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
  64 int ipoib_sendq_size __read_mostly = IPOIB_TX_RING_SIZE;
  65 int ipoib_recvq_size __read_mostly = IPOIB_RX_RING_SIZE;
  66
  67 module_param_named(send_queue_size, ipoib_sendq_size, int, 0444);
  68 MODULE_PARM_DESC(send_queue_size, "Number of descriptors in send queue");
  69 module_param_named(recv_queue_size, ipoib_recvq_size, int, 0444);
  70 MODULE_PARM_DESC(recv_queue_size, "Number of descriptors in receive queue");
  71
  72 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
  73 int ipoib_debug_level;
  74
  75 module_param_named(debug_level, ipoib_debug_level, int, 0644);
  76 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
  77 #endif
  78
  79 struct ipoib_path_iter {
  80         struct net_device *dev;
  81         struct ipoib_path  path;
  82 };
  83
  84 static const u8 ipv4_bcast_addr[] = {
  85         0x00, 0xff, 0xff, 0xff,
  86         0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
  87         0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff
  88 };
  89
  90 struct workqueue_struct *ipoib_workqueue;
  91
  92 struct ib_sa_client ipoib_sa_client;
  93
  94 static void ipoib_add_one(struct ib_device *device);
  95 static void ipoib_remove_one(struct ib_device *device, void *client_data);
  96 static void ipoib_neigh_reclaim(struct rcu_head *rp);
  97 static struct net_device *ipoib_get_net_dev_by_params(
  98                 struct ib_device *dev, u8 port, u16 pkey,
  99                 const union ib_gid *gid, const struct sockaddr *addr,
 100                 void *client_data);
 101 static int ipoib_set_mac(struct net_device *dev, void *addr);
 102 static int ipoib_ioctl(struct net_device *dev, struct ifreq *ifr,
 103                        int cmd);
 104
 105 static struct ib_client ipoib_client = {
 106         .name   = "ipoib",
 107         .add    = ipoib_add_one,
 108         .remove = ipoib_remove_one,
 109         .get_net_dev_by_params = ipoib_get_net_dev_by_params,
 110 };
 111
 112 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 113 static int ipoib_netdev_event(struct notifier_block *this,
 114                               unsigned long event, void *ptr)
 115 {
 116         struct netdev_notifier_info *ni = ptr;
 117         struct net_device *dev = ni->dev;
 118
 119         if (dev->netdev_ops->ndo_open != ipoib_open)
 120                 return NOTIFY_DONE;
 121
 122         switch (event) {
 123         case NETDEV_REGISTER:
 124                 ipoib_create_debug_files(dev);
 125                 break;
 126         case NETDEV_CHANGENAME:
 127                 ipoib_delete_debug_files(dev);
 128                 ipoib_create_debug_files(dev);
 129                 break;
 130         case NETDEV_UNREGISTER:
 131                 ipoib_delete_debug_files(dev);
 132                 break;
 133         }
 134
 135         return NOTIFY_DONE;
 136 }
 137 #endif
 138
 139 int ipoib_open(struct net_device *dev)
 140 {
 141         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 142
 143         ipoib_dbg(priv, "bringing up interface\n");
 144
 145         netif_carrier_off(dev);
 146
 147         set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 148
 149         priv->sm_fullmember_sendonly_support = false;
 150
 151         if (ipoib_ib_dev_open(dev)) {
 152                 if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags))
 153                         return 0;
 154                 goto err_disable;
 155         }
 156
 157         ipoib_ib_dev_up(dev);
 158
 159         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 160                 struct ipoib_dev_priv *cpriv;
 161
 162                 /* Bring up any child interfaces too */
 163                 down_read(&priv->vlan_rwsem);
 164                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 165                         int flags;
 166
 167                         flags = cpriv->dev->flags;
 168                         if (flags & IFF_UP)
 169                                 continue;
 170
 171                         dev_change_flags(cpriv->dev, flags | IFF_UP);
 172                 }
 173                 up_read(&priv->vlan_rwsem);
 174         }
 175
 176         netif_start_queue(dev);
 177
 178         return 0;
 179
 180 err_disable:
 181         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 182
 183         return -EINVAL;
 184 }
 185
 186 static int ipoib_stop(struct net_device *dev)
 187 {
 188         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 189
 190         ipoib_dbg(priv, "stopping interface\n");
 191
 192         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 193
 194         netif_stop_queue(dev);
 195
 196         ipoib_ib_dev_down(dev);
 197         ipoib_ib_dev_stop(dev);
 198
 199         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 200                 struct ipoib_dev_priv *cpriv;
 201
 202                 /* Bring down any child interfaces too */
 203                 down_read(&priv->vlan_rwsem);
 204                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 205                         int flags;
 206
 207                         flags = cpriv->dev->flags;
 208                         if (!(flags & IFF_UP))
 209                                 continue;
 210
 211                         dev_change_flags(cpriv->dev, flags & ~IFF_UP);
 212                 }
 213                 up_read(&priv->vlan_rwsem);
 214         }
 215
 216         return 0;
 217 }
 218
 219 static void ipoib_uninit(struct net_device *dev)
 220 {
 221         ipoib_dev_cleanup(dev);
 222 }
 223
 224 static netdev_features_t ipoib_fix_features(struct net_device *dev, netdev_features_t features)
 225 {
 226         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 227
 228         if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
 229                 features &= ~(NETIF_F_IP_CSUM | NETIF_F_TSO);
 230
 231         return features;
 232 }
 233
 234 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
 235 {
 236         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 237         int ret = 0;
 238
 239         /* dev->mtu > 2K ==> connected mode */
 240         if (ipoib_cm_admin_enabled(dev)) {
 241                 if (new_mtu > ipoib_cm_max_mtu(dev))
 242                         return -EINVAL;
 243
 244                 if (new_mtu > priv->mcast_mtu)
 245                         ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
 246                                    priv->mcast_mtu);
 247
 248                 dev->mtu = new_mtu;
 249                 return 0;
 250         }
 251
 252         if (new_mtu > IPOIB_UD_MTU(priv->max_ib_mtu))
 253                 return -EINVAL;
 254
 255         priv->admin_mtu = new_mtu;
 256
 257         if (priv->mcast_mtu < priv->admin_mtu)
 258                 ipoib_dbg(priv, "MTU must be smaller than the underlying "
 259                                 "link layer MTU - 4 (%u)\n", priv->mcast_mtu);
 260
 261         new_mtu = min(priv->mcast_mtu, priv->admin_mtu);
 262
 263         if (priv->rn_ops->ndo_change_mtu) {
 264                 bool carrier_status = netif_carrier_ok(dev);
 265
 266                 netif_carrier_off(dev);
 267
 268                 /* notify lower level on the real mtu */
 269                 ret = priv->rn_ops->ndo_change_mtu(dev, new_mtu);
 270
 271                 if (carrier_status)
 272                         netif_carrier_on(dev);
 273         } else {
 274                 dev->mtu = new_mtu;
 275         }
 276
 277         return ret;
 278 }
 279
 280 static void ipoib_get_stats(struct net_device *dev,
 281                             struct rtnl_link_stats64 *stats)
 282 {
 283         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 284
 285         if (priv->rn_ops->ndo_get_stats64)
 286                 priv->rn_ops->ndo_get_stats64(dev, stats);
 287         else
 288                 netdev_stats_to_stats64(stats, &dev->stats);
 289 }
 290
 291 /* Called with an RCU read lock taken */
 292 static bool ipoib_is_dev_match_addr_rcu(const struct sockaddr *addr,
 293                                         struct net_device *dev)
 294 {
 295         struct net *net = dev_net(dev);
 296         struct in_device *in_dev;
 297         struct sockaddr_in *addr_in = (struct sockaddr_in *)addr;
 298         struct sockaddr_in6 *addr_in6 = (struct sockaddr_in6 *)addr;
 299         __be32 ret_addr;
 300
 301         switch (addr->sa_family) {
 302         case AF_INET:
 303                 in_dev = in_dev_get(dev);
 304                 if (!in_dev)
 305                         return false;
 306
 307                 ret_addr = inet_confirm_addr(net, in_dev, 0,
 308                                              addr_in->sin_addr.s_addr,
 309                                              RT_SCOPE_HOST);
 310                 in_dev_put(in_dev);
 311                 if (ret_addr)
 312                         return true;
 313
 314                 break;
 315         case AF_INET6:
 316                 if (IS_ENABLED(CONFIG_IPV6) &&
 317                     ipv6_chk_addr(net, &addr_in6->sin6_addr, dev, 1))
 318                         return true;
 319
 320                 break;
 321         }
 322         return false;
 323 }
 324
 325 /**
 326  * Find the master net_device on top of the given net_device.
 327  * @dev: base IPoIB net_device
 328  *
 329  * Returns the master net_device with a reference held, or the same net_device
 330  * if no master exists.
 331  */
 332 static struct net_device *ipoib_get_master_net_dev(struct net_device *dev)
 333 {
 334         struct net_device *master;
 335
 336         rcu_read_lock();
 337         master = netdev_master_upper_dev_get_rcu(dev);
 338         if (master)
 339                 dev_hold(master);
 340         rcu_read_unlock();
 341
 342         if (master)
 343                 return master;
 344
 345         dev_hold(dev);
 346         return dev;
 347 }
 348
 349 struct ipoib_walk_data {
 350         const struct sockaddr *addr;
 351         struct net_device *result;
 352 };
 353
 354 static int ipoib_upper_walk(struct net_device *upper, void *_data)
 355 {
 356         struct ipoib_walk_data *data = _data;
 357         int ret = 0;
 358
 359         if (ipoib_is_dev_match_addr_rcu(data->addr, upper)) {
 360                 dev_hold(upper);
 361                 data->result = upper;
 362                 ret = 1;
 363         }
 364
 365         return ret;
 366 }
 367
 368 /**
 369  * Find a net_device matching the given address, which is an upper device of
 370  * the given net_device.
 371  * @addr: IP address to look for.
 372  * @dev: base IPoIB net_device
 373  *
 374  * If found, returns the net_device with a reference held. Otherwise return
 375  * NULL.
 376  */
 377 static struct net_device *ipoib_get_net_dev_match_addr(
 378                 const struct sockaddr *addr, struct net_device *dev)
 379 {
 380         struct ipoib_walk_data data = {
 381                 .addr = addr,
 382         };
 383
 384         rcu_read_lock();
 385         if (ipoib_is_dev_match_addr_rcu(addr, dev)) {
 386                 dev_hold(dev);
 387                 data.result = dev;
 388                 goto out;
 389         }
 390
 391         netdev_walk_all_upper_dev_rcu(dev, ipoib_upper_walk, &data);
 392 out:
 393         rcu_read_unlock();
 394         return data.result;
 395 }
 396
 397 /* returns the number of IPoIB netdevs on top a given ipoib device matching a
 398  * pkey_index and address, if one exists.
