drivers/infiniband/ulp/ipoib/ipoib_main.c

   1 /*
   2  * Copyright (c) 2004 Topspin Communications.  All rights reserved.
   3  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
   4  * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
   5  *
   6  * This software is available to you under a choice of one of two
   7  * licenses.  You may choose to be licensed under the terms of the GNU
   8  * General Public License (GPL) Version 2, available from the file
   9  * COPYING in the main directory of this source tree, or the
  10  * OpenIB.org BSD license below:
  11  *
  12  *     Redistribution and use in source and binary forms, with or
  13  *     without modification, are permitted provided that the following
  14  *     conditions are met:
  15  *
  16  *      - Redistributions of source code must retain the above
  17  *        copyright notice, this list of conditions and the following
  18  *        disclaimer.
  19  *
  20  *      - Redistributions in binary form must reproduce the above
  21  *        copyright notice, this list of conditions and the following
  22  *        disclaimer in the documentation and/or other materials
  23  *        provided with the distribution.
  24  *
  25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  32  * SOFTWARE.
  33  */
  34
  35 #include "ipoib.h"
  36
  37 #include <linux/module.h>
  38
  39 #include <linux/init.h>
  40 #include <linux/slab.h>
  41 #include <linux/kernel.h>
  42 #include <linux/vmalloc.h>
  43
  44 #include <linux/if_arp.h>       /* For ARPHRD_xxx */
  45
  46 #include <linux/ip.h>
  47 #include <linux/in.h>
  48
  49 #include <linux/jhash.h>
  50 #include <net/arp.h>
  51 #include <net/addrconf.h>
  52 #include <linux/inetdevice.h>
  53 #include <rdma/ib_cache.h>
  54 #include <linux/pci.h>
  55
  56 #define DRV_VERSION "1.0.0"
  57
  58 const char ipoib_driver_version[] = DRV_VERSION;
  59
  60 MODULE_AUTHOR("Roland Dreier");
  61 MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
  62 MODULE_LICENSE("Dual BSD/GPL");
  63 MODULE_VERSION(DRV_VERSION);
  64
  65 int ipoib_sendq_size __read_mostly = IPOIB_TX_RING_SIZE;
  66 int ipoib_recvq_size __read_mostly = IPOIB_RX_RING_SIZE;
  67
  68 module_param_named(send_queue_size, ipoib_sendq_size, int, 0444);
  69 MODULE_PARM_DESC(send_queue_size, "Number of descriptors in send queue");
  70 module_param_named(recv_queue_size, ipoib_recvq_size, int, 0444);
  71 MODULE_PARM_DESC(recv_queue_size, "Number of descriptors in receive queue");
  72
  73 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
  74 int ipoib_debug_level;
  75
  76 module_param_named(debug_level, ipoib_debug_level, int, 0644);
  77 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
  78 #endif
  79
  80 struct ipoib_path_iter {
  81         struct net_device *dev;
  82         struct ipoib_path  path;
  83 };
  84
  85 static const u8 ipv4_bcast_addr[] = {
  86         0x00, 0xff, 0xff, 0xff,
  87         0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
  88         0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff
  89 };
  90
  91 struct workqueue_struct *ipoib_workqueue;
  92
  93 struct ib_sa_client ipoib_sa_client;
  94
  95 static void ipoib_add_one(struct ib_device *device);
  96 static void ipoib_remove_one(struct ib_device *device, void *client_data);
  97 static void ipoib_neigh_reclaim(struct rcu_head *rp);
  98 static struct net_device *ipoib_get_net_dev_by_params(
  99                 struct ib_device *dev, u8 port, u16 pkey,
 100                 const union ib_gid *gid, const struct sockaddr *addr,
 101                 void *client_data);
 102
 103 static struct ib_client ipoib_client = {
 104         .name   = "ipoib",
 105         .add    = ipoib_add_one,
 106         .remove = ipoib_remove_one,
 107         .get_net_dev_by_params = ipoib_get_net_dev_by_params,
 108 };
 109
 110 int ipoib_open(struct net_device *dev)
 111 {
 112         struct ipoib_dev_priv *priv = netdev_priv(dev);
 113
 114         ipoib_dbg(priv, "bringing up interface\n");
 115
 116         netif_carrier_off(dev);
 117
 118         set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 119
 120         if (ipoib_ib_dev_open(dev)) {
 121                 if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags))
 122                         return 0;
 123                 goto err_disable;
 124         }
 125
 126         if (ipoib_ib_dev_up(dev))
 127                 goto err_stop;
 128
 129         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 130                 struct ipoib_dev_priv *cpriv;
 131
 132                 /* Bring up any child interfaces too */
 133                 down_read(&priv->vlan_rwsem);
 134                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 135                         int flags;
 136
 137                         flags = cpriv->dev->flags;
 138                         if (flags & IFF_UP)
 139                                 continue;
 140
 141                         dev_change_flags(cpriv->dev, flags | IFF_UP);
 142                 }
 143                 up_read(&priv->vlan_rwsem);
 144         }
 145
 146         netif_start_queue(dev);
 147
 148         return 0;
 149
 150 err_stop:
 151         ipoib_ib_dev_stop(dev);
 152
 153 err_disable:
 154         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 155
 156         return -EINVAL;
 157 }
 158
 159 static int ipoib_stop(struct net_device *dev)
 160 {
 161         struct ipoib_dev_priv *priv = netdev_priv(dev);
 162
 163         ipoib_dbg(priv, "stopping interface\n");
 164
 165         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 166
 167         netif_stop_queue(dev);
 168
 169         ipoib_ib_dev_down(dev);
 170         ipoib_ib_dev_stop(dev);
 171
 172         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 173                 struct ipoib_dev_priv *cpriv;
 174
 175                 /* Bring down any child interfaces too */
 176                 down_read(&priv->vlan_rwsem);
 177                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 178                         int flags;
 179
 180                         flags = cpriv->dev->flags;
 181                         if (!(flags & IFF_UP))
 182                                 continue;
 183
 184                         dev_change_flags(cpriv->dev, flags & ~IFF_UP);
 185                 }
 186                 up_read(&priv->vlan_rwsem);
 187         }
 188
 189         return 0;
 190 }
 191
 192 static void ipoib_uninit(struct net_device *dev)
 193 {
 194         ipoib_dev_cleanup(dev);
 195 }
 196
 197 static netdev_features_t ipoib_fix_features(struct net_device *dev, netdev_features_t features)
 198 {
 199         struct ipoib_dev_priv *priv = netdev_priv(dev);
 200
 201         if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
 202                 features &= ~(NETIF_F_IP_CSUM | NETIF_F_TSO);
 203
 204         return features;
 205 }
 206
 207 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
 208 {
 209         struct ipoib_dev_priv *priv = netdev_priv(dev);
 210
 211         /* dev->mtu > 2K ==> connected mode */
 212         if (ipoib_cm_admin_enabled(dev)) {
 213                 if (new_mtu > ipoib_cm_max_mtu(dev))
 214                         return -EINVAL;
 215
 216                 if (new_mtu > priv->mcast_mtu)
 217                         ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
 218                                    priv->mcast_mtu);
 219
 220                 dev->mtu = new_mtu;
 221                 return 0;
 222         }
 223
 224         if (new_mtu > IPOIB_UD_MTU(priv->max_ib_mtu))
 225                 return -EINVAL;
 226
 227         priv->admin_mtu = new_mtu;
 228
 229         dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
 230
 231         return 0;
 232 }
 233
 234 /* Called with an RCU read lock taken */
 235 static bool ipoib_is_dev_match_addr_rcu(const struct sockaddr *addr,
 236                                         struct net_device *dev)
 237 {
 238         struct net *net = dev_net(dev);
 239         struct in_device *in_dev;
 240         struct sockaddr_in *addr_in = (struct sockaddr_in *)addr;
 241         struct sockaddr_in6 *addr_in6 = (struct sockaddr_in6 *)addr;
 242         __be32 ret_addr;
 243
 244         switch (addr->sa_family) {
 245         case AF_INET:
 246                 in_dev = in_dev_get(dev);
 247                 if (!in_dev)
 248                         return false;
 249
 250                 ret_addr = inet_confirm_addr(net, in_dev, 0,
 251                                              addr_in->sin_addr.s_addr,
 252                                              RT_SCOPE_HOST);
 253                 in_dev_put(in_dev);
 254                 if (ret_addr)
 255                         return true;
 256
 257                 break;
 258         case AF_INET6:
 259                 if (IS_ENABLED(CONFIG_IPV6) &&
 260                     ipv6_chk_addr(net, &addr_in6->sin6_addr, dev, 1))
 261                         return true;
 262
 263                 break;
 264         }
 265         return false;
 266 }
 267
 268 /**
 269  * Find the master net_device on top of the given net_device.
 270  * @dev: base IPoIB net_device
 271  *
 272  * Returns the master net_device with a reference held, or the same net_device
 273  * if no master exists.
 274  */
 275 static struct net_device *ipoib_get_master_net_dev(struct net_device *dev)
 276 {
 277         struct net_device *master;
 278
 279         rcu_read_lock();
 280         master = netdev_master_upper_dev_get_rcu(dev);
 281         if (master)
 282                 dev_hold(master);
 283         rcu_read_unlock();
 284
 285         if (master)
 286                 return master;
 287
 288         dev_hold(dev);
 289         return dev;
 290 }
 291
 292 /**
 293  * Find a net_device matching the given address, which is an upper device of
 294  * the given net_device.
