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 static int ipoib_set_mac(struct net_device *dev, void *addr);
 103
 104 static struct ib_client ipoib_client = {
 105         .name   = "ipoib",
 106         .add    = ipoib_add_one,
 107         .remove = ipoib_remove_one,
 108         .get_net_dev_by_params = ipoib_get_net_dev_by_params,
 109 };
 110
 111 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 112 static int ipoib_netdev_event(struct notifier_block *this,
 113                               unsigned long event, void *ptr)
 114 {
 115         struct netdev_notifier_info *ni = ptr;
 116         struct net_device *dev = ni->dev;
 117
 118         if (dev->netdev_ops->ndo_open != ipoib_open)
 119                 return NOTIFY_DONE;
 120
 121         switch (event) {
 122         case NETDEV_REGISTER:
 123                 ipoib_create_debug_files(dev);
 124                 break;
 125         case NETDEV_CHANGENAME:
 126                 ipoib_delete_debug_files(dev);
 127                 ipoib_create_debug_files(dev);
 128                 break;
 129         case NETDEV_UNREGISTER:
 130                 ipoib_delete_debug_files(dev);
 131                 break;
 132         }
 133
 134         return NOTIFY_DONE;
 135 }
 136 #endif
 137
 138 int ipoib_open(struct net_device *dev)
 139 {
 140         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 141
 142         ipoib_dbg(priv, "bringing up interface\n");
 143
 144         netif_carrier_off(dev);
 145
 146         set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 147
 148         priv->sm_fullmember_sendonly_support = false;
 149
 150         if (ipoib_ib_dev_open(dev)) {
 151                 if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags))
 152                         return 0;
 153                 goto err_disable;
 154         }
 155
 156         ipoib_ib_dev_up(dev);
 157
 158         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 159                 struct ipoib_dev_priv *cpriv;
 160
 161                 /* Bring up any child interfaces too */
 162                 down_read(&priv->vlan_rwsem);
 163                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 164                         int flags;
 165
 166                         flags = cpriv->dev->flags;
 167                         if (flags & IFF_UP)
 168                                 continue;
 169
 170                         dev_change_flags(cpriv->dev, flags | IFF_UP);
 171                 }
 172                 up_read(&priv->vlan_rwsem);
 173         }
 174
 175         netif_start_queue(dev);
 176
 177         return 0;
 178
 179 err_disable:
 180         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 181
 182         return -EINVAL;
 183 }
 184
 185 static int ipoib_stop(struct net_device *dev)
 186 {
 187         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 188
 189         ipoib_dbg(priv, "stopping interface\n");
 190
 191         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 192
 193         netif_stop_queue(dev);
 194
 195         ipoib_ib_dev_down(dev);
 196         ipoib_ib_dev_stop(dev);
 197
 198         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 199                 struct ipoib_dev_priv *cpriv;
 200
 201                 /* Bring down any child interfaces too */
 202                 down_read(&priv->vlan_rwsem);
 203                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 204                         int flags;
 205
 206                         flags = cpriv->dev->flags;
 207                         if (!(flags & IFF_UP))
 208                                 continue;
 209
 210                         dev_change_flags(cpriv->dev, flags & ~IFF_UP);
 211                 }
 212                 up_read(&priv->vlan_rwsem);
 213         }
 214
 215         return 0;
 216 }
 217
 218 static void ipoib_uninit(struct net_device *dev)
 219 {
 220         ipoib_dev_cleanup(dev);
 221 }
 222
 223 static netdev_features_t ipoib_fix_features(struct net_device *dev, netdev_features_t features)
 224 {
 225         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 226
 227         if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
 228                 features &= ~(NETIF_F_IP_CSUM | NETIF_F_TSO);
 229
 230         return features;
 231 }
 232
 233 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
 234 {
 235         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 236
 237         /* dev->mtu > 2K ==> connected mode */
 238         if (ipoib_cm_admin_enabled(dev)) {
 239                 if (new_mtu > ipoib_cm_max_mtu(dev))
 240                         return -EINVAL;
 241
 242                 if (new_mtu > priv->mcast_mtu)
 243                         ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
 244                                    priv->mcast_mtu);
 245
 246                 dev->mtu = new_mtu;
 247                 return 0;
 248         }
 249
 250         if (new_mtu > IPOIB_UD_MTU(priv->max_ib_mtu))
 251                 return -EINVAL;
 252
 253         priv->admin_mtu = new_mtu;
 254
 255         if (priv->mcast_mtu < priv->admin_mtu)
 256                 ipoib_dbg(priv, "MTU must be smaller than the underlying "
 257                                 "link layer MTU - 4 (%u)\n", priv->mcast_mtu);
 258
 259         dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
 260
 261         return 0;
 262 }
 263
 264 /* Called with an RCU read lock taken */
 265 static bool ipoib_is_dev_match_addr_rcu(const struct sockaddr *addr,
 266                                         struct net_device *dev)
 267 {
 268         struct net *net = dev_net(dev);
 269         struct in_device *in_dev;
 270         struct sockaddr_in *addr_in = (struct sockaddr_in *)addr;
 271         struct sockaddr_in6 *addr_in6 = (struct sockaddr_in6 *)addr;
 272         __be32 ret_addr;
 273
 274         switch (addr->sa_family) {
 275         case AF_INET:
 276                 in_dev = in_dev_get(dev);
 277                 if (!in_dev)
 278                         return false;
 279
 280                 ret_addr = inet_confirm_addr(net, in_dev, 0,
 281                                              addr_in->sin_addr.s_addr,
 282                                              RT_SCOPE_HOST);
 283                 in_dev_put(in_dev);
 284                 if (ret_addr)
 285                         return true;
 286
 287                 break;
 288         case AF_INET6:
 289                 if (IS_ENABLED(CONFIG_IPV6) &&
 290                     ipv6_chk_addr(net, &addr_in6->sin6_addr, dev, 1))
 291                         return true;
 292
 293                 break;
 294         }
 295         return false;
 296 }
 297
 298 /**
 299  * Find the master net_device on top of the given net_device.
 300  * @dev: base IPoIB net_device
 301  *
 302  * Returns the master net_device with a reference held, or the same net_device
 303  * if no master exists.
 304  */
 305 static struct net_device *ipoib_get_master_net_dev(struct net_device *dev)
 306 {
 307         struct net_device *master;
 308
 309         rcu_read_lock();
 310         master = netdev_master_upper_dev_get_rcu(dev);
 311         if (master)
 312                 dev_hold(master);
 313         rcu_read_unlock();
 314
 315         if (master)
 316                 return master;
 317
 318         dev_hold(dev);
 319         return dev;
 320 }
 321
 322 struct ipoib_walk_data {
 323         const struct sockaddr *addr;
 324         struct net_device *result;
 325 };
 326
 327 static int ipoib_upper_walk(struct net_device *upper, void *_data)
 328 {
 329         struct ipoib_walk_data *data = _data;
 330         int ret = 0;
 331
 332         if (ipoib_is_dev_match_addr_rcu(data->addr, upper)) {
 333                 dev_hold(upper);
 334                 data->result = upper;
 335                 ret = 1;
 336         }
 337
 338         return ret;
 339 }
 340
 341 /**
 342  * Find a net_device matching the given address, which is an upper device of
 343  * the given net_device.
 344  * @addr: IP address to look for.
 345  * @dev: base IPoIB net_device
 346  *
 347  * If found, returns the net_device with a reference held. Otherwise return
 348  * NULL.
 349  */
 350 static struct net_device *ipoib_get_net_dev_match_addr(
 351                 const struct sockaddr *addr, struct net_device *dev)
 352 {
 353         struct ipoib_walk_data data = {
 354                 .addr = addr,
 355         };
 356
 357         rcu_read_lock();
 358         if (ipoib_is_dev_match_addr_rcu(addr, dev)) {
 359                 dev_hold(dev);
 360                 data.result = dev;
 361                 goto out;
 362         }
 363
 364         netdev_walk_all_upper_dev_rcu(dev, ipoib_upper_walk, &data);
 365 out:
 366         rcu_read_unlock();
 367         return data.result;
 368 }
 369
 370 /* returns the number of IPoIB netdevs on top a given ipoib device matching a
 371  * pkey_index and address, if one exists.
