drivers/infiniband/ulp/ipoib/ipoib_main.c

   1 /*
   2  * Copyright (c) 2004 Topspin Communications.  All rights reserved.
   3  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
   4  * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
   5  *
   6  * This software is available to you under a choice of one of two
   7  * licenses.  You may choose to be licensed under the terms of the GNU
   8  * General Public License (GPL) Version 2, available from the file
   9  * COPYING in the main directory of this source tree, or the
  10  * OpenIB.org BSD license below:
  11  *
  12  *     Redistribution and use in source and binary forms, with or
  13  *     without modification, are permitted provided that the following
  14  *     conditions are met:
  15  *
  16  *      - Redistributions of source code must retain the above
  17  *        copyright notice, this list of conditions and the following
  18  *        disclaimer.
  19  *
  20  *      - Redistributions in binary form must reproduce the above
  21  *        copyright notice, this list of conditions and the following
  22  *        disclaimer in the documentation and/or other materials
  23  *        provided with the distribution.
  24  *
  25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  32  * SOFTWARE.
  33  */
  34
  35 #include "ipoib.h"
  36
  37 #include <linux/module.h>
  38
  39 #include <linux/init.h>
  40 #include <linux/slab.h>
  41 #include <linux/kernel.h>
  42 #include <linux/vmalloc.h>
  43
  44 #include <linux/if_arp.h>       /* For ARPHRD_xxx */
  45
  46 #include <linux/ip.h>
  47 #include <linux/in.h>
  48
  49 #include <linux/jhash.h>
  50 #include <net/arp.h>
  51 #include <net/addrconf.h>
  52 #include <linux/inetdevice.h>
  53 #include <rdma/ib_cache.h>
  54 #include <linux/pci.h>
  55
  56 #define DRV_VERSION "1.0.0"
  57
  58 const char ipoib_driver_version[] = DRV_VERSION;
  59
  60 MODULE_AUTHOR("Roland Dreier");
  61 MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
  62 MODULE_LICENSE("Dual BSD/GPL");
  63 MODULE_VERSION(DRV_VERSION);
  64
  65 int ipoib_sendq_size __read_mostly = IPOIB_TX_RING_SIZE;
  66 int ipoib_recvq_size __read_mostly = IPOIB_RX_RING_SIZE;
  67
  68 module_param_named(send_queue_size, ipoib_sendq_size, int, 0444);
  69 MODULE_PARM_DESC(send_queue_size, "Number of descriptors in send queue");
  70 module_param_named(recv_queue_size, ipoib_recvq_size, int, 0444);
  71 MODULE_PARM_DESC(recv_queue_size, "Number of descriptors in receive queue");
  72
  73 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
  74 int ipoib_debug_level;
  75
  76 module_param_named(debug_level, ipoib_debug_level, int, 0644);
  77 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
  78 #endif
  79
  80 struct ipoib_path_iter {
  81         struct net_device *dev;
  82         struct ipoib_path  path;
  83 };
  84
  85 static const u8 ipv4_bcast_addr[] = {
  86         0x00, 0xff, 0xff, 0xff,
  87         0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
  88         0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff
  89 };
  90
  91 struct workqueue_struct *ipoib_workqueue;
  92
  93 struct ib_sa_client ipoib_sa_client;
  94
  95 static void ipoib_add_one(struct ib_device *device);
  96 static void ipoib_remove_one(struct ib_device *device, void *client_data);
  97 static void ipoib_neigh_reclaim(struct rcu_head *rp);
  98 static struct net_device *ipoib_get_net_dev_by_params(
  99                 struct ib_device *dev, u8 port, u16 pkey,
 100                 const union ib_gid *gid, const struct sockaddr *addr,
 101                 void *client_data);
 102 static int ipoib_set_mac(struct net_device *dev, void *addr);
 103
 104 static struct ib_client ipoib_client = {
 105         .name   = "ipoib",
 106         .add    = ipoib_add_one,
 107         .remove = ipoib_remove_one,
 108         .get_net_dev_by_params = ipoib_get_net_dev_by_params,
 109 };
 110
 111 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 112 static int ipoib_netdev_event(struct notifier_block *this,
 113                               unsigned long event, void *ptr)
 114 {
 115         struct netdev_notifier_info *ni = ptr;
 116         struct net_device *dev = ni->dev;
 117
 118         if (dev->netdev_ops->ndo_open != ipoib_open)
 119                 return NOTIFY_DONE;
 120
 121         switch (event) {
 122         case NETDEV_REGISTER:
 123                 ipoib_create_debug_files(dev);
 124                 break;
 125         case NETDEV_CHANGENAME:
 126                 ipoib_delete_debug_files(dev);
 127                 ipoib_create_debug_files(dev);
 128                 break;
 129         case NETDEV_UNREGISTER:
 130                 ipoib_delete_debug_files(dev);
 131                 break;
 132         }
 133
 134         return NOTIFY_DONE;
 135 }
 136 #endif
 137
 138 int ipoib_open(struct net_device *dev)
 139 {
 140         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 141
 142         ipoib_dbg(priv, "bringing up interface\n");
 143
 144         netif_carrier_off(dev);
 145
 146         set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 147
 148         priv->sm_fullmember_sendonly_support = false;
 149
 150         if (ipoib_ib_dev_open(dev)) {
 151                 if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags))
 152                         return 0;
 153                 goto err_disable;
 154         }
 155
 156         ipoib_ib_dev_up(dev);
 157
 158         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 159                 struct ipoib_dev_priv *cpriv;
 160
 161                 /* Bring up any child interfaces too */
 162                 down_read(&priv->vlan_rwsem);
 163                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 164                         int flags;
 165
 166                         flags = cpriv->dev->flags;
 167                         if (flags & IFF_UP)
 168                                 continue;
 169
 170                         dev_change_flags(cpriv->dev, flags | IFF_UP);
 171                 }
 172                 up_read(&priv->vlan_rwsem);
 173         }
 174
 175         netif_start_queue(dev);
 176
 177         return 0;
 178
 179 err_disable:
 180         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 181
 182         return -EINVAL;
 183 }
 184
 185 static int ipoib_stop(struct net_device *dev)
 186 {
 187         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 188
 189         ipoib_dbg(priv, "stopping interface\n");
 190
 191         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 192
 193         netif_stop_queue(dev);
 194
 195         ipoib_ib_dev_down(dev);
 196         ipoib_ib_dev_stop(dev);
 197
 198         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 199                 struct ipoib_dev_priv *cpriv;
 200
 201                 /* Bring down any child interfaces too */
 202                 down_read(&priv->vlan_rwsem);
 203                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 204                         int flags;
 205
 206                         flags = cpriv->dev->flags;
 207                         if (!(flags & IFF_UP))
 208                                 continue;
 209
 210                         dev_change_flags(cpriv->dev, flags & ~IFF_UP);
 211                 }
 212                 up_read(&priv->vlan_rwsem);
 213         }
 214
 215         return 0;
 216 }
 217
 218 static void ipoib_uninit(struct net_device *dev)
 219 {
 220         ipoib_dev_cleanup(dev);
 221 }
 222
 223 static netdev_features_t ipoib_fix_features(struct net_device *dev, netdev_features_t features)
 224 {
 225         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 226
 227         if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
 228                 features &= ~(NETIF_F_IP_CSUM | NETIF_F_TSO);
 229
 230         return features;
 231 }
 232
 233 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
 234 {
 235         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 236
 237         /* dev->mtu > 2K ==> connected mode */
 238         if (ipoib_cm_admin_enabled(dev)) {
 239                 if (new_mtu > ipoib_cm_max_mtu(dev))
 240                         return -EINVAL;
 241
 242                 if (new_mtu > priv->mcast_mtu)
 243                         ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
 244                                    priv->mcast_mtu);
 245
 246                 dev->mtu = new_mtu;
 247                 return 0;
 248         }
 249
 250         if (new_mtu > IPOIB_UD_MTU(priv->max_ib_mtu))
 251                 return -EINVAL;
 252
 253         priv->admin_mtu = new_mtu;
 254
 255         if (priv->mcast_mtu < priv->admin_mtu)
 256                 ipoib_dbg(priv, "MTU must be smaller than the underlying "
 257                                 "link layer MTU - 4 (%u)\n", priv->mcast_mtu);
 258
 259         dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
 260
 261         return 0;
 262 }
 263
 264 /* Called with an RCU read lock taken */
 265 static bool ipoib_is_dev_match_addr_rcu(const struct sockaddr *addr,
 266                                         struct net_device *dev)
 267 {
 268         struct net *net = dev_net(dev);
 269         struct in_device *in_dev;
 270         struct sockaddr_in *addr_in = (struct sockaddr_in *)addr;
 271         struct sockaddr_in6 *addr_in6 = (struct sockaddr_in6 *)addr;
 272         __be32 ret_addr;
 273
 274         switch (addr->sa_family) {
 275         case AF_INET:
 276                 in_dev = in_dev_get(dev);
 277                 if (!in_dev)
 278                         return false;
 279
 280                 ret_addr = inet_confirm_addr(net, in_dev, 0,
 281                                              addr_in->sin_addr.s_addr,
 282                                              RT_SCOPE_HOST);
 283                 in_dev_put(in_dev);
 284                 if (ret_addr)
 285                         return true;
 286
 287                 break;
 288         case AF_INET6:
 289                 if (IS_ENABLED(CONFIG_IPV6) &&
 290                     ipv6_chk_addr(net, &addr_in6->sin6_addr, dev, 1))
 291                         return true;
 292
 293                 break;
 294         }
 295         return false;
 296 }
 297
 298 /**
 299  * Find the master net_device on top of the given net_device.
 300  * @dev: base IPoIB net_device
 301  *
 302  * Returns the master net_device with a reference held, or the same net_device
 303  * if no master exists.
 304  */
 305 static struct net_device *ipoib_get_master_net_dev(struct net_device *dev)
 306 {
 307         struct net_device *master;
 308
 309         rcu_read_lock();
 310         master = netdev_master_upper_dev_get_rcu(dev);
 311         if (master)
 312                 dev_hold(master);
 313         rcu_read_unlock();
 314
 315         if (master)
 316                 return master;
 317
 318         dev_hold(dev);
 319         return dev;
 320 }
 321
 322 struct ipoib_walk_data {
 323         const struct sockaddr *addr;
 324         struct net_device *result;
 325 };
 326
 327 static int ipoib_upper_walk(struct net_device *upper, void *_data)
 328 {
 329         struct ipoib_walk_data *data = _data;
 330         int ret = 0;
 331
 332         if (ipoib_is_dev_match_addr_rcu(data->addr, upper)) {
 333                 dev_hold(upper);
 334                 data->result = upper;
 335                 ret = 1;
 336         }
 337
 338         return ret;
 339 }
 340
 341 /**
 342  * Find a net_device matching the given address, which is an upper device of
 343  * the given net_device.
