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 unsigned int rds_ib_mr_1m_pool_size = RDS_MR_1M_POOL_SIZE;
  49 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         if (rds_ibdev->vector_load)
 115                 kfree(rds_ibdev->vector_load);
 116
 117         kfree(rds_ibdev);
 118 }
 119
 120 void rds_ib_dev_put(struct rds_ib_device *rds_ibdev)
 121 {
 122         BUG_ON(atomic_read(&rds_ibdev->refcount) <= 0);
 123         if (atomic_dec_and_test(&rds_ibdev->refcount))
 124                 queue_work(rds_wq, &rds_ibdev->free_work);
 125 }
 126
 127 static void rds_ib_add_one(struct ib_device *device)
 128 {
 129         struct rds_ib_device *rds_ibdev;
 130
 131         /* Only handle IB (no iWARP) devices */
 132         if (device->node_type != RDMA_NODE_IB_CA)
 133                 return;
 134
 135         rds_ibdev = kzalloc_node(sizeof(struct rds_ib_device), GFP_KERNEL,
 136                                  ibdev_to_node(device));
 137         if (!rds_ibdev)
 138                 return;
 139
 140         spin_lock_init(&rds_ibdev->spinlock);
 141         atomic_set(&rds_ibdev->refcount, 1);
 142         INIT_WORK(&rds_ibdev->free_work, rds_ib_dev_free);
 143
 144         rds_ibdev->max_wrs = device->attrs.max_qp_wr;
 145         rds_ibdev->max_sge = min(device->attrs.max_sge, RDS_IB_MAX_SGE);
 146
 147         rds_ibdev->has_fr = (device->attrs.device_cap_flags &
 148                                   IB_DEVICE_MEM_MGT_EXTENSIONS);
 149         rds_ibdev->has_fmr = (device->alloc_fmr && device->dealloc_fmr &&
 150                             device->map_phys_fmr && device->unmap_fmr);
 151         rds_ibdev->use_fastreg = (rds_ibdev->has_fr && !rds_ibdev->has_fmr);
 152
 153         rds_ibdev->fmr_max_remaps = device->attrs.max_map_per_fmr?: 32;
 154         rds_ibdev->max_1m_mrs = device->attrs.max_mr ?
 155                 min_t(unsigned int, (device->attrs.max_mr / 2),
 156                       rds_ib_mr_1m_pool_size) : rds_ib_mr_1m_pool_size;
 157
 158         rds_ibdev->max_8k_mrs = device->attrs.max_mr ?
 159                 min_t(unsigned int, ((device->attrs.max_mr / 2) * RDS_MR_8K_SCALE),
 160                       rds_ib_mr_8k_pool_size) : rds_ib_mr_8k_pool_size;
 161
 162         rds_ibdev->max_initiator_depth = device->attrs.max_qp_init_rd_atom;
 163         rds_ibdev->max_responder_resources = device->attrs.max_qp_rd_atom;
 164
 165         rds_ibdev->vector_load = kzalloc(sizeof(int) * device->num_comp_vectors,
 166                                          GFP_KERNEL);
 167         if (!rds_ibdev->vector_load) {
 168                 pr_err("RDS/IB: %s failed to allocate vector memory\n",
 169                         __func__);
 170                 goto put_dev;
 171         }
 172
 173         rds_ibdev->dev = device;
 174         rds_ibdev->pd = ib_alloc_pd(device, 0);
 175         if (IS_ERR(rds_ibdev->pd)) {
 176                 rds_ibdev->pd = NULL;
 177                 goto put_dev;
 178         }
 179
 180         rds_ibdev->mr_1m_pool =
 181                 rds_ib_create_mr_pool(rds_ibdev, RDS_IB_MR_1M_POOL);
 182         if (IS_ERR(rds_ibdev->mr_1m_pool)) {
 183                 rds_ibdev->mr_1m_pool = NULL;
 184                 goto put_dev;
 185         }
 186
 187         rds_ibdev->mr_8k_pool =
 188                 rds_ib_create_mr_pool(rds_ibdev, RDS_IB_MR_8K_POOL);
 189         if (IS_ERR(rds_ibdev->mr_8k_pool)) {
