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