 344  * @addr: IP address to look for.
 345  * @dev: base IPoIB net_device
 346  *
 347  * If found, returns the net_device with a reference held. Otherwise return
 348  * NULL.
 349  */
 350 static struct net_device *ipoib_get_net_dev_match_addr(
 351                 const struct sockaddr *addr, struct net_device *dev)
 352 {
 353         struct ipoib_walk_data data = {
 354                 .addr = addr,
 355         };
 356
 357         rcu_read_lock();
 358         if (ipoib_is_dev_match_addr_rcu(addr, dev)) {
 359                 dev_hold(dev);
 360                 data.result = dev;
 361                 goto out;
 362         }
 363
 364         netdev_walk_all_upper_dev_rcu(dev, ipoib_upper_walk, &data);
 365 out:
 366         rcu_read_unlock();
 367         return data.result;
 368 }
 369
 370 /* returns the number of IPoIB netdevs on top a given ipoib device matching a
 371  * pkey_index and address, if one exists.
 372  *
 373  * @found_net_dev: contains a matching net_device if the return value >= 1,
 374  * with a reference held. */
 375 static int ipoib_match_gid_pkey_addr(struct ipoib_dev_priv *priv,
 376                                      const union ib_gid *gid,
 377                                      u16 pkey_index,
 378                                      const struct sockaddr *addr,
 379                                      int nesting,
 380                                      struct net_device **found_net_dev)
 381 {
 382         struct ipoib_dev_priv *child_priv;
 383         struct net_device *net_dev = NULL;
 384         int matches = 0;
 385
 386         if (priv->pkey_index == pkey_index &&
 387             (!gid || !memcmp(gid, &priv->local_gid, sizeof(*gid)))) {
 388                 if (!addr) {
 389                         net_dev = ipoib_get_master_net_dev(priv->dev);
 390                 } else {
 391                         /* Verify the net_device matches the IP address, as
 392                          * IPoIB child devices currently share a GID. */
 393                         net_dev = ipoib_get_net_dev_match_addr(addr, priv->dev);
 394                 }
 395                 if (net_dev) {
 396                         if (!*found_net_dev)
 397                                 *found_net_dev = net_dev;
 398                         else
 399                                 dev_put(net_dev);
 400                         ++matches;
 401                 }
 402         }
 403
 404         /* Check child interfaces */
 405         down_read_nested(&priv->vlan_rwsem, nesting);
 406         list_for_each_entry(child_priv, &priv->child_intfs, list) {
 407                 matches += ipoib_match_gid_pkey_addr(child_priv, gid,
 408                                                     pkey_index, addr,
 409                                                     nesting + 1,
 410                                                     found_net_dev);
 411                 if (matches > 1)
 412                         break;
 413         }
 414         up_read(&priv->vlan_rwsem);
 415
 416         return matches;
 417 }
 418
 419 /* Returns the number of matching net_devs found (between 0 and 2). Also
 420  * return the matching net_device in the @net_dev parameter, holding a
 421  * reference to the net_device, if the number of matches >= 1 */
 422 static int __ipoib_get_net_dev_by_params(struct list_head *dev_list, u8 port,
 423                                          u16 pkey_index,
 424                                          const union ib_gid *gid,
 425                                          const struct sockaddr *addr,
 426                                          struct net_device **net_dev)
 427 {
 428         struct ipoib_dev_priv *priv;
 429         int matches = 0;
 430
 431         *net_dev = NULL;
 432
 433         list_for_each_entry(priv, dev_list, list) {
 434                 if (priv->port != port)
 435                         continue;
 436
 437                 matches += ipoib_match_gid_pkey_addr(priv, gid, pkey_index,
 438                                                      addr, 0, net_dev);
 439                 if (matches > 1)
 440                         break;
 441         }
 442
 443         return matches;
 444 }
 445
 446 static struct net_device *ipoib_get_net_dev_by_params(
 447                 struct ib_device *dev, u8 port, u16 pkey,
 448                 const union ib_gid *gid, const struct sockaddr *addr,
 449                 void *client_data)
 450 {
 451         struct net_device *net_dev;
 452         struct list_head *dev_list = client_data;
 453         u16 pkey_index;
 454         int matches;
 455         int ret;
 456
 457         if (!rdma_protocol_ib(dev, port))
 458                 return NULL;
 459
 460         ret = ib_find_cached_pkey(dev, port, pkey, &pkey_index);
 461         if (ret)
 462                 return NULL;
 463
 464         if (!dev_list)
 465                 return NULL;
 466
 467         /* See if we can find a unique device matching the L2 parameters */
 468         matches = __ipoib_get_net_dev_by_params(dev_list, port, pkey_index,
 469                                                 gid, NULL, &net_dev);
 470
 471         switch (matches) {
 472         case 0:
 473                 return NULL;
 474         case 1:
 475                 return net_dev;
 476         }
 477
 478         dev_put(net_dev);
 479
 480         /* Couldn't find a unique device with L2 parameters only. Use L3
 481          * address to uniquely match the net device */
 482         matches = __ipoib_get_net_dev_by_params(dev_list, port, pkey_index,
 483                                                 gid, addr, &net_dev);
 484         switch (matches) {
 485         case 0:
 486                 return NULL;
 487         default:
 488                 dev_warn_ratelimited(&dev->dev,
 489                                      "duplicate IP address detected\n");
 490                 /* Fall through */
 491         case 1:
 492                 return net_dev;
 493         }
 494 }
 495
 496 int ipoib_set_mode(struct net_device *dev, const char *buf)
 497 {
 498         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 499
 500         if ((test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags) &&
 501              !strcmp(buf, "connected\n")) ||
 502              (!test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags) &&
 503              !strcmp(buf, "datagram\n"))) {
 504                 return 0;
 505         }
 506
 507         /* flush paths if we switch modes so that connections are restarted */
 508         if (IPOIB_CM_SUPPORTED(dev->dev_addr) && !strcmp(buf, "connected\n")) {
 509                 set_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
 510                 ipoib_warn(priv, "enabling connected mode "
 511                            "will cause multicast packet drops\n");
 512                 netdev_update_features(dev);
 513                 dev_set_mtu(dev, ipoib_cm_max_mtu(dev));
 514                 rtnl_unlock();
 515                 priv->tx_wr.wr.send_flags &= ~IB_SEND_IP_CSUM;
 516
 517                 ipoib_flush_paths(dev);
 518                 return (!rtnl_trylock()) ? -EBUSY : 0;
 519         }
 520
 521         if (!strcmp(buf, "datagram\n")) {
 522                 clear_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
 523                 netdev_update_features(dev);
 524                 dev_set_mtu(dev, min(priv->mcast_mtu, dev->mtu));
 525                 rtnl_unlock();
 526                 ipoib_flush_paths(dev);
 527                 return (!rtnl_trylock()) ? -EBUSY : 0;
 528         }
 529
 530         return -EINVAL;
 531 }
 532
 533 struct ipoib_path *__path_find(struct net_device *dev, void *gid)
 534 {
 535         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 536         struct rb_node *n = priv->path_tree.rb_node;
 537         struct ipoib_path *path;
 538         int ret;
 539
 540         while (n) {
 541                 path = rb_entry(n, struct ipoib_path, rb_node);
 542
 543                 ret = memcmp(gid, path->pathrec.dgid.raw,
 544                              sizeof (union ib_gid));
 545
 546                 if (ret < 0)
 547                         n = n->rb_left;
 548                 else if (ret > 0)
 549                         n = n->rb_right;
 550                 else
 551                         return path;
 552         }
 553
 554         return NULL;
 555 }
 556
 557 static int __path_add(struct net_device *dev, struct ipoib_path *path)
 558 {
 559         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 560         struct rb_node **n = &priv->path_tree.rb_node;
 561         struct rb_node *pn = NULL;
 562         struct ipoib_path *tpath;
 563         int ret;
 564
 565         while (*n) {
 566                 pn = *n;
 567                 tpath = rb_entry(pn, struct ipoib_path, rb_node);
 568
 569                 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
 570                              sizeof (union ib_gid));
 571                 if (ret < 0)
 572                         n = &pn->rb_left;
 573                 else if (ret > 0)
 574                         n = &pn->rb_right;
 575                 else
 576                         return -EEXIST;
 577         }
 578
 579         rb_link_node(&path->rb_node, pn, n);
 580         rb_insert_color(&path->rb_node, &priv->path_tree);
 581
 582         list_add_tail(&path->list, &priv->path_list);
 583
 584         return 0;
 585 }
 586
 587 static void path_free(struct net_device *dev, struct ipoib_path *path)
 588 {
 589         struct sk_buff *skb;
 590
 591         while ((skb = __skb_dequeue(&path->queue)))
 592                 dev_kfree_skb_irq(skb);
 593
 594         ipoib_dbg(ipoib_priv(dev), "path_free\n");
 595
 596         /* remove all neigh connected to this path */
 597         ipoib_del_neighs_by_gid(dev, path->pathrec.dgid.raw);
 598
 599         if (path->ah)
 600                 ipoib_put_ah(path->ah);
 601
 602         kfree(path);
 603 }
 604
 605 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 606
 607 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
 608 {
 609         struct ipoib_path_iter *iter;
 610
 611         iter = kmalloc(sizeof *iter, GFP_KERNEL);
 612         if (!iter)
 613                 return NULL;
 614
 615         iter->dev = dev;
 616         memset(iter->path.pathrec.dgid.raw, 0, 16);
 617
 618         if (ipoib_path_iter_next(iter)) {
 619                 kfree(iter);
 620                 return NULL;
 621         }
 622
 623         return iter;
 624 }
 625
 626 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
 627 {
 628         struct ipoib_dev_priv *priv = ipoib_priv(iter->dev);
 629         struct rb_node *n;