 399  *
 400  * @found_net_dev: contains a matching net_device if the return value >= 1,
 401  * with a reference held. */
 402 static int ipoib_match_gid_pkey_addr(struct ipoib_dev_priv *priv,
 403                                      const union ib_gid *gid,
 404                                      u16 pkey_index,
 405                                      const struct sockaddr *addr,
 406                                      int nesting,
 407                                      struct net_device **found_net_dev)
 408 {
 409         struct ipoib_dev_priv *child_priv;
 410         struct net_device *net_dev = NULL;
 411         int matches = 0;
 412
 413         if (priv->pkey_index == pkey_index &&
 414             (!gid || !memcmp(gid, &priv->local_gid, sizeof(*gid)))) {
 415                 if (!addr) {
 416                         net_dev = ipoib_get_master_net_dev(priv->dev);
 417                 } else {
 418                         /* Verify the net_device matches the IP address, as
 419                          * IPoIB child devices currently share a GID. */
 420                         net_dev = ipoib_get_net_dev_match_addr(addr, priv->dev);
 421                 }
 422                 if (net_dev) {
 423                         if (!*found_net_dev)
 424                                 *found_net_dev = net_dev;
 425                         else
 426                                 dev_put(net_dev);
 427                         ++matches;
 428                 }
 429         }
 430
 431         /* Check child interfaces */
 432         down_read_nested(&priv->vlan_rwsem, nesting);
 433         list_for_each_entry(child_priv, &priv->child_intfs, list) {
 434                 matches += ipoib_match_gid_pkey_addr(child_priv, gid,
 435                                                     pkey_index, addr,
 436                                                     nesting + 1,
 437                                                     found_net_dev);
 438                 if (matches > 1)
 439                         break;
 440         }
 441         up_read(&priv->vlan_rwsem);
 442
 443         return matches;
 444 }
 445
 446 /* Returns the number of matching net_devs found (between 0 and 2). Also
 447  * return the matching net_device in the @net_dev parameter, holding a
 448  * reference to the net_device, if the number of matches >= 1 */
 449 static int __ipoib_get_net_dev_by_params(struct list_head *dev_list, u8 port,
 450                                          u16 pkey_index,
 451                                          const union ib_gid *gid,
 452                                          const struct sockaddr *addr,
 453                                          struct net_device **net_dev)
 454 {
 455         struct ipoib_dev_priv *priv;
 456         int matches = 0;
 457
 458         *net_dev = NULL;
 459
 460         list_for_each_entry(priv, dev_list, list) {
 461                 if (priv->port != port)
 462                         continue;
 463
 464                 matches += ipoib_match_gid_pkey_addr(priv, gid, pkey_index,
 465                                                      addr, 0, net_dev);
 466                 if (matches > 1)
 467                         break;
 468         }
 469
 470         return matches;
 471 }
 472
 473 static struct net_device *ipoib_get_net_dev_by_params(
 474                 struct ib_device *dev, u8 port, u16 pkey,
 475                 const union ib_gid *gid, const struct sockaddr *addr,
 476                 void *client_data)
 477 {
 478         struct net_device *net_dev;
 479         struct list_head *dev_list = client_data;
 480         u16 pkey_index;
 481         int matches;
 482         int ret;
 483
 484         if (!rdma_protocol_ib(dev, port))
 485                 return NULL;
 486
 487         ret = ib_find_cached_pkey(dev, port, pkey, &pkey_index);
 488         if (ret)
 489                 return NULL;
 490
 491         if (!dev_list)
 492                 return NULL;
 493
 494         /* See if we can find a unique device matching the L2 parameters */
 495         matches = __ipoib_get_net_dev_by_params(dev_list, port, pkey_index,
 496                                                 gid, NULL, &net_dev);
 497
 498         switch (matches) {
 499         case 0:
 500                 return NULL;
 501         case 1:
 502                 return net_dev;
 503         }
 504
 505         dev_put(net_dev);
 506
 507         /* Couldn't find a unique device with L2 parameters only. Use L3
 508          * address to uniquely match the net device */
 509         matches = __ipoib_get_net_dev_by_params(dev_list, port, pkey_index,
 510                                                 gid, addr, &net_dev);
 511         switch (matches) {
 512         case 0:
 513                 return NULL;
 514         default:
 515                 dev_warn_ratelimited(&dev->dev,
 516                                      "duplicate IP address detected\n");
 517                 /* Fall through */
 518         case 1:
 519                 return net_dev;
 520         }
 521 }
 522
 523 int ipoib_set_mode(struct net_device *dev, const char *buf)
 524 {
 525         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 526
 527         if ((test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags) &&
 528              !strcmp(buf, "connected\n")) ||
 529              (!test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags) &&
 530              !strcmp(buf, "datagram\n"))) {
 531                 return 0;
 532         }
 533
 534         /* flush paths if we switch modes so that connections are restarted */
 535         if (IPOIB_CM_SUPPORTED(dev->dev_addr) && !strcmp(buf, "connected\n")) {
 536                 set_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
 537                 ipoib_warn(priv, "enabling connected mode "
 538                            "will cause multicast packet drops\n");
 539                 netdev_update_features(dev);
 540                 dev_set_mtu(dev, ipoib_cm_max_mtu(dev));
 541                 rtnl_unlock();
 542                 priv->tx_wr.wr.send_flags &= ~IB_SEND_IP_CSUM;
 543
 544                 ipoib_flush_paths(dev);
 545                 return (!rtnl_trylock()) ? -EBUSY : 0;
 546         }
 547
 548         if (!strcmp(buf, "datagram\n")) {
 549                 clear_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
 550                 netdev_update_features(dev);
 551                 dev_set_mtu(dev, min(priv->mcast_mtu, dev->mtu));
 552                 rtnl_unlock();
 553                 ipoib_flush_paths(dev);
 554                 return (!rtnl_trylock()) ? -EBUSY : 0;
 555         }
 556
 557         return -EINVAL;
 558 }
 559
 560 struct ipoib_path *__path_find(struct net_device *dev, void *gid)
 561 {
 562         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 563         struct rb_node *n = priv->path_tree.rb_node;
 564         struct ipoib_path *path;
 565         int ret;
 566
 567         while (n) {
 568                 path = rb_entry(n, struct ipoib_path, rb_node);
 569
 570                 ret = memcmp(gid, path->pathrec.dgid.raw,
 571                              sizeof (union ib_gid));
 572
 573                 if (ret < 0)
 574                         n = n->rb_left;
 575                 else if (ret > 0)
 576                         n = n->rb_right;
 577                 else
 578                         return path;
 579         }
 580
 581         return NULL;
 582 }
 583
 584 static int __path_add(struct net_device *dev, struct ipoib_path *path)
 585 {
 586         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 587         struct rb_node **n = &priv->path_tree.rb_node;
 588         struct rb_node *pn = NULL;
 589         struct ipoib_path *tpath;
 590         int ret;
 591
 592         while (*n) {
 593                 pn = *n;
 594                 tpath = rb_entry(pn, struct ipoib_path, rb_node);
 595
 596                 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
 597                              sizeof (union ib_gid));
 598                 if (ret < 0)
 599                         n = &pn->rb_left;
 600                 else if (ret > 0)
 601                         n = &pn->rb_right;
 602                 else
 603                         return -EEXIST;
 604         }
 605
 606         rb_link_node(&path->rb_node, pn, n);
 607         rb_insert_color(&path->rb_node, &priv->path_tree);
 608
 609         list_add_tail(&path->list, &priv->path_list);
 610
 611         return 0;
 612 }
 613
 614 static void path_free(struct net_device *dev, struct ipoib_path *path)
 615 {
 616         struct sk_buff *skb;
 617
 618         while ((skb = __skb_dequeue(&path->queue)))
 619                 dev_kfree_skb_irq(skb);
 620
 621         ipoib_dbg(ipoib_priv(dev), "path_free\n");
 622
 623         /* remove all neigh connected to this path */
 624         ipoib_del_neighs_by_gid(dev, path->pathrec.dgid.raw);
 625
 626         if (path->ah)
 627                 ipoib_put_ah(path->ah);
 628
 629         kfree(path);
 630 }
 631
 632 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 633
 634 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
 635 {
 636         struct ipoib_path_iter *iter;
 637
 638         iter = kmalloc(sizeof *iter, GFP_KERNEL);
 639         if (!iter)
 640                 return NULL;
 641
 642         iter->dev = dev;
 643         memset(iter->path.pathrec.dgid.raw, 0, 16);
 644
 645         if (ipoib_path_iter_next(iter)) {
 646                 kfree(iter);
 647                 return NULL;
 648         }
 649
 650         return iter;
 651 }
 652
 653 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
 654 {
 655         struct ipoib_dev_priv *priv = ipoib_priv(iter->dev);
 656         struct rb_node *n;
 657         struct ipoib_path *path;
 658         int ret = 1;
 659
 660         spin_lock_irq(&priv->lock);
 661
 662         n = rb_first(&priv->path_tree);
 663
 664         while (n) {
 665                 path = rb_entry(n, struct ipoib_path, rb_node);
 666
 667                 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
 668                            sizeof (union ib_gid)) < 0) {
 669                         iter->path = *path;
 670                         ret = 0;
 671                         break;
 672                 }
 673
 674                 n = rb_next(n);
 675         }
 676
 677         spin_unlock_irq(&priv->lock);
 678
 679         return ret;
 680 }
 681
 682 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
 683                           struct ipoib_path *path)
 684 {
 685         *path = iter->path;
 686 }
 687
 688 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
 689
 690 void ipoib_mark_paths_invalid(struct net_device *dev)
 691 {
 692         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 693         struct ipoib_path *path, *tp;
 694
 695         spin_lock_irq(&priv->lock);
 696
 697         list_for_each_entry_safe(path, tp, &priv->path_list, list) {
 698                 ipoib_dbg(priv, "mark path LID 0x%08x GID %pI6 invalid\n",
 699                           be32_to_cpu(sa_path_get_dlid(&path->pathrec)),
 700                           path->pathrec.dgid.raw);
 701                 path->valid =  0;
 702         }
 703
 704         spin_unlock_irq(&priv->lock);
 705 }
 706
 707 static void push_pseudo_header(struct sk_buff *skb, const char *daddr)
 708 {
 709         struct ipoib_pseudo_header *phdr;
 710
 711         phdr = skb_push(skb, sizeof(*phdr));
 712         memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
 713 }
 714
 715 void ipoib_flush_paths(struct net_device *dev)
 716 {
 717         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 718         struct ipoib_path *path, *tp;
 719         LIST_HEAD(remove_list);
 720         unsigned long flags;
 721
 722         netif_tx_lock_bh(dev);
 723         spin_lock_irqsave(&priv->lock, flags);
 724
 725         list_splice_init(&priv->path_list, &remove_list);
 726
 727         list_for_each_entry(path, &remove_list, list)
 728                 rb_erase(&path->rb_node, &priv->path_tree);
 729
 730         list_for_each_entry_safe(path, tp, &remove_list, list) {
 731                 if (path->query)
 732                         ib_sa_cancel_query(path->query_id, path->query);
 733                 spin_unlock_irqrestore(&priv->lock, flags);
 734                 netif_tx_unlock_bh(dev);
 735                 wait_for_completion(&path->done);
 736                 path_free(dev, path);
 737                 netif_tx_lock_bh(dev);