 190                 rds_ibdev->mr_8k_pool = NULL;
 191                 goto put_dev;
 192         }
 193
 194         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",
 195                  device->attrs.max_fmr, rds_ibdev->max_wrs, rds_ibdev->max_sge,
 196                  rds_ibdev->fmr_max_remaps, rds_ibdev->max_1m_mrs,
 197                  rds_ibdev->max_8k_mrs);
 198
 199         pr_info("RDS/IB: %s: %s supported and preferred\n",
 200                 device->name,
 201                 rds_ibdev->use_fastreg ? "FRMR" : "FMR");
 202
 203         INIT_LIST_HEAD(&rds_ibdev->ipaddr_list);
 204         INIT_LIST_HEAD(&rds_ibdev->conn_list);
 205
 206         down_write(&rds_ib_devices_lock);
 207         list_add_tail_rcu(&rds_ibdev->list, &rds_ib_devices);
 208         up_write(&rds_ib_devices_lock);
 209         atomic_inc(&rds_ibdev->refcount);
 210
 211         ib_set_client_data(device, &rds_ib_client, rds_ibdev);
 212         atomic_inc(&rds_ibdev->refcount);
 213
 214         rds_ib_nodev_connect();
 215
 216 put_dev:
 217         rds_ib_dev_put(rds_ibdev);
 218 }
 219
 220 /*
 221  * New connections use this to find the device to associate with the
 222  * connection.  It's not in the fast path so we're not concerned about the
 223  * performance of the IB call.  (As of this writing, it uses an interrupt
 224  * blocking spinlock to serialize walking a per-device list of all registered
 225  * clients.)
 226  *
 227  * RCU is used to handle incoming connections racing with device teardown.
 228  * Rather than use a lock to serialize removal from the client_data and
 229  * getting a new reference, we use an RCU grace period.  The destruction
 230  * path removes the device from client_data and then waits for all RCU
 231  * readers to finish.
 232  *
 233  * A new connection can get NULL from this if its arriving on a
 234  * device that is in the process of being removed.
 235  */
 236 struct rds_ib_device *rds_ib_get_client_data(struct ib_device *device)
 237 {
 238         struct rds_ib_device *rds_ibdev;
 239
 240         rcu_read_lock();
 241         rds_ibdev = ib_get_client_data(device, &rds_ib_client);
 242         if (rds_ibdev)
 243                 atomic_inc(&rds_ibdev->refcount);
 244         rcu_read_unlock();
 245         return rds_ibdev;
 246 }
 247
 248 /*
 249  * The IB stack is letting us know that a device is going away.  This can
 250  * happen if the underlying HCA driver is removed or if PCI hotplug is removing
 251  * the pci function, for example.
 252  *
 253  * This can be called at any time and can be racing with any other RDS path.
 254  */
 255 static void rds_ib_remove_one(struct ib_device *device, void *client_data)
 256 {
 257         struct rds_ib_device *rds_ibdev = client_data;
 258
 259         if (!rds_ibdev)
 260                 return;
 261
 262         rds_ib_dev_shutdown(rds_ibdev);
 263
 264         /* stop connection attempts from getting a reference to this device. */
 265         ib_set_client_data(device, &rds_ib_client, NULL);
 266
 267         down_write(&rds_ib_devices_lock);
 268         list_del_rcu(&rds_ibdev->list);
 269         up_write(&rds_ib_devices_lock);
 270
 271         /*
 272          * This synchronize rcu is waiting for readers of both the ib
 273          * client data and the devices list to finish before we drop
 274          * both of those references.