 630         struct ipoib_path *path;
 631         int ret = 1;
 632
 633         spin_lock_irq(&priv->lock);
 634
 635         n = rb_first(&priv->path_tree);
 636
 637         while (n) {
 638                 path = rb_entry(n, struct ipoib_path, rb_node);
 639
 640                 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
 641                            sizeof (union ib_gid)) < 0) {
 642                         iter->path = *path;
 643                         ret = 0;
 644                         break;
 645                 }
 646
 647                 n = rb_next(n);
 648         }
 649
 650         spin_unlock_irq(&priv->lock);
 651
 652         return ret;
 653 }
 654
 655 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
 656                           struct ipoib_path *path)
 657 {
 658         *path = iter->path;
 659 }
 660
 661 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
 662
 663 void ipoib_mark_paths_invalid(struct net_device *dev)
 664 {
 665         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 666         struct ipoib_path *path, *tp;
 667
 668         spin_lock_irq(&priv->lock);
 669
 670         list_for_each_entry_safe(path, tp, &priv->path_list, list) {
 671                 ipoib_dbg(priv, "mark path LID 0x%04x GID %pI6 invalid\n",
 672                         be16_to_cpu(path->pathrec.dlid),
 673                         path->pathrec.dgid.raw);
 674                 path->valid =  0;
 675         }
 676
 677         spin_unlock_irq(&priv->lock);
 678 }
 679
 680 static void push_pseudo_header(struct sk_buff *skb, const char *daddr)
 681 {
 682         struct ipoib_pseudo_header *phdr;
 683
 684         phdr = (struct ipoib_pseudo_header *)skb_push(skb, sizeof(*phdr));
 685         memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
 686 }
 687
 688 void ipoib_flush_paths(struct net_device *dev)
 689 {
 690         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 691         struct ipoib_path *path, *tp;
 692         LIST_HEAD(remove_list);
 693         unsigned long flags;
 694
 695         netif_tx_lock_bh(dev);
 696         spin_lock_irqsave(&priv->lock, flags);
 697
 698         list_splice_init(&priv->path_list, &remove_list);
 699
 700         list_for_each_entry(path, &remove_list, list)
 701                 rb_erase(&path->rb_node, &priv->path_tree);
 702
 703         list_for_each_entry_safe(path, tp, &remove_list, list) {
 704                 if (path->query)
 705                         ib_sa_cancel_query(path->query_id, path->query);
 706                 spin_unlock_irqrestore(&priv->lock, flags);
 707                 netif_tx_unlock_bh(dev);
 708                 wait_for_completion(&path->done);
 709                 path_free(dev, path);
 710                 netif_tx_lock_bh(dev);
 711                 spin_lock_irqsave(&priv->lock, flags);
 712         }
 713
 714         spin_unlock_irqrestore(&priv->lock, flags);
 715         netif_tx_unlock_bh(dev);
 716 }
 717
 718 static void path_rec_completion(int status,
 719                                 struct ib_sa_path_rec *pathrec,
 720                                 void *path_ptr)
 721 {
 722         struct ipoib_path *path = path_ptr;
 723         struct net_device *dev = path->dev;
 724         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 725         struct ipoib_ah *ah = NULL;
 726         struct ipoib_ah *old_ah = NULL;
 727         struct ipoib_neigh *neigh, *tn;
 728         struct sk_buff_head skqueue;
 729         struct sk_buff *skb;
 730         unsigned long flags;
 731
 732         if (!status)
 733                 ipoib_dbg(priv, "PathRec LID 0x%04x for GID %pI6\n",
 734                           be16_to_cpu(pathrec->dlid), pathrec->dgid.raw);
 735         else
 736                 ipoib_dbg(priv, "PathRec status %d for GID %pI6\n",
 737                           status, path->pathrec.dgid.raw);
 738
 739         skb_queue_head_init(&skqueue);
 740
 741         if (!status) {
 742                 struct rdma_ah_attr av;
 743
 744                 if (!ib_init_ah_from_path(priv->ca, priv->port, pathrec, &av))
 745                         ah = ipoib_create_ah(dev, priv->pd, &av);
 746         }
 747
 748         spin_lock_irqsave(&priv->lock, flags);
 749
 750         if (!IS_ERR_OR_NULL(ah)) {
 751                 path->pathrec = *pathrec;
 752
 753                 old_ah   = path->ah;
 754                 path->ah = ah;
 755
 756                 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
 757                           ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
 758
 759                 while ((skb = __skb_dequeue(&path->queue)))
 760                         __skb_queue_tail(&skqueue, skb);
 761
 762                 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
 763                         if (neigh->ah) {
 764                                 WARN_ON(neigh->ah != old_ah);
 765                                 /*
 766                                  * Dropping the ah reference inside
 767                                  * priv->lock is safe here, because we
 768                                  * will hold one more reference from
 769                                  * the original value of path->ah (ie
 770                                  * old_ah).
 771                                  */
 772                                 ipoib_put_ah(neigh->ah);
 773                         }
 774                         kref_get(&path->ah->ref);
 775                         neigh->ah = path->ah;
 776
 777                         if (ipoib_cm_enabled(dev, neigh->daddr)) {
 778                                 if (!ipoib_cm_get(neigh))
 779                                         ipoib_cm_set(neigh, ipoib_cm_create_tx(dev,
 780                                                                                path,
 781                                                                                neigh));
 782                                 if (!ipoib_cm_get(neigh)) {
 783                                         ipoib_neigh_free(neigh);
 784                                         continue;
 785                                 }
 786                         }
 787
 788                         while ((skb = __skb_dequeue(&neigh->queue)))
 789                                 __skb_queue_tail(&skqueue, skb);
 790                 }
 791                 path->valid = 1;
 792         }
 793
 794         path->query = NULL;
 795         complete(&path->done);
 796
 797         spin_unlock_irqrestore(&priv->lock, flags);
 798
 799         if (IS_ERR_OR_NULL(ah))
 800                 ipoib_del_neighs_by_gid(dev, path->pathrec.dgid.raw);
 801
 802         if (old_ah)
 803                 ipoib_put_ah(old_ah);
 804
 805         while ((skb = __skb_dequeue(&skqueue))) {
 806                 int ret;
 807                 skb->dev = dev;
 808                 ret = dev_queue_xmit(skb);
 809                 if (ret)
 810                         ipoib_warn(priv, "%s: dev_queue_xmit failed to re-queue packet, ret:%d\n",
 811                                    __func__, ret);
 812         }
 813 }
 814
 815 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
 816 {
 817         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 818         struct ipoib_path *path;
 819
 820         if (!priv->broadcast)
 821                 return NULL;
 822
 823         path = kzalloc(sizeof *path, GFP_ATOMIC);
 824         if (!path)
 825                 return NULL;
 826
 827         path->dev = dev;
 828
 829         skb_queue_head_init(&path->queue);
 830
 831         INIT_LIST_HEAD(&path->neigh_list);
 832
 833         memcpy(path->pathrec.dgid.raw, gid, sizeof (union ib_gid));
 834         path->pathrec.sgid          = priv->local_gid;
 835         path->pathrec.pkey          = cpu_to_be16(priv->pkey);
 836         path->pathrec.numb_path     = 1;
 837         path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
 838
 839         return path;
 840 }
 841
 842 static int path_rec_start(struct net_device *dev,
 843                           struct ipoib_path *path)
 844 {
 845         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 846
 847         ipoib_dbg(priv, "Start path record lookup for %pI6\n",
 848                   path->pathrec.dgid.raw);
 849
 850         init_completion(&path->done);
 851
 852         path->query_id =
 853                 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
 854                                    &path->pathrec,
 855                                    IB_SA_PATH_REC_DGID          |
 856                                    IB_SA_PATH_REC_SGID          |
 857                                    IB_SA_PATH_REC_NUMB_PATH     |
 858                                    IB_SA_PATH_REC_TRAFFIC_CLASS |
 859                                    IB_SA_PATH_REC_PKEY,
 860                                    1000, GFP_ATOMIC,
 861                                    path_rec_completion,
 862                                    path, &path->query);
 863         if (path->query_id < 0) {
 864                 ipoib_warn(priv, "ib_sa_path_rec_get failed: %d\n", path->query_id);
 865                 path->query = NULL;
 866                 complete(&path->done);
 867                 return path->query_id;
 868         }
 869
 870         return 0;
 871 }
 872
 873 static void neigh_add_path(struct sk_buff *skb, u8 *daddr,
 874                            struct net_device *dev)
 875 {
 876         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 877         struct rdma_netdev *rn = netdev_priv(dev);
 878         struct ipoib_path *path;
 879         struct ipoib_neigh *neigh;
 880         unsigned long flags;
 881
 882         spin_lock_irqsave(&priv->lock, flags);
 883         neigh = ipoib_neigh_alloc(daddr, dev);
 884         if (!neigh) {
 885                 spin_unlock_irqrestore(&priv->lock, flags);
 886                 ++dev->stats.tx_dropped;
 887                 dev_kfree_skb_any(skb);
 888                 return;
 889         }
 890
 891         path = __path_find(dev, daddr + 4);
 892         if (!path) {
 893                 path = path_rec_create(dev, daddr + 4);
 894                 if (!path)
 895                         goto err_path;
 896
 897                 __path_add(dev, path);
 898         }
 899
 900         list_add_tail(&neigh->list, &path->neigh_list);
 901
 902         if (path->ah) {
 903                 kref_get(&path->ah->ref);
 904                 neigh->ah = path->ah;
 905
 906                 if (ipoib_cm_enabled(dev, neigh->daddr)) {
 907                         if (!ipoib_cm_get(neigh))
 908                                 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