 738                 spin_lock_irqsave(&priv->lock, flags);
 739         }
 740
 741         spin_unlock_irqrestore(&priv->lock, flags);
 742         netif_tx_unlock_bh(dev);
 743 }
 744
 745 static void path_rec_completion(int status,
 746                                 struct sa_path_rec *pathrec,
 747                                 void *path_ptr)
 748 {
 749         struct ipoib_path *path = path_ptr;
 750         struct net_device *dev = path->dev;
 751         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 752         struct ipoib_ah *ah = NULL;
 753         struct ipoib_ah *old_ah = NULL;
 754         struct ipoib_neigh *neigh, *tn;
 755         struct sk_buff_head skqueue;
 756         struct sk_buff *skb;
 757         unsigned long flags;
 758
 759         if (!status)
 760                 ipoib_dbg(priv, "PathRec LID 0x%04x for GID %pI6\n",
 761                           be32_to_cpu(sa_path_get_dlid(pathrec)),
 762                           pathrec->dgid.raw);
 763         else
 764                 ipoib_dbg(priv, "PathRec status %d for GID %pI6\n",
 765                           status, path->pathrec.dgid.raw);
 766
 767         skb_queue_head_init(&skqueue);
 768
 769         if (!status) {
 770                 struct rdma_ah_attr av;
 771
 772                 if (!ib_init_ah_from_path(priv->ca, priv->port, pathrec, &av))
 773                         ah = ipoib_create_ah(dev, priv->pd, &av);
 774         }
 775
 776         spin_lock_irqsave(&priv->lock, flags);
 777
 778         if (!IS_ERR_OR_NULL(ah)) {
 779                 path->pathrec = *pathrec;
 780
 781                 old_ah   = path->ah;
 782                 path->ah = ah;
 783
 784                 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
 785                           ah, be32_to_cpu(sa_path_get_dlid(pathrec)),
 786                           pathrec->sl);
 787
 788                 while ((skb = __skb_dequeue(&path->queue)))
 789                         __skb_queue_tail(&skqueue, skb);
 790
 791                 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
 792                         if (neigh->ah) {
 793                                 WARN_ON(neigh->ah != old_ah);
 794                                 /*
 795                                  * Dropping the ah reference inside
 796                                  * priv->lock is safe here, because we
 797                                  * will hold one more reference from
 798                                  * the original value of path->ah (ie
 799                                  * old_ah).
 800                                  */
 801                                 ipoib_put_ah(neigh->ah);
 802                         }
 803                         kref_get(&path->ah->ref);
 804                         neigh->ah = path->ah;
 805
 806                         if (ipoib_cm_enabled(dev, neigh->daddr)) {
 807                                 if (!ipoib_cm_get(neigh))
 808                                         ipoib_cm_set(neigh, ipoib_cm_create_tx(dev,
 809                                                                                path,
 810                                                                                neigh));
 811                                 if (!ipoib_cm_get(neigh)) {
 812                                         ipoib_neigh_free(neigh);
 813                                         continue;
 814                                 }
 815                         }
 816
 817                         while ((skb = __skb_dequeue(&neigh->queue)))
 818                                 __skb_queue_tail(&skqueue, skb);
 819                 }
 820                 path->valid = 1;
 821         }
 822
 823         path->query = NULL;
 824         complete(&path->done);
 825
 826         spin_unlock_irqrestore(&priv->lock, flags);
 827
 828         if (IS_ERR_OR_NULL(ah))
 829                 ipoib_del_neighs_by_gid(dev, path->pathrec.dgid.raw);
 830
 831         if (old_ah)
 832                 ipoib_put_ah(old_ah);
 833
 834         while ((skb = __skb_dequeue(&skqueue))) {
 835                 int ret;
 836                 skb->dev = dev;
 837                 ret = dev_queue_xmit(skb);
 838                 if (ret)
 839                         ipoib_warn(priv, "%s: dev_queue_xmit failed to re-queue packet, ret:%d\n",
 840                                    __func__, ret);
 841         }
 842 }
 843
 844 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
 845 {
 846         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 847         struct ipoib_path *path;
 848
 849         if (!priv->broadcast)
 850                 return NULL;
 851
 852         path = kzalloc(sizeof *path, GFP_ATOMIC);
 853         if (!path)
 854                 return NULL;
 855
 856         path->dev = dev;
 857
 858         skb_queue_head_init(&path->queue);
 859
 860         INIT_LIST_HEAD(&path->neigh_list);
 861
 862         if (rdma_cap_opa_ah(priv->ca, priv->port))
 863                 path->pathrec.rec_type = SA_PATH_REC_TYPE_OPA;
 864         else
 865                 path->pathrec.rec_type = SA_PATH_REC_TYPE_IB;
 866         memcpy(path->pathrec.dgid.raw, gid, sizeof (union ib_gid));
 867         path->pathrec.sgid          = priv->local_gid;
 868         path->pathrec.pkey          = cpu_to_be16(priv->pkey);
 869         path->pathrec.numb_path     = 1;
 870         path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
 871
 872         return path;
 873 }
 874
 875 static int path_rec_start(struct net_device *dev,
 876                           struct ipoib_path *path)
 877 {
 878         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 879
 880         ipoib_dbg(priv, "Start path record lookup for %pI6\n",
 881                   path->pathrec.dgid.raw);
 882
 883         init_completion(&path->done);
 884
 885         path->query_id =
 886                 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
 887                                    &path->pathrec,
 888                                    IB_SA_PATH_REC_DGID          |
 889                                    IB_SA_PATH_REC_SGID          |
 890                                    IB_SA_PATH_REC_NUMB_PATH     |
 891                                    IB_SA_PATH_REC_TRAFFIC_CLASS |
 892                                    IB_SA_PATH_REC_PKEY,
 893                                    1000, GFP_ATOMIC,
 894                                    path_rec_completion,
 895                                    path, &path->query);
 896         if (path->query_id < 0) {
 897                 ipoib_warn(priv, "ib_sa_path_rec_get failed: %d\n", path->query_id);
 898                 path->query = NULL;
 899                 complete(&path->done);
 900                 return path->query_id;
 901         }
 902
 903         return 0;
 904 }
 905
 906 static void neigh_add_path(struct sk_buff *skb, u8 *daddr,
 907                            struct net_device *dev)
 908 {
 909         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 910         struct rdma_netdev *rn = netdev_priv(dev);
 911         struct ipoib_path *path;
 912         struct ipoib_neigh *neigh;
 913         unsigned long flags;
 914
 915         spin_lock_irqsave(&priv->lock, flags);
 916         neigh = ipoib_neigh_alloc(daddr, dev);
 917         if (!neigh) {
 918                 spin_unlock_irqrestore(&priv->lock, flags);
 919                 ++dev->stats.tx_dropped;
 920                 dev_kfree_skb_any(skb);
 921                 return;
 922         }
 923
 924         path = __path_find(dev, daddr + 4);
 925         if (!path) {
 926                 path = path_rec_create(dev, daddr + 4);
 927                 if (!path)
 928                         goto err_path;
 929
 930                 __path_add(dev, path);
 931         }
 932
 933         list_add_tail(&neigh->list, &path->neigh_list);
 934
 935         if (path->ah) {
 936                 kref_get(&path->ah->ref);
 937                 neigh->ah = path->ah;
 938
 939                 if (ipoib_cm_enabled(dev, neigh->daddr)) {
 940                         if (!ipoib_cm_get(neigh))
 941                                 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
 942                         if (!ipoib_cm_get(neigh)) {
 943                                 ipoib_neigh_free(neigh);
 944                                 goto err_drop;
 945                         }
 946                         if (skb_queue_len(&neigh->queue) <
 947                             IPOIB_MAX_PATH_REC_QUEUE) {
 948                                 push_pseudo_header(skb, neigh->daddr);
 949                                 __skb_queue_tail(&neigh->queue, skb);
 950                         } else {
 951                                 ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
 952                                            skb_queue_len(&neigh->queue));
 953                                 goto err_drop;
 954                         }
 955                 } else {
 956                         spin_unlock_irqrestore(&priv->lock, flags);
 957                         path->ah->last_send = rn->send(dev, skb, path->ah->ah,
 958                                                        IPOIB_QPN(daddr));
 959                         ipoib_neigh_put(neigh);
 960                         return;
 961                 }
 962         } else {
 963                 neigh->ah  = NULL;
 964
 965                 if (!path->query && path_rec_start(dev, path))
 966                         goto err_path;
 967                 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 968                         push_pseudo_header(skb, neigh->daddr);
 969                         __skb_queue_tail(&neigh->queue, skb);
 970                 } else {
 971                         goto err_drop;
 972                 }
 973         }
 974
 975         spin_unlock_irqrestore(&priv->lock, flags);
 976         ipoib_neigh_put(neigh);
 977         return;
 978
 979 err_path:
 980         ipoib_neigh_free(neigh);
 981 err_drop:
 982         ++dev->stats.tx_dropped;
 983         dev_kfree_skb_any(skb);
 984
 985         spin_unlock_irqrestore(&priv->lock, flags);
 986         ipoib_neigh_put(neigh);
 987 }
 988
 989 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
 990                              struct ipoib_pseudo_header *phdr)
 991 {
 992         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 993         struct rdma_netdev *rn = netdev_priv(dev);
 994         struct ipoib_path *path;
 995         unsigned long flags;
 996
 997         spin_lock_irqsave(&priv->lock, flags);
 998
 999         path = __path_find(dev, phdr->hwaddr + 4);
1000         if (!path || !path->valid) {
1001                 int new_path = 0;
1002
1003                 if (!path) {
1004                         path = path_rec_create(dev, phdr->hwaddr + 4);
1005                         new_path = 1;
1006                 }
1007                 if (path) {
1008                         if (skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
1009                                 push_pseudo_header(skb, phdr->hwaddr);
1010                                 __skb_queue_tail(&path->queue, skb);
1011                         } else {
1012                                 ++dev->stats.tx_dropped;
1013                                 dev_kfree_skb_any(skb);
1014                         }
1015
1016                         if (!path->query && path_rec_start(dev, path)) {
1017                                 spin_unlock_irqrestore(&priv->lock, flags);
1018                                 if (new_path)
1019                                         path_free(dev, path);
1020                                 return;
1021                         } else
1022                                 __path_add(dev, path);
1023                 } else {
1024                         ++dev->stats.tx_dropped;
1025                         dev_kfree_skb_any(skb);
1026                 }
1027
1028                 spin_unlock_irqrestore(&priv->lock, flags);
1029                 return;
1030         }
1031
1032         if (path->ah) {
1033                 ipoib_dbg(priv, "Send unicast ARP to %08x\n",
1034                           be32_to_cpu(sa_path_get_dlid(&path->pathrec)));
1035
1036                 spin_unlock_irqrestore(&priv->lock, flags);
1037                 path->ah->last_send = rn->send(dev, skb, path->ah->ah,
1038                                                IPOIB_QPN(phdr->hwaddr));
1039                 return;
1040         } else if ((path->query || !path_rec_start(dev, path)) &&
