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
  55 #define DRV_VERSION "1.0.0"
  56
  57 const char ipoib_driver_version[] = DRV_VERSION;
  58
  59 MODULE_AUTHOR("Roland Dreier");
  60 MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
  61 MODULE_LICENSE("Dual BSD/GPL");
  62
  63 int ipoib_sendq_size __read_mostly = IPOIB_TX_RING_SIZE;
  64 int ipoib_recvq_size __read_mostly = IPOIB_RX_RING_SIZE;
  65
  66 module_param_named(send_queue_size, ipoib_sendq_size, int, 0444);
  67 MODULE_PARM_DESC(send_queue_size, "Number of descriptors in send queue");
  68 module_param_named(recv_queue_size, ipoib_recvq_size, int, 0444);
  69 MODULE_PARM_DESC(recv_queue_size, "Number of descriptors in receive queue");
  70
  71 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
  72 int ipoib_debug_level;
  73
  74 module_param_named(debug_level, ipoib_debug_level, int, 0644);
  75 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
  76 #endif
  77
  78 struct ipoib_path_iter {
  79         struct net_device *dev;
  80         struct ipoib_path  path;
  81 };
  82
  83 static const u8 ipv4_bcast_addr[] = {
  84         0x00, 0xff, 0xff, 0xff,
  85         0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
  86         0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff
  87 };
  88
  89 struct workqueue_struct *ipoib_workqueue;
  90
  91 struct ib_sa_client ipoib_sa_client;
  92
  93 static void ipoib_add_one(struct ib_device *device);
  94 static void ipoib_remove_one(struct ib_device *device, void *client_data);
  95 static void ipoib_neigh_reclaim(struct rcu_head *rp);
  96 static struct net_device *ipoib_get_net_dev_by_params(
  97                 struct ib_device *dev, u8 port, u16 pkey,
  98                 const union ib_gid *gid, const struct sockaddr *addr,
  99                 void *client_data);
 100 static int ipoib_set_mac(struct net_device *dev, void *addr);
 101 static int ipoib_ioctl(struct net_device *dev, struct ifreq *ifr,
 102                        int cmd);
 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 struct ipoib_neigh *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 NULL;
 921         }
 922
 923         /* To avoid race condition, make sure that the
 924          * neigh will be added only once.
 925          */
 926         if (unlikely(!list_empty(&neigh->list))) {
 927                 spin_unlock_irqrestore(&priv->lock, flags);
 928                 return neigh;
 929         }
 930
 931         path = __path_find(dev, daddr + 4);
 932         if (!path) {
 933                 path = path_rec_create(dev, daddr + 4);
 934                 if (!path)
 935                         goto err_path;
 936
 937                 __path_add(dev, path);
 938         }
 939
 940         list_add_tail(&neigh->list, &path->neigh_list);
 941
 942         if (path->ah) {
 943                 kref_get(&path->ah->ref);
 944                 neigh->ah = path->ah;
 945
 946                 if (ipoib_cm_enabled(dev, neigh->daddr)) {
 947                         if (!ipoib_cm_get(neigh))
 948                                 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
 949                         if (!ipoib_cm_get(neigh)) {
 950                                 ipoib_neigh_free(neigh);
 951                                 goto err_drop;
 952                         }
 953                         if (skb_queue_len(&neigh->queue) <
 954                             IPOIB_MAX_PATH_REC_QUEUE) {
 955                                 push_pseudo_header(skb, neigh->daddr);
 956                                 __skb_queue_tail(&neigh->queue, skb);
 957                         } else {
 958                                 ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
 959                                            skb_queue_len(&neigh->queue));
 960                                 goto err_drop;
 961                         }
 962                 } else {
 963                         spin_unlock_irqrestore(&priv->lock, flags);
 964                         path->ah->last_send = rn->send(dev, skb, path->ah->ah,
 965                                                        IPOIB_QPN(daddr));
 966                         ipoib_neigh_put(neigh);
 967                         return NULL;
 968                 }
 969         } else {
 970                 neigh->ah  = NULL;
 971
 972                 if (!path->query && path_rec_start(dev, path))
 973                         goto err_path;
 974                 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 975                         push_pseudo_header(skb, neigh->daddr);
 976                         __skb_queue_tail(&neigh->queue, skb);
 977                 } else {
 978                         goto err_drop;
 979                 }
 980         }
 981
 982         spin_unlock_irqrestore(&priv->lock, flags);
 983         ipoib_neigh_put(neigh);
 984         return NULL;
 985
 986 err_path:
 987         ipoib_neigh_free(neigh);
 988 err_drop:
 989         ++dev->stats.tx_dropped;
 990         dev_kfree_skb_any(skb);
 991
 992         spin_unlock_irqrestore(&priv->lock, flags);
 993         ipoib_neigh_put(neigh);
 994
 995         return NULL;
 996 }
 997
 998 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
 999                              struct ipoib_pseudo_header *phdr)
1000 {
1001         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1002         struct rdma_netdev *rn = netdev_priv(dev);
1003         struct ipoib_path *path;
1004         unsigned long flags;
1005
1006         spin_lock_irqsave(&priv->lock, flags);
1007
1008         path = __path_find(dev, phdr->hwaddr + 4);
1009         if (!path || !path->valid) {
1010                 int new_path = 0;
1011
1012                 if (!path) {
1013                         path = path_rec_create(dev, phdr->hwaddr + 4);
1014                         new_path = 1;
1015                 }
1016                 if (path) {
1017                         if (skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
1018                                 push_pseudo_header(skb, phdr->hwaddr);
1019                                 __skb_queue_tail(&path->queue, skb);
1020                         } else {
1021                                 ++dev->stats.tx_dropped;
1022                                 dev_kfree_skb_any(skb);
1023                         }
1024
1025                         if (!path->query && path_rec_start(dev, path)) {
1026                                 spin_unlock_irqrestore(&priv->lock, flags);
1027                                 if (new_path)
1028                                         path_free(dev, path);
1029                                 return;
1030                         } else
1031                                 __path_add(dev, path);
1032                 } else {
1033                         ++dev->stats.tx_dropped;
1034                         dev_kfree_skb_any(skb);
1035                 }
1036
1037                 spin_unlock_irqrestore(&priv->lock, flags);
1038                 return;
1039         }
1040
1041         if (path->ah) {