 909                         if (!ipoib_cm_get(neigh)) {
 910                                 ipoib_neigh_free(neigh);
 911                                 goto err_drop;
 912                         }
 913                         if (skb_queue_len(&neigh->queue) <
 914                             IPOIB_MAX_PATH_REC_QUEUE) {
 915                                 push_pseudo_header(skb, neigh->daddr);
 916                                 __skb_queue_tail(&neigh->queue, skb);
 917                         } else {
 918                                 ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
 919                                            skb_queue_len(&neigh->queue));
 920                                 goto err_drop;
 921                         }
 922                 } else {
 923                         spin_unlock_irqrestore(&priv->lock, flags);
 924                         path->ah->last_send = rn->send(dev, skb, path->ah->ah,
 925                                                        IPOIB_QPN(daddr));
 926                         ipoib_neigh_put(neigh);
 927                         return;
 928                 }
 929         } else {
 930                 neigh->ah  = NULL;
 931
 932                 if (!path->query && path_rec_start(dev, path))
 933                         goto err_path;
 934                 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 935                         push_pseudo_header(skb, neigh->daddr);
 936                         __skb_queue_tail(&neigh->queue, skb);
 937                 } else {
 938                         goto err_drop;
 939                 }
 940         }
 941
 942         spin_unlock_irqrestore(&priv->lock, flags);
 943         ipoib_neigh_put(neigh);
 944         return;
 945
 946 err_path:
 947         ipoib_neigh_free(neigh);
 948 err_drop:
 949         ++dev->stats.tx_dropped;
 950         dev_kfree_skb_any(skb);
 951
 952         spin_unlock_irqrestore(&priv->lock, flags);
 953         ipoib_neigh_put(neigh);
 954 }
 955
 956 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
 957                              struct ipoib_pseudo_header *phdr)
 958 {
 959         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 960         struct rdma_netdev *rn = netdev_priv(dev);
 961         struct ipoib_path *path;
 962         unsigned long flags;
 963
 964         spin_lock_irqsave(&priv->lock, flags);
 965
 966         path = __path_find(dev, phdr->hwaddr + 4);
 967         if (!path || !path->valid) {
 968                 int new_path = 0;
 969
 970                 if (!path) {
 971                         path = path_rec_create(dev, phdr->hwaddr + 4);
 972                         new_path = 1;
 973                 }
 974                 if (path) {
 975                         if (skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 976                                 push_pseudo_header(skb, phdr->hwaddr);
 977                                 __skb_queue_tail(&path->queue, skb);
 978                         } else {
 979                                 ++dev->stats.tx_dropped;
 980                                 dev_kfree_skb_any(skb);
 981                         }
 982
 983                         if (!path->query && path_rec_start(dev, path)) {
 984                                 spin_unlock_irqrestore(&priv->lock, flags);
 985                                 if (new_path)
 986                                         path_free(dev, path);
 987                                 return;
 988                         } else
 989                                 __path_add(dev, path);
 990                 } else {
 991                         ++dev->stats.tx_dropped;
 992                         dev_kfree_skb_any(skb);
 993                 }
 994
 995                 spin_unlock_irqrestore(&priv->lock, flags);
 996                 return;
 997         }
 998
 999         if (path->ah) {
1000                 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
1001                           be16_to_cpu(path->pathrec.dlid));
1002
1003                 spin_unlock_irqrestore(&priv->lock, flags);
1004                 path->ah->last_send = rn->send(dev, skb, path->ah->ah,
1005                                                IPOIB_QPN(phdr->hwaddr));
1006                 return;
1007         } else if ((path->query || !path_rec_start(dev, path)) &&
1008                    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         spin_unlock_irqrestore(&priv->lock, flags);
1017 }
1018
1019 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
1020 {
1021         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1022         struct rdma_netdev *rn = netdev_priv(dev);
1023         struct ipoib_neigh *neigh;
1024         struct ipoib_pseudo_header *phdr;
1025         struct ipoib_header *header;
1026         unsigned long flags;
1027
1028         phdr = (struct ipoib_pseudo_header *) skb->data;
1029         skb_pull(skb, sizeof(*phdr));
1030         header = (struct ipoib_header *) skb->data;
1031
1032         if (unlikely(phdr->hwaddr[4] == 0xff)) {
1033                 /* multicast, arrange "if" according to probability */
1034                 if ((header->proto != htons(ETH_P_IP)) &&
1035                     (header->proto != htons(ETH_P_IPV6)) &&
1036                     (header->proto != htons(ETH_P_ARP)) &&
1037                     (header->proto != htons(ETH_P_RARP)) &&
1038                     (header->proto != htons(ETH_P_TIPC))) {
1039                         /* ethertype not supported by IPoIB */
1040                         ++dev->stats.tx_dropped;
1041                         dev_kfree_skb_any(skb);
1042                         return NETDEV_TX_OK;
1043                 }
1044                 /* Add in the P_Key for multicast*/
1045                 phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
1046                 phdr->hwaddr[9] = priv->pkey & 0xff;
1047
1048                 neigh = ipoib_neigh_get(dev, phdr->hwaddr);
1049                 if (likely(neigh))
1050                         goto send_using_neigh;
1051                 ipoib_mcast_send(dev, phdr->hwaddr, skb);
1052                 return NETDEV_TX_OK;
1053         }
1054
1055         /* unicast, arrange "switch" according to probability */
1056         switch (header->proto) {
1057         case htons(ETH_P_IP):
1058         case htons(ETH_P_IPV6):
1059         case htons(ETH_P_TIPC):
1060                 neigh = ipoib_neigh_get(dev, phdr->hwaddr);
1061                 if (unlikely(!neigh)) {
1062                         neigh_add_path(skb, phdr->hwaddr, dev);
1063                         return NETDEV_TX_OK;
1064                 }
1065                 break;
1066         case htons(ETH_P_ARP):
1067         case htons(ETH_P_RARP):
1068                 /* for unicast ARP and RARP should always perform path find */
1069                 unicast_arp_send(skb, dev, phdr);
1070                 return NETDEV_TX_OK;
1071         default:
1072                 /* ethertype not supported by IPoIB */
1073                 ++dev->stats.tx_dropped;
1074                 dev_kfree_skb_any(skb);
1075                 return NETDEV_TX_OK;
1076         }
1077
1078 send_using_neigh:
1079         /* note we now hold a ref to neigh */
1080         if (ipoib_cm_get(neigh)) {
1081                 if (ipoib_cm_up(neigh)) {
1082                         ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
1083                         goto unref;
1084                 }
1085         } else if (neigh->ah) {
1086                 neigh->ah->last_send = rn->send(dev, skb, neigh->ah->ah,
1087                                                 IPOIB_QPN(phdr->hwaddr));
1088                 goto unref;
1089         }
1090
1091         if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
1092                 push_pseudo_header(skb, phdr->hwaddr);
1093                 spin_lock_irqsave(&priv->lock, flags);
1094                 __skb_queue_tail(&neigh->queue, skb);
1095                 spin_unlock_irqrestore(&priv->lock, flags);
1096         } else {
1097                 ++dev->stats.tx_dropped;
1098                 dev_kfree_skb_any(skb);
1099         }
1100
1101 unref:
1102         ipoib_neigh_put(neigh);
1103
1104         return NETDEV_TX_OK;
1105 }
1106
1107 static void ipoib_timeout(struct net_device *dev)
1108 {
1109         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1110
1111         ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
1112                    jiffies_to_msecs(jiffies - dev_trans_start(dev)));
1113         ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
1114                    netif_queue_stopped(dev),
1115                    priv->tx_head, priv->tx_tail);
1116         /* XXX reset QP, etc. */
1117 }
1118
1119 static int ipoib_hard_header(struct sk_buff *skb,
1120                              struct net_device *dev,
1121                              unsigned short type,
1122                              const void *daddr, const void *saddr, unsigned len)
1123 {
1124         struct ipoib_header *header;
1125
1126         header = (struct ipoib_header *) skb_push(skb, sizeof *header);
1127
1128         header->proto = htons(type);
1129         header->reserved = 0;
1130
1131         /*
1132          * we don't rely on dst_entry structure,  always stuff the
1133          * destination address into skb hard header so we can figure out where
1134          * to send the packet later.
1135          */
1136         push_pseudo_header(skb, daddr);
1137
1138         return IPOIB_HARD_LEN;
1139 }
1140
1141 static void ipoib_set_mcast_list(struct net_device *dev)
1142 {
1143         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1144
1145         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
1146                 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
1147                 return;
1148         }
1149
1150         queue_work(priv->wq, &priv->restart_task);
1151 }
1152
1153 static int ipoib_get_iflink(const struct net_device *dev)
1154 {
1155         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1156
1157         /* parent interface */
1158         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags))
1159                 return dev->ifindex;
1160
1161         /* child/vlan interface */
1162         return priv->parent->ifindex;
1163 }
1164
1165 static u32 ipoib_addr_hash(struct ipoib_neigh_hash *htbl, u8 *daddr)
1166 {
1167         /*
1168          * Use only the address parts that contributes to spreading
1169          * The subnet prefix is not used as one can not connect to
1170          * same remote port (GUID) using the same remote QPN via two
1171          * different subnets.