 295  * @addr: IP address to look for.
 296  * @dev: base IPoIB net_device
 297  *
 298  * If found, returns the net_device with a reference held. Otherwise return
 299  * NULL.
 300  */
 301 static struct net_device *ipoib_get_net_dev_match_addr(
 302                 const struct sockaddr *addr, struct net_device *dev)
 303 {
 304         struct net_device *upper,
 305                           *result = NULL;
 306         struct list_head *iter;
 307
 308         rcu_read_lock();
 309         if (ipoib_is_dev_match_addr_rcu(addr, dev)) {
 310                 dev_hold(dev);
 311                 result = dev;
 312                 goto out;
 313         }
 314
 315         netdev_for_each_all_upper_dev_rcu(dev, upper, iter) {
 316                 if (ipoib_is_dev_match_addr_rcu(addr, upper)) {
 317                         dev_hold(upper);
 318                         result = upper;
 319                         break;
 320                 }
 321         }
 322 out:
 323         rcu_read_unlock();
 324         return result;
 325 }
 326
 327 /* returns the number of IPoIB netdevs on top a given ipoib device matching a
 328  * pkey_index and address, if one exists.
 329  *
 330  * @found_net_dev: contains a matching net_device if the return value >= 1,
 331  * with a reference held. */
 332 static int ipoib_match_gid_pkey_addr(struct ipoib_dev_priv *priv,
 333                                      const union ib_gid *gid,
 334                                      u16 pkey_index,
 335                                      const struct sockaddr *addr,
 336                                      int nesting,
 337                                      struct net_device **found_net_dev)
 338 {
 339         struct ipoib_dev_priv *child_priv;
 340         struct net_device *net_dev = NULL;
 341         int matches = 0;
 342
 343         if (priv->pkey_index == pkey_index &&
 344             (!gid || !memcmp(gid, &priv->local_gid, sizeof(*gid)))) {
 345                 if (!addr) {
 346                         net_dev = ipoib_get_master_net_dev(priv->dev);
 347                 } else {
 348                         /* Verify the net_device matches the IP address, as
 349                          * IPoIB child devices currently share a GID. */
 350                         net_dev = ipoib_get_net_dev_match_addr(addr, priv->dev);
 351                 }
 352                 if (net_dev) {
 353                         if (!*found_net_dev)
 354                                 *found_net_dev = net_dev;
 355                         else
 356                                 dev_put(net_dev);
 357                         ++matches;
 358                 }
 359         }
 360
 361         /* Check child interfaces */
 362         down_read_nested(&priv->vlan_rwsem, nesting);
 363         list_for_each_entry(child_priv, &priv->child_intfs, list) {
 364                 matches += ipoib_match_gid_pkey_addr(child_priv, gid,
 365                                                     pkey_index, addr,
 366                                                     nesting + 1,
 367                                                     found_net_dev);
 368                 if (matches > 1)
 369                         break;
 370         }
 371         up_read(&priv->vlan_rwsem);
 372
 373         return matches;
 374 }
 375
 376 /* Returns the number of matching net_devs found (between 0 and 2). Also
 377  * return the matching net_device in the @net_dev parameter, holding a
 378  * reference to the net_device, if the number of matches >= 1 */
 379 static int __ipoib_get_net_dev_by_params(struct list_head *dev_list, u8 port,
 380                                          u16 pkey_index,
 381                                          const union ib_gid *gid,
 382                                          const struct sockaddr *addr,
 383                                          struct net_device **net_dev)
 384 {
 385         struct ipoib_dev_priv *priv;
 386         int matches = 0;
 387
 388         *net_dev = NULL;
 389
 390         list_for_each_entry(priv, dev_list, list) {
 391                 if (priv->port != port)
 392                         continue;
 393
 394                 matches += ipoib_match_gid_pkey_addr(priv, gid, pkey_index,
 395                                                      addr, 0, net_dev);
 396                 if (matches > 1)
 397                         break;
 398         }
 399
 400         return matches;
 401 }
 402
 403 static struct net_device *ipoib_get_net_dev_by_params(
 404                 struct ib_device *dev, u8 port, u16 pkey,
 405                 const union ib_gid *gid, const struct sockaddr *addr,
 406                 void *client_data)
 407 {
 408         struct net_device *net_dev;
 409         struct list_head *dev_list = client_data;
 410         u16 pkey_index;
 411         int matches;
 412         int ret;
 413
 414         if (!rdma_protocol_ib(dev, port))
 415                 return NULL;
 416
 417         ret = ib_find_cached_pkey(dev, port, pkey, &pkey_index);
 418         if (ret)
 419                 return NULL;
 420
 421         if (!dev_list)
 422                 return NULL;
 423
 424         /* See if we can find a unique device matching the L2 parameters */
 425         matches = __ipoib_get_net_dev_by_params(dev_list, port, pkey_index,
 426                                                 gid, NULL, &net_dev);
 427
 428         switch (matches) {
 429         case 0:
 430                 return NULL;
 431         case 1:
 432                 return net_dev;
 433         }
 434
 435         dev_put(net_dev);
 436
 437         /* Couldn't find a unique device with L2 parameters only. Use L3
 438          * address to uniquely match the net device */
 439         matches = __ipoib_get_net_dev_by_params(dev_list, port, pkey_index,
 440                                                 gid, addr, &net_dev);
 441         switch (matches) {
 442         case 0:
 443                 return NULL;
 444         default:
 445                 dev_warn_ratelimited(&dev->dev,
 446                                      "duplicate IP address detected\n");
 447                 /* Fall through */
 448         case 1:
 449                 return net_dev;
 450         }
 451 }
 452
 453 int ipoib_set_mode(struct net_device *dev, const char *buf)
 454 {
 455         struct ipoib_dev_priv *priv = netdev_priv(dev);
 456
 457         /* flush paths if we switch modes so that connections are restarted */
 458         if (IPOIB_CM_SUPPORTED(dev->dev_addr) && !strcmp(buf, "connected\n")) {
 459                 set_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
 460                 ipoib_warn(priv, "enabling connected mode "
 461                            "will cause multicast packet drops\n");
 462                 netdev_update_features(dev);
 463                 dev_set_mtu(dev, ipoib_cm_max_mtu(dev));
 464                 rtnl_unlock();
 465                 priv->tx_wr.wr.send_flags &= ~IB_SEND_IP_CSUM;
 466
 467                 ipoib_flush_paths(dev);
 468                 rtnl_lock();
 469                 return 0;
 470         }
 471
 472         if (!strcmp(buf, "datagram\n")) {
 473                 clear_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
 474                 netdev_update_features(dev);
 475                 dev_set_mtu(dev, min(priv->mcast_mtu, dev->mtu));
 476                 rtnl_unlock();
 477                 ipoib_flush_paths(dev);
 478                 rtnl_lock();
 479                 return 0;
 480         }
 481
 482         return -EINVAL;
 483 }
 484
 485 static struct ipoib_path *__path_find(struct net_device *dev, void *gid)
 486 {
 487         struct ipoib_dev_priv *priv = netdev_priv(dev);
 488         struct rb_node *n = priv->path_tree.rb_node;
 489         struct ipoib_path *path;
 490         int ret;
 491
 492         while (n) {
 493                 path = rb_entry(n, struct ipoib_path, rb_node);
 494
 495                 ret = memcmp(gid, path->pathrec.dgid.raw,
 496                              sizeof (union ib_gid));
 497
 498                 if (ret < 0)
 499                         n = n->rb_left;
 500                 else if (ret > 0)
 501                         n = n->rb_right;
 502                 else
 503                         return path;
 504         }
 505
 506         return NULL;
 507 }
 508
 509 static int __path_add(struct net_device *dev, struct ipoib_path *path)
 510 {
 511         struct ipoib_dev_priv *priv = netdev_priv(dev);
 512         struct rb_node **n = &priv->path_tree.rb_node;
 513         struct rb_node *pn = NULL;
 514         struct ipoib_path *tpath;
 515         int ret;
 516
 517         while (*n) {
 518                 pn = *n;
 519                 tpath = rb_entry(pn, struct ipoib_path, rb_node);
 520
 521                 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
 522                              sizeof (union ib_gid));
 523                 if (ret < 0)
 524                         n = &pn->rb_left;
 525                 else if (ret > 0)
 526                         n = &pn->rb_right;
 527                 else
 528                         return -EEXIST;
 529         }
 530
 531         rb_link_node(&path->rb_node, pn, n);
 532         rb_insert_color(&path->rb_node, &priv->path_tree);
 533
 534         list_add_tail(&path->list, &priv->path_list);
 535
 536         return 0;
 537 }
 538
 539 static void path_free(struct net_device *dev, struct ipoib_path *path)
 540 {
 541         struct sk_buff *skb;
 542
 543         while ((skb = __skb_dequeue(&path->queue)))
 544                 dev_kfree_skb_irq(skb);
 545
 546         ipoib_dbg(netdev_priv(dev), "path_free\n");
 547
 548         /* remove all neigh connected to this path */
 549         ipoib_del_neighs_by_gid(dev, path->pathrec.dgid.raw);
 550
 551         if (path->ah)
 552                 ipoib_put_ah(path->ah);
 553
 554         kfree(path);
