds_ibdev)
 259                 return;
 260
 261         rds_ib_dev_shutdown(rds_ibdev);
 262
 263         /* stop connection attempts from getting a reference to this device. */
 264         ib_set_client_data(device, &rds_ib_client, NULL);
 265
 266         down_write(&rds_ib_devices_lock);
 267         list_del_rcu(&rds_ibdev->list);
 268         up_write(&rds_ib_devices_lock);
 269
 270         /*
 271          * This synchronize rcu is waiting for readers of both the ib
 272          * client data and the devices list to finish before we drop
 273          * both of those references.
 274          */
 275         synchronize_rcu();
 276         rds_ib_dev_put(rds_ibdev);
 277         rds_ib_dev_put(rds_ibdev);
 278 }
 279
 280 struct ib_client rds_ib_client = {
 281         .name   = "rds_ib",
 282         .add    = rds_ib_add_one,
 283         .remove = rds_ib_remove_one
 284 };
 285
 286 static int rds_ib_conn_info_visitor(struct rds_connection *conn,
 287                                     void *buffer)
 288 {
 289         struct rds_info_rdma_connection *iinfo = buffer;
 290         struct rds_ib_connection *ic;
 291
 292         /* We will only ever look at IB transports */
 293         if (conn->c_trans != &rds_ib_transport)
 294                 return 0;
 295
 296         iinfo->src_addr = conn->c_laddr;
 297         iinfo->dst_addr = conn->c_faddr;
 298
 299         memset(&iinfo->src_gid, 0, sizeof(iinfo->src_gid));
 300         memset(&iinfo->dst_gid, 0, sizeof(iinfo->dst_gid));
 301         if (rds_conn_state(conn) == RDS_CONN_UP) {
 302                 struct rds_ib_device *rds_ibdev;
 303                 struct rdma_dev_addr *dev_addr;
 304
 305                 ic = conn->c_transport_data;
 306                 dev_addr = &ic->i_cm_id->route.addr.dev_addr;
 307
 308                 rdma_addr_get_sgid(dev_addr, (union ib_gid *) &iinfo->src_gid);
 309                 rdma_addr_get_dgid(dev_addr, (union ib_gid *) &iinfo->dst_gid);
 310
 311                 rds_ibdev = ic->rds_ibdev;
 312                 iinfo->max_send_wr = ic->i_send_ring.w_nr;
 313                 iinfo->max_recv_wr = ic->i_recv_ring.w_nr;
 314                 iinfo->max_send_sge = rds_ibdev->max_sge;
 315                 rds_ib_get_mr_info(rds_ibdev, iinfo);
 316         }
 317         return 1;
 318 }
 319
 320 static void rds_ib_ic_info(struct socket *sock, unsigned int len,
 321                            struct rds_info_iterator *iter,
 322                            struct rds_info_lengths *lens)
 323 {
 324         rds_for_each_conn_info(sock, len, iter, lens,
 325                                 rds_ib_conn_info_visitor,
 326                                 sizeof(struct rds_info_rdma_connection));
 327 }
 328
 329
 330 /*
 331  * Early RDS/IB was built to only bind to an address if there is an IPoIB
 332  * device with that address set.
 333  *
 334  * If it were me, I'd advocate for something more flexible.  Sending and
 335  * receiving should be device-agnostic.  Transports would try and maintain
 336  * connections between peers who have messages queued.  Userspace would be
 337  * allowed to influence which paths have priority.  We could call userspace
 338  * asserting this policy "routing".
 339  */
 340 static int rds_ib_laddr_check(struct net *net, __be32 addr)
 341 {
 342         int ret;
 343         struct rdma_cm_id *cm_id;
 344         struct sockaddr_in sin;
 345
 346         /* Create a CMA ID and try to bind it. This catches both
 347          * IB and iWARP capable NICs.