1042                 ipoib_dbg(priv, "Send unicast ARP to %08x\n",
1043                           be32_to_cpu(sa_path_get_dlid(&path->pathrec)));
1044
1045                 spin_unlock_irqrestore(&priv->lock, flags);
1046                 path->ah->last_send = rn->send(dev, skb, path->ah->ah,
1047                                                IPOIB_QPN(phdr->hwaddr));
1048                 return;
1049         } else if ((path->query || !path_rec_start(dev, path)) &&
1050                    skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
1051                 push_pseudo_header(skb, phdr->hwaddr);
1052                 __skb_queue_tail(&path->queue, skb);
1053         } else {
1054                 ++dev->stats.tx_dropped;
1055                 dev_kfree_skb_any(skb);
1056         }
1057
1058         spin_unlock_irqrestore(&priv->lock, flags);
1059 }
1060
1061 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
1062 {
1063         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1064         struct rdma_netdev *rn = netdev_priv(dev);
1065         struct ipoib_neigh *neigh;
1066         struct ipoib_pseudo_header *phdr;
1067         struct ipoib_header *header;
1068         unsigned long flags;
1069
1070         phdr = (struct ipoib_pseudo_header *) skb->data;
1071         skb_pull(skb, sizeof(*phdr));
1072         header = (struct ipoib_header *) skb->data;
1073
1074         if (unlikely(phdr->hwaddr[4] == 0xff)) {
1075                 /* multicast, arrange "if" according to probability */
1076                 if ((header->proto != htons(ETH_P_IP)) &&
1077                     (header->proto != htons(ETH_P_IPV6)) &&
1078                     (header->proto != htons(ETH_P_ARP)) &&
1079                     (header->proto != htons(ETH_P_RARP)) &&
1080                     (header->proto != htons(ETH_P_TIPC))) {
1081                         /* ethertype not supported by IPoIB */
1082                         ++dev->stats.tx_dropped;
1083                         dev_kfree_skb_any(skb);
1084                         return NETDEV_TX_OK;
1085                 }
1086                 /* Add in the P_Key for multicast*/
1087                 phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
1088                 phdr->hwaddr[9] = priv->pkey & 0xff;
1089
1090                 neigh = ipoib_neigh_get(dev, phdr->hwaddr);
1091                 if (likely(neigh))
1092                         goto send_using_neigh;
1093                 ipoib_mcast_send(dev, phdr->hwaddr, skb);
1094                 return NETDEV_TX_OK;
1095         }
1096
1097         /* unicast, arrange "switch" according to probability */
1098         switch (header->proto) {
1099         case htons(ETH_P_IP):
1100         case htons(ETH_P_IPV6):
1101         case htons(ETH_P_TIPC):
1102                 neigh = ipoib_neigh_get(dev, phdr->hwaddr);
1103                 if (unlikely(!neigh)) {
1104                         neigh = neigh_add_path(skb, phdr->hwaddr, dev);
1105                         if (likely(!neigh))
1106                                 return NETDEV_TX_OK;
1107                 }
1108                 break;
1109         case htons(ETH_P_ARP):
1110         case htons(ETH_P_RARP):
1111                 /* for unicast ARP and RARP should always perform path find */
1112                 unicast_arp_send(skb, dev, phdr);
1113                 return NETDEV_TX_OK;
1114         default:
1115                 /* ethertype not supported by IPoIB */
1116                 ++dev->stats.tx_dropped;
1117                 dev_kfree_skb_any(skb);
1118                 return NETDEV_TX_OK;
1119         }
1120
1121 send_using_neigh:
1122         /* note we now hold a ref to neigh */
1123         if (ipoib_cm_get(neigh)) {
1124                 if (ipoib_cm_up(neigh)) {
1125                         ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
1126                         goto unref;
1127                 }
1128         } else if (neigh->ah) {
1129                 neigh->ah->last_send = rn->send(dev, skb, neigh->ah->ah,
1130                                                 IPOIB_QPN(phdr->hwaddr));
1131                 goto unref;
1132         }
1133
1134         if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
1135                 push_pseudo_header(skb, phdr->hwaddr);
1136                 spin_lock_irqsave(&priv->lock, flags);
1137                 __skb_queue_tail(&neigh->queue, skb);
1138                 spin_unlock_irqrestore(&priv->lock, flags);
1139         } else {
1140                 ++dev->stats.tx_dropped;
1141                 dev_kfree_skb_any(skb);
1142         }
1143
1144 unref:
1145         ipoib_neigh_put(neigh);
1146
1147         return NETDEV_TX_OK;
1148 }
1149
1150 static void ipoib_timeout(struct net_device *dev)
1151 {
1152         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1153
1154         ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
1155                    jiffies_to_msecs(jiffies - dev_trans_start(dev)));
1156         ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
1157                    netif_queue_stopped(dev),
1158                    priv->tx_head, priv->tx_tail);
1159         /* XXX reset QP, etc. */
1160 }
1161
1162 static int ipoib_hard_header(struct sk_buff *skb,
1163                              struct net_device *dev,
1164                              unsigned short type,
1165                              const void *daddr, const void *saddr, unsigned len)
1166 {
1167         struct ipoib_header *header;
1168
1169         header = skb_push(skb, sizeof *header);
1170
1171         header->proto = htons(type);
1172         header->reserved = 0;
1173
1174         /*
1175          * we don't rely on dst_entry structure,  always stuff the
1176          * destination address into skb hard header so we can figure out where
1177          * to send the packet later.
1178          */
1179         push_pseudo_header(skb, daddr);
1180
1181         return IPOIB_HARD_LEN;
1182 }
1183
1184 static void ipoib_set_mcast_list(struct net_device *dev)
1185 {
1186         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1187
1188         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
1189                 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
1190                 return;
1191         }
1192
1193         queue_work(priv->wq, &priv->restart_task);
1194 }
1195
1196 static int ipoib_get_iflink(const struct net_device *dev)
1197 {
1198         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1199
1200         /* parent interface */
1201         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags))
1202                 return dev->ifindex;
1203
1204         /* child/vlan interface */
1205         return priv->parent->ifindex;
1206 }
1207
1208 static u32 ipoib_addr_hash(struct ipoib_neigh_hash *htbl, u8 *daddr)
1209 {
1210         /*
1211          * Use only the address parts that contributes to spreading
1212          * The subnet prefix is not used as one can not connect to
1213          * same remote port (GUID) using the same remote QPN via two
1214          * different subnets.