1172          */
1173          /* qpn octets[1:4) & port GUID octets[12:20) */
1174         u32 *d32 = (u32 *) daddr;
1175         u32 hv;
1176
1177         hv = jhash_3words(d32[3], d32[4], IPOIB_QPN_MASK & d32[0], 0);
1178         return hv & htbl->mask;
1179 }
1180
1181 struct ipoib_neigh *ipoib_neigh_get(struct net_device *dev, u8 *daddr)
1182 {
1183         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1184         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1185         struct ipoib_neigh_hash *htbl;
1186         struct ipoib_neigh *neigh = NULL;
1187         u32 hash_val;
1188
1189         rcu_read_lock_bh();
1190
1191         htbl = rcu_dereference_bh(ntbl->htbl);
1192
1193         if (!htbl)
1194                 goto out_unlock;
1195
1196         hash_val = ipoib_addr_hash(htbl, daddr);
1197         for (neigh = rcu_dereference_bh(htbl->buckets[hash_val]);
1198              neigh != NULL;
1199              neigh = rcu_dereference_bh(neigh->hnext)) {
1200                 if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) {
1201                         /* found, take one ref on behalf of the caller */
1202                         if (!atomic_inc_not_zero(&neigh->refcnt)) {
1203                                 /* deleted */
1204                                 neigh = NULL;
1205                                 goto out_unlock;
1206                         }
1207
1208                         if (likely(skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE))
1209                                 neigh->alive = jiffies;
1210                         goto out_unlock;
1211                 }
1212         }
1213
1214 out_unlock:
1215         rcu_read_unlock_bh();
1216         return neigh;
1217 }
1218
1219 static void __ipoib_reap_neigh(struct ipoib_dev_priv *priv)
1220 {
1221         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1222         struct ipoib_neigh_hash *htbl;
1223         unsigned long neigh_obsolete;
1224         unsigned long dt;
1225         unsigned long flags;
1226         int i;
1227         LIST_HEAD(remove_list);
1228
1229         if (test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
1230                 return;
1231
1232         spin_lock_irqsave(&priv->lock, flags);
1233
1234         htbl = rcu_dereference_protected(ntbl->htbl,
1235                                          lockdep_is_held(&priv->lock));
1236
1237         if (!htbl)
1238                 goto out_unlock;
1239
1240         /* neigh is obsolete if it was idle for two GC periods */
1241         dt = 2 * arp_tbl.gc_interval;
1242         neigh_obsolete = jiffies - dt;
1243         /* handle possible race condition */
1244         if (test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
1245                 goto out_unlock;
1246
1247         for (i = 0; i < htbl->size; i++) {
1248                 struct ipoib_neigh *neigh;
1249                 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1250
1251                 while ((neigh = rcu_dereference_protected(*np,
1252                                                           lockdep_is_held(&priv->lock))) != NULL) {
1253                         /* was the neigh idle for two GC periods */
1254                         if (time_after(neigh_obsolete, neigh->alive)) {
1255
1256                                 ipoib_check_and_add_mcast_sendonly(priv, neigh->daddr + 4, &remove_list);
1257
1258                                 rcu_assign_pointer(*np,
1259                                                    rcu_dereference_protected(neigh->hnext,
1260                                                                              lockdep_is_held(&priv->lock)));
1261                                 /* remove from path/mc list */
1262                                 list_del_init(&neigh->list);
1263                                 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1264                         } else {
1265                                 np = &neigh->hnext;
1266                         }
1267
1268                 }
1269         }
1270
1271 out_unlock:
1272         spin_unlock_irqrestore(&priv->lock, flags);
1273         ipoib_mcast_remove_list(&remove_list);
1274 }
1275
1276 static void ipoib_reap_neigh(struct work_struct *work)
1277 {
1278         struct ipoib_dev_priv *priv =
1279                 container_of(work, struct ipoib_dev_priv, neigh_reap_task.work);
1280
1281         __ipoib_reap_neigh(priv);
1282
1283         if (!test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
1284                 queue_delayed_work(priv->wq, &priv->neigh_reap_task,
1285                                    arp_tbl.gc_interval);
1286 }
1287
1288
1289 static struct ipoib_neigh *ipoib_neigh_ctor(u8 *daddr,
1290                                       struct net_device *dev)
1291 {
1292         struct ipoib_neigh *neigh;
1293
1294         neigh = kzalloc(sizeof *neigh, GFP_ATOMIC);
1295         if (!neigh)
1296                 return NULL;
1297
1298         neigh->dev = dev;
1299         memcpy(&neigh->daddr, daddr, sizeof(neigh->daddr));
1300         skb_queue_head_init(&neigh->queue);
1301         INIT_LIST_HEAD(&neigh->list);
1302         ipoib_cm_set(neigh, NULL);
1303         /* one ref on behalf of the caller */
1304         atomic_set(&neigh->refcnt, 1);
1305
1306         return neigh;
1307 }
1308
1309 struct ipoib_neigh *ipoib_neigh_alloc(u8 *daddr,
1310                                       struct net_device *dev)
1311 {
1312         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1313         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1314         struct ipoib_neigh_hash *htbl;
1315         struct ipoib_neigh *neigh;
1316         u32 hash_val;
1317
1318         htbl = rcu_dereference_protected(ntbl->htbl,
1319                                          lockdep_is_held(&priv->lock));
1320         if (!htbl) {
1321                 neigh = NULL;
1322                 goto out_unlock;
1323         }
1324
1325         /* need to add a new neigh, but maybe some other thread succeeded?
1326          * recalc hash, maybe hash resize took place so we do a search
1327          */
1328         hash_val = ipoib_addr_hash(htbl, daddr);
1329         for (neigh = rcu_dereference_protected(htbl->buckets[hash_val],
1330                                                lockdep_is_held(&priv->lock));
1331              neigh != NULL;
1332              neigh = rcu_dereference_protected(neigh->hnext,
1333                                                lockdep_is_held(&priv->lock))) {
1334                 if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) {
1335                         /* found, take one ref on behalf of the caller */
1336                         if (!atomic_inc_not_zero(&neigh->refcnt)) {
1337                                 /* deleted */
1338                                 neigh = NULL;
1339                                 break;
1340                         }
1341                         neigh->alive = jiffies;
1342                         goto out_unlock;
1343                 }
1344         }
1345
1346         neigh = ipoib_neigh_ctor(daddr, dev);
1347         if (!neigh)
1348                 goto out_unlock;
1349
1350         /* one ref on behalf of the hash table */
1351         atomic_inc(&neigh->refcnt);
1352         neigh->alive = jiffies;
1353         /* put in hash */
1354         rcu_assign_pointer(neigh->hnext,
1355                            rcu_dereference_protected(htbl->buckets[hash_val],
1356                                                      lockdep_is_held(&priv->lock)));
1357         rcu_assign_pointer(htbl->buckets[hash_val], neigh);
1358         atomic_inc(&ntbl->entries);
1359
1360 out_unlock:
1361
1362         return neigh;
1363 }
1364
1365 void ipoib_neigh_dtor(struct ipoib_neigh *neigh)
1366 {
1367         /* neigh reference count was dropprd to zero */
1368         struct net_device *dev = neigh->dev;
1369         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1370         struct sk_buff *skb;
1371         if (neigh->ah)
1372                 ipoib_put_ah(neigh->ah);
1373         while ((skb = __skb_dequeue(&neigh->queue))) {
1374                 ++dev->stats.tx_dropped;
1375                 dev_kfree_skb_any(skb);
1376         }
1377         if (ipoib_cm_get(neigh))
1378                 ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
1379         ipoib_dbg(ipoib_priv(dev),
1380                   "neigh free for %06x %pI6\n",
1381                   IPOIB_QPN(neigh->daddr),
1382                   neigh->daddr + 4);
1383         kfree(neigh);
1384         if (atomic_dec_and_test(&priv->ntbl.entries)) {
1385                 if (test_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags))
1386                         complete(&priv->ntbl.flushed);
1387         }
1388 }
1389
1390 static void ipoib_neigh_reclaim(struct rcu_head *rp)
1391 {
1392         /* Called as a result of removal from hash table */
1393         struct ipoib_neigh *neigh = container_of(rp, struct ipoib_neigh, rcu);
1394         /* note TX context may hold another ref */
1395         ipoib_neigh_put(neigh);
1396 }
1397
1398 void ipoib_neigh_free(struct ipoib_neigh *neigh)
1399 {
1400         struct net_device *dev = neigh->dev;
1401         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1402         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1403         struct ipoib_neigh_hash *htbl;
1404         struct ipoib_neigh __rcu **np;
1405         struct ipoib_neigh *n;
1406         u32 hash_val;
1407
1408         htbl = rcu_dereference_protected(ntbl->htbl,
1409                                         lockdep_is_held(&priv->lock));
1410         if (!htbl)
1411                 return;
1412
1413         hash_val = ipoib_addr_hash(htbl, neigh->daddr);
1414         np = &htbl->buckets[hash_val];
1415         for (n = rcu_dereference_protected(*np,
1416                                             lockdep_is_held(&priv->lock));
1417              n != NULL;
1418              n = rcu_dereference_protected(*np,
1419                                         lockdep_is_held(&priv->lock))) {
1420                 if (n == neigh) {
1421                         /* found */
1422                         rcu_assign_pointer(*np,
1423                                            rcu_dereference_protected(neigh->hnext,
1424                                                                      lockdep_is_held(&priv->lock)));
1425                         /* remove from parent list */
1426                         list_del_init(&neigh->list);
1427                         call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1428                         return;
1429                 } else {
1430                         np = &n->hnext;
1431                 }
1432         }
1433 }
1434
1435 static int ipoib_neigh_hash_init(struct ipoib_dev_priv *priv)
1436 {
1437         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1438         struct ipoib_neigh_hash *htbl;
1439         struct ipoib_neigh __rcu **buckets;
1440         u32 size;
1441
1442         clear_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags);
1443         ntbl->htbl = NULL;
1444         htbl = kzalloc(sizeof(*htbl), GFP_KERNEL);
1445         if (!htbl)
1446                 return -ENOMEM;
1447         set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1448         size = roundup_pow_of_two(arp_tbl.gc_thresh3);
1449         buckets = kzalloc(size * sizeof(*buckets), GFP_KERNEL);
1450         if (!buckets) {
1451                 kfree(htbl);
1452                 return -ENOMEM;