 555 }
 556
 557 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 558
 559 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
 560 {
 561         struct ipoib_path_iter *iter;
 562
 563         iter = kmalloc(sizeof *iter, GFP_KERNEL);
 564         if (!iter)
 565                 return NULL;
 566
 567         iter->dev = dev;
 568         memset(iter->path.pathrec.dgid.raw, 0, 16);
 569
 570         if (ipoib_path_iter_next(iter)) {
 571                 kfree(iter);
 572                 return NULL;
 573         }
 574
 575         return iter;
 576 }
 577
 578 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
 579 {
 580         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
 581         struct rb_node *n;
 582         struct ipoib_path *path;
 583         int ret = 1;
 584
 585         spin_lock_irq(&priv->lock);
 586
 587         n = rb_first(&priv->path_tree);
 588
 589         while (n) {
 590                 path = rb_entry(n, struct ipoib_path, rb_node);
 591
 592                 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
 593                            sizeof (union ib_gid)) < 0) {
 594                         iter->path = *path;
 595                         ret = 0;
 596                         break;
 597                 }
 598
 599                 n = rb_next(n);
 600         }
 601
 602         spin_unlock_irq(&priv->lock);
 603
 604         return ret;
 605 }
 606
 607 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
 608                           struct ipoib_path *path)
 609 {
 610         *path = iter->path;
 611 }
 612
 613 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
 614
 615 void ipoib_mark_paths_invalid(struct net_device *dev)
 616 {
 617         struct ipoib_dev_priv *priv = netdev_priv(dev);
 618         struct ipoib_path *path, *tp;
 619
 620         spin_lock_irq(&priv->lock);
 621
 622         list_for_each_entry_safe(path, tp, &priv->path_list, list) {
 623                 ipoib_dbg(priv, "mark path LID 0x%04x GID %pI6 invalid\n",
 624                         be16_to_cpu(path->pathrec.dlid),
 625                         path->pathrec.dgid.raw);
 626                 path->valid =  0;
 627         }
 628
 629         spin_unlock_irq(&priv->lock);
 630 }
 631
 632 void ipoib_flush_paths(struct net_device *dev)
 633 {
 634         struct ipoib_dev_priv *priv = netdev_priv(dev);
 635         struct ipoib_path *path, *tp;
 636         LIST_HEAD(remove_list);
 637         unsigned long flags;
 638
 639         netif_tx_lock_bh(dev);
 640         spin_lock_irqsave(&priv->lock, flags);
 641
 642         list_splice_init(&priv->path_list, &remove_list);
 643
 644         list_for_each_entry(path, &remove_list, list)
 645                 rb_erase(&path->rb_node, &priv->path_tree);
 646
 647         list_for_each_entry_safe(path, tp, &remove_list, list) {
 648                 if (path->query)
 649                         ib_sa_cancel_query(path->query_id, path->query);
 650                 spin_unlock_irqrestore(&priv->lock, flags);
 651                 netif_tx_unlock_bh(dev);
 652                 wait_for_completion(&path->done);
 653                 path_free(dev, path);
 654                 netif_tx_lock_bh(dev);
 655                 spin_lock_irqsave(&priv->lock, flags);
 656         }
 657
 658         spin_unlock_irqrestore(&priv->lock, flags);
 659         netif_tx_unlock_bh(dev);
 660 }
 661
 662 static void path_rec_completion(int status,
 663                                 struct ib_sa_path_rec *pathrec,
 664                                 void *path_ptr)
 665 {
 666         struct ipoib_path *path = path_ptr;
 667         struct net_device *dev = path->dev;
 668         struct ipoib_dev_priv *priv = netdev_priv(dev);
 669         struct ipoib_ah *ah = NULL;
 670         struct ipoib_ah *old_ah = NULL;
 671         struct ipoib_neigh *neigh, *tn;
 672         struct sk_buff_head skqueue;
 673         struct sk_buff *skb;
 674         unsigned long flags;
 675
 676         if (!status)
 677                 ipoib_dbg(priv, "PathRec LID 0x%04x for GID %pI6\n",
 678                           be16_to_cpu(pathrec->dlid), pathrec->dgid.raw);
 679         else
 680                 ipoib_dbg(priv, "PathRec status %d for GID %pI6\n",
 681                           status, path->pathrec.dgid.raw);
 682
 683         skb_queue_head_init(&skqueue);
 684
 685         if (!status) {
 686                 struct ib_ah_attr av;
 687
 688                 if (!ib_init_ah_from_path(priv->ca, priv->port, pathrec, &av))
 689                         ah = ipoib_create_ah(dev, priv->pd, &av);
 690         }
 691
 692         spin_lock_irqsave(&priv->lock, flags);
 693
 694         if (!IS_ERR_OR_NULL(ah)) {
 695                 path->pathrec = *pathrec;
 696
 697                 old_ah   = path->ah;
 698                 path->ah = ah;
 699
 700                 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
 701                           ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
 702
 703                 while ((skb = __skb_dequeue(&path->queue)))
 704                         __skb_queue_tail(&skqueue, skb);
 705
 706                 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
 707                         if (neigh->ah) {
 708                                 WARN_ON(neigh->ah != old_ah);
 709                                 /*
 710                                  * Dropping the ah reference inside
 711                                  * priv->lock is safe here, because we
 712                                  * will hold one more reference from
 713                                  * the original value of path->ah (ie
 714                                  * old_ah).
 715                                  */
 716                                 ipoib_put_ah(neigh->ah);
 717                         }
 718                         kref_get(&path->ah->ref);
 719                         neigh->ah = path->ah;
 720
 721                         if (ipoib_cm_enabled(dev, neigh->daddr)) {
 722                                 if (!ipoib_cm_get(neigh))
 723                                         ipoib_cm_set(neigh, ipoib_cm_create_tx(dev,
 724                                                                                path,
 725                                                                                neigh));
 726                                 if (!ipoib_cm_get(neigh)) {
 727                                         ipoib_neigh_free(neigh);
 728                                         continue;
 729                                 }
 730                         }
 731
 732                         while ((skb = __skb_dequeue(&neigh->queue)))
 733                                 __skb_queue_tail(&skqueue, skb);
 734                 }
 735                 path->valid = 1;
 736         }
 737
 738         path->query = NULL;
 739         complete(&path->done);
 740
 741         spin_unlock_irqrestore(&priv->lock, flags);
 742
 743         if (IS_ERR_OR_NULL(ah))
 744                 ipoib_del_neighs_by_gid(dev, path->pathrec.dgid.raw);
 745
 746         if (old_ah)
 747                 ipoib_put_ah(old_ah);
 748
 749         while ((skb = __skb_dequeue(&skqueue))) {
 750                 skb->dev = dev;
 751                 if (dev_queue_xmit(skb))
 752                         ipoib_warn(priv, "dev_queue_xmit failed "
 753                                    "to requeue packet\n");
 754         }
 755 }
 756
 757 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
 758 {
 759         struct ipoib_dev_priv *priv = netdev_priv(dev);
 760         struct ipoib_path *path;
 761
 762         if (!priv->broadcast)
 763                 return NULL;
 764
 765         path = kzalloc(sizeof *path, GFP_ATOMIC);
 766         if (!path)
 767                 return NULL;
 768
 769         path->dev = dev;
 770
 771         skb_queue_head_init(&path->queue);
 772
 773         INIT_LIST_HEAD(&path->neigh_list);
 774
 775         memcpy(path->pathrec.dgid.raw, gid, sizeof (union ib_gid));
 776         path->pathrec.sgid          = priv->local_gid;
 777         path->pathrec.pkey          = cpu_to_be16(priv->pkey);
 778         path->pathrec.numb_path     = 1;
 779         path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
 780
 781         return path;
 782 }
 783
 784 static int path_rec_start(struct net_device *dev,
 785                           struct ipoib_path *path)
 786 {
 787         struct ipoib_dev_priv *priv = netdev_priv(dev);
 788
 789         ipoib_dbg(priv, "Start path record lookup for %pI6\n",
 790                   path->pathrec.dgid.raw);
 791
 792         init_completion(&path->done);
 793
 794         path->query_id =
 795                 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
 796                                    &path->pathrec,
 797                                    IB_SA_PATH_REC_DGID          |
 798                                    IB_SA_PATH_REC_SGID          |
 799                                    IB_SA_PATH_REC_NUMB_PATH     |
 800                                    IB_SA_PATH_REC_TRAFFIC_CLASS |
 801                                    IB_SA_PATH_REC_PKEY,
 802                                    1000, GFP_ATOMIC,
 803                                    path_rec_completion,
 804                                    path, &path->query);
 805         if (path->query_id < 0) {
 806                 ipoib_warn(priv, "ib_sa_path_rec_get failed: %d\n", path->query_id);
 807                 path->query = NULL;
 808                 complete(&path->done);
 809                 return path->query_id;
 810         }
 811
 812         return 0;
 813 }
 814
 815 static void neigh_add_path(struct sk_buff *skb, u8 *daddr,
 816                            struct net_device *dev)
 817 {
 818         struct ipoib_dev_priv *priv = netdev_priv(dev);
 819         struct ipoib_path *path;