1215          */
1216          /* qpn octets[1:4) & port GUID octets[12:20) */
1217         u32 *d32 = (u32 *) daddr;
1218         u32 hv;
1219
1220         hv = jhash_3words(d32[3], d32[4], IPOIB_QPN_MASK & d32[0], 0);
1221         return hv & htbl->mask;
1222 }
1223
1224 struct ipoib_neigh *ipoib_neigh_get(struct net_device *dev, u8 *daddr)
1225 {
1226         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1227         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1228         struct ipoib_neigh_hash *htbl;
1229         struct ipoib_neigh *neigh = NULL;
1230         u32 hash_val;
1231
1232         rcu_read_lock_bh();
1233
1234         htbl = rcu_dereference_bh(ntbl->htbl);
1235
1236         if (!htbl)
1237                 goto out_unlock;
1238
1239         hash_val = ipoib_addr_hash(htbl, daddr);
1240         for (neigh = rcu_dereference_bh(htbl->buckets[hash_val]);
1241              neigh != NULL;
1242              neigh = rcu_dereference_bh(neigh->hnext)) {
1243                 if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) {
1244                         /* found, take one ref on behalf of the caller */
1245                         if (!atomic_inc_not_zero(&neigh->refcnt)) {
1246                                 /* deleted */
1247                                 neigh = NULL;
1248                                 goto out_unlock;
1249                         }
1250
1251                         if (likely(skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE))
1252                                 neigh->alive = jiffies;
1253                         goto out_unlock;
1254                 }
1255         }
1256
1257 out_unlock:
1258         rcu_read_unlock_bh();
1259         return neigh;
1260 }
1261
1262 static void __ipoib_reap_neigh(struct ipoib_dev_priv *priv)
1263 {
1264         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1265         struct ipoib_neigh_hash *htbl;
1266         unsigned long neigh_obsolete;
1267         unsigned long dt;
1268         unsigned long flags;
1269         int i;
1270         LIST_HEAD(remove_list);
1271
1272         if (test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
1273                 return;
1274
1275         spin_lock_irqsave(&priv->lock, flags);
1276
1277         htbl = rcu_dereference_protected(ntbl->htbl,
1278                                          lockdep_is_held(&priv->lock));
1279
1280         if (!htbl)
1281                 goto out_unlock;
1282
1283         /* neigh is obsolete if it was idle for two GC periods */
1284         dt = 2 * arp_tbl.gc_interval;
1285         neigh_obsolete = jiffies - dt;
1286         /* handle possible race condition */
1287         if (test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
1288                 goto out_unlock;
1289
1290         for (i = 0; i < htbl->size; i++) {
1291                 struct ipoib_neigh *neigh;
1292                 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1293
1294                 while ((neigh = rcu_dereference_protected(*np,
1295                                                           lockdep_is_held(&priv->lock))) != NULL) {
1296                         /* was the neigh idle for two GC periods */
1297                         if (time_after(neigh_obsolete, neigh->alive)) {
1298
1299                                 ipoib_check_and_add_mcast_sendonly(priv, neigh->daddr + 4, &remove_list);
1300
1301                                 rcu_assign_pointer(*np,
1302                                                    rcu_dereference_protected(neigh->hnext,
1303                                                                              lockdep_is_held(&priv->lock)));
1304                                 /* remove from path/mc list */
1305                                 list_del_init(&neigh->list);
1306                                 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1307                         } else {
1308                                 np = &neigh->hnext;
1309                         }
1310
1311                 }
1312         }
1313
1314 out_unlock:
1315         spin_unlock_irqrestore(&priv->lock, flags);
1316         ipoib_mcast_remove_list(&remove_list);
1317 }
1318
1319 static void ipoib_reap_neigh(struct work_struct *work)
1320 {
1321         struct ipoib_dev_priv *priv =
1322                 container_of(work, struct ipoib_dev_priv, neigh_reap_task.work);
1323
1324         __ipoib_reap_neigh(priv);
1325
1326         if (!test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
1327                 queue_delayed_work(priv->wq, &priv->neigh_reap_task,
1328                                    arp_tbl.gc_interval);
1329 }
1330
1331
1332 static struct ipoib_neigh *ipoib_neigh_ctor(u8 *daddr,
1333                                       struct net_device *dev)
1334 {
1335         struct ipoib_neigh *neigh;
1336
1337         neigh = kzalloc(sizeof *neigh, GFP_ATOMIC);
1338         if (!neigh)
1339                 return NULL;
1340
1341         neigh->dev = dev;
1342         memcpy(&neigh->daddr, daddr, sizeof(neigh->daddr));
1343         skb_queue_head_init(&neigh->queue);
1344         INIT_LIST_HEAD(&neigh->list);
1345         ipoib_cm_set(neigh, NULL);
1346         /* one ref on behalf of the caller */
1347         atomic_set(&neigh->refcnt, 1);
1348
1349         return neigh;
1350 }
1351
1352 struct ipoib_neigh *ipoib_neigh_alloc(u8 *daddr,
1353                                       struct net_device *dev)
1354 {
1355         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1356         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1357         struct ipoib_neigh_hash *htbl;
1358         struct ipoib_neigh *neigh;
1359         u32 hash_val;
1360
1361         htbl = rcu_dereference_protected(ntbl->htbl,
1362                                          lockdep_is_held(&priv->lock));
1363         if (!htbl) {
1364                 neigh = NULL;
1365                 goto out_unlock;
1366         }
1367
1368         /* need to add a new neigh, but maybe some other thread succeeded?