1453         }
1454         htbl->size = size;
1455         htbl->mask = (size - 1);
1456         htbl->buckets = buckets;
1457         RCU_INIT_POINTER(ntbl->htbl, htbl);
1458         htbl->ntbl = ntbl;
1459         atomic_set(&ntbl->entries, 0);
1460
1461         /* start garbage collection */
1462         clear_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1463         queue_delayed_work(priv->wq, &priv->neigh_reap_task,
1464                            arp_tbl.gc_interval);
1465
1466         return 0;
1467 }
1468
1469 static void neigh_hash_free_rcu(struct rcu_head *head)
1470 {
1471         struct ipoib_neigh_hash *htbl = container_of(head,
1472                                                     struct ipoib_neigh_hash,
1473                                                     rcu);
1474         struct ipoib_neigh __rcu **buckets = htbl->buckets;
1475         struct ipoib_neigh_table *ntbl = htbl->ntbl;
1476
1477         kfree(buckets);
1478         kfree(htbl);
1479         complete(&ntbl->deleted);
1480 }
1481
1482 void ipoib_del_neighs_by_gid(struct net_device *dev, u8 *gid)
1483 {
1484         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1485         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1486         struct ipoib_neigh_hash *htbl;
1487         unsigned long flags;
1488         int i;
1489
1490         /* remove all neigh connected to a given path or mcast */
1491         spin_lock_irqsave(&priv->lock, flags);
1492
1493         htbl = rcu_dereference_protected(ntbl->htbl,
1494                                          lockdep_is_held(&priv->lock));
1495
1496         if (!htbl)
1497                 goto out_unlock;
1498
1499         for (i = 0; i < htbl->size; i++) {
1500                 struct ipoib_neigh *neigh;
1501                 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1502
1503                 while ((neigh = rcu_dereference_protected(*np,
1504                                                           lockdep_is_held(&priv->lock))) != NULL) {
1505                         /* delete neighs belong to this parent */
1506                         if (!memcmp(gid, neigh->daddr + 4, sizeof (union ib_gid))) {
1507                                 rcu_assign_pointer(*np,
1508                                                    rcu_dereference_protected(neigh->hnext,
1509                                                                              lockdep_is_held(&priv->lock)));
1510                                 /* remove from parent list */
1511                                 list_del_init(&neigh->list);
1512                                 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1513                         } else {
1514                                 np = &neigh->hnext;
1515                         }
1516
1517                 }
1518         }
1519 out_unlock:
1520         spin_unlock_irqrestore(&priv->lock, flags);
1521 }
1522
1523 static void ipoib_flush_neighs(struct ipoib_dev_priv *priv)
1524 {
1525         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1526         struct ipoib_neigh_hash *htbl;
1527         unsigned long flags;
1528         int i, wait_flushed = 0;
1529
1530         init_completion(&priv->ntbl.flushed);
1531
1532         spin_lock_irqsave(&priv->lock, flags);
1533
1534         htbl = rcu_dereference_protected(ntbl->htbl,
1535                                         lockdep_is_held(&priv->lock));
1536         if (!htbl)
1537                 goto out_unlock;
1538
1539         wait_flushed = atomic_read(&priv->ntbl.entries);
1540         if (!wait_flushed)
1541                 goto free_htbl;
1542
1543         for (i = 0; i < htbl->size; i++) {
1544                 struct ipoib_neigh *neigh;
1545                 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1546
1547                 while ((neigh = rcu_dereference_protected(*np,
1548                                        lockdep_is_held(&priv->lock))) != NULL) {
1549                         rcu_assign_pointer(*np,
1550                                            rcu_dereference_protected(neigh->hnext,
1551                                                                      lockdep_is_held(&priv->lock)));
1552                         /* remove from path/mc list */
1553                         list_del_init(&neigh->list);
1554                         call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1555                 }
1556         }
1557
1558 free_htbl:
1559         rcu_assign_pointer(ntbl->htbl, NULL);
1560         call_rcu(&htbl->rcu, neigh_hash_free_rcu);
1561
1562 out_unlock:
1563         spin_unlock_irqrestore(&priv->lock, flags);
1564         if (wait_flushed)
1565                 wait_for_completion(&priv->ntbl.flushed);
1566 }
1567
1568 static void ipoib_neigh_hash_uninit(struct net_device *dev)
1569 {
1570         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1571         int stopped;
1572
1573         ipoib_dbg(priv, "ipoib_neigh_hash_uninit\n");
1574         init_completion(&priv->ntbl.deleted);
1575         set_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags);
1576
1577         /* Stop GC if called at init fail need to cancel work */
1578         stopped = test_and_set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1579         if (!stopped)
1580                 cancel_delayed_work(&priv->neigh_reap_task);
1581
1582         ipoib_flush_neighs(priv);
1583
1584         wait_for_completion(&priv->ntbl.deleted);
1585 }
1586
1587 void ipoib_dev_uninit_default(struct net_device *dev)
1588 {
1589         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1590
1591         ipoib_transport_dev_cleanup(dev);
1592
1593         ipoib_cm_dev_cleanup(dev);
1594
1595         kfree(priv->rx_ring);
1596         vfree(priv->tx_ring);
1597
1598         priv->rx_ring = NULL;
1599         priv->tx_ring = NULL;
1600 }
1601
1602 static int ipoib_dev_init_default(struct net_device *dev)
1603 {
1604         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1605
1606         netif_napi_add(dev, &priv->napi, ipoib_poll, NAPI_POLL_WEIGHT);
1607
1608         /* Allocate RX/TX "rings" to hold queued skbs */
1609         priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
1610                                 GFP_KERNEL);
1611         if (!priv->rx_ring)
1612                 goto out;
1613
1614         priv->tx_ring = vzalloc(ipoib_sendq_size * sizeof *priv->tx_ring);
1615         if (!priv->tx_ring) {
1616                 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
1617                        priv->ca->name, ipoib_sendq_size);
1618                 goto out_rx_ring_cleanup;
1619         }
1620
1621         /* priv->tx_head, tx_tail & tx_outstanding are already 0 */
1622
1623         if (ipoib_transport_dev_init(dev, priv->ca)) {
1624                 pr_warn("%s: ipoib_transport_dev_init failed\n",
1625                         priv->ca->name);
1626                 goto out_tx_ring_cleanup;
1627         }
1628
1629         /* after qp created set dev address */
1630         priv->dev->dev_addr[1] = (priv->qp->qp_num >> 16) & 0xff;
1631         priv->dev->dev_addr[2] = (priv->qp->qp_num >>  8) & 0xff;
1632         priv->dev->dev_addr[3] = (priv->qp->qp_num) & 0xff;
1633
1634         setup_timer(&priv->poll_timer, ipoib_ib_tx_timer_func,
1635                     (unsigned long)dev);
1636
1637         return 0;
1638
1639 out_tx_ring_cleanup:
1640         vfree(priv->tx_ring);
1641
1642 out_rx_ring_cleanup:
1643         kfree(priv->rx_ring);
1644
1645 out:
1646         return -ENOMEM;
1647 }
1648
1649 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
1650 {
1651         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1652         int ret = -ENOMEM;
1653
1654         priv->ca = ca;
1655         priv->port = port;
1656         priv->qp = NULL;
1657
1658         /*
1659          * the various IPoIB tasks assume they will never race against
1660          * themselves, so always use a single thread workqueue
1661          */
1662         priv->wq = alloc_ordered_workqueue("ipoib_wq", WQ_MEM_RECLAIM);
1663         if (!priv->wq) {
1664                 pr_warn("%s: failed to allocate device WQ\n", dev->name);
1665                 goto out;
1666         }
1667
1668         /* create pd, which used both for control and datapath*/
1669         priv->pd = ib_alloc_pd(priv->ca, 0);
1670         if (IS_ERR(priv->pd)) {
1671                 pr_warn("%s: failed to allocate PD\n", ca->name);
1672                 goto clean_wq;
1673         }
1674
1675         ret = priv->rn_ops->ndo_init(dev);
1676         if (ret) {
1677                 pr_warn("%s failed to init HW resource\n", dev->name);
1678                 goto out_free_pd;
1679         }
1680
1681         if (ipoib_neigh_hash_init(priv) < 0) {
1682                 pr_warn("%s failed to init neigh hash\n", dev->name);
1683                 goto out_dev_uninit;
1684         }
1685
1686         if (dev->flags & IFF_UP) {
1687                 if (ipoib_ib_dev_open(dev)) {
1688                         pr_warn("%s failed to open device\n", dev->name);
1689                         ret = -ENODEV;
1690                         goto out_dev_uninit;
1691                 }
1692         }
1693
1694         return 0;
1695
1696 out_dev_uninit:
1697         ipoib_ib_dev_cleanup(dev);
1698
1699 out_free_pd:
1700         if (priv->pd) {
1701                 ib_dealloc_pd(priv->pd);
1702                 priv->pd = NULL;
1703         }
1704
1705 clean_wq:
1706         if (priv->wq) {
1707                 destroy_workqueue(priv->wq);
1708                 priv->wq = NULL;
1709         }
1710
1711 out:
1712         return ret;
1713 }
1714
1715 void ipoib_dev_cleanup(struct net_device *dev)
1716 {
1717         struct ipoib_dev_priv *priv = ipoib_priv(dev), *cpriv, *tcpriv;
1718         LIST_HEAD(head);
1719
1720         ASSERT_RTNL();
1721
1722         /* Delete any child interfaces first */
1723         list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
1724                 /* Stop GC on child */
1725                 set_bit(IPOIB_STOP_NEIGH_GC, &cpriv->flags);
1726                 cancel_delayed_work(&cpriv->neigh_reap_task);
1727                 unregister_netdevice_queue(cpriv->dev, &head);
1728         }
1729         unregister_netdevice_many(&head);
1730
1731         ipoib_neigh_hash_uninit(dev);
1732
1733         ipoib_ib_dev_cleanup(dev);
1734
1735         /* no more works over the priv->wq */
1736         if (priv->wq) {
1737                 flush_workqueue(priv->wq);
1738                 destroy_workqueue(priv->wq);
1739                 priv->wq = NULL;
1740         }
1741 }
1742
1743 static int ipoib_set_vf_link_state(struct net_device *dev, int vf, int link_state)
1744 {
1745         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1746
1747         return ib_set_vf_link_state(priv->ca, vf, priv->port, link_state);
1748 }
1749
1750 static int ipoib_get_vf_config(struct net_device *dev, int vf,
1751                                struct ifla_vf_info *ivf)
1752 {
1753         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1754         int err;
1755
1756         err = ib_get_vf_config(priv->ca, vf, priv->port, ivf);
1757         if (err)
1758                 return err;
1759
1760         ivf->vf = vf;
1761
1762         return 0;
1763 }
1764
1765 static int ipoib_set_vf_guid(struct net_device *dev, int vf, u64 guid, int type)
1766 {
1767         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1768
1769         if (type != IFLA_VF_IB_NODE_GUID && type != IFLA_VF_IB_PORT_GUID)
1770                 return -EINVAL;
1771