1041                    skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
1042                 push_pseudo_header(skb, phdr->hwaddr);
1043                 __skb_queue_tail(&path->queue, skb);
1044         } else {
1045                 ++dev->stats.tx_dropped;
1046                 dev_kfree_skb_any(skb);
1047         }
1048
1049         spin_unlock_irqrestore(&priv->lock, flags);
1050 }
1051
1052 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
1053 {
1054         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1055         struct rdma_netdev *rn = netdev_priv(dev);
1056         struct ipoib_neigh *neigh;
1057         struct ipoib_pseudo_header *phdr;
1058         struct ipoib_header *header;
1059         unsigned long flags;
1060
1061         phdr = (struct ipoib_pseudo_header *) skb->data;
1062         skb_pull(skb, sizeof(*phdr));
1063         header = (struct ipoib_header *) skb->data;
1064
1065         if (unlikely(phdr->hwaddr[4] == 0xff)) {
1066                 /* multicast, arrange "if" according to probability */
1067                 if ((header->proto != htons(ETH_P_IP)) &&
1068                     (header->proto != htons(ETH_P_IPV6)) &&
1069                     (header->proto != htons(ETH_P_ARP)) &&
1070                     (header->proto != htons(ETH_P_RARP)) &&
1071                     (header->proto != htons(ETH_P_TIPC))) {
1072                         /* ethertype not supported by IPoIB */
1073                         ++dev->stats.tx_dropped;
1074                         dev_kfree_skb_any(skb);
1075                         return NETDEV_TX_OK;
1076                 }
1077                 /* Add in the P_Key for multicast*/
1078                 phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
1079                 phdr->hwaddr[9] = priv->pkey & 0xff;
1080
1081                 neigh = ipoib_neigh_get(dev, phdr->hwaddr);
1082                 if (likely(neigh))
1083                         goto send_using_neigh;
1084                 ipoib_mcast_send(dev, phdr->hwaddr, skb);
1085                 return NETDEV_TX_OK;
1086         }
1087
1088         /* unicast, arrange "switch" according to probability */
1089         switch (header->proto) {
1090         case htons(ETH_P_IP):
1091         case htons(ETH_P_IPV6):
1092         case htons(ETH_P_TIPC):
1093                 neigh = ipoib_neigh_get(dev, phdr->hwaddr);
1094                 if (unlikely(!neigh)) {
1095                         neigh_add_path(skb, phdr->hwaddr, dev);
1096                         return NETDEV_TX_OK;
1097                 }
1098                 break;
1099         case htons(ETH_P_ARP):
1100         case htons(ETH_P_RARP):
1101                 /* for unicast ARP and RARP should always perform path find */
1102                 unicast_arp_send(skb, dev, phdr);
1103                 return NETDEV_TX_OK;
1104         default:
1105                 /* ethertype not supported by IPoIB */
1106                 ++dev->stats.tx_dropped;
1107                 dev_kfree_skb_any(skb);
1108                 return NETDEV_TX_OK;
1109         }
1110
1111 send_using_neigh:
1112         /* note we now hold a ref to neigh */
1113         if (ipoib_cm_get(neigh)) {
1114                 if (ipoib_cm_up(neigh)) {
1115                         ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
1116                         goto unref;
1117                 }
1118         } else if (neigh->ah) {
1119                 neigh->ah->last_send = rn->send(dev, skb, neigh->ah->ah,
1120                                                 IPOIB_QPN(phdr->hwaddr));
1121                 goto unref;
1122         }
1123
1124         if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
1125                 push_pseudo_header(skb, phdr->hwaddr);
1126                 spin_lock_irqsave(&priv->lock, flags);
1127                 __skb_queue_tail(&neigh->queue, skb);
1128                 spin_unlock_irqrestore(&priv->lock, flags);
1129         } else {
1130                 ++dev->stats.tx_dropped;
1131                 dev_kfree_skb_any(skb);
1132         }
1133
1134 unref:
1135         ipoib_neigh_put(neigh);
1136
1137         return NETDEV_TX_OK;
1138 }
1139
1140 static void ipoib_timeout(struct net_device *dev)
1141 {
1142         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1143
1144         ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
1145                    jiffies_to_msecs(jiffies - dev_trans_start(dev)));
1146         ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
1147                    netif_queue_stopped(dev),
1148                    priv->tx_head, priv->tx_tail);
1149         /* XXX reset QP, etc. */
1150 }
1151
1152 static int ipoib_hard_header(struct sk_buff *skb,
1153                              struct net_device *dev,
1154                              unsigned short type,
1155                              const void *daddr, const void *saddr, unsigned len)
1156 {
1157         struct ipoib_header *header;
1158
1159         header = skb_push(skb, sizeof *header);
1160
1161         header->proto = htons(type);
1162         header->reserved = 0;
1163
1164         /*
1165          * we don't rely on dst_entry structure,  always stuff the
1166          * destination address into skb hard header so we can figure out where
1167          * to send the packet later.
1168          */
1169         push_pseudo_header(skb, daddr);
1170
1171         return IPOIB_HARD_LEN;
1172 }
1173
1174 static void ipoib_set_mcast_list(struct net_device *dev)
1175 {
1176         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1177
1178         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
1179                 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
1180                 return;
1181         }
1182
1183         queue_work(priv->wq, &priv->restart_task);
1184 }
1185
1186 static int ipoib_get_iflink(const struct net_device *dev)
1187 {
1188         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1189
1190         /* parent interface */
1191         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags))
1192                 return dev->ifindex;
1193
1194         /* child/vlan interface */
1195         return priv->parent->ifindex;
1196 }
1197
1198 static u32 ipoib_addr_hash(struct ipoib_neigh_hash *htbl, u8 *daddr)
1199 {
1200         /*
1201          * Use only the address parts that contributes to spreading
1202          * The subnet prefix is not used as one can not connect to
1203          * same remote port (GUID) using the same remote QPN via two
1204          * different subnets.
1205          */
1206          /* qpn octets[1:4) & port GUID octets[12:20) */
1207         u32 *d32 = (u32 *) daddr;
1208         u32 hv;
1209
1210         hv = jhash_3words(d32[3], d32[4], IPOIB_QPN_MASK & d32[0], 0);
1211         return hv & htbl->mask;
1212 }
1213
1214 struct ipoib_neigh *ipoib_neigh_get(struct net_device *dev, u8 *daddr)
1215 {
1216         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1217         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1218         struct ipoib_neigh_hash *htbl;
1219         struct ipoib_neigh *neigh = NULL;
1220         u32 hash_val;
1221
1222         rcu_read_lock_bh();
1223
1224         htbl = rcu_dereference_bh(ntbl->htbl);
1225
1226         if (!htbl)
1227                 goto out_unlock;
1228
1229         hash_val = ipoib_addr_hash(htbl, daddr);
1230         for (neigh = rcu_dereference_bh(htbl->buckets[hash_val]);
1231              neigh != NULL;
1232              neigh = rcu_dereference_bh(neigh->hnext)) {
1233                 if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) {
1234                         /* found, take one ref on behalf of the caller */
1235                         if (!atomic_inc_not_zero(&neigh->refcnt)) {
1236                                 /* deleted */
1237                                 neigh = NULL;
1238                                 goto out_unlock;
1239                         }
1240
1241                         if (likely(skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE))
1242                                 neigh->alive = jiffies;
1243                         goto out_unlock;
1244                 }
1245         }
1246
1247 out_unlock:
1248         rcu_read_unlock_bh();
1249         return neigh;
1250 }
1251
1252 static void __ipoib_reap_neigh(struct ipoib_dev_priv *priv)
1253 {
1254         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1255         struct ipoib_neigh_hash *htbl;
1256         unsigned long neigh_obsolete;
1257         unsigned long dt;
1258         unsigned long flags;
1259         int i;
1260         LIST_HEAD(remove_list);
1261
1262         if (test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
1263                 return;
1264
1265         spin_lock_irqsave(&priv->lock, flags);
1266
1267         htbl = rcu_dereference_protected(ntbl->htbl,
1268                                          lockdep_is_held(&priv->lock));
1269
1270         if (!htbl)
1271                 goto out_unlock;
1272
1273         /* neigh is obsolete if it was idle for two GC periods */
1274         dt = 2 * arp_tbl.gc_interval;
1275         neigh_obsolete = jiffies - dt;
1276         /* handle possible race condition */
1277         if (test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
1278                 goto out_unlock;
1279
1280         for (i = 0; i < htbl->size; i++) {
1281                 struct ipoib_neigh *neigh;
1282                 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1283
1284                 while ((neigh = rcu_dereference_protected(*np,
1285                                                           lockdep_is_held(&priv->lock))) != NULL) {
1286                         /* was the neigh idle for two GC periods */
1287                         if (time_after(neigh_obsolete, neigh->alive)) {
1288
1289                                 ipoib_check_and_add_mcast_sendonly(priv, neigh->daddr + 4, &remove_list);
1290
1291                                 rcu_assign_pointer(*np,
1292                                                    rcu_dereference_protected(neigh->hnext,
1293                                                                              lockdep_is_held(&priv->lock)));
1294                                 /* remove from path/mc list */
1295                                 list_del_init(&neigh->list);
1296                                 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1297                         } else {
1298                                 np = &neigh->hnext;
1299                         }
1300
1301                 }
1302         }
1303
1304 out_unlock:
1305         spin_unlock_irqrestore(&priv->lock, flags);
1306         ipoib_mcast_remove_list(&remove_list);
1307 }
1308
1309 static void ipoib_reap_neigh(struct work_struct *work)
1310 {
1311         struct ipoib_dev_priv *priv =
1312                 container_of(work, struct ipoib_dev_priv, neigh_reap_task.work);
1313
1314         __ipoib_reap_neigh(priv);
1315
1316         if (!test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
1317                 queue_delayed_work(priv->wq, &priv->neigh_reap_task,
1318                                    arp_tbl.gc_interval);
1319 }
1320
1321
1322 static struct ipoib_neigh *ipoib_neigh_ctor(u8 *daddr,
1323                                       struct net_device *dev)
1324 {
1325         struct ipoib_neigh *neigh;
1326
1327         neigh = kzalloc(sizeof *neigh, GFP_ATOMIC);
1328         if (!neigh)
1329                 return NULL;
1330
1331         neigh->dev = dev;
1332         memcpy(&neigh->daddr, daddr, sizeof(neigh->daddr));
1333         skb_queue_head_init(&neigh->queue);
1334         INIT_LIST_HEAD(&neigh->list);
1335         ipoib_cm_set(neigh, NULL);
1336         /* one ref on behalf of the caller */
1337         atomic_set(&neigh->refcnt, 1);
1338
1339         return neigh;
1340 }
1341
1342 struct ipoib_neigh *ipoib_neigh_alloc(u8 *daddr,
1343                                       struct net_device *dev)
1344 {
1345         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1346         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1347         struct ipoib_neigh_hash *htbl;
1348         struct ipoib_neigh *neigh;
1349         u32 hash_val;
1350
1351         htbl = rcu_dereference_protected(ntbl->htbl,
1352                                          lockdep_is_held(&priv->lock));
1353         if (!htbl) {
1354                 neigh = NULL;
1355                 goto out_unlock;
1356         }
1357
1358         /* need to add a new neigh, but maybe some other thread succeeded?