1369          * recalc hash, maybe hash resize took place so we do a search
1370          */
1371         hash_val = ipoib_addr_hash(htbl, daddr);
1372         for (neigh = rcu_dereference_protected(htbl->buckets[hash_val],
1373                                                lockdep_is_held(&priv->lock));
1374              neigh != NULL;
1375              neigh = rcu_dereference_protected(neigh->hnext,
1376                                                lockdep_is_held(&priv->lock))) {
1377                 if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) {
1378                         /* found, take one ref on behalf of the caller */
1379                         if (!atomic_inc_not_zero(&neigh->refcnt)) {
1380                                 /* deleted */
1381                                 neigh = NULL;
1382                                 break;
1383                         }
1384                         neigh->alive = jiffies;
1385                         goto out_unlock;
1386                 }
1387         }
1388
1389         neigh = ipoib_neigh_ctor(daddr, dev);
1390         if (!neigh)
1391                 goto out_unlock;
1392
1393         /* one ref on behalf of the hash table */
1394         atomic_inc(&neigh->refcnt);
1395         neigh->alive = jiffies;
1396         /* put in hash */
1397         rcu_assign_pointer(neigh->hnext,
1398                            rcu_dereference_protected(htbl->buckets[hash_val],
1399                                                      lockdep_is_held(&priv->lock)));
1400         rcu_assign_pointer(htbl->buckets[hash_val], neigh);
1401         atomic_inc(&ntbl->entries);
1402
1403 out_unlock:
1404
1405         return neigh;
1406 }
1407
1408 void ipoib_neigh_dtor(struct ipoib_neigh *neigh)
1409 {
1410         /* neigh reference count was dropprd to zero */
1411         struct net_device *dev = neigh->dev;
1412         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1413         struct sk_buff *skb;
1414         if (neigh->ah)
1415                 ipoib_put_ah(neigh->ah);
1416         while ((skb = __skb_dequeue(&neigh->queue))) {
1417                 ++dev->stats.tx_dropped;
1418                 dev_kfree_skb_any(skb);
1419         }
1420         if (ipoib_cm_get(neigh))
1421                 ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
1422         ipoib_dbg(ipoib_priv(dev),
1423                   "neigh free for %06x %pI6\n",
1424                   IPOIB_QPN(neigh->daddr),
1425                   neigh->daddr + 4);
1426         kfree(neigh);
1427         if (atomic_dec_and_test(&priv->ntbl.entries)) {
1428                 if (test_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags))
1429                         complete(&priv->ntbl.flushed);
1430         }
1431 }
1432
1433 static void ipoib_neigh_reclaim(struct rcu_head *rp)
1434 {
1435         /* Called as a result of removal from hash table */
1436         struct ipoib_neigh *neigh = container_of(rp, struct ipoib_neigh, rcu);
1437         /* note TX context may hold another ref */
1438         ipoib_neigh_put(neigh);
1439 }
1440
1441 void ipoib_neigh_free(struct ipoib_neigh *neigh)
1442 {
1443         struct net_device *dev = neigh->dev;
1444         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1445         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1446         struct ipoib_neigh_hash *htbl;
1447         struct ipoib_neigh __rcu **np;
1448         struct ipoib_neigh *n;
1449         u32 hash_val;
1450
1451         htbl = rcu_dereference_protected(ntbl->htbl,
1452                                         lockdep_is_held(&priv->lock));
1453         if (!htbl)
1454                 return;
1455
1456         hash_val = ipoib_addr_hash(htbl, neigh->daddr);
1457         np = &htbl->buckets[hash_val];
1458         for (n = rcu_dereference_protected(*np,
1459                                             lockdep_is_held(&priv->lock));
1460              n != NULL;
1461              n = rcu_dereference_protected(*np,
1462                                         lockdep_is_held(&priv->lock))) {
1463                 if (n == neigh) {
1464                         /* found */
1465                         rcu_assign_pointer(*np,
1466                                            rcu_dereference_protected(neigh->hnext,
1467                                                                      lockdep_is_held(&priv->lock)));
1468                         /* remove from parent list */
1469                         list_del_init(&neigh->list);
1470                         call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1471                         return;
1472                 } else {
1473                         np = &n->hnext;
1474                 }
1475         }
1476 }
1477
1478 static int ipoib_neigh_hash_init(struct ipoib_dev_priv *priv)
1479 {
1480         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1481         struct ipoib_neigh_hash *htbl;
1482         struct ipoib_neigh __rcu **buckets;
1483         u32 size;
1484
1485         clear_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags);
1486         ntbl->htbl = NULL;
1487         htbl = kzalloc(sizeof(*htbl), GFP_KERNEL);
1488         if (!htbl)
1489                 return -ENOMEM;
1490         set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1491         size = roundup_pow_of_two(arp_tbl.gc_thresh3);
1492         buckets = kzalloc(size * sizeof(*buckets), GFP_KERNEL);
1493         if (!buckets) {
1494                 kfree(htbl);
1495                 return -ENOMEM;
1496         }
1497         htbl->size = size;
1498         htbl->mask = (size - 1);
1499         htbl->buckets = buckets;
1500         RCU_INIT_POINTER(ntbl->htbl, htbl);
1501         htbl->ntbl = ntbl;
1502         atomic_set(&ntbl->entries, 0);
1503
1504         /* start garbage collection */
1505         clear_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1506         queue_delayed_work(priv->wq, &priv->neigh_reap_task,
1507                            arp_tbl.gc_interval);
1508
1509         return 0;
1510 }
1511
1512 static void neigh_hash_free_rcu(struct rcu_head *head)
1513 {
1514         struct ipoib_neigh_hash *htbl = container_of(head,
1515                                                     struct ipoib_neigh_hash,
1516                                                     rcu);
1517         struct ipoib_neigh __rcu **buckets = htbl->buckets;
1518         struct ipoib_neigh_table *ntbl = htbl->ntbl;
1519
1520         kfree(buckets);
1521         kfree(htbl);
1522         complete(&ntbl->deleted);
1523 }
1524
1525 void ipoib_del_neighs_by_gid(struct net_device *dev, u8 *gid)
1526 {
1527         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1528         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1529         struct ipoib_neigh_hash *htbl;
1530         unsigned long flags;
1531         int i;
1532
1533         /* remove all neigh connected to a given path or mcast */
1534         spin_lock_irqsave(&priv->lock, flags);
1535
1536         htbl = rcu_dereference_protected(ntbl->htbl,