1772         return ib_set_vf_guid(priv->ca, vf, priv->port, guid, type);
1773 }
1774
1775 static int ipoib_get_vf_stats(struct net_device *dev, int vf,
1776                               struct ifla_vf_stats *vf_stats)
1777 {
1778         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1779
1780         return ib_get_vf_stats(priv->ca, vf, priv->port, vf_stats);
1781 }
1782
1783 static const struct header_ops ipoib_header_ops = {
1784         .create = ipoib_hard_header,
1785 };
1786
1787 static const struct net_device_ops ipoib_netdev_ops_pf = {
1788         .ndo_uninit              = ipoib_uninit,
1789         .ndo_open                = ipoib_open,
1790         .ndo_stop                = ipoib_stop,
1791         .ndo_change_mtu          = ipoib_change_mtu,
1792         .ndo_fix_features        = ipoib_fix_features,
1793         .ndo_start_xmit          = ipoib_start_xmit,
1794         .ndo_tx_timeout          = ipoib_timeout,
1795         .ndo_set_rx_mode         = ipoib_set_mcast_list,
1796         .ndo_get_iflink          = ipoib_get_iflink,
1797         .ndo_set_vf_link_state   = ipoib_set_vf_link_state,
1798         .ndo_get_vf_config       = ipoib_get_vf_config,
1799         .ndo_get_vf_stats        = ipoib_get_vf_stats,
1800         .ndo_set_vf_guid         = ipoib_set_vf_guid,
1801         .ndo_set_mac_address     = ipoib_set_mac,
1802 };
1803
1804 static const struct net_device_ops ipoib_netdev_ops_vf = {
1805         .ndo_uninit              = ipoib_uninit,
1806         .ndo_open                = ipoib_open,
1807         .ndo_stop                = ipoib_stop,
1808         .ndo_change_mtu          = ipoib_change_mtu,
1809         .ndo_fix_features        = ipoib_fix_features,
1810         .ndo_start_xmit          = ipoib_start_xmit,
1811         .ndo_tx_timeout          = ipoib_timeout,
1812         .ndo_set_rx_mode         = ipoib_set_mcast_list,
1813         .ndo_get_iflink          = ipoib_get_iflink,
1814 };
1815
1816 void ipoib_setup_common(struct net_device *dev)
1817 {
1818         dev->header_ops          = &ipoib_header_ops;
1819
1820         ipoib_set_ethtool_ops(dev);
1821
1822         dev->watchdog_timeo      = HZ;
1823
1824         dev->flags              |= IFF_BROADCAST | IFF_MULTICAST;
1825
1826         dev->hard_header_len     = IPOIB_HARD_LEN;
1827         dev->addr_len            = INFINIBAND_ALEN;
1828         dev->type                = ARPHRD_INFINIBAND;
1829         dev->tx_queue_len        = ipoib_sendq_size * 2;
1830         dev->features            = (NETIF_F_VLAN_CHALLENGED     |
1831                                     NETIF_F_HIGHDMA);
1832         netif_keep_dst(dev);
1833
1834         memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
1835 }
1836
1837 static void ipoib_build_priv(struct net_device *dev)
1838 {
1839         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1840
1841         priv->dev = dev;
1842         spin_lock_init(&priv->lock);
1843         init_rwsem(&priv->vlan_rwsem);
1844
1845         INIT_LIST_HEAD(&priv->path_list);
1846         INIT_LIST_HEAD(&priv->child_intfs);
1847         INIT_LIST_HEAD(&priv->dead_ahs);
1848         INIT_LIST_HEAD(&priv->multicast_list);
1849
1850         INIT_DELAYED_WORK(&priv->mcast_task,   ipoib_mcast_join_task);
1851         INIT_WORK(&priv->carrier_on_task, ipoib_mcast_carrier_on_task);
1852         INIT_WORK(&priv->flush_light,   ipoib_ib_dev_flush_light);
1853         INIT_WORK(&priv->flush_normal,   ipoib_ib_dev_flush_normal);
1854         INIT_WORK(&priv->flush_heavy,   ipoib_ib_dev_flush_heavy);
1855         INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
1856         INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
1857         INIT_DELAYED_WORK(&priv->neigh_reap_task, ipoib_reap_neigh);
1858 }
1859
1860 static const struct net_device_ops ipoib_netdev_default_pf = {
1861         .ndo_init                = ipoib_dev_init_default,
1862         .ndo_uninit              = ipoib_dev_uninit_default,
1863         .ndo_open                = ipoib_ib_dev_open_default,
1864         .ndo_stop                = ipoib_ib_dev_stop_default,
1865 };
1866
1867 static struct net_device
1868 *ipoib_create_netdev_default(struct ib_device *hca,
1869                              const char *name,
1870                              unsigned char name_assign_type,
1871                              void (*setup)(struct net_device *))
1872 {
1873         struct net_device *dev;
1874         struct rdma_netdev *rn;
1875
1876         dev = alloc_netdev((int)sizeof(struct rdma_netdev),
1877                            name,
1878                            name_assign_type, setup);
1879         if (!dev)
1880                 return NULL;
1881
1882         rn = netdev_priv(dev);
1883
1884         rn->send = ipoib_send;
1885         rn->attach_mcast = ipoib_mcast_attach;
1886         rn->detach_mcast = ipoib_mcast_detach;
1887         rn->hca = hca;
1888
1889         dev->netdev_ops = &ipoib_netdev_default_pf;
1890
1891         return dev;
1892 }
1893
1894 static struct net_device *ipoib_get_netdev(struct ib_device *hca, u8 port,
1895                                            const char *name)
1896 {
1897         struct net_device *dev;
1898
1899         if (hca->alloc_rdma_netdev) {
1900                 dev = hca->alloc_rdma_netdev(hca, port,
1901                                              RDMA_NETDEV_IPOIB, name,
1902                                              NET_NAME_UNKNOWN,
1903                                              ipoib_setup_common);
1904                 if (IS_ERR_OR_NULL(dev) && PTR_ERR(dev) != -EOPNOTSUPP)
1905                         return NULL;
1906         }
1907
1908         if (!hca->alloc_rdma_netdev || PTR_ERR(dev) == -EOPNOTSUPP)
1909                 dev = ipoib_create_netdev_default(hca, name, NET_NAME_UNKNOWN,
1910                                                   ipoib_setup_common);
1911
1912         return dev;
1913 }
1914
1915 struct ipoib_dev_priv *ipoib_intf_alloc(struct ib_device *hca, u8 port,
1916                                         const char *name)
1917 {
1918         struct net_device *dev;
1919         struct ipoib_dev_priv *priv;
1920         struct rdma_netdev *rn;
1921
1922         priv = kzalloc(sizeof(*priv), GFP_KERNEL);
1923         if (!priv)
1924                 return NULL;
1925
1926         dev = ipoib_get_netdev(hca, port, name);
1927         if (!dev)
1928                 goto free_priv;
1929
1930         priv->rn_ops = dev->netdev_ops;
1931
1932         /* fixme : should be after the query_cap */
1933         if (priv->hca_caps & IB_DEVICE_VIRTUAL_FUNCTION)
1934                 dev->netdev_ops = &ipoib_netdev_ops_vf;
1935         else
1936                 dev->netdev_ops = &ipoib_netdev_ops_pf;
1937
1938         rn = netdev_priv(dev);
1939         rn->clnt_priv = priv;
1940         ipoib_build_priv(dev);
1941
1942         return priv;
1943 free_priv:
1944         kfree(priv);
1945         return NULL;
1946 }
1947
1948 static ssize_t show_pkey(struct device *dev,
1949                          struct device_attribute *attr, char *buf)
1950 {
1951         struct net_device *ndev = to_net_dev(dev);
1952         struct ipoib_dev_priv *priv = ipoib_priv(ndev);
1953
1954         return sprintf(buf, "0x%04x\n", priv->pkey);
1955 }
1956 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1957
1958 static ssize_t show_umcast(struct device *dev,
1959                            struct device_attribute *attr, char *buf)
1960 {
1961         struct net_device *ndev = to_net_dev(dev);
1962         struct ipoib_dev_priv *priv = ipoib_priv(ndev);
1963
1964         return sprintf(buf, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
1965 }
1966
1967 void ipoib_set_umcast(struct net_device *ndev, int umcast_val)
1968 {
1969         struct ipoib_dev_priv *priv = ipoib_priv(ndev);
1970
1971         if (umcast_val > 0) {
1972                 set_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1973                 ipoib_warn(priv, "ignoring multicast groups joined directly "
1974                                 "by userspace\n");
1975         } else
1976                 clear_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1977 }
1978
1979 static ssize_t set_umcast(struct device *dev,
1980                           struct device_attribute *attr,
1981                           const char *buf, size_t count)
1982 {
1983         unsigned long umcast_val = simple_strtoul(buf, NULL, 0);
1984
1985         ipoib_set_umcast(to_net_dev(dev), umcast_val);
1986
1987         return count;
1988 }
1989 static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast);
1990
1991 int ipoib_add_umcast_attr(struct net_device *dev)
1992 {
1993         return device_create_file(&dev->dev, &dev_attr_umcast);
1994 }
1995
1996 static void set_base_guid(struct ipoib_dev_priv *priv, union ib_gid *gid)
1997 {
1998         struct ipoib_dev_priv *child_priv;
1999         struct net_device *netdev = priv->dev;
2000
2001         netif_addr_lock_bh(netdev);
2002
2003         memcpy(&priv->local_gid.global.interface_id,
2004                &gid->global.interface_id,
2005                sizeof(gid->global.interface_id));
2006         memcpy(netdev->dev_addr + 4, &priv->local_gid, sizeof(priv->local_gid));
2007         clear_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags);
2008
2009         netif_addr_unlock_bh(netdev);
2010
2011         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
2012                 down_read(&priv->vlan_rwsem);
2013                 list_for_each_entry(child_priv, &priv->child_intfs, list)
2014                         set_base_guid(child_priv, gid);
2015                 up_read(&priv->vlan_rwsem);
2016         }
2017 }
2018
2019 static int ipoib_check_lladdr(struct net_device *dev,
2020                               struct sockaddr_storage *ss)
2021 {
2022         union ib_gid *gid = (union ib_gid *)(ss->__data + 4);
2023         int ret = 0;
2024
2025         netif_addr_lock_bh(dev);
2026
2027         /* Make sure the QPN, reserved and subnet prefix match the current
2028          * lladdr, it also makes sure the lladdr is unicast.
2029          */
2030         if (memcmp(dev->dev_addr, ss->__data,
2031                    4 + sizeof(gid->global.subnet_prefix)) ||
2032             gid->global.interface_id == 0)
2033                 ret = -EINVAL;
2034
2035         netif_addr_unlock_bh(dev);
2036
2037         return ret;
2038 }
2039
2040 static int ipoib_set_mac(struct net_device *dev, void *addr)
2041 {
2042         struct ipoib_dev_priv *priv = ipoib_priv(dev);
2043         struct sockaddr_storage *ss = addr;
2044         int ret;
2045
2046         if (!(dev->priv_flags & IFF_LIVE_ADDR_CHANGE) && netif_running(dev))
2047                 return -EBUSY;
2048
2049         ret = ipoib_check_lladdr(dev, ss);
2050         if (ret)
2051                 return ret;
2052
2053         set_base_guid(priv, (union ib_gid *)(ss->__data + 4));
2054
2055         queue_work(ipoib_workqueue, &priv->flush_light);
2056
2057         return 0;
2058 }
2059
2060 static ssize_t create_child(struct device *dev,
2061                             struct device_attribute *attr,
2062                             const char *buf, size_t count)
2063 {
2064         int pkey;
2065         int ret;
2066
2067         if (sscanf(buf, "%i", &pkey) != 1)
2068                 return -EINVAL;
2069
2070         if (pkey <= 0 || pkey > 0xffff || pkey == 0x8000)
2071                 return -EINVAL;
2072
2073         /*
2074          * Set the full membership bit, so that we join the right
2075          * broadcast group, etc.