 348          */
 349         cm_id = rdma_create_id(&init_net, NULL, NULL, RDMA_PS_TCP, IB_QPT_RC);
 350         if (IS_ERR(cm_id))
 351                 return PTR_ERR(cm_id);
 352
 353         memset(&sin, 0, sizeof(sin));
 354         sin.sin_family = AF_INET;
 355         sin.sin_addr.s_addr = addr;
 356
 357         /* rdma_bind_addr will only succeed for IB & iWARP devices */
 358         ret = rdma_bind_addr(cm_id, (struct sockaddr *)&sin);
 359         /* due to this, we will claim to support iWARP devices unless we
 360            check node_type. */
 361         if (ret || !cm_id->device ||
 362             cm_id->device->node_type != RDMA_NODE_IB_CA)
 363                 ret = -EADDRNOTAVAIL;
 364
 365         rdsdebug("addr %pI4 ret %d node type %d\n",
 366                 &addr, ret,
 367                 cm_id->device ? cm_id->device->node_type : -1);
 368
 369         rdma_destroy_id(cm_id);
 370
 371         return ret;
 372 }
 373
 374 static void rds_ib_unregister_client(void)
 375 {
 376         ib_unregister_client(&rds_ib_client);
 377         /* wait for rds_ib_dev_free() to complete */
 378         flush_workqueue(rds_wq);
 379 }
 380
 381 void rds_ib_exit(void)
 382 {
 383         rds_info_deregister_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info);
 384         rds_ib_unregister_client();
 385         rds_ib_destroy_nodev_conns();
 386         rds_ib_sysctl_exit();
 387         rds_ib_recv_exit();
 388         rds_trans_unregister(&rds_ib_transport);
 389         rds_ib_mr_exit();
 390 }
 391
 392 struct rds_transport rds_ib_transport = {
 393         .laddr_check            = rds_ib_laddr_check,
 394         .xmit_path_complete     = rds_ib_xmit_path_complete,
 395         .xmit                   = rds_ib_xmit,
 396         .xmit_rdma              = rds_ib_xmit_rdma,
 397         .xmit_atomic            = rds_ib_xmit_atomic,
 398         .recv_path              = rds_ib_recv_path,
 399         .conn_alloc             = rds_ib_conn_alloc,
 400         .conn_free              = rds_ib_conn_free,
 401         .conn_path_connect      = rds_ib_conn_path_connect,
 402         .conn_path_shutdown     = rds_ib_conn_path_shutdown,
 403         .inc_copy_to_user       = rds_ib_inc_copy_to_user,
 404         .inc_free               = rds_ib_inc_free,
 405         .cm_initiate_connect    = rds_ib_cm_initiate_connect,
 406         .cm_handle_connect      = rds_ib_cm_handle_connect,
 407         .cm_connect_complete    = rds_ib_cm_connect_complete,
 408         .stats_info_copy        = rds_ib_stats_info_copy,
 409         .exit                   = rds_ib_exit,
 410         .get_mr                 = rds_ib_get_mr,
 411         .sync_mr                = rds_ib_sync_mr,
 412         .free_mr                = rds_ib_free_mr,
 413         .flush_mrs              = rds_ib_flush_mrs,
 414         .t_owner                = THIS_MODULE,
 415         .t_name                 = "infiniband",
 416         .t_type                 = RDS_TRANS_IB
 417 };
 418
 419 int rds_ib_init(void)
 420 {
 421         int ret;
 422
 423         INIT_LIST_HEAD(&rds_ib_devices);
 424
 425         ret = rds_ib_mr_init();
 426         if (ret)
 427                 goto out;
 428
 429         ret = ib_register_client(&rds_ib_client);
 430         if (ret)
 431                 goto out_mr_exit;
 432
 433         ret = rds_ib_sysctl_init();
 434         if (ret)
 435                 goto out_ibreg;
 436
 437         ret = rds_ib_recv_init();
 438         if (ret)
 439                 goto out_sysctl;
 440
 441         rds_trans_register(&rds_ib_transport);
 442
 443         rds_info_register_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info);
 444
 445         goto out;
 446
 447 out_sysctl:
 448         rds_ib_sysctl_exit();
 449 out_ibreg:
 450         rds_ib_unregister_client();
 451 out_mr_exit:
 452         rds_ib_mr_exit();
 453 out:
 454         return ret;
 455 }
 456
 457 MODULE_LICENSE("GPL");
 458