1537                                          lockdep_is_held(&priv->lock));
1538
1539         if (!htbl)
1540                 goto out_unlock;
1541
1542         for (i = 0; i < htbl->size; i++) {
1543                 struct ipoib_neigh *neigh;
1544                 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1545
1546                 while ((neigh = rcu_dereference_protected(*np,
1547                                                           lockdep_is_held(&priv->lock))) != NULL) {
1548                         /* delete neighs belong to this parent */
1549                         if (!memcmp(gid, neigh->daddr + 4, sizeof (union ib_gid))) {
1550                                 rcu_assign_pointer(*np,
1551                                                    rcu_dereference_protected(neigh->hnext,
1552                                                                              lockdep_is_held(&priv->lock)));
1553                                 /* remove from parent list */
1554                                 list_del_init(&neigh->list);
1555                                 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1556                         } else {
1557                                 np = &neigh->hnext;
1558                         }
1559
1560                 }
1561         }
1562 out_unlock:
1563         spin_unlock_irqrestore(&priv->lock, flags);
1564 }
1565
1566 static void ipoib_flush_neighs(struct ipoib_dev_priv *priv)
1567 {
1568         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1569         struct ipoib_neigh_hash *htbl;
1570         unsigned long flags;
1571         int i, wait_flushed = 0;
1572
1573         init_completion(&priv->ntbl.flushed);
1574         set_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags);
1575
1576         spin_lock_irqsave(&priv->lock, flags);
1577
1578         htbl = rcu_dereference_protected(ntbl->htbl,
1579                                         lockdep_is_held(&priv->lock));
1580         if (!htbl)
1581                 goto out_unlock;
1582
1583         wait_flushed = atomic_read(&priv->ntbl.entries);
1584         if (!wait_flushed)
1585                 goto free_htbl;
1586
1587         for (i = 0; i < htbl->size; i++) {
1588                 struct ipoib_neigh *neigh;
1589                 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1590
1591                 while ((neigh = rcu_dereference_protected(*np,
1592                                        lockdep_is_held(&priv->lock))) != NULL) {
1593                         rcu_assign_pointer(*np,
1594                                            rcu_dereference_protected(neigh->hnext,
1595                                                                      lockdep_is_held(&priv->lock)));
1596                         /* remove from path/mc list */
1597                         list_del_init(&neigh->list);
1598                         call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1599                 }
1600         }
1601
1602 free_htbl:
1603         rcu_assign_pointer(ntbl->htbl, NULL);
1604         call_rcu(&htbl->rcu, neigh_hash_free_rcu);
1605
1606 out_unlock:
1607         spin_unlock_irqrestore(&priv->lock, flags);
1608         if (wait_flushed)
1609                 wait_for_completion(&priv->ntbl.flushed);
1610 }
1611
1612 static void ipoib_neigh_hash_uninit(struct net_device *dev)
1613 {
1614         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1615         int stopped;
1616
1617         ipoib_dbg(priv, "ipoib_neigh_hash_uninit\n");
1618         init_completion(&priv->ntbl.deleted);
1619
1620         /* Stop GC if called at init fail need to cancel work */
1621         stopped = test_and_set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1622         if (!stopped)
1623                 cancel_delayed_work(&priv->neigh_reap_task);
1624
1625         ipoib_flush_neighs(priv);
1626
1627         wait_for_completion(&priv->ntbl.deleted);
1628 }
1629
1630 static void ipoib_napi_add(struct net_device *dev)
1631 {
1632         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1633
1634         netif_napi_add(dev, &priv->recv_napi, ipoib_rx_poll, IPOIB_NUM_WC);
1635         netif_napi_add(dev, &priv->send_napi, ipoib_tx_poll, MAX_SEND_CQE);
1636 }
1637
1638 static void ipoib_napi_del(struct net_device *dev)
1639 {
1640         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1641
1642         netif_napi_del(&priv->recv_napi);
1643         netif_napi_del(&priv->send_napi);
1644 }
1645
1646 static void ipoib_dev_uninit_default(struct net_device *dev)
1647 {
1648         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1649
1650         ipoib_transport_dev_cleanup(dev);
1651
1652         ipoib_napi_del(dev);
1653
1654         ipoib_cm_dev_cleanup(dev);
1655
1656         kfree(priv->rx_ring);
1657         vfree(priv->tx_ring);
1658
1659         priv->rx_ring = NULL;
1660         priv->tx_ring = NULL;
1661 }
1662
1663 static int ipoib_dev_init_default(struct net_device *dev)
1664 {
1665         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1666
1667         ipoib_napi_add(dev);
1668
1669         /* Allocate RX/TX "rings" to hold queued skbs */
1670         priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
1671                                 GFP_KERNEL);
1672         if (!priv->rx_ring)
1673                 goto out;
1674
1675         priv->tx_ring = vzalloc(ipoib_sendq_size * sizeof *priv->tx_ring);
1676         if (!priv->tx_ring) {
1677                 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
1678                        priv->ca->name, ipoib_sendq_size);
1679                 goto out_rx_ring_cleanup;
1680         }
1681
1682         /* priv->tx_head, tx_tail & tx_outstanding are already 0 */
1683
1684         if (ipoib_transport_dev_init(dev, priv->ca)) {
1685                 pr_warn("%s: ipoib_transport_dev_init failed\n",
1686                         priv->ca->name);
1687                 goto out_tx_ring_cleanup;
1688         }
1689
1690         /* after qp created set dev address */
1691         priv->dev->dev_addr[1] = (priv->qp->qp_num >> 16) & 0xff;
1692         priv->dev->dev_addr[2] = (priv->qp->qp_num >>  8) & 0xff;
1693         priv->dev->dev_addr[3] = (priv->qp->qp_num) & 0xff;
1694
1695         return 0;
1696
1697 out_tx_ring_cleanup:
1698         vfree(priv->tx_ring);
1699
1700 out_rx_ring_cleanup:
1701         kfree(priv->rx_ring);
1702
1703 out:
1704         ipoib_napi_del(dev);
1705         return -ENOMEM;
1706 }
1707
1708 static int ipoib_ioctl(struct net_device *dev, struct ifreq *ifr,
1709                        int cmd)
1710 {
1711         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1712
1713         if (!priv->rn_ops->ndo_do_ioctl)
1714                 return -EOPNOTSUPP;
1715
1716         return priv->rn_ops->ndo_do_ioctl(dev, ifr, cmd);
1717 }
1718
1719 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
1720 {
1721         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1722         int ret = -ENOMEM;
1723
1724         priv->ca = ca;
1725         priv->port = port;
1726         priv->qp = NULL;
1727
1728         /*