 820         struct ipoib_neigh *neigh;
 821         unsigned long flags;
 822
 823         spin_lock_irqsave(&priv->lock, flags);
 824         neigh = ipoib_neigh_alloc(daddr, dev);
 825         if (!neigh) {
 826                 spin_unlock_irqrestore(&priv->lock, flags);
 827                 ++dev->stats.tx_dropped;
 828                 dev_kfree_skb_any(skb);
 829                 return;
 830         }
 831
 832         path = __path_find(dev, daddr + 4);
 833         if (!path) {
 834                 path = path_rec_create(dev, daddr + 4);
 835                 if (!path)
 836                         goto err_path;
 837
 838                 __path_add(dev, path);
 839         }
 840
 841         list_add_tail(&neigh->list, &path->neigh_list);
 842
 843         if (path->ah) {
 844                 kref_get(&path->ah->ref);
 845                 neigh->ah = path->ah;
 846
 847                 if (ipoib_cm_enabled(dev, neigh->daddr)) {
 848                         if (!ipoib_cm_get(neigh))
 849                                 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
 850                         if (!ipoib_cm_get(neigh)) {
 851                                 ipoib_neigh_free(neigh);
 852                                 goto err_drop;
 853                         }
 854                         if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE)
 855                                 __skb_queue_tail(&neigh->queue, skb);
 856                         else {
 857                                 ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
 858                                            skb_queue_len(&neigh->queue));
 859                                 goto err_drop;
 860                         }
 861                 } else {
 862                         spin_unlock_irqrestore(&priv->lock, flags);
 863                         ipoib_send(dev, skb, path->ah, IPOIB_QPN(daddr));
 864                         ipoib_neigh_put(neigh);
 865                         return;
 866                 }
 867         } else {
 868                 neigh->ah  = NULL;
 869
 870                 if (!path->query && path_rec_start(dev, path))
 871                         goto err_path;
 872                 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE)
 873                         __skb_queue_tail(&neigh->queue, skb);
 874                 else
 875                         goto err_drop;
 876         }
 877
 878         spin_unlock_irqrestore(&priv->lock, flags);
 879         ipoib_neigh_put(neigh);
 880         return;
 881
 882 err_path:
 883         ipoib_neigh_free(neigh);
 884 err_drop:
 885         ++dev->stats.tx_dropped;
 886         dev_kfree_skb_any(skb);
 887
 888         spin_unlock_irqrestore(&priv->lock, flags);
 889         ipoib_neigh_put(neigh);
 890 }
 891
 892 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
 893                              struct ipoib_cb *cb)
 894 {
 895         struct ipoib_dev_priv *priv = netdev_priv(dev);
 896         struct ipoib_path *path;
 897         unsigned long flags;
 898
 899         spin_lock_irqsave(&priv->lock, flags);
 900
 901         path = __path_find(dev, cb->hwaddr + 4);
 902         if (!path || !path->valid) {
 903                 int new_path = 0;
 904
 905                 if (!path) {
 906                         path = path_rec_create(dev, cb->hwaddr + 4);
 907                         new_path = 1;
 908                 }
 909                 if (path) {
 910                         if (skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 911                                 __skb_queue_tail(&path->queue, skb);
 912                         } else {
 913                                 ++dev->stats.tx_dropped;
 914                                 dev_kfree_skb_any(skb);
 915                         }
 916
 917                         if (!path->query && path_rec_start(dev, path)) {
 918                                 spin_unlock_irqrestore(&priv->lock, flags);
 919                                 if (new_path)
 920                                         path_free(dev, path);
 921                                 return;
 922                         } else
 923                                 __path_add(dev, path);
 924                 } else {
 925                         ++dev->stats.tx_dropped;
 926                         dev_kfree_skb_any(skb);
 927                 }
 928
 929                 spin_unlock_irqrestore(&priv->lock, flags);
 930                 return;
 931         }
 932
 933         if (path->ah) {
 934                 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
 935                           be16_to_cpu(path->pathrec.dlid));
 936
 937                 spin_unlock_irqrestore(&priv->lock, flags);
 938                 ipoib_send(dev, skb, path->ah, IPOIB_QPN(cb->hwaddr));
 939                 return;
 940         } else if ((path->query || !path_rec_start(dev, path)) &&
 941                    skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 942                 __skb_queue_tail(&path->queue, skb);
 943         } else {
 944                 ++dev->stats.tx_dropped;
 945                 dev_kfree_skb_any(skb);
 946         }
 947
 948         spin_unlock_irqrestore(&priv->lock, flags);
 949 }
 950
 951 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
 952 {
 953         struct ipoib_dev_priv *priv = netdev_priv(dev);
 954         struct ipoib_neigh *neigh;
 955         struct ipoib_cb *cb = ipoib_skb_cb(skb);
 956         struct ipoib_header *header;
 957         unsigned long flags;
 958
 959         header = (struct ipoib_header *) skb->data;
 960
 961         if (unlikely(cb->hwaddr[4] == 0xff)) {
 962                 /* multicast, arrange "if" according to probability */
 963                 if ((header->proto != htons(ETH_P_IP)) &&
 964                     (header->proto != htons(ETH_P_IPV6)) &&
 965                     (header->proto != htons(ETH_P_ARP)) &&
 966                     (header->proto != htons(ETH_P_RARP)) &&
 967                     (header->proto != htons(ETH_P_TIPC))) {
 968                         /* ethertype not supported by IPoIB */
 969                         ++dev->stats.tx_dropped;
 970                         dev_kfree_skb_any(skb);
 971                         return NETDEV_TX_OK;
 972                 }
 973                 /* Add in the P_Key for multicast*/
 974                 cb->hwaddr[8] = (priv->pkey >> 8) & 0xff;
 975                 cb->hwaddr[9] = priv->pkey & 0xff;
 976
 977                 neigh = ipoib_neigh_get(dev, cb->hwaddr);
 978                 if (likely(neigh))
 979                         goto send_using_neigh;
 980                 ipoib_mcast_send(dev, cb->hwaddr, skb);
 981                 return NETDEV_TX_OK;
 982         }
 983
 984         /* unicast, arrange "switch" according to probability */
 985         switch (header->proto) {
 986         case htons(ETH_P_IP):
 987         case htons(ETH_P_IPV6):
 988         case htons(ETH_P_TIPC):
 989                 neigh = ipoib_neigh_get(dev, cb->hwaddr);
 990                 if (unlikely(!neigh)) {
 991                         neigh_add_path(skb, cb->hwaddr, dev);
 992                         return NETDEV_TX_OK;
 993                 }
 994                 break;
 995         case htons(ETH_P_ARP):
 996         case htons(ETH_P_RARP):
 997                 /* for unicast ARP and RARP should always perform path find */
 998                 unicast_arp_send(skb, dev, cb);
 999                 return NETDEV_TX_OK;
1000         default:
1001                 /* ethertype not supported by IPoIB */
1002                 ++dev->stats.tx_dropped;
1003                 dev_kfree_skb_any(skb);
1004                 return NETDEV_TX_OK;
1005         }
1006
1007 send_using_neigh:
1008         /* note we now hold a ref to neigh */
1009         if (ipoib_cm_get(neigh)) {
1010                 if (ipoib_cm_up(neigh)) {
1011                         ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
1012                         goto unref;
1013                 }
1014         } else if (neigh->ah) {
1015                 ipoib_send(dev, skb, neigh->ah, IPOIB_QPN(cb->hwaddr));
1016                 goto unref;
1017         }
1018
1019         if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
1020                 spin_lock_irqsave(&priv->lock, flags);
1021                 __skb_queue_tail(&neigh->queue, skb);
1022                 spin_unlock_irqrestore(&priv->lock, flags);
1023         } else {
1024                 ++dev->stats.tx_dropped;
1025                 dev_kfree_skb_any(skb);
1026         }
1027
1028 unref:
1029         ipoib_neigh_put(neigh);
1030
1031         return NETDEV_TX_OK;
1032 }
1033
1034 static void ipoib_timeout(struct net_device *dev)
1035 {
1036         struct ipoib_dev_priv *priv = netdev_priv(dev);
1037
1038         ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
1039                    jiffies_to_msecs(jiffies - dev_trans_start(dev)));
1040         ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
1041                    netif_queue_stopped(dev),
1042                    priv->tx_head, priv->tx_tail);
1043         /* XXX reset QP, etc. */
1044 }
1045
1046 static int ipoib_hard_header(struct sk_buff *skb,
1047                              struct net_device *dev,
1048                              unsigned short type,
1049                              const void *daddr, const void *saddr, unsigned len)
1050 {
1051         struct ipoib_header *header;
1052         struct ipoib_cb *cb = ipoib_skb_cb(skb);
1053
1054         header = (struct ipoib_header *) skb_push(skb, sizeof *header);
1055
1056         header->proto = htons(type);
1057         header->reserved = 0;
1058
1059         /*
1060          * we don't rely on dst_entry structure,  always stuff the
1061          * destination address into skb->cb so we can figure out where
1062          * to send the packet later.