 372  *
 373  * @found_net_dev: contains a matching net_device if the return value >= 1,
 374  * with a reference held. */
 375 static int ipoib_match_gid_pkey_addr(struct ipoib_dev_priv *priv,
 376                                      const union ib_gid *gid,
 377                                      u16 pkey_index,
 378                                      const struct sockaddr *addr,
 379                                      int nesting,
 380                                      struct net_device **found_net_dev)
 381 {
 382         struct ipoib_dev_priv *child_priv;
 383         struct net_device *net_dev = NULL;
 384         int matches = 0;
 385
 386         if (priv->pkey_index == pkey_index &&
 387             (!gid || !memcmp(gid, &priv->local_gid, sizeof(*gid)))) {
 388                 if (!addr) {
 389                         net_dev = ipoib_get_master_net_dev(priv->dev);
 390                 } else {
 391                         /* Verify the net_device matches the IP address, as
 392                          * IPoIB child devices currently share a GID. */
 393                         net_dev = ipoib_get_net_dev_match_addr(addr, priv->dev);
 394                 }
 395                 if (net_dev) {
 396                         if (!*found_net_dev)
 397                                 *found_net_dev = net_dev;
 398                         else
 399                                 dev_put(net_dev);
 400                         ++matches;
 401                 }
 402         }
 403
 404         /* Check child interfaces */
 405         down_read_nested(&priv->vlan_rwsem, nesting);
 406         list_for_each_entry(child_priv, &priv->child_intfs, list) {
 407                 matches += ipoib_match_gid_pkey_addr(child_priv, gid,
 408                                                     pkey_index, addr,
 409                                                     nesting + 1,
 410                                                     found_net_dev);
 411                 if (matches > 1)
 412                         break;
 413         }
 414         up_read(&priv->vlan_rwsem);
 415
 416         return matches;
 417 }
 418
 419 /* Returns the number of matching net_devs found (between 0 and 2). Also
 420  * return the matching net_device in the @net_dev parameter, holding a
 421  * reference to the net_device, if the number of matches >= 1 */
 422 static int __ipoib_get_net_dev_by_params(struct list_head *dev_list, u8 port,
 423                                          u16 pkey_index,
 424                                          const union ib_gid *gid,
 425                                          const struct sockaddr *addr,
 426                                          struct net_device **net_dev)
 427 {
 428         struct ipoib_dev_priv *priv;
 429         int matches = 0;
 430
 431         *net_dev = NULL;
 432
 433         list_for_each_entry(priv, dev_list, list) {
 434                 if (priv->port != port)
 435                         continue;
 436
 437                 matches += ipoib_match_gid_pkey_addr(priv, gid, pkey_index,
 438                                                      addr, 0, net_dev);
 439                 if (matches > 1)
 440                         break;
 441         }
 442
 443         return matches;
 444 }
 445
 446 static struct net_device *ipoib_get_net_dev_by_params(
 447                 struct ib_device *dev, u8 port, u16 pkey,
 448                 const union ib_gid *gid, const struct sockaddr *addr,
 449                 void *client_data)
 450 {
 451         struct net_device *net_dev;
 452         struct list_head *dev_list = client_data;
 453         u16 pkey_index;
 454         int matches;
 455         int ret;
 456
 457         if (!rdma_protocol_ib(dev, port))
 458                 return NULL;
 459
 460         ret = ib_find_cached_pkey(dev, port, pkey, &pkey_index);
 461         if (ret)
 462                 return NULL;
 463
 464         if (!dev_list)
 465                 return NULL;
 466
 467         /* See if we can find a unique device matching the L2 parameters */
 468         matches = __ipoib_get_net_dev_by_params(dev_list, port, pkey_index,
 469                                                 gid, NULL, &net_dev);
 470
 471         switch (matches) {
 472         case 0:
 473                 return NULL;
 474         case 1:
 475                 return net_dev;
 476         }
 477
 478         dev_put(net_dev);
 479
 480         /* Couldn't find a unique device with L2 parameters only. Use L3
 481          * address to uniquely match the net device */
 482         matches = __ipoib_get_net_dev_by_params(dev_list, port, pkey_index,
 483                                                 gid, addr, &net_dev);
 484         switch (matches) {
 485         case 0:
 486                 return NULL;
 487         default:
 488                 dev_warn_ratelimited(&dev->dev,
 489                                      "duplicate IP address detected\n");
 490                 /* Fall through */
 491         case 1:
 492                 return net_dev;
 493         }
 494 }
 495
 496 int ipoib_set_mode(struct net_device *dev, const char *buf)
 497 {
 498         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 499
 500         if ((test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags) &&
 501              !strcmp(buf, "connected\n")) ||
 502              (!test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags) &&
 503              !strcmp(buf, "datagram\n"))) {
 504                 return 0;
 505         }
 506
 507         /* flush paths if we switch modes so that connections are restarted */
 508         if (IPOIB_CM_SUPPORTED(dev->dev_addr) && !strcmp(buf, "connected\n")) {
 509                 set_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
 510                 ipoib_warn(priv, "enabling connected mode "
 511                            "will cause multicast packet drops\n");
 512                 netdev_update_features(dev);
 513                 dev_set_mtu(dev, ipoib_cm_max_mtu(dev));
 514                 rtnl_unlock();
 515                 priv->tx_wr.wr.send_flags &= ~IB_SEND_IP_CSUM;
 516
 517                 ipoib_flush_paths(dev);
 518                 return (!rtnl_trylock()) ? -EBUSY : 0;
 519         }
 520
 521         if (!strcmp(buf, "datagram\n")) {
 522                 clear_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
 523                 netdev_update_features(dev);
 524                 dev_set_mtu(dev, min(priv->mcast_mtu, dev->mtu));
 525                 rtnl_unlock();
 526                 ipoib_flush_paths(dev);
 527                 return (!rtnl_trylock()) ? -EBUSY : 0;
 528         }
 529
 530         return -EINVAL;
 531 }
 532
 533 struct ipoib_path *__path_find(struct net_device *dev, void *gid)
 534 {
 535         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 536         struct rb_node *n = priv->path_tree.rb_node;
 537         struct ipoib_path *path;
 538         int ret;
 539
 540         while (n) {
 541                 path = rb_entry(n, struct ipoib_path, rb_node);
 542
 543                 ret = memcmp(gid, path->pathrec.dgid.raw,
 544                              sizeof (union ib_gid));
 545
 546                 if (ret < 0)
 547                         n = n->rb_left;
 548                 else if (ret > 0)
 549                         n = n->rb_right;
 550                 else
 551                         return path;
 552         }
 553
 554         return NULL;
 555 }
 556
 557 static int __path_add(struct net_device *dev, struct ipoib_path *path)
 558 {
 559         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 560         struct rb_node **n = &priv->path_tree.rb_node;
 561         struct rb_node *pn = NULL;
 562         struct ipoib_path *tpath;
 563         int ret;
 564
 565         while (*n) {
 566                 pn = *n;
 567                 tpath = rb_entry(pn, struct ipoib_path, rb_node);
 568
 569                 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
 570                              sizeof (union ib_gid));
 571                 if (ret < 0)
 572                         n = &pn->rb_left;
 573                 else if (ret > 0)
 574                         n = &pn->rb_right;
 575                 else
 576                         return -EEXIST;
 577         }
 578
 579         rb_link_node(&path->rb_node, pn, n);
 580         rb_insert_color(&path->rb_node, &priv->path_tree);
 581
 582         list_add_tail(&path->list, &priv->path_list);
 583
 584         return 0;
 585 }
 586
 587 static void path_free(struct net_device *dev, struct ipoib_path *path)
 588 {
 589         struct sk_buff *skb;
 590
 591         while ((skb = __skb_dequeue(&path->queue)))
 592                 dev_kfree_skb_irq(skb);
 593
 594         ipoib_dbg(ipoib_priv(dev), "path_free\n");
 595
 596         /* remove all neigh connected to this path */
 597         ipoib_del_neighs_by_gid(dev, path->pathrec.dgid.raw);
 598
 599         if (path->ah)
 600                 ipoib_put_ah(path->ah);
 601
 602         kfree(path);
 603 }
 604
 605 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 606
 607 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
 608 {
 609         struct ipoib_path_iter *iter;
 610
 611         iter = kmalloc(sizeof *iter, GFP_KERNEL);
 612         if (!iter)
 613                 return NULL;
 614
 615         iter->dev = dev;
 616         memset(iter->path.pathrec.dgid.raw, 0, 16);
 617
 618         if (ipoib_path_iter_next(iter)) {
 619                 kfree(iter);
 620                 return NULL;
 621         }
 622
 623         return iter;
 624 }
 625
 626 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
 627 {
 628         struct ipoib_dev_priv *priv = ipoib_priv(iter->dev);
 629         struct rb_node *n;
 630         struct ipoib_path *path;
 631         int ret = 1;
 632
 633         spin_lock_irq(&priv->lock);
 634
 635         n = rb_first(&priv->path_tree);
 636
 637         while (n) {
 638                 path = rb_entry(n, struct ipoib_path, rb_node);
 639
 640                 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
 641                            sizeof (union ib_gid)) < 0) {
 642                         iter->path = *path;
 643                         ret = 0;
 644                         break;
 645                 }
 646
 647                 n = rb_next(n);
 648         }
 649
 650         spin_unlock_irq(&priv->lock);
 651
 652         return ret;
 653 }
 654
 655 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
 656                           struct ipoib_path *path)
 657 {
 658         *path = iter->path;
 659 }
 660
 661 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
 662
 663 void ipoib_mark_paths_invalid(struct net_device *dev)
 664 {
 665         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 666         struct ipoib_path *path, *tp;
 667
 668         spin_lock_irq(&priv->lock);
 669
 670         list_for_each_entry_safe(path, tp, &priv->path_list, list) {
 671                 ipoib_dbg(priv, "mark path LID 0x%08x GID %pI6 invalid\n",
 672                           be32_to_cpu(sa_path_get_dlid(&path->pathrec)),
 673                           path->pathrec.dgid.raw);
 674                 path->valid =  0;
 675         }
 676
 677         spin_unlock_irq(&priv->lock);
 678 }
 679
 680 static void push_pseudo_header(struct sk_buff *skb, const char *daddr)
 681 {
 682         struct ipoib_pseudo_header *phdr;
 683
 684         phdr = (struct ipoib_pseudo_header *)skb_push(skb, sizeof(*phdr));
 685         memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
 686 }
 687
 688 void ipoib_flush_paths(struct net_device *dev)
 689 {
 690         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 691         struct ipoib_path *path, *tp;
 692         LIST_HEAD(remove_list);
 693         unsigned long flags;
 694
 695         netif_tx_lock_bh(dev);
 696         spin_lock_irqsave(&priv->lock, flags);
 697
 698         list_splice_init(&priv->path_list, &remove_list);
 699
 700         list_for_each_entry(path, &remove_list, list)
 701                 rb_erase(&path->rb_node, &priv->path_tree);
 702
 703         list_for_each_entry_safe(path, tp, &remove_list, list) {
 704                 if (path->query)
 705                         ib_sa_cancel_query(path->query_id, path->query);
 706                 spin_unlock_irqrestore(&priv->lock, flags);
 707                 netif_tx_unlock_bh(dev);
 708                 wait_for_completion(&path->done);
 709                 path_free(dev, path);
 710                 netif_tx_lock_bh(dev);
 711                 spin_lock_irqsave(&priv->lock, flags);
 712         }
 713
 714         spin_unlock_irqrestore(&priv->lock, flags);
 715         netif_tx_unlock_bh(dev);
 716 }
 717
 718 static void path_rec_completion(int status,
 719                                 struct sa_path_rec *pathrec,
 720                                 void *path_ptr)
 721 {
 722         struct ipoib_path *path = path_ptr;
 723         struct net_device *dev = path->dev;
 724         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 725         struct ipoib_ah *ah = NULL;
 726         struct ipoib_ah *old_ah = NULL;
 727         struct ipoib_neigh *neigh, *tn;
 728         struct sk_buff_head skqueue;
 729         struct sk_buff *skb;
 730         unsigned long flags;
 731
 732         if (!status)
 733                 ipoib_dbg(priv, "PathRec LID 0x%04x for GID %pI6\n",
 734                           be32_to_cpu(sa_path_get_dlid(pathrec)),
 735                           pathrec->dgid.raw);
 736         else
 737                 ipoib_dbg(priv, "PathRec status %d for GID %pI6\n",
 738                           status, path->pathrec.dgid.raw);
 739
 740         skb_queue_head_init(&skqueue);
 741
 742         if (!status) {
 743                 struct rdma_ah_attr av;
 744
 745                 if (!ib_init_ah_from_path(priv->ca, priv->port, pathrec, &av))
 746                         ah = ipoib_create_ah(dev, priv->pd, &av);
 747         }
 748
 749         spin_lock_irqsave(&priv->lock, flags);
 750
 751         if (!IS_ERR_OR_NULL(ah)) {
 752                 path->pathrec = *pathrec;
 753
 754                 old_ah   = path->ah;
 755                 path->ah = ah;
 756
 757                 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
 758                           ah, be32_to_cpu(sa_path_get_dlid(pathrec)),
 759                           pathrec->sl);
 760
 761                 while ((skb = __skb_dequeue(&path->queue)))
 762                         __skb_queue_tail(&skqueue, skb);
 763
 764                 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
 765                         if (neigh->ah) {
 766                                 WARN_ON(neigh->ah != old_ah);
 767                                 /*
 768                                  * Dropping the ah reference inside
 769                                  * priv->lock is safe here, because we
 770                                  * will hold one more reference from
 771                                  * the original value of path->ah (ie
 772                                  * old_ah).