 275          */
 276         synchronize_rcu();
 277         rds_ib_dev_put(rds_ibdev);
 278         rds_ib_dev_put(rds_ibdev);
 279 }
 280
 281 struct ib_client rds_ib_client = {
 282         .name   = "rds_ib",
 283         .add    = rds_ib_add_one,
 284         .remove = rds_ib_remove_one
 285 };
 286
 287 static int rds_ib_conn_info_visitor(struct rds_connection *conn,
 288                                     void *buffer)
 289 {
 290         struct rds_info_rdma_connection *iinfo = buffer;
 291         struct rds_ib_connection *ic;
 292
 293         /* We will only ever look at IB transports */
 294         if (conn->c_trans != &rds_ib_transport)
 295                 return 0;
 296
 297         iinfo->src_addr = conn->c_laddr;
 298         iinfo->dst_addr = conn->c_faddr;
 299
 300         memset(&iinfo->src_gid, 0, sizeof(iinfo->src_gid));
 301         memset(&iinfo->dst_gid, 0, sizeof(iinfo->dst_gid));
 302         if (rds_conn_state(conn) == RDS_CONN_UP) {
 303                 struct rds_ib_device *rds_ibdev;
 304                 struct rdma_dev_addr *dev_addr;
 305
 306                 ic = conn->c_transport_data;
 307                 dev_addr = &ic->i_cm_id->route.addr.dev_addr;
 308
 309                 rdma_addr_get_sgid(dev_addr, (union ib_gid *) &iinfo->src_gid);
 310                 rdma_addr_get_dgid(dev_addr, (union ib_gid *) &iinfo->dst_gid);
 311
 312                 rds_ibdev = ic->rds_ibdev;
 313                 iinfo->max_send_wr = ic->i_send_ring.w_nr;
 314                 iinfo->max_recv_wr = ic->i_recv_ring.w_nr;
 315                 iinfo->max_send_sge = rds_ibdev->max_sge;
 316                 rds_ib_get_mr_info(rds_ibdev, iinfo);
 317         }
 318         return 1;
 319 }
 320
 321 static void rds_ib_ic_info(struct socket *sock, unsigned int len,
 322                            struct rds_info_iterator *iter,
 323                            struct rds_info_lengths *lens)
 324 {
 325         rds_for_each_conn_info(sock, len, iter, lens,
 326                                 rds_ib_conn_info_visitor,
 327                                 sizeof(struct rds_info_rdma_connection));
 328 }
 329
 330
 331 /*
 332  * Early RDS/IB was built to only bind to an address if there is an IPoIB
 333  * device with that address set.
 334  *
 335  * If it were me, I'd advocate for something more flexible.  Sending and
 336  * receiving should be device-agnostic.  Transports would try and maintain
 337  * connections between peers who have messages queued.  Userspace would be
 338  * allowed to influence which paths have priority.  We could call userspace
 339  * asserting this policy "routing".
 340  */
 341 static int rds_ib_laddr_check(struct net *net, __be32 addr)
 342 {
 343         int ret;
 344         struct rdma_cm_id *cm_id;
 345         struct sockaddr_in sin;
 346
 347         /* Create a CMA ID and try to bind it. This catches both
 348          * IB and iWARP capable NICs.