1729          * the various IPoIB tasks assume they will never race against
1730          * themselves, so always use a single thread workqueue
1731          */
1732         priv->wq = alloc_ordered_workqueue("ipoib_wq", WQ_MEM_RECLAIM);
1733         if (!priv->wq) {
1734                 pr_warn("%s: failed to allocate device WQ\n", dev->name);
1735                 goto out;
1736         }
1737
1738         /* create pd, which used both for control and datapath*/
1739         priv->pd = ib_alloc_pd(priv->ca, 0);
1740         if (IS_ERR(priv->pd)) {
1741                 pr_warn("%s: failed to allocate PD\n", ca->name);
1742                 goto clean_wq;
1743         }
1744
1745         ret = priv->rn_ops->ndo_init(dev);
1746         if (ret) {
1747                 pr_warn("%s failed to init HW resource\n", dev->name);
1748                 goto out_free_pd;
1749         }
1750
1751         if (ipoib_neigh_hash_init(priv) < 0) {
1752                 pr_warn("%s failed to init neigh hash\n", dev->name);
1753                 goto out_dev_uninit;
1754         }
1755
1756         if (dev->flags & IFF_UP) {
1757                 if (ipoib_ib_dev_open(dev)) {
1758                         pr_warn("%s failed to open device\n", dev->name);
1759                         ret = -ENODEV;
1760                         goto out_dev_uninit;
1761                 }
1762         }
1763
1764         return 0;
1765
1766 out_dev_uninit:
1767         ipoib_ib_dev_cleanup(dev);
1768
1769 out_free_pd:
1770         if (priv->pd) {
1771                 ib_dealloc_pd(priv->pd);
1772                 priv->pd = NULL;
1773         }
1774
1775 clean_wq:
1776         if (priv->wq) {
1777                 destroy_workqueue(priv->wq);
1778                 priv->wq = NULL;
1779         }
1780
1781 out:
1782         return ret;
1783 }
1784
1785 void ipoib_dev_cleanup(struct net_device *dev)
1786 {
1787         struct ipoib_dev_priv *priv = ipoib_priv(dev), *cpriv, *tcpriv;
1788         LIST_HEAD(head);
1789
1790         ASSERT_RTNL();
1791
1792         /* Delete any child interfaces first */
1793         list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
1794                 /* Stop GC on child */
1795                 set_bit(IPOIB_STOP_NEIGH_GC, &cpriv->flags);
1796                 cancel_delayed_work(&cpriv->neigh_reap_task);
1797                 unregister_netdevice_queue(cpriv->dev, &head);
1798         }
1799         unregister_netdevice_many(&head);
1800
1801         ipoib_neigh_hash_uninit(dev);
1802
1803         ipoib_ib_dev_cleanup(dev);
1804
1805         /* no more works over the priv->wq */
1806         if (priv->wq) {
1807                 flush_workqueue(priv->wq);
1808                 destroy_workqueue(priv->wq);
1809                 priv->wq = NULL;
1810         }
1811 }
1812
1813 static int ipoib_set_vf_link_state(struct net_device *dev, int vf, int link_state)
1814 {
1815         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1816
1817         return ib_set_vf_link_state(priv->ca, vf, priv->port, link_state);
1818 }
1819
1820 static int ipoib_get_vf_config(struct net_device *dev, int vf,
1821                                struct ifla_vf_info *ivf)
1822 {
1823         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1824         int err;
1825
1826         err = ib_get_vf_config(priv->ca, vf, priv->port, ivf);
1827         if (err)
1828                 return err;
1829
1830         ivf->vf = vf;
1831
1832         return 0;
1833 }
1834
1835 static int ipoib_set_vf_guid(struct net_device *dev, int vf, u64 guid, int type)
1836 {
1837         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1838
1839         if (type != IFLA_VF_IB_NODE_GUID && type != IFLA_VF_IB_PORT_GUID)
1840                 return -EINVAL;
1841
1842         return ib_set_vf_guid(priv->ca, vf, priv->port, guid, type);
1843 }
1844
1845 static int ipoib_get_vf_stats(struct net_device *dev, int vf,
1846                               struct ifla_vf_stats *vf_stats)
1847 {
1848         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1849
1850         return ib_get_vf_stats(priv->ca, vf, priv->port, vf_stats);
1851 }
1852
1853 static const struct header_ops ipoib_header_ops = {
1854         .create = ipoib_hard_header,
1855 };
1856
1857 static const struct net_device_ops ipoib_netdev_ops_pf = {
1858         .ndo_uninit              = ipoib_uninit,
1859         .ndo_open                = ipoib_open,
1860         .ndo_stop                = ipoib_stop,
1861         .ndo_change_mtu          = ipoib_change_mtu,
1862         .ndo_fix_features        = ipoib_fix_features,
1863         .ndo_start_xmit          = ipoib_start_xmit,
1864         .ndo_tx_timeout          = ipoib_timeout,
1865         .ndo_set_rx_mode         = ipoib_set_mcast_list,
1866         .ndo_get_iflink          = ipoib_get_iflink,
1867         .ndo_set_vf_link_state   = ipoib_set_vf_link_state,
1868         .ndo_get_vf_config       = ipoib_get_vf_config,
1869         .ndo_get_vf_stats        = ipoib_get_vf_stats,
1870         .ndo_set_vf_guid         = ipoib_set_vf_guid,
1871         .ndo_set_mac_address     = ipoib_set_mac,
1872         .ndo_get_stats64         = ipoib_get_stats,
1873         .ndo_do_ioctl            = ipoib_ioctl,
1874 };
1875
1876 static const struct net_device_ops ipoib_netdev_ops_vf = {
1877         .ndo_uninit              = ipoib_uninit,
1878         .ndo_open                = ipoib_open,
1879         .ndo_stop                = ipoib_stop,
1880         .ndo_change_mtu          = ipoib_change_mtu,
1881         .ndo_fix_features        = ipoib_fix_features,
1882         .ndo_start_xmit          = ipoib_start_xmit,
1883         .ndo_tx_timeout          = ipoib_timeout,
1884         .ndo_set_rx_mode         = ipoib_set_mcast_list,
1885         .ndo_get_iflink          = ipoib_get_iflink,
1886         .ndo_get_stats64         = ipoib_get_stats,
1887         .ndo_do_ioctl            = ipoib_ioctl,
1888 };
1889
1890 void ipoib_setup_common(struct net_device *dev)
1891 {
1892         dev->header_ops          = &ipoib_header_ops;
1893
1894         ipoib_set_ethtool_ops(dev);
1895
1896         dev->watchdog_timeo      = HZ;
1897
1898         dev->flags              |= IFF_BROADCAST | IFF_MULTICAST;
1899
1900         dev->hard_header_len     = IPOIB_HARD_LEN;
1901         dev->addr_len            = INFINIBAND_ALEN;
1902         dev->type                = ARPHRD_INFINIBAND;
1903         dev->tx_queue_len        = ipoib_sendq_size * 2;
1904         dev->features            = (NETIF_F_VLAN_CHALLENGED     |
1905                                     NETIF_F_HIGHDMA);
1906         netif_keep_dst(dev);
1907
1908         memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
1909 }
1910
1911 static void ipoib_build_priv(struct net_device *dev)
1912 {
1913         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1914
1915         priv->dev = dev;
1916         spin_lock_init(&priv->lock);
1917         init_rwsem(&priv->vlan_rwsem);
1918         mutex_init(&priv->mcast_mutex);
1919         mutex_init(&priv->sysfs_mutex);
1920
1921         INIT_LIST_HEAD(&priv->path_list);