 773                                  */
 774                                 ipoib_put_ah(neigh->ah);
 775                         }
 776                         kref_get(&path->ah->ref);
 777                         neigh->ah = path->ah;
 778
 779                         if (ipoib_cm_enabled(dev, neigh->daddr)) {
 780                                 if (!ipoib_cm_get(neigh))
 781                                         ipoib_cm_set(neigh, ipoib_cm_create_tx(dev,
 782                                                                                path,
 783                                                                                neigh));
 784                                 if (!ipoib_cm_get(neigh)) {
 785                                         ipoib_neigh_free(neigh);
 786                                         continue;
 787                                 }
 788                         }
 789
 790                         while ((skb = __skb_dequeue(&neigh->queue)))
 791                                 __skb_queue_tail(&skqueue, skb);
 792                 }
 793                 path->valid = 1;
 794         }
 795
 796         path->query = NULL;
 797         complete(&path->done);
 798
 799         spin_unlock_irqrestore(&priv->lock, flags);
 800
 801         if (IS_ERR_OR_NULL(ah))
 802                 ipoib_del_neighs_by_gid(dev, path->pathrec.dgid.raw);
 803
 804         if (old_ah)
 805                 ipoib_put_ah(old_ah);
 806
 807         while ((skb = __skb_dequeue(&skqueue))) {
 808                 int ret;
 809                 skb->dev = dev;
 810                 ret = dev_queue_xmit(skb);
 811                 if (ret)
 812                         ipoib_warn(priv, "%s: dev_queue_xmit failed to re-queue packet, ret:%d\n",
 813                                    __func__, ret);
 814         }
 815 }
 816
 817 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
 818 {
 819         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 820         struct ipoib_path *path;
 821
 822         if (!priv->broadcast)
 823                 return NULL;
 824
 825         path = kzalloc(sizeof *path, GFP_ATOMIC);
 826         if (!path)
 827                 return NULL;
 828
 829         path->dev = dev;
 830
 831         skb_queue_head_init(&path->queue);
 832
 833         INIT_LIST_HEAD(&path->neigh_list);
 834
 835         if (rdma_cap_opa_ah(priv->ca, priv->port))
 836                 path->pathrec.rec_type = SA_PATH_REC_TYPE_OPA;
 837         else
 838                 path->pathrec.rec_type = SA_PATH_REC_TYPE_IB;
 839         memcpy(path->pathrec.dgid.raw, gid, sizeof (union ib_gid));
 840         path->pathrec.sgid          = priv->local_gid;
 841         path->pathrec.pkey          = cpu_to_be16(priv->pkey);
 842         path->pathrec.numb_path     = 1;
 843         path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
 844
 845         return path;
 846 }
 847
 848 static int path_rec_start(struct net_device *dev,
 849                           struct ipoib_path *path)
 850 {
 851         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 852
 853         ipoib_dbg(priv, "Start path record lookup for %pI6\n",
 854                   path->pathrec.dgid.raw);
 855
 856         init_completion(&path->done);
 857
 858         path->query_id =
 859                 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
 860                                    &path->pathrec,
 861                                    IB_SA_PATH_REC_DGID          |
 862                                    IB_SA_PATH_REC_SGID          |
 863                                    IB_SA_PATH_REC_NUMB_PATH     |
 864                                    IB_SA_PATH_REC_TRAFFIC_CLASS |
 865                                    IB_SA_PATH_REC_PKEY,
 866                                    1000, GFP_ATOMIC,
 867                                    path_rec_completion,
 868                                    path, &path->query);
 869         if (path->query_id < 0) {
 870                 ipoib_warn(priv, "ib_sa_path_rec_get failed: %d\n", path->query_id);
 871                 path->query = NULL;
 872                 complete(&path->done);
 873                 return path->query_id;
 874         }
 875
 876         return 0;
 877 }
 878
 879 static void neigh_add_path(struct sk_buff *skb, u8 *daddr,
 880                            struct net_device *dev)
 881 {
 882         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 883         struct rdma_netdev *rn = netdev_priv(dev);
 884         struct ipoib_path *path;
 885         struct ipoib_neigh *neigh;
 886         unsigned long flags;
 887
 888         spin_lock_irqsave(&priv->lock, flags);
 889         neigh = ipoib_neigh_alloc(daddr, dev);
 890         if (!neigh) {
 891                 spin_unlock_irqrestore(&priv->lock, flags);
 892                 ++dev->stats.tx_dropped;
 893                 dev_kfree_skb_any(skb);
 894                 return;
 895         }
 896
 897         path = __path_find(dev, daddr + 4);
 898         if (!path) {
 899                 path = path_rec_create(dev, daddr + 4);
 900                 if (!path)
 901                         goto err_path;
 902
 903                 __path_add(dev, path);
 904         }
 905
 906         list_add_tail(&neigh->list, &path->neigh_list);
 907
 908         if (path->ah) {
 909                 kref_get(&path->ah->ref);
 910                 neigh->ah = path->ah;
 911
 912                 if (ipoib_cm_enabled(dev, neigh->daddr)) {
 913                         if (!ipoib_cm_get(neigh))
 914                                 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
 915                         if (!ipoib_cm_get(neigh)) {
 916                                 ipoib_neigh_free(neigh);
 917                                 goto err_drop;
 918                         }
 919                         if (skb_queue_len(&neigh->queue) <
 920                             IPOIB_MAX_PATH_REC_QUEUE) {
 921                                 push_pseudo_header(skb, neigh->daddr);
 922                                 __skb_queue_tail(&neigh->queue, skb);
 923                         } else {
 924                                 ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
 925                                            skb_queue_len(&neigh->queue));
 926                                 goto err_drop;
 927                         }
 928                 } else {
 929                         spin_unlock_irqrestore(&priv->lock, flags);
 930                         path->ah->last_send = rn->send(dev, skb, path->ah->ah,
 931                                                        IPOIB_QPN(daddr));
 932                         ipoib_neigh_put(neigh);
 933                         return;
 934                 }
 935         } else {
 936                 neigh->ah  = NULL;
 937
 938                 if (!path->query && path_rec_start(dev, path))
 939                         goto err_path;
 940                 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 941                         push_pseudo_header(skb, neigh->daddr);
 942                         __skb_queue_tail(&neigh->queue, skb);
 943                 } else {
 944                         goto err_drop;
 945                 }
 946         }
 947
 948         spin_unlock_irqrestore(&priv->lock, flags);
 949         ipoib_neigh_put(neigh);
 950         return;
 951
 952 err_path:
 953         ipoib_neigh_free(neigh);
 954 err_drop:
 955         ++dev->stats.tx_dropped;
 956         dev_kfree_skb_any(skb);
 957
 958         spin_unlock_irqrestore(&priv->lock, flags);
 959         ipoib_neigh_put(neigh);
 960 }
 961
 962 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
 963                              struct ipoib_pseudo_header *phdr)
 964 {
 965         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 966         struct rdma_netdev *rn = netdev_priv(dev);
 967         struct ipoib_path *path;
 968         unsigned long flags;
 969
 970         spin_lock_irqsave(&priv->lock, flags);
 971
 972         path = __path_find(dev, phdr->hwaddr + 4);
 973         if (!path || !path->valid) {
 974                 int new_path = 0;
 975
 976                 if (!path) {
 977                         path = path_rec_create(dev, phdr->hwaddr + 4);
 978                         new_path = 1;
 979                 }
 980                 if (path) {
 981                         if (skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 982                                 push_pseudo_header(skb, phdr->hwaddr);
 983                                 __skb_queue_tail(&path->queue, skb);
 984                         } else {
 985                                 ++dev->stats.tx_dropped;
 986                                 dev_kfree_skb_any(skb);
 987                         }
 988
 989                         if (!path->query && path_rec_start(dev, path)) {
 990                                 spin_unlock_irqrestore(&priv->lock, flags);
 991                                 if (new_path)
 992                                         path_free(dev, path);
 993                                 return;
 994                         } else
 995                                 __path_add(dev, path);
 996                 } else {
 997                         ++dev->stats.tx_dropped;
 998                         dev_kfree_skb_any(skb);
 999                 }
1000
1001                 spin_unlock_irqrestore(&priv->lock, flags);
1002                 return;
1003         }
1004
1005         if (path->ah) {
1006                 ipoib_dbg(priv, "Send unicast ARP to %08x\n",
1007                           be32_to_cpu(sa_path_get_dlid(&path->pathrec)));
1008
1009                 spin_unlock_irqrestore(&priv->lock, flags);
1010                 path->ah->last_send = rn->send(dev, skb, path->ah->ah,
1011                                                IPOIB_QPN(phdr->hwaddr));
1012                 return;
1013         } else if ((path->query || !path_rec_start(dev, path)) &&
1014                    skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
1015                 push_pseudo_header(skb, phdr->hwaddr);
1016                 __skb_queue_tail(&path->queue, skb);
1017         } else {
1018                 ++dev->stats.tx_dropped;
1019                 dev_kfree_skb_any(skb);
1020         }
1021
1022         spin_unlock_irqrestore(&priv->lock, flags);
1023 }
1024
1025 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
1026 {
1027         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1028         struct rdma_netdev *rn = netdev_priv(dev);
1029         struct ipoib_neigh *neigh;
1030         struct ipoib_pseudo_header *phdr;
1031         struct ipoib_header *header;
1032         unsigned long flags;
1033
1034         phdr = (struct ipoib_pseudo_header *) skb->data;
1035         skb_pull(skb, sizeof(*phdr));
1036         header = (struct ipoib_header *) skb->data;
1037
1038         if (unlikely(phdr->hwaddr[4] == 0xff)) {
1039                 /* multicast, arrange "if" according to probability */
1040                 if ((header->proto != htons(ETH_P_IP)) &&
1041                     (header->proto != htons(ETH_P_IPV6)) &&
1042                     (header->proto != htons(ETH_P_ARP)) &&
1043                     (header->proto != htons(ETH_P_RARP)) &&