1063          */
1064         memcpy(cb->hwaddr, daddr, INFINIBAND_ALEN);
1065
1066         return sizeof *header;
1067 }
1068
1069 static void ipoib_set_mcast_list(struct net_device *dev)
1070 {
1071         struct ipoib_dev_priv *priv = netdev_priv(dev);
1072
1073         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
1074                 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
1075                 return;
1076         }
1077
1078         queue_work(priv->wq, &priv->restart_task);
1079 }
1080
1081 static int ipoib_get_iflink(const struct net_device *dev)
1082 {
1083         struct ipoib_dev_priv *priv = netdev_priv(dev);
1084
1085         /* parent interface */
1086         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags))
1087                 return dev->ifindex;
1088
1089         /* child/vlan interface */
1090         return priv->parent->ifindex;
1091 }
1092
1093 static u32 ipoib_addr_hash(struct ipoib_neigh_hash *htbl, u8 *daddr)
1094 {
1095         /*
1096          * Use only the address parts that contributes to spreading
1097          * The subnet prefix is not used as one can not connect to
1098          * same remote port (GUID) using the same remote QPN via two
1099          * different subnets.
1100          */
1101          /* qpn octets[1:4) & port GUID octets[12:20) */
1102         u32 *d32 = (u32 *) daddr;
1103         u32 hv;
1104
1105         hv = jhash_3words(d32[3], d32[4], IPOIB_QPN_MASK & d32[0], 0);
1106         return hv & htbl->mask;
1107 }
1108
1109 struct ipoib_neigh *ipoib_neigh_get(struct net_device *dev, u8 *daddr)
1110 {
1111         struct ipoib_dev_priv *priv = netdev_priv(dev);
1112         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1113         struct ipoib_neigh_hash *htbl;
1114         struct ipoib_neigh *neigh = NULL;
1115         u32 hash_val;
1116
1117         rcu_read_lock_bh();
1118
1119         htbl = rcu_dereference_bh(ntbl->htbl);
1120
1121         if (!htbl)
1122                 goto out_unlock;
1123
1124         hash_val = ipoib_addr_hash(htbl, daddr);
1125         for (neigh = rcu_dereference_bh(htbl->buckets[hash_val]);
1126              neigh != NULL;
1127              neigh = rcu_dereference_bh(neigh->hnext)) {
1128                 if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) {
1129                         /* found, take one ref on behalf of the caller */
1130                         if (!atomic_inc_not_zero(&neigh->refcnt)) {
1131                                 /* deleted */
1132                                 neigh = NULL;
1133                                 goto out_unlock;
1134                         }
1135                         neigh->alive = jiffies;
1136                         goto out_unlock;
1137                 }
1138         }
1139
1140 out_unlock:
1141         rcu_read_unlock_bh();
1142         return neigh;
1143 }
1144
1145 static void __ipoib_reap_neigh(struct ipoib_dev_priv *priv)
1146 {
1147         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1148         struct ipoib_neigh_hash *htbl;
1149         unsigned long neigh_obsolete;
1150         unsigned long dt;
1151         unsigned long flags;
1152         int i;
1153         LIST_HEAD(remove_list);
1154
1155         if (test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
1156                 return;
1157
1158         spin_lock_irqsave(&priv->lock, flags);
1159
1160         htbl = rcu_dereference_protected(ntbl->htbl,
1161                                          lockdep_is_held(&priv->lock));
1162
1163         if (!htbl)
1164                 goto out_unlock;
1165
1166         /* neigh is obsolete if it was idle for two GC periods */
1167         dt = 2 * arp_tbl.gc_interval;
1168         neigh_obsolete = jiffies - dt;
1169         /* handle possible race condition */
1170         if (test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
1171                 goto out_unlock;
1172
1173         for (i = 0; i < htbl->size; i++) {
1174                 struct ipoib_neigh *neigh;
1175                 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1176
1177                 while ((neigh = rcu_dereference_protected(*np,
1178                                                           lockdep_is_held(&priv->lock))) != NULL) {
1179                         /* was the neigh idle for two GC periods */
1180                         if (time_after(neigh_obsolete, neigh->alive)) {
1181
1182                                 ipoib_check_and_add_mcast_sendonly(priv, neigh->daddr + 4, &remove_list);
1183
1184                                 rcu_assign_pointer(*np,
1185                                                    rcu_dereference_protected(neigh->hnext,
1186                                                                              lockdep_is_held(&priv->lock)));
1187                                 /* remove from path/mc list */
1188                                 list_del(&neigh->list);
1189                                 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1190                         } else {
1191                                 np = &neigh->hnext;
1192                         }
1193
1194                 }
1195         }
1196
1197 out_unlock:
1198         spin_unlock_irqrestore(&priv->lock, flags);
1199         ipoib_mcast_remove_list(&remove_list);
1200 }
1201
1202 static void ipoib_reap_neigh(struct work_struct *work)
1203 {
1204         struct ipoib_dev_priv *priv =
1205                 container_of(work, struct ipoib_dev_priv, neigh_reap_task.work);
1206
1207         __ipoib_reap_neigh(priv);
1208
1209         if (!test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
1210                 queue_delayed_work(priv->wq, &priv->neigh_reap_task,
1211                                    arp_tbl.gc_interval);
1212 }
1213
1214
1215 static struct ipoib_neigh *ipoib_neigh_ctor(u8 *daddr,
1216                                       struct net_device *dev)
1217 {
1218         struct ipoib_neigh *neigh;
1219
1220         neigh = kzalloc(sizeof *neigh, GFP_ATOMIC);
1221         if (!neigh)
1222                 return NULL;
1223
1224         neigh->dev = dev;
1225         memcpy(&neigh->daddr, daddr, sizeof(neigh->daddr));
1226         skb_queue_head_init(&neigh->queue);
1227         INIT_LIST_HEAD(&neigh->list);
1228         ipoib_cm_set(neigh, NULL);
1229         /* one ref on behalf of the caller */
1230         atomic_set(&neigh->refcnt, 1);
1231
1232         return neigh;
1233 }
1234
1235 struct ipoib_neigh *ipoib_neigh_alloc(u8 *daddr,
1236                                       struct net_device *dev)
1237 {
1238         struct ipoib_dev_priv *priv = netdev_priv(dev);
1239         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1240         struct ipoib_neigh_hash *htbl;
1241         struct ipoib_neigh *neigh;
1242         u32 hash_val;
1243
1244         htbl = rcu_dereference_protected(ntbl->htbl,
1245                                          lockdep_is_held(&priv->lock));
1246         if (!htbl) {
1247                 neigh = NULL;
1248                 goto out_unlock;
1249         }
1250
1251         /* need to add a new neigh, but maybe some other thread succeeded?
1252          * recalc hash, maybe hash resize took place so we do a search
1253          */
1254         hash_val = ipoib_addr_hash(htbl, daddr);
1255         for (neigh = rcu_dereference_protected(htbl->buckets[hash_val],
1256                                                lockdep_is_held(&priv->lock));
1257              neigh != NULL;
1258              neigh = rcu_dereference_protected(neigh->hnext,
1259                                                lockdep_is_held(&priv->lock))) {
1260                 if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) {
1261                         /* found, take one ref on behalf of the caller */
1262                         if (!atomic_inc_not_zero(&neigh->refcnt)) {
1263                                 /* deleted */
1264                                 neigh = NULL;
1265                                 break;
1266                         }
1267                         neigh->alive = jiffies;
1268                         goto out_unlock;
1269                 }
1270         }
1271
1272         neigh = ipoib_neigh_ctor(daddr, dev);
1273         if (!neigh)
1274                 goto out_unlock;
1275
1276         /* one ref on behalf of the hash table */
1277         atomic_inc(&neigh->refcnt);
1278         neigh->alive = jiffies;
1279         /* put in hash */
1280         rcu_assign_pointer(neigh->hnext,
1281                            rcu_dereference_protected(htbl->buckets[hash_val],
1282                                                      lockdep_is_held(&priv->lock)));
1283         rcu_assign_pointer(htbl->buckets[hash_val], neigh);
1284         atomic_inc(&ntbl->entries);
1285
1286 out_unlock:
1287
1288         return neigh;
1289 }
1290
1291 void ipoib_neigh_dtor(struct ipoib_neigh *neigh)
1292 {
1293         /* neigh reference count was dropprd to zero */
1294         struct net_device *dev = neigh->dev;
1295         struct ipoib_dev_priv *priv = netdev_priv(dev);
1296         struct sk_buff *skb;
1297         if (neigh->ah)
1298                 ipoib_put_ah(neigh->ah);
1299         while ((skb = __skb_dequeue(&neigh->queue))) {
1300                 ++dev->stats.tx_dropped;
1301                 dev_kfree_skb_any(skb);
1302         }
1303         if (ipoib_cm_get(neigh))