1922         INIT_LIST_HEAD(&priv->child_intfs);
1923         INIT_LIST_HEAD(&priv->dead_ahs);
1924         INIT_LIST_HEAD(&priv->multicast_list);
1925
1926         INIT_DELAYED_WORK(&priv->mcast_task,   ipoib_mcast_join_task);
1927         INIT_WORK(&priv->carrier_on_task, ipoib_mcast_carrier_on_task);
1928         INIT_WORK(&priv->flush_light,   ipoib_ib_dev_flush_light);
1929         INIT_WORK(&priv->flush_normal,   ipoib_ib_dev_flush_normal);
1930         INIT_WORK(&priv->flush_heavy,   ipoib_ib_dev_flush_heavy);
1931         INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
1932         INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
1933         INIT_DELAYED_WORK(&priv->neigh_reap_task, ipoib_reap_neigh);
1934 }
1935
1936 static const struct net_device_ops ipoib_netdev_default_pf = {
1937         .ndo_init                = ipoib_dev_init_default,
1938         .ndo_uninit              = ipoib_dev_uninit_default,
1939         .ndo_open                = ipoib_ib_dev_open_default,
1940         .ndo_stop                = ipoib_ib_dev_stop_default,
1941 };
1942
1943 static struct net_device
1944 *ipoib_create_netdev_default(struct ib_device *hca,
1945                              const char *name,
1946                              unsigned char name_assign_type,
1947                              void (*setup)(struct net_device *))
1948 {
1949         struct net_device *dev;
1950         struct rdma_netdev *rn;
1951
1952         dev = alloc_netdev((int)sizeof(struct rdma_netdev),
1953                            name,
1954                            name_assign_type, setup);
1955         if (!dev)
1956                 return NULL;
1957
1958         rn = netdev_priv(dev);
1959
1960         rn->send = ipoib_send;
1961         rn->attach_mcast = ipoib_mcast_attach;
1962         rn->detach_mcast = ipoib_mcast_detach;
1963         rn->free_rdma_netdev = free_netdev;
1964         rn->hca = hca;
1965
1966         dev->netdev_ops = &ipoib_netdev_default_pf;
1967
1968         return dev;
1969 }
1970
1971 static struct net_device *ipoib_get_netdev(struct ib_device *hca, u8 port,
1972                                            const char *name)
1973 {
1974         struct net_device *dev;
1975
1976         if (hca->alloc_rdma_netdev) {
1977                 dev = hca->alloc_rdma_netdev(hca, port,
1978                                              RDMA_NETDEV_IPOIB, name,
1979                                              NET_NAME_UNKNOWN,
1980                                              ipoib_setup_common);
1981                 if (IS_ERR_OR_NULL(dev) && PTR_ERR(dev) != -EOPNOTSUPP)
1982                         return NULL;
1983         }
1984
1985         if (!hca->alloc_rdma_netdev || PTR_ERR(dev) == -EOPNOTSUPP)
1986                 dev = ipoib_create_netdev_default(hca, name, NET_NAME_UNKNOWN,
1987                                                   ipoib_setup_common);
1988
1989         return dev;
1990 }
1991
1992 struct ipoib_dev_priv *ipoib_intf_alloc(struct ib_device *hca, u8 port,
1993                                         const char *name)
1994 {
1995         struct net_device *dev;
1996         struct ipoib_dev_priv *priv;
1997         struct rdma_netdev *rn;
1998
1999         priv = kzalloc(sizeof(*priv), GFP_KERNEL);
2000         if (!priv)
2001                 return NULL;
2002
2003         dev = ipoib_get_netdev(hca, port, name);
2004         if (!dev)
2005                 goto free_priv;
2006
2007         priv->rn_ops = dev->netdev_ops;
2008
2009         /* fixme : should be after the query_cap */
2010         if (priv->hca_caps & IB_DEVICE_VIRTUAL_FUNCTION)
2011                 dev->netdev_ops = &ipoib_netdev_ops_vf;
2012         else
2013                 dev->netdev_ops = &ipoib_netdev_ops_pf;
2014
2015         rn = netdev_priv(dev);
2016         rn->clnt_priv = priv;
2017         ipoib_build_priv(dev);
2018
2019         return priv;
2020 free_priv:
2021         kfree(priv);
2022         return NULL;
2023 }
2024
2025 static ssize_t show_pkey(struct device *dev,
2026                          struct device_attribute *attr, char *buf)
2027 {
2028         struct net_device *ndev = to_net_dev(dev);
2029         struct ipoib_dev_priv *priv = ipoib_priv(ndev);
2030
2031         return sprintf(buf, "0x%04x\n", priv->pkey);
2032 }
2033 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
2034
2035 static ssize_t show_umcast(struct device *dev,
2036                            struct device_attribute *attr, char *buf)
2037 {
2038         struct net_device *ndev = to_net_dev(dev);
2039         struct ipoib_dev_priv *priv = ipoib_priv(ndev);
2040
2041         return sprintf(buf, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
2042 }
2043
2044 void ipoib_set_umcast(struct net_device *ndev, int umcast_val)
2045 {
2046         struct ipoib_dev_priv *priv = ipoib_priv(ndev);
2047
2048         if (umcast_val > 0) {
2049                 set_bit(IPOIB_FLAG_UMCAST, &priv->flags);
2050                 ipoib_warn(priv, "ignoring multicast groups joined directly "
2051                                 "by userspace\n");
2052         } else
2053                 clear_bit(IPOIB_FLAG_UMCAST, &priv->flags);
2054 }
2055
2056 static ssize_t set_umcast(struct device *dev,
2057                           struct device_attribute *attr,
2058                           const char *buf, size_t count)
2059 {
2060         unsigned long umcast_val = simple_strtoul(buf, NULL, 0);
2061
2062         ipoib_set_umcast(to_net_dev(dev), umcast_val);
2063
2064         return count;
2065 }
2066 static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast);
2067
2068 int ipoib_add_umcast_attr(struct net_device *dev)
2069 {
2070         return device_create_file(&dev->dev, &dev_attr_umcast);
2071 }
2072
2073 static void set_base_guid(struct ipoib_dev_priv *priv, union ib_gid *gid)
2074 {
2075         struct ipoib_dev_priv *child_priv;
2076         struct net_device *netdev = priv->dev;
2077
2078         netif_addr_lock_bh(netdev);
2079
2080         memcpy(&priv->local_gid.global.interface_id,
2081                &gid->global.interface_id,
2082                sizeof(gid->global.interface_id));
2083         memcpy(netdev->dev_addr + 4, &priv->local_gid, sizeof(priv->local_gid));
2084         clear_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags);
2085
2086         netif_addr_unlock_bh(netdev);
2087
2088         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
2089                 down_read(&priv->vlan_rwsem);
2090                 list_for_each_entry(child_priv, &priv->child_intfs, list)
2091                         set_base_guid(child_priv, gid);
2092                 up_read(&priv->vlan_rwsem);
2093         }
2094 }
2095
2096 static int ipoib_check_lladdr(struct net_device *dev,
2097                               struct sockaddr_storage *ss)
2098 {
2099         union ib_gid *gid = (union ib_gid *)(ss->__data + 4);
2100         int ret = 0;
2101
2102         netif_addr_lock_bh(dev);
2103
2104         /* Make sure the QPN, reserved and subnet prefix match the current
2105          * lladdr, it also makes sure the lladdr is unicast.