1044                     (header->proto != htons(ETH_P_TIPC))) {
1045                         /* ethertype not supported by IPoIB */
1046                         ++dev->stats.tx_dropped;
1047                         dev_kfree_skb_any(skb);
1048                         return NETDEV_TX_OK;
1049                 }
1050                 /* Add in the P_Key for multicast*/
1051                 phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
1052                 phdr->hwaddr[9] = priv->pkey & 0xff;
1053
1054                 neigh = ipoib_neigh_get(dev, phdr->hwaddr);
1055                 if (likely(neigh))
1056                         goto send_using_neigh;
1057                 ipoib_mcast_send(dev, phdr->hwaddr, skb);
1058                 return NETDEV_TX_OK;
1059         }
1060
1061         /* unicast, arrange "switch" according to probability */
1062         switch (header->proto) {
1063         case htons(ETH_P_IP):
1064         case htons(ETH_P_IPV6):
1065         case htons(ETH_P_TIPC):
1066                 neigh = ipoib_neigh_get(dev, phdr->hwaddr);
1067                 if (unlikely(!neigh)) {
1068                         neigh_add_path(skb, phdr->hwaddr, dev);
1069                         return NETDEV_TX_OK;
1070                 }
1071                 break;
1072         case htons(ETH_P_ARP):
1073         case htons(ETH_P_RARP):
1074                 /* for unicast ARP and RARP should always perform path find */
1075                 unicast_arp_send(skb, dev, phdr);
1076                 return NETDEV_TX_OK;
1077         default:
1078                 /* ethertype not supported by IPoIB */
1079                 ++dev->stats.tx_dropped;
1080                 dev_kfree_skb_any(skb);
1081                 return NETDEV_TX_OK;
1082         }
1083
1084 send_using_neigh:
1085         /* note we now hold a ref to neigh */
1086         if (ipoib_cm_get(neigh)) {
1087                 if (ipoib_cm_up(neigh)) {
1088                         ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
1089                         goto unref;
1090                 }
1091         } else if (neigh->ah) {
1092                 neigh->ah->last_send = rn->send(dev, skb, neigh->ah->ah,
1093                                                 IPOIB_QPN(phdr->hwaddr));
1094                 goto unref;
1095         }
1096
1097         if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
1098                 push_pseudo_header(skb, phdr->hwaddr);
1099                 spin_lock_irqsave(&priv->lock, flags);
1100                 __skb_queue_tail(&neigh->queue, skb);
1101                 spin_unlock_irqrestore(&priv->lock, flags);
1102         } else {
1103                 ++dev->stats.tx_dropped;
1104                 dev_kfree_skb_any(skb);
1105         }
1106
1107 unref:
1108         ipoib_neigh_put(neigh);
1109
1110         return NETDEV_TX_OK;
1111 }
1112
1113 static void ipoib_timeout(struct net_device *dev)
1114 {
1115         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1116
1117         ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
1118                    jiffies_to_msecs(jiffies - dev_trans_start(dev)));
1119         ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
1120                    netif_queue_stopped(dev),
1121                    priv->tx_head, priv->tx_tail);
1122         /* XXX reset QP, etc. */
1123 }
1124
1125 static int ipoib_hard_header(struct sk_buff *skb,
1126                              struct net_device *dev,
1127                              unsigned short type,
1128                              const void *daddr, const void *saddr, unsigned len)
1129 {
1130         struct ipoib_header *header;
1131
1132         header = (struct ipoib_header *) skb_push(skb, sizeof *header);
1133
1134         header->proto = htons(type);
1135         header->reserved = 0;
1136
1137         /*
1138          * we don't rely on dst_entry structure,  always stuff the
1139          * destination address into skb hard header so we can figure out where
1140          * to send the packet later.
1141          */
1142         push_pseudo_header(skb, daddr);
1143
1144         return IPOIB_HARD_LEN;
1145 }
1146
1147 static void ipoib_set_mcast_list(struct net_device *dev)
1148 {
1149         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1150
1151         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
1152                 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
1153                 return;
1154         }
1155
1156         queue_work(priv->wq, &priv->restart_task);
1157 }
1158
1159 static int ipoib_get_iflink(const struct net_device *dev)
1160 {
1161         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1162
1163         /* parent interface */
1164         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags))
1165                 return dev->ifindex;
1166
1167         /* child/vlan interface */
1168         return priv->parent->ifindex;
1169 }
1170
1171 static u32 ipoib_addr_hash(struct ipoib_neigh_hash *htbl, u8 *daddr)
1172 {
1173         /*
1174          * Use only the address parts that contributes to spreading
1175          * The subnet prefix is not used as one can not connect to
1176          * same remote port (GUID) using the same remote QPN via two
1177          * different subnets.
1178          */
1179          /* qpn octets[1:4) & port GUID octets[12:20) */
1180         u32 *d32 = (u32 *) daddr;
1181         u32 hv;
1182
1183         hv = jhash_3words(d32[3], d32[4], IPOIB_QPN_MASK & d32[0], 0);
1184         return hv & htbl->mask;
1185 }
1186
1187 struct ipoib_neigh *ipoib_neigh_get(struct net_device *dev, u8 *daddr)
1188 {
1189         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1190         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1191         struct ipoib_neigh_hash *htbl;
1192         struct ipoib_neigh *neigh = NULL;
1193         u32 hash_val;
1194
1195         rcu_read_lock_bh();
1196
1197         htbl = rcu_dereference_bh(ntbl->htbl);
1198
1199         if (!htbl)
1200                 goto out_unlock;
1201
1202         hash_val = ipoib_addr_hash(htbl, daddr);
1203         for (neigh = rcu_dereference_bh(htbl->buckets[hash_val]);
1204              neigh != NULL;
1205              neigh = rcu_dereference_bh(neigh->hnext)) {
1206                 if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) {
1207                         /* found, take one ref on behalf of the caller */
1208                         if (!atomic_inc_not_zero(&neigh->refcnt)) {
1209                                 /* deleted */
1210                                 neigh = NULL;
1211                                 goto out_unlock;
1212                         }
1213
1214                         if (likely(skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE))
1215                                 neigh->alive = jiffies;
1216                         goto out_unlock;
1217                 }
1218         }
1219
1220 out_unlock:
1221         rcu_read_unlock_bh();
1222         return neigh;
1223 }
1224
1225 static void __ipoib_reap_neigh(struct ipoib_dev_priv *priv)
1226 {
1227         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1228         struct ipoib_neigh_hash *htbl;
1229         unsigned long neigh_obsolete;
1230         unsigned long dt;
1231         unsigned long flags;
1232         int i;
1233         LIST_HEAD(remove_list);
1234
1235         if (test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
1236                 return;
1237
1238         spin_lock_irqsave(&priv->lock, flags);
1239
1240         htbl = rcu_dereference_protected(ntbl->htbl,
1241                                          lockdep_is_held(&priv->lock));
1242
1243         if (!htbl)
1244                 goto out_unlock;
1245
1246         /* neigh is obsolete if it was idle for two GC periods */
1247         dt = 2 * arp_tbl.gc_interval;
1248         neigh_obsolete = jiffies - dt;
1249         /* handle possible race condition */
1250         if (test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
1251                 goto out_unlock;
1252
1253         for (i = 0; i < htbl->size; i++) {
1254                 struct ipoib_neigh *neigh;
1255                 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1256
1257                 while ((neigh = rcu_dereference_protected(*np,
1258                                                           lockdep_is_held(&priv->lock))) != NULL) {
1259                         /* was the neigh idle for two GC periods */
1260                         if (time_after(neigh_obsolete, neigh->alive)) {
1261
1262                                 ipoib_check_and_add_mcast_sendonly(priv, neigh->daddr + 4, &remove_list);
1263
1264                                 rcu_assign_pointer(*np,
1265                                                    rcu_dereference_protected(neigh->hnext,
1266                                                                              lockdep_is_held(&priv->lock)));
1267                                 /* remove from path/mc list */
1268                                 list_del_init(&neigh->list);
1269                                 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1270                         } else {
1271                                 np = &neigh->hnext;
1272                         }
1273
1274                 }
1275         }
1276
1277 out_unlock:
1278         spin_unlock_irqrestore(&priv->lock, flags);
1279         ipoib_mcast_remove_list(&remove_list);
1280 }
1281
1282 static void ipoib_reap_neigh(struct work_struct *work)
1283 {
1284         struct ipoib_dev_priv *priv =
1285                 container_of(work, struct ipoib_dev_priv, neigh_reap_task.work);
1286
1287         __ipoib_reap_neigh(priv);
1288
1289         if (!test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
1290                 queue_delayed_work(priv->wq, &priv->neigh_reap_task,
1291                                    arp_tbl.gc_interval);
1292 }
1293
1294
1295 static struct ipoib_neigh *ipoib_neigh_ctor(u8 *daddr,
1296                                       struct net_device *dev)
1297 {
1298         struct ipoib_neigh *neigh;
1299
1300         neigh = kzalloc(sizeof *neigh, GFP_ATOMIC);
1301         if (!neigh)
1302                 return NULL;
1303
1304         neigh->dev = dev;
1305         memcpy(&neigh->daddr, daddr, sizeof(neigh->daddr));
1306         skb_queue_head_init(&neigh->queue);
1307         INIT_LIST_HEAD(&neigh->list);
1308         ipoib_cm_set(neigh, NULL);
1309         /* one ref on behalf of the caller */
1310         atomic_set(&neigh->refcnt, 1);
1311
1312         return neigh;
1313 }
1314
1315 struct ipoib_neigh *ipoib_neigh_alloc(u8 *daddr,
1316                                       struct net_device *dev)
1317 {
1318         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1319         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1320         struct ipoib_neigh_hash *htbl;
1321         struct ipoib_neigh *neigh;
1322         u32 hash_val;
1323
1324         htbl = rcu_dereference_protected(ntbl->htbl,
1325                                          lockdep_is_held(&priv->lock));
1326         if (!htbl) {
1327                 neigh = NULL;
1328                 goto out_unlock;
1329         }
1330
1331         /* need to add a new neigh, but maybe some other thread succeeded?