 349          */
 350         cm_id = rdma_create_id(&init_net, NULL, NULL, RDMA_PS_TCP, IB_QPT_RC);
 351         if (IS_ERR(cm_id))
 352                 return PTR_ERR(cm_id);
 353
 354         memset(&sin, 0, sizeof(sin));
 355         sin.sin_family = AF_INET;
 356         sin.sin_addr.s_addr = addr;
 357
 358         /* rdma_bind_addr will only succeed for IB & iWARP devices */
 359         ret = rdma_bind_addr(cm_id, (struct sockaddr *)&sin);
 360         /* due to this, we will claim to support iWARP devices unless we
 361            check node_type. */
 362         if (ret || !cm_id->device ||
 363             cm_id->device->node_type != RDMA_NODE_IB_CA)
 364                 ret = -EADDRNOTAVAIL;
 365
 366         rdsdebug("addr %pI4 ret %d node type %d\n",
 367                 &addr, ret,
 368                 cm_id->device ? cm_id->device->node_type : -1);
 369
 370         rdma_destroy_id(cm_id);
 371
 372         return ret;
 373 }
 374
 375 static void rds_ib_unregister_client(void)
 376 {
 377         ib_unregister_client(&rds_ib_client);
 378         /* wait for rds_ib_dev_free() to complete */
 379         flush_workqueue(rds_wq);
 380 }
 381
 382 void rds_ib_exit(void)
 383 {
 384         rds_info_deregister_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info);
 385         rds_ib_unregister_client();
 386         rds_ib_destroy_nodev_conns();
 387         rds_ib_sysctl_exit();
 388         rds_ib_recv_exit();
 389         rds_trans_unregister(&rds_ib_transport);
 390         rds_ib_mr_exit();
 391 }
 392
 393 struct rds_transport rds_ib_transport = {
 394         .laddr_check            = rds_ib_laddr_check,
 395         .xmit_path_complete     = rds_ib_xmit_path_complete,
 396         .xmit                   = rds_ib_xmit,
 397         .xmit_rdma              = rds_ib_xmit_rdma,
 398         .xmit_atomic            = rds_ib_xmit_atomic,
 399         .recv_path              = rds_ib_recv_path,
 400         .conn_alloc             = rds_ib_conn_alloc,
 401         .conn_free              = rds_ib_conn_free,
 402         .conn_path_connect      = rds_ib_conn_path_connect,
 403         .conn_path_shutdown     = rds_ib_conn_path_shutdown,
 404         .inc_copy_to_user       = rds_ib_inc_copy_to_user,
 405         .inc_free               = rds_ib_inc_free,
 406         .cm_initiate_connect    = rds_ib_cm_initiate_connect,
 407         .cm_handle_connect      = rds_ib_cm_handle_connect,
 408         .cm_connect_complete    = rds_ib_cm_connect_complete,
 409         .stats_info_copy        = rds_ib_stats_info_copy,
 410         .exit                   = rds_ib_exit,
 411         .get_mr                 = rds_ib_get_mr,
 412         .sync_mr                = rds_ib_sync_mr,
 413         .free_mr                = rds_ib_free_mr,
 414         .flush_mrs              = rds_ib_flush_mrs,
 415         .t_owner                = THIS_MODULE,
 416         .t_name                 = "infiniband",
 417         .t_type                 = RDS_TRANS_IB
 418 };
 419
 420 int rds_ib_init(void)
 421 {
 422         int ret;
 423
 424         INIT_LIST_HEAD(&rds_ib_devices);
 425
 426         ret = rds_ib_mr_init();
 427         if (ret)
 428                 goto out;
 429
 430         ret = ib_register_client(&rds_ib_client);
 431         if (ret)
 432                 goto out_mr_exit;
 433
 434         ret = rds_ib_sysctl_init();
 435         if (ret)
 436                 goto out_ibreg;
 437
 438         ret = rds_ib_recv_init();
 439         if (ret)
 440                 goto out_sysctl;
 441
 442         ret = rds_trans_register(&rds_ib_transport);
 443         if (ret)
 444                 goto out_recv;
 445
 446         rds_info_register_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info);
 447
 448         goto out;
 449
 450 out_recv:
 451         rds_ib_recv_exit();
 452 out_sysctl:
 453         rds_ib_sysctl_exit();
 454 out_ibreg:
 455         rds_ib_unregister_client();
 456 out_mr_exit:
 457         rds_ib_mr_exit();
 458 out:
 459         return ret;
 460 }
 461
 462 MODULE_LICENSE("GPL");
 463