1304                 ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
1305         ipoib_dbg(netdev_priv(dev),
1306                   "neigh free for %06x %pI6\n",
1307                   IPOIB_QPN(neigh->daddr),
1308                   neigh->daddr + 4);
1309         kfree(neigh);
1310         if (atomic_dec_and_test(&priv->ntbl.entries)) {
1311                 if (test_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags))
1312                         complete(&priv->ntbl.flushed);
1313         }
1314 }
1315
1316 static void ipoib_neigh_reclaim(struct rcu_head *rp)
1317 {
1318         /* Called as a result of removal from hash table */
1319         struct ipoib_neigh *neigh = container_of(rp, struct ipoib_neigh, rcu);
1320         /* note TX context may hold another ref */
1321         ipoib_neigh_put(neigh);
1322 }
1323
1324 void ipoib_neigh_free(struct ipoib_neigh *neigh)
1325 {
1326         struct net_device *dev = neigh->dev;
1327         struct ipoib_dev_priv *priv = netdev_priv(dev);
1328         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1329         struct ipoib_neigh_hash *htbl;
1330         struct ipoib_neigh __rcu **np;
1331         struct ipoib_neigh *n;
1332         u32 hash_val;
1333
1334         htbl = rcu_dereference_protected(ntbl->htbl,
1335                                         lockdep_is_held(&priv->lock));
1336         if (!htbl)
1337                 return;
1338
1339         hash_val = ipoib_addr_hash(htbl, neigh->daddr);
1340         np = &htbl->buckets[hash_val];
1341         for (n = rcu_dereference_protected(*np,
1342                                             lockdep_is_held(&priv->lock));
1343              n != NULL;
1344              n = rcu_dereference_protected(*np,
1345                                         lockdep_is_held(&priv->lock))) {
1346                 if (n == neigh) {
1347                         /* found */
1348                         rcu_assign_pointer(*np,
1349                                            rcu_dereference_protected(neigh->hnext,
1350                                                                      lockdep_is_held(&priv->lock)));
1351                         /* remove from parent list */
1352                         list_del(&neigh->list);
1353                         call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1354                         return;
1355                 } else {
1356                         np = &n->hnext;
1357                 }
1358         }
1359 }
1360
1361 static int ipoib_neigh_hash_init(struct ipoib_dev_priv *priv)
1362 {
1363         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1364         struct ipoib_neigh_hash *htbl;
1365         struct ipoib_neigh __rcu **buckets;
1366         u32 size;
1367
1368         clear_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags);
1369         ntbl->htbl = NULL;
1370         htbl = kzalloc(sizeof(*htbl), GFP_KERNEL);
1371         if (!htbl)
1372                 return -ENOMEM;
1373         set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1374         size = roundup_pow_of_two(arp_tbl.gc_thresh3);
1375         buckets = kzalloc(size * sizeof(*buckets), GFP_KERNEL);
1376         if (!buckets) {
1377                 kfree(htbl);
1378                 return -ENOMEM;
1379         }
1380         htbl->size = size;
1381         htbl->mask = (size - 1);
1382         htbl->buckets = buckets;
1383         RCU_INIT_POINTER(ntbl->htbl, htbl);
1384         htbl->ntbl = ntbl;
1385         atomic_set(&ntbl->entries, 0);
1386
1387         /* start garbage collection */
1388         clear_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1389         queue_delayed_work(priv->wq, &priv->neigh_reap_task,
1390                            arp_tbl.gc_interval);
1391
1392         return 0;
1393 }
1394
1395 static void neigh_hash_free_rcu(struct rcu_head *head)
1396 {
1397         struct ipoib_neigh_hash *htbl = container_of(head,
1398                                                     struct ipoib_neigh_hash,
1399                                                     rcu);
1400         struct ipoib_neigh __rcu **buckets = htbl->buckets;
1401         struct ipoib_neigh_table *ntbl = htbl->ntbl;
1402
1403         kfree(buckets);
1404         kfree(htbl);
1405         complete(&ntbl->deleted);
1406 }
1407
1408 void ipoib_del_neighs_by_gid(struct net_device *dev, u8 *gid)
1409 {
1410         struct ipoib_dev_priv *priv = netdev_priv(dev);
1411         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1412         struct ipoib_neigh_hash *htbl;
1413         unsigned long flags;
1414         int i;
1415
1416         /* remove all neigh connected to a given path or mcast */
1417         spin_lock_irqsave(&priv->lock, flags);
1418
1419         htbl = rcu_dereference_protected(ntbl->htbl,
1420                                          lockdep_is_held(&priv->lock));
1421
1422         if (!htbl)
1423                 goto out_unlock;
1424
1425         for (i = 0; i < htbl->size; i++) {
1426                 struct ipoib_neigh *neigh;
1427                 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1428
1429                 while ((neigh = rcu_dereference_protected(*np,
1430                                                           lockdep_is_held(&priv->lock))) != NULL) {
1431                         /* delete neighs belong to this parent */
1432                         if (!memcmp(gid, neigh->daddr + 4, sizeof (union ib_gid))) {
1433                                 rcu_assign_pointer(*np,
1434                                                    rcu_dereference_protected(neigh->hnext,
1435                                                                              lockdep_is_held(&priv->lock)));
1436                                 /* remove from parent list */
1437                                 list_del(&neigh->list);
1438                                 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1439                         } else {
1440                                 np = &neigh->hnext;
1441                         }
1442
1443                 }
1444         }
1445 out_unlock:
1446         spin_unlock_irqrestore(&priv->lock, flags);
1447 }
1448
1449 static void ipoib_flush_neighs(struct ipoib_dev_priv *priv)
1450 {
1451         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1452         struct ipoib_neigh_hash *htbl;
1453         unsigned long flags;
1454         int i, wait_flushed = 0;
1455
1456         init_completion(&priv->ntbl.flushed);
1457
1458         spin_lock_irqsave(&priv->lock, flags);
1459
1460         htbl = rcu_dereference_protected(ntbl->htbl,
1461                                         lockdep_is_held(&priv->lock));
1462         if (!htbl)
1463                 goto out_unlock;
1464
1465         wait_flushed = atomic_read(&priv->ntbl.entries);
1466         if (!wait_flushed)
1467                 goto free_htbl;
1468
1469         for (i = 0; i < htbl->size; i++) {
1470                 struct ipoib_neigh *neigh;
1471                 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1472
1473                 while ((neigh = rcu_dereference_protected(*np,
1474                                        lockdep_is_held(&priv->lock))) != NULL) {
1475                         rcu_assign_pointer(*np,
1476                                            rcu_dereference_protected(neigh->hnext,
1477                                                                      lockdep_is_held(&priv->lock)));
1478                         /* remove from path/mc list */
1479                         list_del(&neigh->list);
1480                         call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1481                 }
1482         }
1483
1484 free_htbl:
1485         rcu_assign_pointer(ntbl->htbl, NULL);
1486         call_rcu(&htbl->rcu, neigh_hash_free_rcu);
1487
1488 out_unlock:
1489         spin_unlock_irqrestore(&priv->lock, flags);
1490         if (wait_flushed)
1491                 wait_for_completion(&priv->ntbl.flushed);
1492 }
1493
1494 static void ipoib_neigh_hash_uninit(struct net_device *dev)
1495 {
1496         struct ipoib_dev_priv *priv = netdev_priv(dev);
1497         int stopped;
1498
1499         ipoib_dbg(priv, "ipoib_neigh_hash_uninit\n");
1500         init_completion(&priv->ntbl.deleted);
1501         set_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags);
1502
1503         /* Stop GC if called at init fail need to cancel work */
1504         stopped = test_and_set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1505         if (!stopped)
1506                 cancel_delayed_work(&priv->neigh_reap_task);
1507
1508         ipoib_flush_neighs(priv);
1509
1510         wait_for_completion(&priv->ntbl.deleted);
1511 }
1512
1513
1514 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
1515 {
1516         struct ipoib_dev_priv *priv = netdev_priv(dev);
1517
1518         /* Allocate RX/TX "rings" to hold queued skbs */
1519         priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
1520                                 GFP_KERNEL);
1521         if (!priv->rx_ring) {
1522                 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
1523                        ca->name, ipoib_recvq_size);
1524                 goto out;
1525         }
1526
1527         priv->tx_ring = vzalloc(ipoib_sendq_size * sizeof *priv->tx_ring);
1528         if (!priv->tx_ring) {