1332          * recalc hash, maybe hash resize took place so we do a search
1333          */
1334         hash_val = ipoib_addr_hash(htbl, daddr);
1335         for (neigh = rcu_dereference_protected(htbl->buckets[hash_val],
1336                                                lockdep_is_held(&priv->lock));
1337              neigh != NULL;
1338              neigh = rcu_dereference_protected(neigh->hnext,
1339                                                lockdep_is_held(&priv->lock))) {
1340                 if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) {
1341                         /* found, take one ref on behalf of the caller */
1342                         if (!atomic_inc_not_zero(&neigh->refcnt)) {
1343                                 /* deleted */
1344                                 neigh = NULL;
1345                                 break;
1346                         }
1347                         neigh->alive = jiffies;
1348                         goto out_unlock;
1349                 }
1350         }
1351
1352         neigh = ipoib_neigh_ctor(daddr, dev);
1353         if (!neigh)
1354                 goto out_unlock;
1355
1356         /* one ref on behalf of the hash table */
1357         atomic_inc(&neigh->refcnt);
1358         neigh->alive = jiffies;
1359         /* put in hash */
1360         rcu_assign_pointer(neigh->hnext,
1361                            rcu_dereference_protected(htbl->buckets[hash_val],
1362                                                      lockdep_is_held(&priv->lock)));
1363         rcu_assign_pointer(htbl->buckets[hash_val], neigh);
1364         atomic_inc(&ntbl->entries);
1365
1366 out_unlock:
1367
1368         return neigh;
1369 }
1370
1371 void ipoib_neigh_dtor(struct ipoib_neigh *neigh)
1372 {
1373         /* neigh reference count was dropprd to zero */
1374         struct net_device *dev = neigh->dev;
1375         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1376         struct sk_buff *skb;
1377         if (neigh->ah)
1378                 ipoib_put_ah(neigh->ah);
1379         while ((skb = __skb_dequeue(&neigh->queue))) {
1380                 ++dev->stats.tx_dropped;
1381                 dev_kfree_skb_any(skb);
1382         }
1383         if (ipoib_cm_get(neigh))
1384                 ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
1385         ipoib_dbg(ipoib_priv(dev),
1386                   "neigh free for %06x %pI6\n",
1387                   IPOIB_QPN(neigh->daddr),
1388                   neigh->daddr + 4);
1389         kfree(neigh);
1390         if (atomic_dec_and_test(&priv->ntbl.entries)) {
1391                 if (test_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags))
1392                         complete(&priv->ntbl.flushed);
1393         }
1394 }
1395
1396 static void ipoib_neigh_reclaim(struct rcu_head *rp)
1397 {
1398         /* Called as a result of removal from hash table */
1399         struct ipoib_neigh *neigh = container_of(rp, struct ipoib_neigh, rcu);
1400         /* note TX context may hold another ref */
1401         ipoib_neigh_put(neigh);
1402 }
1403
1404 void ipoib_neigh_free(struct ipoib_neigh *neigh)
1405 {
1406         struct net_device *dev = neigh->dev;
1407         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1408         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1409         struct ipoib_neigh_hash *htbl;
1410         struct ipoib_neigh __rcu **np;
1411         struct ipoib_neigh *n;
1412         u32 hash_val;
1413
1414         htbl = rcu_dereference_protected(ntbl->htbl,
1415                                         lockdep_is_held(&priv->lock));
1416         if (!htbl)
1417                 return;
1418
1419         hash_val = ipoib_addr_hash(htbl, neigh->daddr);
1420         np = &htbl->buckets[hash_val];
1421         for (n = rcu_dereference_protected(*np,
1422                                             lockdep_is_held(&priv->lock));
1423              n != NULL;
1424              n = rcu_dereference_protected(*np,
1425                                         lockdep_is_held(&priv->lock))) {
1426                 if (n == neigh) {
1427                         /* found */
1428                         rcu_assign_pointer(*np,
1429                                            rcu_dereference_protected(neigh->hnext,
1430                                                                      lockdep_is_held(&priv->lock)));
1431                         /* remove from parent list */
1432                         list_del_init(&neigh->list);
1433                         call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1434                         return;
1435                 } else {
1436                         np = &n->hnext;
1437                 }
1438         }
1439 }
1440
1441 static int ipoib_neigh_hash_init(struct ipoib_dev_priv *priv)
1442 {
1443         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1444         struct ipoib_neigh_hash *htbl;
1445         struct ipoib_neigh __rcu **buckets;
1446         u32 size;
1447
1448         clear_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags);
1449         ntbl->htbl = NULL;
1450         htbl = kzalloc(sizeof(*htbl), GFP_KERNEL);
1451         if (!htbl)
1452                 return -ENOMEM;
1453         set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1454         size = roundup_pow_of_two(arp_tbl.gc_thresh3);
1455         buckets = kzalloc(size * sizeof(*buckets), GFP_KERNEL);
1456         if (!buckets) {
1457                 kfree(htbl);
1458                 return -ENOMEM;
1459         }
1460         htbl->size = size;
1461         htbl->mask = (size - 1);
1462         htbl->buckets = buckets;
1463         RCU_INIT_POINTER(ntbl->htbl, htbl);
1464         htbl->ntbl = ntbl;
1465         atomic_set(&ntbl->entries, 0);
1466
1467         /* start garbage collection */
1468         clear_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1469         queue_delayed_work(priv->wq, &priv->neigh_reap_task,
1470                            arp_tbl.gc_interval);
1471
1472         return 0;
1473 }
1474
1475 static void neigh_hash_free_rcu(struct rcu_head *head)
1476 {
1477         struct ipoib_neigh_hash *htbl = container_of(head,
1478                                                     struct ipoib_neigh_hash,
1479                                                     rcu);
1480         struct ipoib_neigh __rcu **buckets = htbl->buckets;
1481         struct ipoib_neigh_table *ntbl = htbl->ntbl;
1482
1483         kfree(buckets);
1484         kfree(htbl);
1485         complete(&ntbl->deleted);
1486 }
1487
1488 void ipoib_del_neighs_by_gid(struct net_device *dev, u8 *gid)
1489 {
1490         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1491         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1492         struct ipoib_neigh_hash *htbl;
1493         unsigned long flags;
1494         int i;
1495
1496         /* remove all neigh connected to a given path or mcast */
1497         spin_lock_irqsave(&priv->lock, flags);
1498
1499         htbl = rcu_dereference_protected(ntbl->htbl,
1500                                          lockdep_is_held(&priv->lock));
1501
1502         if (!htbl)
1503                 goto out_unlock;
1504
1505         for (i = 0; i < htbl->size; i++) {
1506                 struct ipoib_neigh *neigh;
1507                 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1508
1509                 while ((neigh = rcu_dereference_protected(*np,
1510                                                           lockdep_is_held(&priv->lock))) != NULL) {
1511                         /* delete neighs belong to this parent */
1512                         if (!memcmp(gid, neigh->daddr + 4, sizeof (union ib_gid))) {
1513                                 rcu_assign_pointer(*np,
1514                                                    rcu_dereference_protected(neigh->hnext,
1515                                                                              lockdep_is_held(&priv->lock)));
1516                                 /* remove from parent list */
1517                                 list_del_init(&neigh->list);
1518                                 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1519                         } else {
1520                                 np = &neigh->hnext;
1521                         }
1522
1523                 }
1524         }
1525 out_unlock:
1526         spin_unlock_irqrestore(&priv->lock, flags);
1527 }
1528
1529 static void ipoib_flush_neighs(struct ipoib_dev_priv *priv)
1530 {
1531         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1532         struct ipoib_neigh_hash *htbl;
1533         unsigned long flags;
1534         int i, wait_flushed = 0;
1535
1536         init_completion(&priv->ntbl.flushed);
1537
1538         spin_lock_irqsave(&priv->lock, flags);
1539
1540         htbl = rcu_dereference_protected(ntbl->htbl,
1541                                         lockdep_is_held(&priv->lock));
1542         if (!htbl)
1543                 goto out_unlock;
1544
1545         wait_flushed = atomic_read(&priv->ntbl.entries);
1546         if (!wait_flushed)
1547                 goto free_htbl;
1548
1549         for (i = 0; i < htbl->size; i++) {
1550                 struct ipoib_neigh *neigh;
1551                 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1552
1553                 while ((neigh = rcu_dereference_protected(*np,
1554                                        lockdep_is_held(&priv->lock))) != NULL) {
1555                         rcu_assign_pointer(*np,
1556                                            rcu_dereference_protected(neigh->hnext,