2076          */
2077         pkey |= 0x8000;
2078
2079         ret = ipoib_vlan_add(to_net_dev(dev), pkey);
2080
2081         return ret ? ret : count;
2082 }
2083 static DEVICE_ATTR(create_child, S_IWUSR, NULL, create_child);
2084
2085 static ssize_t delete_child(struct device *dev,
2086                             struct device_attribute *attr,
2087                             const char *buf, size_t count)
2088 {
2089         int pkey;
2090         int ret;
2091
2092         if (sscanf(buf, "%i", &pkey) != 1)
2093                 return -EINVAL;
2094
2095         if (pkey < 0 || pkey > 0xffff)
2096                 return -EINVAL;
2097
2098         ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
2099
2100         return ret ? ret : count;
2101
2102 }
2103 static DEVICE_ATTR(delete_child, S_IWUSR, NULL, delete_child);
2104
2105 int ipoib_add_pkey_attr(struct net_device *dev)
2106 {
2107         return device_create_file(&dev->dev, &dev_attr_pkey);
2108 }
2109
2110 void ipoib_set_dev_features(struct ipoib_dev_priv *priv, struct ib_device *hca)
2111 {
2112         priv->hca_caps = hca->attrs.device_cap_flags;
2113
2114         if (priv->hca_caps & IB_DEVICE_UD_IP_CSUM) {
2115                 priv->dev->hw_features |= NETIF_F_IP_CSUM | NETIF_F_RXCSUM;
2116
2117                 if (priv->hca_caps & IB_DEVICE_UD_TSO)
2118                         priv->dev->hw_features |= NETIF_F_TSO;
2119
2120                 priv->dev->features |= priv->dev->hw_features;
2121         }
2122 }
2123
2124 static struct net_device *ipoib_add_port(const char *format,
2125                                          struct ib_device *hca, u8 port)
2126 {
2127         struct ipoib_dev_priv *priv;
2128         struct ib_port_attr attr;
2129         int result = -ENOMEM;
2130
2131         priv = ipoib_intf_alloc(hca, port, format);
2132         if (!priv)
2133                 goto alloc_mem_failed;
2134
2135         SET_NETDEV_DEV(priv->dev, hca->dev.parent);
2136         priv->dev->dev_id = port - 1;
2137
2138         result = ib_query_port(hca, port, &attr);
2139         if (!result)
2140                 priv->max_ib_mtu = ib_mtu_enum_to_int(attr.max_mtu);
2141         else {
2142                 printk(KERN_WARNING "%s: ib_query_port %d failed\n",
2143                        hca->name, port);
2144                 goto device_init_failed;
2145         }
2146
2147         /* MTU will be reset when mcast join happens */
2148         priv->dev->mtu  = IPOIB_UD_MTU(priv->max_ib_mtu);
2149         priv->mcast_mtu  = priv->admin_mtu = priv->dev->mtu;
2150         priv->dev->max_mtu = IPOIB_CM_MTU;
2151
2152         priv->dev->neigh_priv_len = sizeof(struct ipoib_neigh);
2153
2154         result = ib_query_pkey(hca, port, 0, &priv->pkey);
2155         if (result) {
2156                 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
2157                        hca->name, port, result);
2158                 goto device_init_failed;
2159         }
2160
2161         ipoib_set_dev_features(priv, hca);
2162
2163         /*
2164          * Set the full membership bit, so that we join the right
2165          * broadcast group, etc.
2166          */
2167         priv->pkey |= 0x8000;
2168
2169         priv->dev->broadcast[8] = priv->pkey >> 8;
2170         priv->dev->broadcast[9] = priv->pkey & 0xff;
2171
2172         result = ib_query_gid(hca, port, 0, &priv->local_gid, NULL);
2173         if (result) {
2174                 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
2175                        hca->name, port, result);
2176                 goto device_init_failed;
2177         } else
2178                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
2179         set_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags);
2180
2181         result = ipoib_dev_init(priv->dev, hca, port);
2182         if (result < 0) {
2183                 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
2184                        hca->name, port, result);
2185                 goto device_init_failed;
2186         }
2187
2188         INIT_IB_EVENT_HANDLER(&priv->event_handler,
2189                               priv->ca, ipoib_event);
2190         result = ib_register_event_handler(&priv->event_handler);
2191         if (result < 0) {
2192                 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
2193                        "port %d (ret = %d)\n",
2194                        hca->name, port, result);
2195                 goto event_failed;
2196         }
2197
2198         result = register_netdev(priv->dev);
2199         if (result) {
2200                 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
2201                        hca->name, port, result);
2202                 goto register_failed;
2203         }
2204
2205         if (ipoib_cm_add_mode_attr(priv->dev))
2206                 goto sysfs_failed;
2207         if (ipoib_add_pkey_attr(priv->dev))
2208                 goto sysfs_failed;
2209         if (ipoib_add_umcast_attr(priv->dev))
2210                 goto sysfs_failed;
2211         if (device_create_file(&priv->dev->dev, &dev_attr_create_child))
2212                 goto sysfs_failed;
2213         if (device_create_file(&priv->dev->dev, &dev_attr_delete_child))
2214                 goto sysfs_failed;
2215
2216         return priv->dev;
2217
2218 sysfs_failed:
2219         unregister_netdev(priv->dev);
2220
2221 register_failed:
2222         ib_unregister_event_handler(&priv->event_handler);
2223         flush_workqueue(ipoib_workqueue);
2224         /* Stop GC if started before flush */
2225         set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
2226         cancel_delayed_work(&priv->neigh_reap_task);
2227         flush_workqueue(priv->wq);
2228
2229 event_failed:
2230         ipoib_dev_cleanup(priv->dev);
2231
2232 device_init_failed:
2233         free_netdev(priv->dev);
2234
2235 alloc_mem_failed:
2236         return ERR_PTR(result);
2237 }
2238
2239 static void ipoib_add_one(struct ib_device *device)
2240 {
2241         struct list_head *dev_list;
2242         struct net_device *dev;
2243         struct ipoib_dev_priv *priv;
2244         int p;
2245         int count = 0;
2246
2247         dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
2248         if (!dev_list)
2249                 return;
2250
2251         INIT_LIST_HEAD(dev_list);
2252
2253         for (p = rdma_start_port(device); p <= rdma_end_port(device); ++p) {
2254                 if (!rdma_protocol_ib(device, p))
2255                         continue;
2256                 dev = ipoib_add_port("ib%d", device, p);
2257                 if (!IS_ERR(dev)) {
2258                         priv = ipoib_priv(dev);
2259                         list_add_tail(&priv->list, dev_list);
2260                         count++;
2261                 }
2262         }
2263
2264         if (!count) {
2265                 kfree(dev_list);
2266                 return;
2267         }
2268
2269         ib_set_client_data(device, &ipoib_client, dev_list);
2270 }
2271
2272 static void ipoib_remove_one(struct ib_device *device, void *client_data)
2273 {
2274         struct ipoib_dev_priv *priv, *tmp;
2275         struct list_head *dev_list = client_data;
2276
2277         if (!dev_list)
2278                 return;
2279
2280         list_for_each_entry_safe(priv, tmp, dev_list, list) {
2281                 ib_unregister_event_handler(&priv->event_handler);
2282                 flush_workqueue(ipoib_workqueue);
2283
2284                 /* mark interface in the middle of destruction */
2285                 set_bit(IPOIB_FLAG_GOING_DOWN, &priv->flags);
2286
2287                 rtnl_lock();
2288                 dev_change_flags(priv->dev, priv->dev->flags & ~IFF_UP);
2289                 rtnl_unlock();
2290
2291                 /* Stop GC */
2292                 set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
2293                 cancel_delayed_work(&priv->neigh_reap_task);
2294                 flush_workqueue(priv->wq);
2295
2296                 unregister_netdev(priv->dev);
2297                 free_netdev(priv->dev);
2298                 kfree(priv);
2299         }
2300
2301         kfree(dev_list);
2302 }
2303
2304 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
2305 static struct notifier_block ipoib_netdev_notifier = {
2306         .notifier_call = ipoib_netdev_event,
2307 };
2308 #endif
2309
2310 static int __init ipoib_init_module(void)
2311 {
2312         int ret;
2313
2314         ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
2315         ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
2316         ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
2317
2318         ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
2319         ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
2320         ipoib_sendq_size = max3(ipoib_sendq_size, 2 * MAX_SEND_CQE, IPOIB_MIN_QUEUE_SIZE);
2321 #ifdef CONFIG_INFINIBAND_IPOIB_CM
2322         ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
2323 #endif
2324
2325         /*
2326          * When copying small received packets, we only copy from the
2327          * linear data part of the SKB, so we rely on this condition.
2328          */
2329         BUILD_BUG_ON(IPOIB_CM_COPYBREAK > IPOIB_CM_HEAD_SIZE);
2330
2331         ret = ipoib_register_debugfs();
2332         if (ret)
2333                 return ret;
2334
2335         /*
2336          * We create a global workqueue here that is used for all flush
2337          * operations.  However, if you attempt to flush a workqueue
2338          * from a task on that same workqueue, it deadlocks the system.
2339          * We want to be able to flush the tasks associated with a
2340          * specific net device, so we also create a workqueue for each
2341          * netdevice.  We queue up the tasks for that device only on
2342          * its private workqueue, and we only queue up flush events
2343          * on our global flush workqueue.  This avoids the deadlocks.
2344          */
2345         ipoib_workqueue = alloc_ordered_workqueue("ipoib_flush",
2346                                                   WQ_MEM_RECLAIM);
2347         if (!ipoib_workqueue) {
2348                 ret = -ENOMEM;
2349                 goto err_fs;
2350         }
2351
2352         ib_sa_register_client(&ipoib_sa_client);
2353
2354         ret = ib_register_client(&ipoib_client);
2355         if (ret)
2356                 goto err_sa;
2357
2358         ret = ipoib_netlink_init();
2359         if (ret)
2360                 goto err_client;
2361
2362 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
2363         register_netdevice_notifier(&ipoib_netdev_notifier);
2364 #endif
2365         return 0;
2366
2367 err_client:
2368         ib_unregister_client(&ipoib_client);
2369
2370 err_sa:
2371         ib_sa_unregister_client(&ipoib_sa_client);
2372         destroy_workqueue(ipoib_workqueue);
2373
2374 err_fs:
2375         ipoib_unregister_debugfs();
2376
2377         return ret;
2378 }
2379
2380 static void __exit ipoib_cleanup_module(void)
2381 {
2382 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
2383         unregister_netdevice_notifier(&ipoib_netdev_notifier);
2384 #endif
2385         ipoib_netlink_fini();
2386         ib_unregister_client(&ipoib_client);
2387         ib_sa_unregister_client(&ipoib_sa_client);
2388         ipoib_unregister_debugfs();
2389         destroy_workqueue(ipoib_workqueue);
2390 }
2391
2392 module_init(ipoib_init_module);
2393 module_exit(ipoib_cleanup_module);