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