1557                                                                      lockdep_is_held(&priv->lock)));
1558                         /* remove from path/mc list */
1559                         list_del_init(&neigh->list);
1560                         call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1561                 }
1562         }
1563
1564 free_htbl:
1565         rcu_assign_pointer(ntbl->htbl, NULL);
1566         call_rcu(&htbl->rcu, neigh_hash_free_rcu);
1567
1568 out_unlock:
1569         spin_unlock_irqrestore(&priv->lock, flags);
1570         if (wait_flushed)
1571                 wait_for_completion(&priv->ntbl.flushed);
1572 }
1573
1574 static void ipoib_neigh_hash_uninit(struct net_device *dev)
1575 {
1576         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1577         int stopped;
1578
1579         ipoib_dbg(priv, "ipoib_neigh_hash_uninit\n");
1580         init_completion(&priv->ntbl.deleted);
1581         set_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags);
1582
1583         /* Stop GC if called at init fail need to cancel work */
1584         stopped = test_and_set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1585         if (!stopped)
1586                 cancel_delayed_work(&priv->neigh_reap_task);
1587
1588         ipoib_flush_neighs(priv);
1589
1590         wait_for_completion(&priv->ntbl.deleted);
1591 }
1592
1593 static void ipoib_dev_uninit_default(struct net_device *dev)
1594 {
1595         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1596
1597         ipoib_transport_dev_cleanup(dev);
1598
1599         netif_napi_del(&priv->napi);
1600
1601         ipoib_cm_dev_cleanup(dev);
1602
1603         kfree(priv->rx_ring);
1604         vfree(priv->tx_ring);
1605
1606         priv->rx_ring = NULL;
1607         priv->tx_ring = NULL;
1608 }
1609
1610 static int ipoib_dev_init_default(struct net_device *dev)
1611 {
1612         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1613
1614         netif_napi_add(dev, &priv->napi, ipoib_poll, NAPI_POLL_WEIGHT);
1615
1616         /* Allocate RX/TX "rings" to hold queued skbs */
1617         priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
1618                                 GFP_KERNEL);
1619         if (!priv->rx_ring)
1620                 goto out;
1621
1622         priv->tx_ring = vzalloc(ipoib_sendq_size * sizeof *priv->tx_ring);
1623         if (!priv->tx_ring) {
1624                 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
1625                        priv->ca->name, ipoib_sendq_size);
1626                 goto out_rx_ring_cleanup;
1627         }
1628
1629         /* priv->tx_head, tx_tail & tx_outstanding are already 0 */
1630
1631         if (ipoib_transport_dev_init(dev, priv->ca)) {
1632                 pr_warn("%s: ipoib_transport_dev_init failed\n",
1633                         priv->ca->name);
1634                 goto out_tx_ring_cleanup;
1635         }
1636
1637         /* after qp created set dev address */
1638         priv->dev->dev_addr[1] = (priv->qp->qp_num >> 16) & 0xff;
1639         priv->dev->dev_addr[2] = (priv->qp->qp_num >>  8) & 0xff;
1640         priv->dev->dev_addr[3] = (priv->qp->qp_num) & 0xff;
1641
1642         setup_timer(&priv->poll_timer, ipoib_ib_tx_timer_func,
1643                     (unsigned long)dev);
1644
1645         return 0;
1646
1647 out_tx_ring_cleanup:
1648         vfree(priv->tx_ring);
1649
1650 out_rx_ring_cleanup:
1651         kfree(priv->rx_ring);
1652
1653 out:
1654         netif_napi_del(&priv->napi);
1655         return -ENOMEM;
1656 }
1657
1658 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
1659 {
1660         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1661         int ret = -ENOMEM;
1662
1663         priv->ca = ca;
1664         priv->port = port;
1665         priv->qp = NULL;
1666
1667         /*
1668          * the various IPoIB tasks assume they will never race against
1669          * themselves, so always use a single thread workqueue
1670          */
1671         priv->wq = alloc_ordered_workqueue("ipoib_wq", WQ_MEM_RECLAIM);
1672         if (!priv->wq) {
1673                 pr_warn("%s: failed to allocate device WQ\n", dev->name);
1674                 goto out;
1675         }
1676
1677         /* create pd, which used both for control and datapath*/
1678         priv->pd = ib_alloc_pd(priv->ca, 0);
1679         if (IS_ERR(priv->pd)) {
1680                 pr_warn("%s: failed to allocate PD\n", ca->name);
1681                 goto clean_wq;
1682         }
1683
1684         ret = priv->rn_ops->ndo_init(dev);
1685         if (ret) {
1686                 pr_warn("%s failed to init HW resource\n", dev->name);
1687                 goto out_free_pd;
1688         }
1689
1690         if (ipoib_neigh_hash_init(priv) < 0) {
1691                 pr_warn("%s failed to init neigh hash\n", dev->name);
1692                 goto out_dev_uninit;
1693         }
1694
1695         if (dev->flags & IFF_UP) {
1696                 if (ipoib_ib_dev_open(dev)) {
1697                         pr_warn("%s failed to open device\n", dev->name);
1698                         ret = -ENODEV;
1699                         goto out_dev_uninit;
1700                 }
1701         }
1702
1703         return 0;
1704
1705 out_dev_uninit:
1706         ipoib_ib_dev_cleanup(dev);
1707
1708 out_free_pd:
1709         if (priv->pd) {
1710                 ib_dealloc_pd(priv->pd);
1711                 priv->pd = NULL;
1712         }
1713
1714 clean_wq:
1715         if (priv->wq) {
1716                 destroy_workqueue(priv->wq);
1717                 priv->wq = NULL;
1718         }
1719
1720 out:
1721         return ret;
1722 }
1723
1724 void ipoib_dev_cleanup(struct net_device *dev)
1725 {
1726         struct ipoib_dev_priv *priv = ipoib_priv(dev), *cpriv, *tcpriv;
1727         LIST_HEAD(head);
1728
1729         ASSERT_RTNL();
1730
1731         /* Delete any child interfaces first */
1732         list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
1733                 /* Stop GC on child */
1734                 set_bit(IPOIB_STOP_NEIGH_GC, &cpriv->flags);
1735                 cancel_delayed_work(&cpriv->neigh_reap_task);
1736                 unregister_netdevice_queue(cpriv->dev, &head);
1737         }
1738         unregister_netdevice_many(&head);
1739
1740         ipoib_neigh_hash_uninit(dev);
1741
1742         ipoib_ib_dev_cleanup(dev);
1743
1744         /* no more works over the priv->wq */
1745         if (priv->wq) {
1746                 flush_workqueue(priv->wq);
1747                 destroy_workqueue(priv->wq);
1748                 priv->wq = NULL;
1749         }
1750 }
1751
1752 static int ipoib_set_vf_link_state(struct net_device *dev, int vf, int link_state)
1753 {
1754         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1755
1756         return ib_set_vf_link_state(priv->ca, vf, priv->port, link_state);
1757 }
1758
1759 static int ipoib_get_vf_config(struct net_device *dev, int vf,
1760                                struct ifla_vf_info *ivf)
1761 {
1762         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1763         int err;
1764
1765         err = ib_get_vf_config(priv->ca, vf, priv->port, ivf);
1766         if (err)
1767                 return err;
1768
1769         ivf->vf = vf;
1770
1771         return 0;
1772 }
1773
1774 static int ipoib_set_vf_guid(struct net_device *dev, int vf, u64 guid, int type)
1775 {
1776         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1777
1778         if (type != IFLA_VF_IB_NODE_GUID && type != IFLA_VF_IB_PORT_GUID)
1779                 return -EINVAL;
1780
1781         return ib_set_vf_guid(priv->ca, vf, priv->port, guid, type);
1782 }
1783
1784 static int ipoib_get_vf_stats(struct net_device *dev, int vf,
1785                               struct ifla_vf_stats *vf_stats)
1786 {
1787         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1788
1789         return ib_get_vf_stats(priv->ca, vf, priv->port, vf_stats);
1790 }
1791
1792 static const struct header_ops ipoib_header_ops = {
1793         .create = ipoib_hard_header,
1794 };
1795
1796 static const struct net_device_ops ipoib_netdev_ops_pf = {
1797         .ndo_uninit              = ipoib_uninit,
1798         .ndo_open                = ipoib_open,
1799         .ndo_stop                = ipoib_stop,
1800         .ndo_change_mtu          = ipoib_change_mtu,
1801         .ndo_fix_features        = ipoib_fix_features,
1802         .ndo_start_xmit          = ipoib_start_xmit,
1803         .ndo_tx_timeout          = ipoib_timeout,
1804         .ndo_set_rx_mode         = ipoib_set_mcast_list,
1805         .ndo_get_iflink          = ipoib_get_iflink,
1806         .ndo_set_vf_link_state   = ipoib_set_vf_link_state,
1807         .ndo_get_vf_config       = ipoib_get_vf_config,
1808         .ndo_get_vf_stats        = ipoib_get_vf_stats,
1809         .ndo_set_vf_guid         = ipoib_set_vf_guid,
1810         .ndo_set_mac_address     = ipoib_set_mac,
1811 };
1812
1813 static const struct net_device_ops ipoib_netdev_ops_vf = {
1814         .ndo_uninit              = ipoib_uninit,
1815         .ndo_open                = ipoib_open,
1816         .ndo_stop                = ipoib_stop,
1817         .ndo_change_mtu          = ipoib_change_mtu,
1818         .ndo_fix_features        = ipoib_fix_features,
1819         .ndo_start_xmit          = ipoib_start_xmit,
1820         .ndo_tx_timeout          = ipoib_timeout,
1821         .ndo_set_rx_mode         = ipoib_set_mcast_list,
1822         .ndo_get_iflink          = ipoib_get_iflink,
1823 };
1824
1825 void ipoib_setup_common(struct net_device *dev)
1826 {