1359          * recalc hash, maybe hash resize took place so we do a search
1360          */
1361         hash_val = ipoib_addr_hash(htbl, daddr);
1362         for (neigh = rcu_dereference_protected(htbl->buckets[hash_val],
1363                                                lockdep_is_held(&priv->lock));
1364              neigh != NULL;
1365              neigh = rcu_dereference_protected(neigh->hnext,
1366                                                lockdep_is_held(&priv->lock))) {
1367                 if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) {
1368                         /* found, take one ref on behalf of the caller */
1369                         if (!atomic_inc_not_zero(&neigh->refcnt)) {
1370                                 /* deleted */
1371                                 neigh = NULL;
1372                                 break;
1373                         }
1374                         neigh->alive = jiffies;
1375                         goto out_unlock;
1376                 }
1377         }
1378
1379         neigh = ipoib_neigh_ctor(daddr, dev);
1380         if (!neigh)
1381                 goto out_unlock;
1382
1383         /* one ref on behalf of the hash table */
1384         atomic_inc(&neigh->refcnt);
1385         neigh->alive = jiffies;
1386         /* put in hash */
1387         rcu_assign_pointer(neigh->hnext,
1388                            rcu_dereference_protected(htbl->buckets[hash_val],
1389                                                      lockdep_is_held(&priv->lock)));
1390         rcu_assign_pointer(htbl->buckets[hash_val], neigh);
1391         atomic_inc(&ntbl->entries);
1392
1393 out_unlock:
1394
1395         return neigh;
1396 }
1397
1398 void ipoib_neigh_dtor(struct ipoib_neigh *neigh)
1399 {
1400         /* neigh reference count was dropprd to zero */
1401         struct net_device *dev = neigh->dev;
1402         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1403         struct sk_buff *skb;
1404         if (neigh->ah)
1405                 ipoib_put_ah(neigh->ah);
1406         while ((skb = __skb_dequeue(&neigh->queue))) {
1407                 ++dev->stats.tx_dropped;
1408                 dev_kfree_skb_any(skb);
1409         }
1410         if (ipoib_cm_get(neigh))
1411                 ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
1412         ipoib_dbg(ipoib_priv(dev),
1413                   "neigh free for %06x %pI6\n",
1414                   IPOIB_QPN(neigh->daddr),
1415                   neigh->daddr + 4);
1416         kfree(neigh);
1417         if (atomic_dec_and_test(&priv->ntbl.entries)) {
1418                 if (test_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags))
1419                         complete(&priv->ntbl.flushed);
1420         }
1421 }
1422
1423 static void ipoib_neigh_reclaim(struct rcu_head *rp)
1424 {
1425         /* Called as a result of removal from hash table */
1426         struct ipoib_neigh *neigh = container_of(rp, struct ipoib_neigh, rcu);
1427         /* note TX context may hold another ref */
1428         ipoib_neigh_put(neigh);
1429 }
1430
1431 void ipoib_neigh_free(struct ipoib_neigh *neigh)
1432 {
1433         struct net_device *dev = neigh->dev;
1434         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1435         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1436         struct ipoib_neigh_hash *htbl;
1437         struct ipoib_neigh __rcu **np;
1438         struct ipoib_neigh *n;
1439         u32 hash_val;
1440
1441         htbl = rcu_dereference_protected(ntbl->htbl,
1442                                         lockdep_is_held(&priv->lock));
1443         if (!htbl)
1444                 return;
1445
1446         hash_val = ipoib_addr_hash(htbl, neigh->daddr);
1447         np = &htbl->buckets[hash_val];
1448         for (n = rcu_dereference_protected(*np,
1449                                             lockdep_is_held(&priv->lock));
1450              n != NULL;
1451              n = rcu_dereference_protected(*np,
1452                                         lockdep_is_held(&priv->lock))) {
1453                 if (n == neigh) {
1454                         /* found */
1455                         rcu_assign_pointer(*np,
1456                                            rcu_dereference_protected(neigh->hnext,
1457                                                                      lockdep_is_held(&priv->lock)));
1458                         /* remove from parent list */
1459                         list_del_init(&neigh->list);
1460                         call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1461                         return;
1462                 } else {
1463                         np = &n->hnext;
1464                 }
1465         }
1466 }
1467
1468 static int ipoib_neigh_hash_init(struct ipoib_dev_priv *priv)
1469 {
1470         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1471         struct ipoib_neigh_hash *htbl;
1472         struct ipoib_neigh __rcu **buckets;
1473         u32 size;
1474
1475         clear_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags);
1476         ntbl->htbl = NULL;
1477         htbl = kzalloc(sizeof(*htbl), GFP_KERNEL);
1478         if (!htbl)
1479                 return -ENOMEM;
1480         set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1481         size = roundup_pow_of_two(arp_tbl.gc_thresh3);
1482         buckets = kzalloc(size * sizeof(*buckets), GFP_KERNEL);
1483         if (!buckets) {
1484                 kfree(htbl);
1485                 return -ENOMEM;
1486         }
1487         htbl->size = size;
1488         htbl->mask = (size - 1);
1489         htbl->buckets = buckets;
1490         RCU_INIT_POINTER(ntbl->htbl, htbl);
1491         htbl->ntbl = ntbl;
1492         atomic_set(&ntbl->entries, 0);
1493
1494         /* start garbage collection */
1495         clear_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1496         queue_delayed_work(priv->wq, &priv->neigh_reap_task,
1497                            arp_tbl.gc_interval);
1498
1499         return 0;
1500 }
1501
1502 static void neigh_hash_free_rcu(struct rcu_head *head)
1503 {
1504         struct ipoib_neigh_hash *htbl = container_of(head,
1505                                                     struct ipoib_neigh_hash,
1506                                                     rcu);
1507         struct ipoib_neigh __rcu **buckets = htbl->buckets;
1508         struct ipoib_neigh_table *ntbl = htbl->ntbl;
1509
1510         kfree(buckets);
1511         kfree(htbl);
1512         complete(&ntbl->deleted);
1513 }
1514
1515 void ipoib_del_neighs_by_gid(struct net_device *dev, u8 *gid)
1516 {
1517         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1518         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1519         struct ipoib_neigh_hash *htbl;
1520         unsigned long flags;
1521         int i;
1522
1523         /* remove all neigh connected to a given path or mcast */
1524         spin_lock_irqsave(&priv->lock, flags);
1525
1526         htbl = rcu_dereference_protected(ntbl->htbl,
1527                                          lockdep_is_held(&priv->lock));
1528
1529         if (!htbl)
1530                 goto out_unlock;
1531
1532         for (i = 0; i < htbl->size; i++) {
1533                 struct ipoib_neigh *neigh;
1534                 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1535
1536                 while ((neigh = rcu_dereference_protected(*np,
1537                                                           lockdep_is_held(&priv->lock))) != NULL) {
1538                         /* delete neighs belong to this parent */
1539                         if (!memcmp(gid, neigh->daddr + 4, sizeof (union ib_gid))) {
1540                                 rcu_assign_pointer(*np,
1541                                                    rcu_dereference_protected(neigh->hnext,
1542                                                                              lockdep_is_held(&priv->lock)));
1543                                 /* remove from parent list */
1544                                 list_del_init(&neigh->list);
1545                                 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1546                         } else {
1547                                 np = &neigh->hnext;
1548                         }
1549
1550                 }
1551         }
1552 out_unlock:
1553         spin_unlock_irqrestore(&priv->lock, flags);
1554 }
1555
1556 static void ipoib_flush_neighs(struct ipoib_dev_priv *priv)
1557 {
1558         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1559         struct ipoib_neigh_hash *htbl;
1560         unsigned long flags;
1561         int i, wait_flushed = 0;
1562
1563         init_completion(&priv->ntbl.flushed);
1564         set_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags);
1565
1566         spin_lock_irqsave(&priv->lock, flags);
1567
1568         htbl = rcu_dereference_protected(ntbl->htbl,
1569                                         lockdep_is_held(&priv->lock));
1570         if (!htbl)
1571                 goto out_unlock;
1572
1573         wait_flushed = atomic_read(&priv->ntbl.entries);
1574         if (!wait_flushed)
1575                 goto free_htbl;
1576
1577         for (i = 0; i < htbl->size; i++) {
1578                 struct ipoib_neigh *neigh;
1579                 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1580
1581                 while ((neigh = rcu_dereference_protected(*np,
1582                                        lockdep_is_held(&priv->lock))) != NULL) {
1583                         rcu_assign_pointer(*np,
1584                                            rcu_dereference_protected(neigh->hnext,
1585                                                                      lockdep_is_held(&priv->lock)));