1529                 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
1530                        ca->name, ipoib_sendq_size);
1531                 goto out_rx_ring_cleanup;
1532         }
1533
1534         /* priv->tx_head, tx_tail & tx_outstanding are already 0 */
1535
1536         if (ipoib_ib_dev_init(dev, ca, port))
1537                 goto out_tx_ring_cleanup;
1538
1539         /*
1540          * Must be after ipoib_ib_dev_init so we can allocate a per
1541          * device wq there and use it here
1542          */
1543         if (ipoib_neigh_hash_init(priv) < 0)
1544                 goto out_dev_uninit;
1545
1546         return 0;
1547
1548 out_dev_uninit:
1549         ipoib_ib_dev_cleanup(dev);
1550
1551 out_tx_ring_cleanup:
1552         vfree(priv->tx_ring);
1553
1554 out_rx_ring_cleanup:
1555         kfree(priv->rx_ring);
1556
1557 out:
1558         return -ENOMEM;
1559 }
1560
1561 void ipoib_dev_cleanup(struct net_device *dev)
1562 {
1563         struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
1564         LIST_HEAD(head);
1565
1566         ASSERT_RTNL();
1567
1568         ipoib_delete_debug_files(dev);
1569
1570         /* Delete any child interfaces first */
1571         list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
1572                 /* Stop GC on child */
1573                 set_bit(IPOIB_STOP_NEIGH_GC, &cpriv->flags);
1574                 cancel_delayed_work(&cpriv->neigh_reap_task);
1575                 unregister_netdevice_queue(cpriv->dev, &head);
1576         }
1577         unregister_netdevice_many(&head);
1578
1579         /*
1580          * Must be before ipoib_ib_dev_cleanup or we delete an in use
1581          * work queue
1582          */
1583         ipoib_neigh_hash_uninit(dev);
1584
1585         ipoib_ib_dev_cleanup(dev);
1586
1587         kfree(priv->rx_ring);
1588         vfree(priv->tx_ring);
1589
1590         priv->rx_ring = NULL;
1591         priv->tx_ring = NULL;
1592 }
1593
1594 static int ipoib_set_vf_link_state(struct net_device *dev, int vf, int link_state)
1595 {
1596         struct ipoib_dev_priv *priv = netdev_priv(dev);
1597
1598         return ib_set_vf_link_state(priv->ca, vf, priv->port, link_state);
1599 }
1600
1601 static int ipoib_get_vf_config(struct net_device *dev, int vf,
1602                                struct ifla_vf_info *ivf)
1603 {
1604         struct ipoib_dev_priv *priv = netdev_priv(dev);
1605         int err;
1606
1607         err = ib_get_vf_config(priv->ca, vf, priv->port, ivf);
1608         if (err)
1609                 return err;
1610
1611         ivf->vf = vf;
1612
1613         return 0;
1614 }
1615
1616 static int ipoib_set_vf_guid(struct net_device *dev, int vf, u64 guid, int type)
1617 {
1618         struct ipoib_dev_priv *priv = netdev_priv(dev);
1619
1620         if (type != IFLA_VF_IB_NODE_GUID && type != IFLA_VF_IB_PORT_GUID)
1621                 return -EINVAL;
1622
1623         return ib_set_vf_guid(priv->ca, vf, priv->port, guid, type);
1624 }
1625
1626 static int ipoib_get_vf_stats(struct net_device *dev, int vf,
1627                               struct ifla_vf_stats *vf_stats)
1628 {
1629         struct ipoib_dev_priv *priv = netdev_priv(dev);
1630
1631         return ib_get_vf_stats(priv->ca, vf, priv->port, vf_stats);
1632 }
1633
1634 static const struct header_ops ipoib_header_ops = {
1635         .create = ipoib_hard_header,
1636 };
1637
1638 static const struct net_device_ops ipoib_netdev_ops_pf = {
1639         .ndo_uninit              = ipoib_uninit,
1640         .ndo_open                = ipoib_open,
1641         .ndo_stop                = ipoib_stop,
1642         .ndo_change_mtu          = ipoib_change_mtu,
1643         .ndo_fix_features        = ipoib_fix_features,
1644         .ndo_start_xmit          = ipoib_start_xmit,
1645         .ndo_tx_timeout          = ipoib_timeout,
1646         .ndo_set_rx_mode         = ipoib_set_mcast_list,
1647         .ndo_get_iflink          = ipoib_get_iflink,
1648         .ndo_set_vf_link_state   = ipoib_set_vf_link_state,
1649         .ndo_get_vf_config       = ipoib_get_vf_config,
1650         .ndo_get_vf_stats        = ipoib_get_vf_stats,
1651         .ndo_set_vf_guid         = ipoib_set_vf_guid,
1652 };
1653
1654 static const struct net_device_ops ipoib_netdev_ops_vf = {
1655         .ndo_uninit              = ipoib_uninit,
1656         .ndo_open                = ipoib_open,
1657         .ndo_stop                = ipoib_stop,
1658         .ndo_change_mtu          = ipoib_change_mtu,
1659         .ndo_fix_features        = ipoib_fix_features,
1660         .ndo_start_xmit          = ipoib_start_xmit,
1661         .ndo_tx_timeout          = ipoib_timeout,
1662         .ndo_set_rx_mode         = ipoib_set_mcast_list,
1663         .ndo_get_iflink          = ipoib_get_iflink,
1664 };
1665
1666 void ipoib_setup(struct net_device *dev)
1667 {
1668         struct ipoib_dev_priv *priv = netdev_priv(dev);
1669
1670         if (priv->hca_caps & IB_DEVICE_VIRTUAL_FUNCTION)
1671                 dev->netdev_ops = &ipoib_netdev_ops_vf;
1672         else
1673                 dev->netdev_ops = &ipoib_netdev_ops_pf;
1674
1675         dev->header_ops          = &ipoib_header_ops;
1676
1677         ipoib_set_ethtool_ops(dev);
1678
1679         netif_napi_add(dev, &priv->napi, ipoib_poll, NAPI_POLL_WEIGHT);
1680
1681         dev->watchdog_timeo      = HZ;
1682
1683         dev->flags              |= IFF_BROADCAST | IFF_MULTICAST;
1684
1685         dev->hard_header_len     = IPOIB_ENCAP_LEN;
1686         dev->addr_len            = INFINIBAND_ALEN;
1687         dev->type                = ARPHRD_INFINIBAND;
1688         dev->tx_queue_len        = ipoib_sendq_size * 2;
1689         dev->features            = (NETIF_F_VLAN_CHALLENGED     |
1690                                     NETIF_F_HIGHDMA);
1691         netif_keep_dst(dev);
1692
1693         memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
1694
1695         priv->dev = dev;
1696
1697         spin_lock_init(&priv->lock);
1698
1699         init_rwsem(&priv->vlan_rwsem);
1700
1701         INIT_LIST_HEAD(&priv->path_list);
1702         INIT_LIST_HEAD(&priv->child_intfs);
1703         INIT_LIST_HEAD(&priv->dead_ahs);
1704         INIT_LIST_HEAD(&priv->multicast_list);
1705
1706         INIT_DELAYED_WORK(&priv->mcast_task,   ipoib_mcast_join_task);
1707         INIT_WORK(&priv->carrier_on_task, ipoib_mcast_carrier_on_task);
1708         INIT_WORK(&priv->flush_light,   ipoib_ib_dev_flush_light);
1709         INIT_WORK(&priv->flush_normal,   ipoib_ib_dev_flush_normal);
1710         INIT_WORK(&priv->flush_heavy,   ipoib_ib_dev_flush_heavy);
1711         INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
1712         INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
1713         INIT_DELAYED_WORK(&priv->neigh_reap_task, ipoib_reap_neigh);
1714 }
1715
1716 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
1717 {
1718         struct net_device *dev;
1719
1720         dev = alloc_netdev((int)sizeof(struct ipoib_dev_priv), name,
1721                            NET_NAME_UNKNOWN, ipoib_setup);
1722         if (!dev)
1723                 return NULL;
1724
1725         return netdev_priv(dev);
1726 }
1727
1728 static ssize_t show_pkey(struct device *dev,
1729                          struct device_attribute *attr, char *buf)
1730 {
1731         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1732
1733         return sprintf(buf, "0x%04x\n", priv->pkey);
1734 }
1735 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1736
1737 static ssize_t show_umcast(struct device *dev,
1738                            struct device_attribute *attr, char *buf)
1739 {
1740         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1741
1742         return sprintf(buf, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
1743 }
1744
1745 void ipoib_set_umcast(struct net_device *ndev, int umcast_val)
1746 {
1747         struct ipoib_dev_priv *priv = netdev_priv(ndev);
1748
1749         if (umcast_val > 0) {
1750                 set_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1751                 ipoib_warn(priv, "ignoring multicast groups joined directly "
1752                                 "by userspace\n");
1753         } else
1754                 clear_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1755 }
1756
1757 static ssize_t set_umcast(struct device *dev,
1758                           struct device_attribute *attr,
1759                           const char *buf, size_t count)
1760 {
1761         unsigned long umcast_val = simple_strtoul(buf, NULL, 0);
1762
1763         ipoib_set_umcast(to_net_dev(dev), umcast_val);
1764
1765         return count;
1766 }
1767 static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast);
1768
1769 int ipoib_add_umcast_attr(struct net_device *dev)
1770 {
1771         return device_create_file(&dev->dev, &dev_attr_umcast);
1772 }
1773
1774 static ssize_t create_child(struct device *dev,
1775                             struct device_attribute *attr,
1776                             const char *buf, size_t count)
1777 {
1778         int pkey;
1779         int ret;
1780
1781         if (sscanf(buf, "%i", &pkey) != 1)
1782                 return -EINVAL;
1783
1784         if (pkey <= 0 || pkey > 0xffff || pkey == 0x8000)
1785                 return -EINVAL;
1786
1787         /*
1788          * Set the full membership bit, so that we join the right
1789          * broadcast group, etc.