1827         dev->header_ops          = &ipoib_header_ops;
1828
1829         ipoib_set_ethtool_ops(dev);
1830
1831         dev->watchdog_timeo      = HZ;
1832
1833         dev->flags              |= IFF_BROADCAST | IFF_MULTICAST;
1834
1835         dev->hard_header_len     = IPOIB_HARD_LEN;
1836         dev->addr_len            = INFINIBAND_ALEN;
1837         dev->type                = ARPHRD_INFINIBAND;
1838         dev->tx_queue_len        = ipoib_sendq_size * 2;
1839         dev->features            = (NETIF_F_VLAN_CHALLENGED     |
1840                                     NETIF_F_HIGHDMA);
1841         netif_keep_dst(dev);
1842
1843         memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
1844 }
1845
1846 static void ipoib_build_priv(struct net_device *dev)
1847 {
1848         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1849
1850         priv->dev = dev;
1851         spin_lock_init(&priv->lock);
1852         init_rwsem(&priv->vlan_rwsem);
1853
1854         INIT_LIST_HEAD(&priv->path_list);
1855         INIT_LIST_HEAD(&priv->child_intfs);
1856         INIT_LIST_HEAD(&priv->dead_ahs);
1857         INIT_LIST_HEAD(&priv->multicast_list);
1858
1859         INIT_DELAYED_WORK(&priv->mcast_task,   ipoib_mcast_join_task);
1860         INIT_WORK(&priv->carrier_on_task, ipoib_mcast_carrier_on_task);
1861         INIT_WORK(&priv->flush_light,   ipoib_ib_dev_flush_light);
1862         INIT_WORK(&priv->flush_normal,   ipoib_ib_dev_flush_normal);
1863         INIT_WORK(&priv->flush_heavy,   ipoib_ib_dev_flush_heavy);
1864         INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
1865         INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
1866         INIT_DELAYED_WORK(&priv->neigh_reap_task, ipoib_reap_neigh);
1867 }
1868
1869 static const struct net_device_ops ipoib_netdev_default_pf = {
1870         .ndo_init                = ipoib_dev_init_default,
1871         .ndo_uninit              = ipoib_dev_uninit_default,
1872         .ndo_open                = ipoib_ib_dev_open_default,
1873         .ndo_stop                = ipoib_ib_dev_stop_default,
1874 };
1875
1876 static struct net_device
1877 *ipoib_create_netdev_default(struct ib_device *hca,
1878                              const char *name,
1879                              unsigned char name_assign_type,
1880                              void (*setup)(struct net_device *))
1881 {
1882         struct net_device *dev;
1883         struct rdma_netdev *rn;
1884
1885         dev = alloc_netdev((int)sizeof(struct rdma_netdev),
1886                            name,
1887                            name_assign_type, setup);
1888         if (!dev)
1889                 return NULL;
1890
1891         rn = netdev_priv(dev);
1892
1893         rn->send = ipoib_send;
1894         rn->attach_mcast = ipoib_mcast_attach;
1895         rn->detach_mcast = ipoib_mcast_detach;
1896         rn->free_rdma_netdev = free_netdev;
1897         rn->hca = hca;
1898
1899         dev->netdev_ops = &ipoib_netdev_default_pf;
1900
1901         return dev;
1902 }
1903
1904 static struct net_device *ipoib_get_netdev(struct ib_device *hca, u8 port,
1905                                            const char *name)
1906 {
1907         struct net_device *dev;
1908
1909         if (hca->alloc_rdma_netdev) {
1910                 dev = hca->alloc_rdma_netdev(hca, port,
1911                                              RDMA_NETDEV_IPOIB, name,
1912                                              NET_NAME_UNKNOWN,
1913                                              ipoib_setup_common);
1914                 if (IS_ERR_OR_NULL(dev) && PTR_ERR(dev) != -EOPNOTSUPP)
1915                         return NULL;
1916         }
1917
1918         if (!hca->alloc_rdma_netdev || PTR_ERR(dev) == -EOPNOTSUPP)
1919                 dev = ipoib_create_netdev_default(hca, name, NET_NAME_UNKNOWN,
1920                                                   ipoib_setup_common);
1921
1922         return dev;
1923 }
1924
1925 struct ipoib_dev_priv *ipoib_intf_alloc(struct ib_device *hca, u8 port,
1926                                         const char *name)
1927 {
1928         struct net_device *dev;
1929         struct ipoib_dev_priv *priv;
1930         struct rdma_netdev *rn;
1931
1932         priv = kzalloc(sizeof(*priv), GFP_KERNEL);
1933         if (!priv)
1934                 return NULL;
1935
1936         dev = ipoib_get_netdev(hca, port, name);
1937         if (!dev)
1938                 goto free_priv;
1939
1940         priv->rn_ops = dev->netdev_ops;
1941
1942         /* fixme : should be after the query_cap */
1943         if (priv->hca_caps & IB_DEVICE_VIRTUAL_FUNCTION)
1944                 dev->netdev_ops = &ipoib_netdev_ops_vf;
1945         else
1946                 dev->netdev_ops = &ipoib_netdev_ops_pf;
1947
1948         rn = netdev_priv(dev);
1949         rn->clnt_priv = priv;
1950         ipoib_build_priv(dev);
1951
1952         return priv;
1953 free_priv:
1954         kfree(priv);
1955         return NULL;
1956 }
1957
1958 static ssize_t show_pkey(struct device *dev,
1959                          struct device_attribute *attr, char *buf)
1960 {
1961         struct net_device *ndev = to_net_dev(dev);
1962         struct ipoib_dev_priv *priv = ipoib_priv(ndev);
1963
1964         return sprintf(buf, "0x%04x\n", priv->pkey);
1965 }
1966 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1967
1968 static ssize_t show_umcast(struct device *dev,
1969                            struct device_attribute *attr, char *buf)
1970 {
1971         struct net_device *ndev = to_net_dev(dev);
1972         struct ipoib_dev_priv *priv = ipoib_priv(ndev);
1973
1974         return sprintf(buf, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
1975 }
1976
1977 void ipoib_set_umcast(struct net_device *ndev, int umcast_val)
1978 {
1979         struct ipoib_dev_priv *priv = ipoib_priv(ndev);
1980
1981         if (umcast_val > 0) {
1982                 set_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1983                 ipoib_warn(priv, "ignoring multicast groups joined directly "
1984                                 "by userspace\n");
1985         } else
1986                 clear_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1987 }
1988
1989 static ssize_t set_umcast(struct device *dev,
1990                           struct device_attribute *attr,
1991                           const char *buf, size_t count)
1992 {
1993         unsigned long umcast_val = simple_strtoul(buf, NULL, 0);
1994
1995         ipoib_set_umcast(to_net_dev(dev), umcast_val);
1996
1997         return count;
1998 }
1999 static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast);
2000
2001 int ipoib_add_umcast_attr(struct net_device *dev)
2002 {
2003         return device_create_file(&dev->dev, &dev_attr_umcast);
2004 }
2005
2006 static void set_base_guid(struct ipoib_dev_priv *priv, union ib_gid *gid)
2007 {
2008         struct ipoib_dev_priv *child_priv;
2009         struct net_device *netdev = priv->dev;
2010
2011         netif_addr_lock_bh(netdev);
2012
2013         memcpy(&priv->local_gid.global.interface_id,
2014                &gid->global.interface_id,
2015                sizeof(gid->global.interface_id));
2016         memcpy(netdev->dev_addr + 4, &priv->local_gid, sizeof(priv->local_gid));
2017         clear_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags);
2018
2019         netif_addr_unlock_bh(netdev);
2020
2021         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
2022                 down_read(&priv->vlan_rwsem);
2023                 list_for_each_entry(child_priv, &priv->child_intfs, list)
2024                         set_base_guid(child_priv, gid);
2025                 up_read(&priv->vlan_rwsem);
2026         }
2027 }
2028
2029 static int ipoib_check_lladdr(struct net_device *dev,
2030                               struct sockaddr_storage *ss)
2031 {
2032         union ib_gid *gid = (union ib_gid *)(ss->__data + 4);
2033         int ret = 0;
2034
2035         netif_addr_lock_bh(dev);
2036
2037         /* Make sure the QPN, reserved and subnet prefix match the current
2038          * lladdr, it also makes sure the lladdr is unicast.
2039          */
2040         if (memcmp(dev->dev_addr, ss->__data,
2041                    4 + sizeof(gid->global.subnet_prefix)) ||
2042             gid->global.interface_id == 0)
2043                 ret = -EINVAL;
2044
2045         netif_addr_unlock_bh(dev);
2046
2047         return ret;
2048 }
2049
2050 static int ipoib_set_mac(struct net_device *dev, void *addr)
2051 {
2052         struct ipoib_dev_priv *priv = ipoib_priv(dev);
2053         struct sockaddr_storage *ss = addr;
2054         int ret;
2055
2056         if (!(dev->priv_flags & IFF_LIVE_ADDR_CHANGE) && netif_running(dev))
2057                 return -EBUSY;
2058
2059         ret = ipoib_check_lladdr(dev, ss);
2060         if (ret)
2061                 return ret;
2062
2063         set_base_guid(priv, (union ib_gid *)(ss->__data + 4));
2064
2065         queue_work(ipoib_workqueue, &priv->flush_light);
2066
2067         return 0;
2068 }
2069
2070 static ssize_t create_child(struct device *dev,
2071                             struct device_attribute *attr,
2072                             const char *buf, size_t count)
2073 {
2074         int pkey;
2075         int ret;
2076
2077         if (sscanf(buf, "%i", &pkey) != 1)
2078                 return -EINVAL;
2079
2080         if (pkey <= 0 || pkey > 0xffff || pkey == 0x8000)
2081                 return -EINVAL;
2082
2083         /*
2084          * Set the full membership bit, so that we join the right
2085          * broadcast group, etc.