1586                         /* remove from path/mc list */
1587                         list_del_init(&neigh->list);
1588                         call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1589                 }
1590         }
1591
1592 free_htbl:
1593         rcu_assign_pointer(ntbl->htbl, NULL);
1594         call_rcu(&htbl->rcu, neigh_hash_free_rcu);
1595
1596 out_unlock:
1597         spin_unlock_irqrestore(&priv->lock, flags);
1598         if (wait_flushed)
1599                 wait_for_completion(&priv->ntbl.flushed);
1600 }
1601
1602 static void ipoib_neigh_hash_uninit(struct net_device *dev)
1603 {
1604         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1605         int stopped;
1606
1607         ipoib_dbg(priv, "ipoib_neigh_hash_uninit\n");
1608         init_completion(&priv->ntbl.deleted);
1609
1610         /* Stop GC if called at init fail need to cancel work */
1611         stopped = test_and_set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1612         if (!stopped)
1613                 cancel_delayed_work(&priv->neigh_reap_task);
1614
1615         ipoib_flush_neighs(priv);
1616
1617         wait_for_completion(&priv->ntbl.deleted);
1618 }
1619
1620 static void ipoib_dev_uninit_default(struct net_device *dev)
1621 {
1622         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1623
1624         ipoib_transport_dev_cleanup(dev);
1625
1626         netif_napi_del(&priv->napi);
1627
1628         ipoib_cm_dev_cleanup(dev);
1629
1630         kfree(priv->rx_ring);
1631         vfree(priv->tx_ring);
1632
1633         priv->rx_ring = NULL;
1634         priv->tx_ring = NULL;
1635 }
1636
1637 static int ipoib_dev_init_default(struct net_device *dev)
1638 {
1639         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1640
1641         netif_napi_add(dev, &priv->napi, ipoib_poll, NAPI_POLL_WEIGHT);
1642
1643         /* Allocate RX/TX "rings" to hold queued skbs */
1644         priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
1645                                 GFP_KERNEL);
1646         if (!priv->rx_ring)
1647                 goto out;
1648
1649         priv->tx_ring = vzalloc(ipoib_sendq_size * sizeof *priv->tx_ring);
1650         if (!priv->tx_ring) {
1651                 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
1652                        priv->ca->name, ipoib_sendq_size);
1653                 goto out_rx_ring_cleanup;
1654         }
1655
1656         /* priv->tx_head, tx_tail & tx_outstanding are already 0 */
1657
1658         if (ipoib_transport_dev_init(dev, priv->ca)) {
1659                 pr_warn("%s: ipoib_transport_dev_init failed\n",
1660                         priv->ca->name);
1661                 goto out_tx_ring_cleanup;
1662         }
1663
1664         /* after qp created set dev address */
1665         priv->dev->dev_addr[1] = (priv->qp->qp_num >> 16) & 0xff;
1666         priv->dev->dev_addr[2] = (priv->qp->qp_num >>  8) & 0xff;
1667         priv->dev->dev_addr[3] = (priv->qp->qp_num) & 0xff;
1668
1669         setup_timer(&priv->poll_timer, ipoib_ib_tx_timer_func,
1670                     (unsigned long)dev);
1671
1672         return 0;
1673
1674 out_tx_ring_cleanup:
1675         vfree(priv->tx_ring);
1676
1677 out_rx_ring_cleanup:
1678         kfree(priv->rx_ring);
1679
1680 out:
1681         netif_napi_del(&priv->napi);
1682         return -ENOMEM;
1683 }
1684
1685 static int ipoib_ioctl(struct net_device *dev, struct ifreq *ifr,
1686                        int cmd)
1687 {
1688         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1689
1690         if (!priv->rn_ops->ndo_do_ioctl)
1691                 return -EOPNOTSUPP;
1692
1693         return priv->rn_ops->ndo_do_ioctl(dev, ifr, cmd);
1694 }
1695
1696 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
1697 {
1698         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1699         int ret = -ENOMEM;
1700
1701         priv->ca = ca;
1702         priv->port = port;
1703         priv->qp = NULL;
1704
1705         /*
1706          * the various IPoIB tasks assume they will never race against
1707          * themselves, so always use a single thread workqueue
1708          */
1709         priv->wq = alloc_ordered_workqueue("ipoib_wq", WQ_MEM_RECLAIM);
1710         if (!priv->wq) {
1711                 pr_warn("%s: failed to allocate device WQ\n", dev->name);
1712                 goto out;
1713         }
1714
1715         /* create pd, which used both for control and datapath*/
1716         priv->pd = ib_alloc_pd(priv->ca, 0);
1717         if (IS_ERR(priv->pd)) {
1718                 pr_warn("%s: failed to allocate PD\n", ca->name);
1719                 goto clean_wq;
1720         }
1721
1722         ret = priv->rn_ops->ndo_init(dev);
1723         if (ret) {
1724                 pr_warn("%s failed to init HW resource\n", dev->name);
1725                 goto out_free_pd;
1726         }
1727
1728         if (ipoib_neigh_hash_init(priv) < 0) {
1729                 pr_warn("%s failed to init neigh hash\n", dev->name);
1730                 goto out_dev_uninit;
1731         }
1732
1733         if (dev->flags & IFF_UP) {
1734                 if (ipoib_ib_dev_open(dev)) {
1735                         pr_warn("%s failed to open device\n", dev->name);
1736                         ret = -ENODEV;
1737                         goto out_dev_uninit;
1738                 }
1739         }
1740
1741         return 0;
1742
1743 out_dev_uninit:
1744         ipoib_ib_dev_cleanup(dev);
1745
1746 out_free_pd:
1747         if (priv->pd) {
1748                 ib_dealloc_pd(priv->pd);
1749                 priv->pd = NULL;
1750         }
1751
1752 clean_wq:
1753         if (priv->wq) {
1754                 destroy_workqueue(priv->wq);
1755                 priv->wq = NULL;
1756         }
1757
1758 out:
1759         return ret;
1760 }
1761
1762 void ipoib_dev_cleanup(struct net_device *dev)
1763 {
1764         struct ipoib_dev_priv *priv = ipoib_priv(dev), *cpriv, *tcpriv;
1765         LIST_HEAD(head);
1766
1767         ASSERT_RTNL();
1768
1769         /* Delete any child interfaces first */
1770         list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
1771                 /* Stop GC on child */
1772                 set_bit(IPOIB_STOP_NEIGH_GC, &cpriv->flags);
1773                 cancel_delayed_work(&cpriv->neigh_reap_task);
1774                 unregister_netdevice_queue(cpriv->dev, &head);
1775         }
1776         unregister_netdevice_many(&head);
1777
1778         ipoib_neigh_hash_uninit(dev);
1779
1780         ipoib_ib_dev_cleanup(dev);
1781
1782         /* no more works over the priv->wq */
1783         if (priv->wq) {
1784                 flush_workqueue(priv->wq);
1785                 destroy_workqueue(priv->wq);
1786                 priv->wq = NULL;
1787         }
1788 }
1789
1790 static int ipoib_set_vf_link_state(struct net_device *dev, int vf, int link_state)
1791 {
1792         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1793
1794         return ib_set_vf_link_state(priv->ca, vf, priv->port, link_state);
1795 }
1796
1797 static int ipoib_get_vf_config(struct net_device *dev, int vf,
1798                                struct ifla_vf_info *ivf)
1799 {
1800         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1801         int err;
1802
1803         err = ib_get_vf_config(priv->ca, vf, priv->port, ivf);
1804         if (err)
1805                 return err;
1806
1807         ivf->vf = vf;
1808
1809         return 0;
1810 }
1811
1812 static int ipoib_set_vf_guid(struct net_device *dev, int vf, u64 guid, int type)
1813 {
1814         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1815
1816         if (type != IFLA_VF_IB_NODE_GUID && type != IFLA_VF_IB_PORT_GUID)
1817                 return -EINVAL;
1818
1819         return ib_set_vf_guid(priv->ca, vf, priv->port, guid, type);
1820 }
1821
1822 static int ipoib_get_vf_stats(struct net_device *dev, int vf,
1823                               struct ifla_vf_stats *vf_stats)
1824 {
1825         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1826
1827         return ib_get_vf_stats(priv->ca, vf, priv->port, vf_stats);
1828 }
1829
1830 static const struct header_ops ipoib_header_ops = {
1831         .create = ipoib_hard_header,
1832 };
1833
1834 static const struct net_device_ops ipoib_netdev_ops_pf = {
1835         .ndo_uninit              = ipoib_uninit,
1836         .ndo_open                = ipoib_open,
1837         .ndo_stop                = ipoib_stop,
1838         .ndo_change_mtu          = ipoib_change_mtu,
1839         .ndo_fix_features        = ipoib_fix_features,
1840         .ndo_start_xmit          = ipoib_start_xmit,
1841         .ndo_tx_timeout          = ipoib_timeout,
1842         .ndo_set_rx_mode         = ipoib_set_mcast_list,
1843         .ndo_get_iflink          = ipoib_get_iflink,
1844         .ndo_set_vf_link_state   = ipoib_set_vf_link_state,
1845         .ndo_get_vf_config       = ipoib_get_vf_config,
1846         .ndo_get_vf_stats        = ipoib_get_vf_stats,
1847         .ndo_set_vf_guid         = ipoib_set_vf_guid,
1848         .ndo_set_mac_address     = ipoib_set_mac,
1849         .ndo_get_stats64         = ipoib_get_stats,
1850         .ndo_do_ioctl            = ipoib_ioctl,
1851 };
1852
1853 static const struct net_device_ops ipoib_netdev_ops_vf = {
1854         .ndo_uninit              = ipoib_uninit,
1855         .ndo_open                = ipoib_open,
1856         .ndo_stop                = ipoib_stop,
1857         .ndo_change_mtu          = ipoib_change_mtu,
1858         .ndo_fix_features        = ipoib_fix_features,
1859         .ndo_start_xmit          = ipoib_start_xmit,
1860         .ndo_tx_timeout          = ipoib_timeout,
1861         .ndo_set_rx_mode         = ipoib_set_mcast_list,
1862         .ndo_get_iflink          = ipoib_get_iflink,
1863         .ndo_get_stats64         = ipoib_get_stats,