1790          */
1791         pkey |= 0x8000;
1792
1793         ret = ipoib_vlan_add(to_net_dev(dev), pkey);
1794
1795         return ret ? ret : count;
1796 }
1797 static DEVICE_ATTR(create_child, S_IWUSR, NULL, create_child);
1798
1799 static ssize_t delete_child(struct device *dev,
1800                             struct device_attribute *attr,
1801                             const char *buf, size_t count)
1802 {
1803         int pkey;
1804         int ret;
1805
1806         if (sscanf(buf, "%i", &pkey) != 1)
1807                 return -EINVAL;
1808
1809         if (pkey < 0 || pkey > 0xffff)
1810                 return -EINVAL;
1811
1812         ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
1813
1814         return ret ? ret : count;
1815
1816 }
1817 static DEVICE_ATTR(delete_child, S_IWUSR, NULL, delete_child);
1818
1819 int ipoib_add_pkey_attr(struct net_device *dev)
1820 {
1821         return device_create_file(&dev->dev, &dev_attr_pkey);
1822 }
1823
1824 int ipoib_set_dev_features(struct ipoib_dev_priv *priv, struct ib_device *hca)
1825 {
1826         priv->hca_caps = hca->attrs.device_cap_flags;
1827
1828         if (priv->hca_caps & IB_DEVICE_UD_IP_CSUM) {
1829                 priv->dev->hw_features = NETIF_F_SG |
1830                         NETIF_F_IP_CSUM | NETIF_F_RXCSUM;
1831
1832                 if (priv->hca_caps & IB_DEVICE_UD_TSO)
1833                         priv->dev->hw_features |= NETIF_F_TSO;
1834
1835                 priv->dev->features |= priv->dev->hw_features;
1836         }
1837
1838         return 0;
1839 }
1840
1841 static struct net_device *ipoib_add_port(const char *format,
1842                                          struct ib_device *hca, u8 port)
1843 {
1844         struct ipoib_dev_priv *priv;
1845         struct ib_port_attr attr;
1846         int result = -ENOMEM;
1847
1848         priv = ipoib_intf_alloc(format);
1849         if (!priv)
1850                 goto alloc_mem_failed;
1851
1852         SET_NETDEV_DEV(priv->dev, hca->dma_device);
1853         priv->dev->dev_id = port - 1;
1854
1855         result = ib_query_port(hca, port, &attr);
1856         if (!result)
1857                 priv->max_ib_mtu = ib_mtu_enum_to_int(attr.max_mtu);
1858         else {
1859                 printk(KERN_WARNING "%s: ib_query_port %d failed\n",
1860                        hca->name, port);
1861                 goto device_init_failed;
1862         }
1863
1864         /* MTU will be reset when mcast join happens */
1865         priv->dev->mtu  = IPOIB_UD_MTU(priv->max_ib_mtu);
1866         priv->mcast_mtu  = priv->admin_mtu = priv->dev->mtu;
1867
1868         priv->dev->neigh_priv_len = sizeof(struct ipoib_neigh);
1869
1870         result = ib_query_pkey(hca, port, 0, &priv->pkey);
1871         if (result) {
1872                 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
1873                        hca->name, port, result);
1874                 goto device_init_failed;
1875         }
1876
1877         result = ipoib_set_dev_features(priv, hca);
1878         if (result)
1879                 goto device_init_failed;
1880
1881         /*
1882          * Set the full membership bit, so that we join the right
1883          * broadcast group, etc.
1884          */
1885         priv->pkey |= 0x8000;
1886
1887         priv->dev->broadcast[8] = priv->pkey >> 8;
1888         priv->dev->broadcast[9] = priv->pkey & 0xff;
1889
1890         result = ib_query_gid(hca, port, 0, &priv->local_gid, NULL);
1891         if (result) {
1892                 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1893                        hca->name, port, result);
1894                 goto device_init_failed;
1895         } else
1896                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1897
1898         result = ipoib_dev_init(priv->dev, hca, port);
1899         if (result < 0) {
1900                 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1901                        hca->name, port, result);
1902                 goto device_init_failed;
1903         }
1904
1905         INIT_IB_EVENT_HANDLER(&priv->event_handler,
1906                               priv->ca, ipoib_event);
1907         result = ib_register_event_handler(&priv->event_handler);
1908         if (result < 0) {
1909                 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1910                        "port %d (ret = %d)\n",
1911                        hca->name, port, result);
1912                 goto event_failed;
1913         }
1914
1915         result = register_netdev(priv->dev);
1916         if (result) {
1917                 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1918                        hca->name, port, result);
1919                 goto register_failed;
1920         }
1921
1922         ipoib_create_debug_files(priv->dev);
1923
1924         if (ipoib_cm_add_mode_attr(priv->dev))
1925                 goto sysfs_failed;
1926         if (ipoib_add_pkey_attr(priv->dev))
1927                 goto sysfs_failed;
1928         if (ipoib_add_umcast_attr(priv->dev))
1929                 goto sysfs_failed;
1930         if (device_create_file(&priv->dev->dev, &dev_attr_create_child))
1931                 goto sysfs_failed;
1932         if (device_create_file(&priv->dev->dev, &dev_attr_delete_child))
1933                 goto sysfs_failed;
1934
1935         return priv->dev;
1936
1937 sysfs_failed:
1938         ipoib_delete_debug_files(priv->dev);
1939         unregister_netdev(priv->dev);
1940
1941 register_failed:
1942         ib_unregister_event_handler(&priv->event_handler);
1943         flush_workqueue(ipoib_workqueue);
1944         /* Stop GC if started before flush */
1945         set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1946         cancel_delayed_work(&priv->neigh_reap_task);
1947         flush_workqueue(priv->wq);
1948
1949 event_failed:
1950         ipoib_dev_cleanup(priv->dev);
1951
1952 device_init_failed:
1953         free_netdev(priv->dev);
1954
1955 alloc_mem_failed:
1956         return ERR_PTR(result);
1957 }
1958
1959 static void ipoib_add_one(struct ib_device *device)
1960 {
1961         struct list_head *dev_list;
1962         struct net_device *dev;
1963         struct ipoib_dev_priv *priv;
1964         int p;
1965         int count = 0;
1966
1967         dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1968         if (!dev_list)
1969                 return;
1970
1971         INIT_LIST_HEAD(dev_list);
1972
1973         for (p = rdma_start_port(device); p <= rdma_end_port(device); ++p) {
1974                 if (!rdma_protocol_ib(device, p))
1975                         continue;
1976                 dev = ipoib_add_port("ib%d", device, p);
1977                 if (!IS_ERR(dev)) {
1978                         priv = netdev_priv(dev);
1979                         list_add_tail(&priv->list, dev_list);
1980                         count++;
1981                 }
1982         }
1983
1984         if (!count) {
1985                 kfree(dev_list);
1986                 return;
1987         }
1988
1989         ib_set_client_data(device, &ipoib_client, dev_list);
1990 }
1991
1992 static void ipoib_remove_one(struct ib_device *device, void *client_data)
1993 {
1994         struct ipoib_dev_priv *priv, *tmp;
1995         struct list_head *dev_list = client_data;
1996
1997         if (!dev_list)
1998                 return;
1999
2000         list_for_each_entry_safe(priv, tmp, dev_list, list) {
2001                 ib_unregister_event_handler(&priv->event_handler);
2002                 flush_workqueue(ipoib_workqueue);
2003
2004                 rtnl_lock();
2005                 dev_change_flags(priv->dev, priv->dev->flags & ~IFF_UP);
2006                 rtnl_unlock();
2007
2008                 /* Stop GC */
2009                 set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
2010                 cancel_delayed_work(&priv->neigh_reap_task);
2011                 flush_workqueue(priv->wq);
2012
2013                 unregister_netdev(priv->dev);
2014                 free_netdev(priv->dev);
2015         }
2016
2017         kfree(dev_list);
2018 }
2019
2020 static int __init ipoib_init_module(void)
2021 {
2022         int ret;
2023
2024         ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
2025         ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
2026         ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
2027
2028         ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
2029         ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
2030         ipoib_sendq_size = max3(ipoib_sendq_size, 2 * MAX_SEND_CQE, IPOIB_MIN_QUEUE_SIZE);
2031 #ifdef CONFIG_INFINIBAND_IPOIB_CM
2032         ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
2033 #endif
2034
2035         /*
2036          * When copying small received packets, we only copy from the
2037          * linear data part of the SKB, so we rely on this condition.
2038          */
2039         BUILD_BUG_ON(IPOIB_CM_COPYBREAK > IPOIB_CM_HEAD_SIZE);
2040
2041         ret = ipoib_register_debugfs();
2042         if (ret)
2043                 return ret;
2044
2045         /*
2046          * We create a global workqueue here that is used for all flush
2047          * operations.  However, if you attempt to flush a workqueue
2048          * from a task on that same workqueue, it deadlocks the system.
2049          * We want to be able to flush the tasks associated with a
2050          * specific net device, so we also create a workqueue for each
2051          * netdevice.  We queue up the tasks for that device only on
2052          * its private workqueue, and we only queue up flush events
2053          * on our global flush workqueue.  This avoids the deadlocks.
2054          */
2055         ipoib_workqueue = create_singlethread_workqueue("ipoib_flush");
2056         if (!ipoib_workqueue) {
2057                 ret = -ENOMEM;
2058                 goto err_fs;
2059         }
2060
2061         ib_sa_register_client(&ipoib_sa_client);
2062
2063         ret = ib_register_client(&ipoib_client);
2064         if (ret)
2065                 goto err_sa;
2066
2067         ret = ipoib_netlink_init();
2068         if (ret)
2069                 goto err_client;
2070
2071         return 0;
2072
2073 err_client:
2074         ib_unregister_client(&ipoib_client);
2075
2076 err_sa:
2077         ib_sa_unregister_client(&ipoib_sa_client);
2078         destroy_workqueue(ipoib_workqueue);
2079
2080 err_fs:
2081         ipoib_unregister_debugfs();
2082
2083         return ret;
2084 }
2085
2086 static void __exit ipoib_cleanup_module(void)
2087 {
2088         ipoib_netlink_fini();
2089         ib_unregister_client(&ipoib_client);
2090         ib_sa_unregister_client(&ipoib_sa_client);
2091         ipoib_unregister_debugfs();
2092         destroy_workqueue(ipoib_workqueue);
2093 }
2094
2095 module_init(ipoib_init_module);
2096 module_exit(ipoib_cleanup_module);