2086          */
2087         pkey |= 0x8000;
2088
2089         ret = ipoib_vlan_add(to_net_dev(dev), pkey);
2090
2091         return ret ? ret : count;
2092 }
2093 static DEVICE_ATTR(create_child, S_IWUSR, NULL, create_child);
2094
2095 static ssize_t delete_child(struct device *dev,
2096                             struct device_attribute *attr,
2097                             const char *buf, size_t count)
2098 {
2099         int pkey;
2100         int ret;
2101
2102         if (sscanf(buf, "%i", &pkey) != 1)
2103                 return -EINVAL;
2104
2105         if (pkey < 0 || pkey > 0xffff)
2106                 return -EINVAL;
2107
2108         ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
2109
2110         return ret ? ret : count;
2111
2112 }
2113 static DEVICE_ATTR(delete_child, S_IWUSR, NULL, delete_child);
2114
2115 int ipoib_add_pkey_attr(struct net_device *dev)
2116 {
2117         return device_create_file(&dev->dev, &dev_attr_pkey);
2118 }
2119
2120 void ipoib_set_dev_features(struct ipoib_dev_priv *priv, struct ib_device *hca)
2121 {
2122         priv->hca_caps = hca->attrs.device_cap_flags;
2123
2124         if (priv->hca_caps & IB_DEVICE_UD_IP_CSUM) {
2125                 priv->dev->hw_features |= NETIF_F_IP_CSUM | NETIF_F_RXCSUM;
2126
2127                 if (priv->hca_caps & IB_DEVICE_UD_TSO)
2128                         priv->dev->hw_features |= NETIF_F_TSO;
2129
2130                 priv->dev->features |= priv->dev->hw_features;
2131         }
2132 }
2133
2134 static struct net_device *ipoib_add_port(const char *format,
2135                                          struct ib_device *hca, u8 port)
2136 {
2137         struct ipoib_dev_priv *priv;
2138         struct ib_port_attr attr;
2139         int result = -ENOMEM;
2140
2141         priv = ipoib_intf_alloc(hca, port, format);
2142         if (!priv)
2143                 goto alloc_mem_failed;
2144
2145         SET_NETDEV_DEV(priv->dev, hca->dev.parent);
2146         priv->dev->dev_id = port - 1;
2147
2148         result = ib_query_port(hca, port, &attr);
2149         if (!result)
2150                 priv->max_ib_mtu = ib_mtu_enum_to_int(attr.max_mtu);
2151         else {
2152                 printk(KERN_WARNING "%s: ib_query_port %d failed\n",
2153                        hca->name, port);
2154                 goto device_init_failed;
2155         }
2156
2157         /* MTU will be reset when mcast join happens */
2158         priv->dev->mtu  = IPOIB_UD_MTU(priv->max_ib_mtu);
2159         priv->mcast_mtu  = priv->admin_mtu = priv->dev->mtu;
2160         priv->dev->max_mtu = IPOIB_CM_MTU;
2161
2162         priv->dev->neigh_priv_len = sizeof(struct ipoib_neigh);
2163
2164         result = ib_query_pkey(hca, port, 0, &priv->pkey);
2165         if (result) {
2166                 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
2167                        hca->name, port, result);
2168                 goto device_init_failed;
2169         }
2170
2171         ipoib_set_dev_features(priv, hca);
2172
2173         /*
2174          * Set the full membership bit, so that we join the right
2175          * broadcast group, etc.
2176          */
2177         priv->pkey |= 0x8000;
2178
2179         priv->dev->broadcast[8] = priv->pkey >> 8;
2180         priv->dev->broadcast[9] = priv->pkey & 0xff;
2181
2182         result = ib_query_gid(hca, port, 0, &priv->local_gid, NULL);
2183         if (result) {
2184                 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
2185                        hca->name, port, result);
2186                 goto device_init_failed;
2187         } else
2188                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
2189         set_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags);
2190
2191         result = ipoib_dev_init(priv->dev, hca, port);
2192         if (result < 0) {
2193                 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
2194                        hca->name, port, result);
2195                 goto device_init_failed;
2196         }
2197
2198         INIT_IB_EVENT_HANDLER(&priv->event_handler,
2199                               priv->ca, ipoib_event);
2200         result = ib_register_event_handler(&priv->event_handler);
2201         if (result < 0) {
2202                 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
2203                        "port %d (ret = %d)\n",
2204                        hca->name, port, result);
2205                 goto event_failed;
2206         }
2207
2208         result = register_netdev(priv->dev);
2209         if (result) {
2210                 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
2211                        hca->name, port, result);
2212                 goto register_failed;
2213         }
2214
2215         if (ipoib_cm_add_mode_attr(priv->dev))
2216                 goto sysfs_failed;
2217         if (ipoib_add_pkey_attr(priv->dev))
2218                 goto sysfs_failed;
2219         if (ipoib_add_umcast_attr(priv->dev))
2220                 goto sysfs_failed;
2221         if (device_create_file(&priv->dev->dev, &dev_attr_create_child))
2222                 goto sysfs_failed;
2223         if (device_create_file(&priv->dev->dev, &dev_attr_delete_child))
2224                 goto sysfs_failed;
2225
2226         return priv->dev;
2227
2228 sysfs_failed:
2229         unregister_netdev(priv->dev);
2230
2231 register_failed:
2232         ib_unregister_event_handler(&priv->event_handler);
2233         flush_workqueue(ipoib_workqueue);
2234         /* Stop GC if started before flush */
2235         set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
2236         cancel_delayed_work(&priv->neigh_reap_task);
2237         flush_workqueue(priv->wq);
2238
2239 event_failed:
2240         ipoib_dev_cleanup(priv->dev);
2241
2242 device_init_failed:
2243         free_netdev(priv->dev);
2244         kfree(priv);
2245
2246 alloc_mem_failed:
2247         return ERR_PTR(result);
2248 }
2249
2250 static void ipoib_add_one(struct ib_device *device)
2251 {
2252         struct list_head *dev_list;
2253         struct net_device *dev;
2254         struct ipoib_dev_priv *priv;
2255         int p;
2256         int count = 0;
2257
2258         dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
2259         if (!dev_list)
2260                 return;
2261
2262         INIT_LIST_HEAD(dev_list);
2263
2264         for (p = rdma_start_port(device); p <= rdma_end_port(device); ++p) {
2265                 if (!rdma_protocol_ib(device, p))
2266                         continue;
2267                 dev = ipoib_add_port("ib%d", device, p);
2268                 if (!IS_ERR(dev)) {
2269                         priv = ipoib_priv(dev);
2270                         list_add_tail(&priv->list, dev_list);
2271                         count++;
2272                 }
2273         }
2274
2275         if (!count) {
2276                 kfree(dev_list);
2277                 return;
2278         }
2279
2280         ib_set_client_data(device, &ipoib_client, dev_list);
2281 }
2282
2283 static void ipoib_remove_one(struct ib_device *device, void *client_data)
2284 {
2285         struct ipoib_dev_priv *priv, *tmp, *cpriv, *tcpriv;
2286         struct list_head *dev_list = client_data;
2287
2288         if (!dev_list)
2289                 return;
2290
2291         list_for_each_entry_safe(priv, tmp, dev_list, list) {
2292                 struct rdma_netdev *rn = netdev_priv(priv->dev);
2293
2294                 ib_unregister_event_handler(&priv->event_handler);
2295                 flush_workqueue(ipoib_workqueue);
2296
2297                 /* mark interface in the middle of destruction */
2298                 set_bit(IPOIB_FLAG_GOING_DOWN, &priv->flags);
2299
2300                 rtnl_lock();
2301                 dev_change_flags(priv->dev, priv->dev->flags & ~IFF_UP);
2302                 rtnl_unlock();
2303
2304                 /* Stop GC */
2305                 set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
2306                 cancel_delayed_work(&priv->neigh_reap_task);
2307                 flush_workqueue(priv->wq);
2308
2309                 unregister_netdev(priv->dev);
2310                 rn->free_rdma_netdev(priv->dev);
2311
2312                 list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list)
2313                         kfree(cpriv);
2314
2315                 kfree(priv);
2316         }
2317
2318         kfree(dev_list);
2319 }
2320
2321 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
2322 static struct notifier_block ipoib_netdev_notifier = {
2323         .notifier_call = ipoib_netdev_event,
2324 };
2325 #endif
2326
2327 static int __init ipoib_init_module(void)
2328 {
2329         int ret;
2330
2331         ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
2332         ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
2333         ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
2334
2335         ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
2336         ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
2337         ipoib_sendq_size = max3(ipoib_sendq_size, 2 * MAX_SEND_CQE, IPOIB_MIN_QUEUE_SIZE);
2338 #ifdef CONFIG_INFINIBAND_IPOIB_CM
2339         ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
2340 #endif
2341
2342         /*
2343          * When copying small received packets, we only copy from the
2344          * linear data part of the SKB, so we rely on this condition.
2345          */
2346         BUILD_BUG_ON(IPOIB_CM_COPYBREAK > IPOIB_CM_HEAD_SIZE);
2347
2348         ret = ipoib_register_debugfs();
2349         if (ret)
2350                 return ret;
2351
2352         /*
2353          * We create a global workqueue here that is used for all flush
2354          * operations.  However, if you attempt to flush a workqueue
2355          * from a task on that same workqueue, it deadlocks the system.
2356          * We want to be able to flush the tasks associated with a
2357          * specific net device, so we also create a workqueue for each
2358          * netdevice.  We queue up the tasks for that device only on
2359          * its private workqueue, and we only queue up flush events
2360          * on our global flush workqueue.  This avoids the deadlocks.
2361          */
2362         ipoib_workqueue = alloc_ordered_workqueue("ipoib_flush",
2363                                                   WQ_MEM_RECLAIM);
2364         if (!ipoib_workqueue) {
2365                 ret = -ENOMEM;
2366                 goto err_fs;
2367         }
2368
2369         ib_sa_register_client(&ipoib_sa_client);
2370
2371         ret = ib_register_client(&ipoib_client);
2372         if (ret)
2373                 goto err_sa;
2374
2375         ret = ipoib_netlink_init();
2376         if (ret)
2377                 goto err_client;
2378
2379 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
2380         register_netdevice_notifier(&ipoib_netdev_notifier);
2381 #endif
2382         return 0;
2383
2384 err_client:
2385         ib_unregister_client(&ipoib_client);
2386
2387 err_sa:
2388         ib_sa_unregister_client(&ipoib_sa_client);
2389         destroy_workqueue(ipoib_workqueue);
2390
2391 err_fs:
2392         ipoib_unregister_debugfs();
2393
2394         return ret;
2395 }
2396
2397 static void __exit ipoib_cleanup_module(void)
2398 {
2399 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
2400         unregister_netdevice_notifier(&ipoib_netdev_notifier);
2401 #endif
2402         ipoib_netlink_fini();
2403         ib_unregister_client(&ipoib_client);
2404         ib_sa_unregister_client(&ipoib_sa_client);
2405         ipoib_unregister_debugfs();
2406         destroy_workqueue(ipoib_workqueue);
2407 }
2408
2409 module_init(ipoib_init_module);
2410 module_exit(ipoib_cleanup_module);