1864         .ndo_do_ioctl            = ipoib_ioctl,
1865 };
1866
1867 void ipoib_setup_common(struct net_device *dev)
1868 {
1869         dev->header_ops          = &ipoib_header_ops;
1870
1871         ipoib_set_ethtool_ops(dev);
1872
1873         dev->watchdog_timeo      = HZ;
1874
1875         dev->flags              |= IFF_BROADCAST | IFF_MULTICAST;
1876
1877         dev->hard_header_len     = IPOIB_HARD_LEN;
1878         dev->addr_len            = INFINIBAND_ALEN;
1879         dev->type                = ARPHRD_INFINIBAND;
1880         dev->tx_queue_len        = ipoib_sendq_size * 2;
1881         dev->features            = (NETIF_F_VLAN_CHALLENGED     |
1882                                     NETIF_F_HIGHDMA);
1883         netif_keep_dst(dev);
1884
1885         memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
1886 }
1887
1888 static void ipoib_build_priv(struct net_device *dev)
1889 {
1890         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1891
1892         priv->dev = dev;
1893         spin_lock_init(&priv->lock);
1894         init_rwsem(&priv->vlan_rwsem);
1895         mutex_init(&priv->mcast_mutex);
1896         mutex_init(&priv->sysfs_mutex);
1897
1898         INIT_LIST_HEAD(&priv->path_list);
1899         INIT_LIST_HEAD(&priv->child_intfs);
1900         INIT_LIST_HEAD(&priv->dead_ahs);
1901         INIT_LIST_HEAD(&priv->multicast_list);
1902
1903         INIT_DELAYED_WORK(&priv->mcast_task,   ipoib_mcast_join_task);
1904         INIT_WORK(&priv->carrier_on_task, ipoib_mcast_carrier_on_task);
1905         INIT_WORK(&priv->flush_light,   ipoib_ib_dev_flush_light);
1906         INIT_WORK(&priv->flush_normal,   ipoib_ib_dev_flush_normal);
1907         INIT_WORK(&priv->flush_heavy,   ipoib_ib_dev_flush_heavy);
1908         INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
1909         INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
1910         INIT_DELAYED_WORK(&priv->neigh_reap_task, ipoib_reap_neigh);
1911 }
1912
1913 static const struct net_device_ops ipoib_netdev_default_pf = {
1914         .ndo_init                = ipoib_dev_init_default,
1915         .ndo_uninit              = ipoib_dev_uninit_default,
1916         .ndo_open                = ipoib_ib_dev_open_default,
1917         .ndo_stop                = ipoib_ib_dev_stop_default,
1918 };
1919
1920 static struct net_device
1921 *ipoib_create_netdev_default(struct ib_device *hca,
1922                              const char *name,
1923                              unsigned char name_assign_type,
1924                              void (*setup)(struct net_device *))
1925 {
1926         struct net_device *dev;
1927         struct rdma_netdev *rn;
1928
1929         dev = alloc_netdev((int)sizeof(struct rdma_netdev),
1930                            name,
1931                            name_assign_type, setup);
1932         if (!dev)
1933                 return NULL;
1934
1935         rn = netdev_priv(dev);
1936
1937         rn->send = ipoib_send;
1938         rn->attach_mcast = ipoib_mcast_attach;
1939         rn->detach_mcast = ipoib_mcast_detach;
1940         rn->free_rdma_netdev = free_netdev;
1941         rn->hca = hca;
1942
1943         dev->netdev_ops = &ipoib_netdev_default_pf;
1944
1945         return dev;
1946 }
1947
1948 static struct net_device *ipoib_get_netdev(struct ib_device *hca, u8 port,
1949                                            const char *name)
1950 {
1951         struct net_device *dev;
1952
1953         if (hca->alloc_rdma_netdev) {
1954                 dev = hca->alloc_rdma_netdev(hca, port,
1955                                              RDMA_NETDEV_IPOIB, name,
1956                                              NET_NAME_UNKNOWN,
1957                                              ipoib_setup_common);
1958                 if (IS_ERR_OR_NULL(dev) && PTR_ERR(dev) != -EOPNOTSUPP)
1959                         return NULL;
1960         }
1961
1962         if (!hca->alloc_rdma_netdev || PTR_ERR(dev) == -EOPNOTSUPP)
1963                 dev = ipoib_create_netdev_default(hca, name, NET_NAME_UNKNOWN,
1964                                                   ipoib_setup_common);
1965
1966         return dev;
1967 }
1968
1969 struct ipoib_dev_priv *ipoib_intf_alloc(struct ib_device *hca, u8 port,
1970                                         const char *name)
1971 {
1972         struct net_device *dev;
1973         struct ipoib_dev_priv *priv;
1974         struct rdma_netdev *rn;
1975
1976         priv = kzalloc(sizeof(*priv), GFP_KERNEL);
1977         if (!priv)
1978                 return NULL;
1979
1980         dev = ipoib_get_netdev(hca, port, name);
1981         if (!dev)
1982                 goto free_priv;
1983
1984         priv->rn_ops = dev->netdev_ops;
1985
1986         /* fixme : should be after the query_cap */
1987         if (priv->hca_caps & IB_DEVICE_VIRTUAL_FUNCTION)
1988                 dev->netdev_ops = &ipoib_netdev_ops_vf;
1989         else
1990                 dev->netdev_ops = &ipoib_netdev_ops_pf;
1991
1992         rn = netdev_priv(dev);
1993         rn->clnt_priv = priv;
1994         ipoib_build_priv(dev);
1995
1996         return priv;
1997 free_priv:
1998         kfree(priv);
1999         return NULL;
2000 }
2001
2002 static ssize_t show_pkey(struct device *dev,
2003                          struct device_attribute *attr, char *buf)
2004 {
2005         struct net_device *ndev = to_net_dev(dev);
2006         struct ipoib_dev_priv *priv = ipoib_priv(ndev);
2007
2008         return sprintf(buf, "0x%04x\n", priv->pkey);
2009 }
2010 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
2011
2012 static ssize_t show_umcast(struct device *dev,
2013                            struct device_attribute *attr, char *buf)
2014 {
2015         struct net_device *ndev = to_net_dev(dev);
2016         struct ipoib_dev_priv *priv = ipoib_priv(ndev);
2017
2018         return sprintf(buf, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
2019 }
2020
2021 void ipoib_set_umcast(struct net_device *ndev, int umcast_val)
2022 {
2023         struct ipoib_dev_priv *priv = ipoib_priv(ndev);
2024
2025         if (umcast_val > 0) {
2026                 set_bit(IPOIB_FLAG_UMCAST, &priv->flags);
2027                 ipoib_warn(priv, "ignoring multicast groups joined directly "
2028                                 "by userspace\n");
2029         } else
2030                 clear_bit(IPOIB_FLAG_UMCAST, &priv->flags);
2031 }
2032
2033 static ssize_t set_umcast(struct device *dev,
2034                           struct device_attribute *attr,
2035                           const char *buf, size_t count)
2036 {
2037         unsigned long umcast_val = simple_strtoul(buf, NULL, 0);
2038
2039         ipoib_set_umcast(to_net_dev(dev), umcast_val);
2040
2041         return count;
2042 }
2043 static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast);
2044
2045 int ipoib_add_umcast_attr(struct net_device *dev)
2046 {
2047         return device_create_file(&dev->dev, &dev_attr_umcast);
2048 }
2049
2050 static void set_base_guid(struct ipoib_dev_priv *priv, union ib_gid *gid)
2051 {
2052         struct ipoib_dev_priv *child_priv;
2053         struct net_device *netdev = priv->dev;
2054
2055         netif_addr_lock_bh(netdev);
2056
2057         memcpy(&priv->local_gid.global.interface_id,
2058                &gid->global.interface_id,
2059                sizeof(gid->global.interface_id));
2060         memcpy(netdev->dev_addr + 4, &priv->local_gid, sizeof(priv->local_gid));
2061         clear_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags);
2062
2063         netif_addr_unlock_bh(netdev);
2064
2065         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
2066                 down_read(&priv->vlan_rwsem);
2067                 list_for_each_entry(child_priv, &priv->child_intfs, list)
2068                         set_base_guid(child_priv, gid);
2069                 up_read(&priv->vlan_rwsem);
2070         }
2071 }
2072
2073 static int ipoib_check_lladdr(struct net_device *dev,
2074                               struct sockaddr_storage *ss)
2075 {
2076         union ib_gid *gid = (union ib_gid *)(ss->__data + 4);
2077         int ret = 0;
2078
2079         netif_addr_lock_bh(dev);
2080
2081         /* Make sure the QPN, reserved and subnet prefix match the current
2082          * lladdr, it also makes sure the lladdr is unicast.
2083          */
2084         if (memcmp(dev->dev_addr, ss->__data,
2085                    4 + sizeof(gid->global.subnet_prefix)) ||
2086             gid->global.interface_id == 0)
2087                 ret = -EINVAL;
2088
2089         netif_addr_unlock_bh(dev);
2090
2091         return ret;
2092 }
2093
2094 static int ipoib_set_mac(struct net_device *dev, void *addr)
2095 {
2096         struct ipoib_dev_priv *priv = ipoib_priv(dev);
2097         struct sockaddr_storage *ss = addr;
2098         int ret;
2099
2100         if (!(dev->priv_flags & IFF_LIVE_ADDR_CHANGE) && netif_running(dev))
2101                 return -EBUSY;
2102
2103         ret = ipoib_check_lladdr(dev, ss);
2104         if (ret)
2105                 return ret;
2106
2107         set_base_guid(priv, (union ib_gid *)(ss->__data + 4));
2108
2109         queue_work(ipoib_workqueue, &priv->flush_light);
2110
2111         return 0;
2112 }
2113
2114 static ssize_t create_child(struct device *dev,
2115                             struct device_attribute *attr,
2116                             const char *buf, size_t count)
2117 {
2118         int pkey;
2119         int ret;
2120
2121         if (sscanf(buf, "%i", &pkey) != 1)
2122                 return -EINVAL;
2123
2124         if (pkey <= 0 || pkey > 0xffff || pkey == 0x8000)
2125                 return -EINVAL;
2126
2127         /*
2128          * Set the full membership bit, so that we join the right
2129          * broadcast group, etc.
2130          */
2131         pkey |= 0x8000;
2132
2133         ret = ipoib_vlan_add(to_net_dev(dev), pkey);
2134
2135         return ret ? ret : count;
2136 }
2137 static DEVICE_ATTR(create_child, S_IWUSR, NULL, create_child);
2138
2139 static ssize_t delete_child(struct device *dev,
2140                             struct device_attribute *attr,
2141                             const char *buf, size_t count)
2142 {
2143         int pkey;
2144         int ret;
2145
2146         if (sscanf(buf, "%i", &pkey) != 1)
2147                 return -EINVAL;
2148
2149         if (pkey < 0 || pkey > 0xffff)
2150                 return -EINVAL;
2151
2152         ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
2153
2154         return ret ? ret : count;
2155
2156 }
2157 static DEVICE_ATTR(delete_child, S_IWUSR, NULL, delete_child);
2158
2159 int ipoib_add_pkey_attr(struct net_device *dev)
2160 {
2161         return device_create_file(&dev->dev, &dev_attr_pkey);
2162 }
2163
2164 void ipoib_set_dev_features(struct ipoib_dev_priv *priv, struct ib_device *hca)
2165 {
2166         priv->hca_caps = hca->attrs.device_cap_flags;
2167
2168         if (priv->hca_caps & IB_DEVICE_UD_IP_CSUM) {
2169                 priv->dev->hw_features |= NETIF_F_IP_CSUM | NETIF_F_RXCSUM;
2170
2171                 if (priv->hca_caps & IB_DEVICE_UD_TSO)
2172                         priv->dev->hw_features |= NETIF_F_TSO;
2173
2174                 priv->dev->features |= priv->dev->hw_features;
2175         }
2176 }
2177
2178 static struct net_device *ipoib_add_port(const char *format,
2179                                          struct ib_device *hca, u8 port)
2180 {
2181         struct ipoib_dev_priv *priv;
2182         struct ib_port_attr attr;
2183         int result = -ENOMEM;
2184
2185         priv = ipoib_intf_alloc(hca, port, format);
2186         if (!priv)
2187                 goto alloc_mem_failed;
2188
2189         SET_NETDEV_DEV(priv->dev, hca->dev.parent);
2190         priv->dev->dev_id = port - 1;
2191
2192         result = ib_query_port(hca, port, &attr);
2193         if (result) {
2194                 printk(KERN_WARNING "%s: ib_query_port %d failed\n",
2195                        hca->name, port);
2196                 goto device_init_failed;
2197         }
2198
2199         priv->max_ib_mtu = ib_mtu_enum_to_int(attr.max_mtu);
2200
2201         /* MTU will be reset when mcast join happens */
2202         priv->dev->mtu  = IPOIB_UD_MTU(priv->max_ib_mtu);
2203         priv->mcast_mtu  = priv->admin_mtu = priv->dev->mtu;
2204         priv->dev->max_mtu = IPOIB_CM_MTU;
2205
2206         priv->dev->neigh_priv_len = sizeof(struct ipoib_neigh);
2207
2208         result = ib_query_pkey(hca, port, 0, &priv->pkey);
2209         if (result) {
2210                 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
2211                        hca->name, port, result);
2212                 goto device_init_failed;
2213         }
2214
2215         ipoib_set_dev_features(priv, hca);
2216
2217         /*
2218          * Set the full membership bit, so that we join the right
2219          * broadcast group, etc.
2220          */
2221         priv->pkey |= 0x8000;
2222
2223         priv->dev->broadcast[8] = priv->pkey >> 8;
2224         priv->dev->broadcast[9] = priv->pkey & 0xff;
2225
2226         result = ib_query_gid(hca, port, 0, &priv->local_gid, NULL);
2227         if (result) {
2228                 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
2229                        hca->name, port, result);
2230                 goto device_init_failed;
2231         }
2232
2233         memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw,
2234                sizeof(union ib_gid));
2235         set_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags);
2236
2237         result = ipoib_dev_init(priv->dev, hca, port);
2238         if (result) {
2239                 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
2240                        hca->name, port, result);
2241                 goto device_init_failed;
2242         }
2243
2244         INIT_IB_EVENT_HANDLER(&priv->event_handler,
2245                               priv->ca, ipoib_event);
2246         ib_register_event_handler(&priv->event_handler);
2247
2248         result = register_netdev(priv->dev);
2249         if (result) {
2250                 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
2251                        hca->name, port, result);
2252                 goto register_failed;
2253         }
2254
2255         result = -ENOMEM;
2256         if (ipoib_cm_add_mode_attr(priv->dev))
2257                 goto sysfs_failed;
2258         if (ipoib_add_pkey_attr(priv->dev))
2259                 goto sysfs_failed;
2260         if (ipoib_add_umcast_attr(priv->dev))
2261                 goto sysfs_failed;
2262         if (device_create_file(&priv->dev->dev, &dev_attr_create_child))
2263                 goto sysfs_failed;
2264         if (device_create_file(&priv->dev->dev, &dev_attr_delete_child))
2265                 goto sysfs_failed;
2266
2267         return priv->dev;
2268
2269 sysfs_failed:
2270         unregister_netdev(priv->dev);
2271
2272 register_failed:
2273         ib_unregister_event_handler(&priv->event_handler);
2274         flush_workqueue(ipoib_workqueue);
2275         /* Stop GC if started before flush */
2276         set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
2277         cancel_delayed_work(&priv->neigh_reap_task);
2278         flush_workqueue(priv->wq);
2279         ipoib_dev_cleanup(priv->dev);
2280
2281 device_init_failed:
2282         free_netdev(priv->dev);
2283         kfree(priv);
2284
2285 alloc_mem_failed:
2286         return ERR_PTR(result);
2287 }
2288
2289 static void ipoib_add_one(struct ib_device *device)
2290 {
2291         struct list_head *dev_list;
2292         struct net_device *dev;
2293         struct ipoib_dev_priv *priv;
2294         int p;
2295         int count = 0;
2296
2297         dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
2298         if (!dev_list)
2299                 return;
2300
2301         INIT_LIST_HEAD(dev_list);
2302
2303         for (p = rdma_start_port(device); p <= rdma_end_port(device); ++p) {
2304                 if (!rdma_protocol_ib(device, p))
2305                         continue;
2306                 dev = ipoib_add_port("ib%d", device, p);
2307                 if (!IS_ERR(dev)) {
2308                         priv = ipoib_priv(dev);
2309                         list_add_tail(&priv->list, dev_list);
2310                         count++;
2311                 }
2312         }
2313
2314         if (!count) {
2315                 kfree(dev_list);
2316                 return;
2317         }
2318
2319         ib_set_client_data(device, &ipoib_client, dev_list);
2320 }
2321
2322 static void ipoib_remove_one(struct ib_device *device, void *client_data)
2323 {
2324         struct ipoib_dev_priv *priv, *tmp, *cpriv, *tcpriv;
2325         struct list_head *dev_list = client_data;
2326
2327         if (!dev_list)
2328                 return;
2329
2330         list_for_each_entry_safe(priv, tmp, dev_list, list) {
2331                 struct rdma_netdev *rn = netdev_priv(priv->dev);
2332
2333                 ib_unregister_event_handler(&priv->event_handler);
2334                 flush_workqueue(ipoib_workqueue);
2335
2336                 /* mark interface in the middle of destruction */
2337                 set_bit(IPOIB_FLAG_GOING_DOWN, &priv->flags);
2338
2339                 rtnl_lock();
2340                 dev_change_flags(priv->dev, priv->dev->flags & ~IFF_UP);
2341                 rtnl_unlock();
2342
2343                 /* Stop GC */
2344                 set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
2345                 cancel_delayed_work(&priv->neigh_reap_task);
2346                 flush_workqueue(priv->wq);
2347
2348                 /* Wrap rtnl_lock/unlock with mutex to protect sysfs calls */
2349                 mutex_lock(&priv->sysfs_mutex);
2350                 unregister_netdev(priv->dev);
2351                 mutex_unlock(&priv->sysfs_mutex);
2352
2353                 rn->free_rdma_netdev(priv->dev);
2354
2355                 list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list)
2356                         kfree(cpriv);
2357
2358                 kfree(priv);
2359         }
2360
2361         kfree(dev_list);
2362 }
2363
2364 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
2365 static struct notifier_block ipoib_netdev_notifier = {
2366         .notifier_call = ipoib_netdev_event,
2367 };
2368 #endif
2369
2370 static int __init ipoib_init_module(void)
2371 {
2372         int ret;
2373
2374         ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
2375         ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
2376         ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
2377
2378         ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
2379         ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
2380         ipoib_sendq_size = max3(ipoib_sendq_size, 2 * MAX_SEND_CQE, IPOIB_MIN_QUEUE_SIZE);
2381 #ifdef CONFIG_INFINIBAND_IPOIB_CM
2382         ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
2383         ipoib_max_conn_qp = max(ipoib_max_conn_qp, 0);
2384 #endif
2385
2386         /*
2387          * When copying small received packets, we only copy from the
2388          * linear data part of the SKB, so we rely on this condition.
2389          */
2390         BUILD_BUG_ON(IPOIB_CM_COPYBREAK > IPOIB_CM_HEAD_SIZE);
2391
2392         ret = ipoib_register_debugfs();
2393         if (ret)
2394                 return ret;
2395
2396         /*
2397          * We create a global workqueue here that is used for all flush
2398          * operations.  However, if you attempt to flush a workqueue
2399          * from a task on that same workqueue, it deadlocks the system.
2400          * We want to be able to flush the tasks associated with a
2401          * specific net device, so we also create a workqueue for each
2402          * netdevice.  We queue up the tasks for that device only on
2403          * its private workqueue, and we only queue up flush events
2404          * on our global flush workqueue.  This avoids the deadlocks.
2405          */
2406         ipoib_workqueue = alloc_ordered_workqueue("ipoib_flush",
2407                                                   WQ_MEM_RECLAIM);
2408         if (!ipoib_workqueue) {
2409                 ret = -ENOMEM;
2410                 goto err_fs;
2411         }
2412
2413         ib_sa_register_client(&ipoib_sa_client);
2414
2415         ret = ib_register_client(&ipoib_client);
2416         if (ret)
2417                 goto err_sa;
2418
2419         ret = ipoib_netlink_init();
2420         if (ret)
2421                 goto err_client;
2422
2423 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
2424         register_netdevice_notifier(&ipoib_netdev_notifier);
2425 #endif
2426         return 0;
2427
2428 err_client:
2429         ib_unregister_client(&ipoib_client);
2430
2431 err_sa:
2432         ib_sa_unregister_client(&ipoib_sa_client);
2433         destroy_workqueue(ipoib_workqueue);
2434
2435 err_fs:
2436         ipoib_unregister_debugfs();
2437
2438         return ret;
2439 }
2440
2441 static void __exit ipoib_cleanup_module(void)
2442 {
2443 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
2444         unregister_netdevice_notifier(&ipoib_netdev_notifier);
2445 #endif
2446         ipoib_netlink_fini();
2447         ib_unregister_client(&ipoib_client);
2448         ib_sa_unregister_client(&ipoib_sa_client);
2449         ipoib_unregister_debugfs();
2450         destroy_workqueue(ipoib_workqueue);
2451 }
2452
2453 module_init(ipoib_init_module);
2454 module_exit(ipoib_cleanup_module);