2106          */
2107         if (memcmp(dev->dev_addr, ss->__data,
2108                    4 + sizeof(gid->global.subnet_prefix)) ||
2109             gid->global.interface_id == 0)
2110                 ret = -EINVAL;
2111
2112         netif_addr_unlock_bh(dev);
2113
2114         return ret;
2115 }
2116
2117 static int ipoib_set_mac(struct net_device *dev, void *addr)
2118 {
2119         struct ipoib_dev_priv *priv = ipoib_priv(dev);
2120         struct sockaddr_storage *ss = addr;
2121         int ret;
2122
2123         if (!(dev->priv_flags & IFF_LIVE_ADDR_CHANGE) && netif_running(dev))
2124                 return -EBUSY;
2125
2126         ret = ipoib_check_lladdr(dev, ss);
2127         if (ret)
2128                 return ret;
2129
2130         set_base_guid(priv, (union ib_gid *)(ss->__data + 4));
2131
2132         queue_work(ipoib_workqueue, &priv->flush_light);
2133
2134         return 0;
2135 }
2136
2137 static ssize_t create_child(struct device *dev,
2138                             struct device_attribute *attr,
2139                             const char *buf, size_t count)
2140 {
2141         int pkey;
2142         int ret;
2143
2144         if (sscanf(buf, "%i", &pkey) != 1)
2145                 return -EINVAL;
2146
2147         if (pkey <= 0 || pkey > 0xffff || pkey == 0x8000)
2148                 return -EINVAL;
2149
2150         /*
2151          * Set the full membership bit, so that we join the right
2152          * broadcast group, etc.
2153          */
2154         pkey |= 0x8000;
2155
2156         ret = ipoib_vlan_add(to_net_dev(dev), pkey);
2157
2158         return ret ? ret : count;
2159 }
2160 static DEVICE_ATTR(create_child, S_IWUSR, NULL, create_child);
2161
2162 static ssize_t delete_child(struct device *dev,
2163                             struct device_attribute *attr,
2164                             const char *buf, size_t count)
2165 {
2166         int pkey;
2167         int ret;
2168
2169         if (sscanf(buf, "%i", &pkey) != 1)
2170                 return -EINVAL;
2171
2172         if (pkey < 0 || pkey > 0xffff)
2173                 return -EINVAL;
2174
2175         ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
2176
2177         return ret ? ret : count;
2178
2179 }
2180 static DEVICE_ATTR(delete_child, S_IWUSR, NULL, delete_child);
2181
2182 int ipoib_add_pkey_attr(struct net_device *dev)
2183 {
2184         return device_create_file(&dev->dev, &dev_attr_pkey);
2185 }
2186
2187 void ipoib_set_dev_features(struct ipoib_dev_priv *priv, struct ib_device *hca)
2188 {
2189         priv->hca_caps = hca->attrs.device_cap_flags;
2190
2191         if (priv->hca_caps & IB_DEVICE_UD_IP_CSUM) {
2192                 priv->dev->hw_features |= NETIF_F_IP_CSUM | NETIF_F_RXCSUM;
2193
2194                 if (priv->hca_caps & IB_DEVICE_UD_TSO)
2195                         priv->dev->hw_features |= NETIF_F_TSO;
2196
2197                 priv->dev->features |= priv->dev->hw_features;
2198         }
2199 }
2200
2201 static struct net_device *ipoib_add_port(const char *format,
2202                                          struct ib_device *hca, u8 port)
2203 {
2204         struct ipoib_dev_priv *priv;
2205         struct ib_port_attr attr;
2206         struct rdma_netdev *rn;
2207         int result = -ENOMEM;
2208
2209         priv = ipoib_intf_alloc(hca, port, format);
2210         if (!priv)
2211                 goto alloc_mem_failed;
2212
2213         SET_NETDEV_DEV(priv->dev, hca->dev.parent);
2214         priv->dev->dev_id = port - 1;
2215
2216         result = ib_query_port(hca, port, &attr);
2217         if (result) {
2218                 printk(KERN_WARNING "%s: ib_query_port %d failed\n",
2219                        hca->name, port);
2220                 goto device_init_failed;
2221         }
2222
2223         priv->max_ib_mtu = ib_mtu_enum_to_int(attr.max_mtu);
2224
2225         /* MTU will be reset when mcast join happens */
2226         priv->dev->mtu  = IPOIB_UD_MTU(priv->max_ib_mtu);
2227         priv->mcast_mtu  = priv->admin_mtu = priv->dev->mtu;
2228         priv->dev->max_mtu = IPOIB_CM_MTU;
2229
2230         priv->dev->neigh_priv_len = sizeof(struct ipoib_neigh);
2231
2232         result = ib_query_pkey(hca, port, 0, &priv->pkey);
2233         if (result) {
2234                 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
2235                        hca->name, port, result);
2236                 goto device_init_failed;
2237         }
2238
2239         ipoib_set_dev_features(priv, hca);
2240
2241         /*
2242          * Set the full membership bit, so that we join the right
2243          * broadcast group, etc.
2244          */
2245         priv->pkey |= 0x8000;
2246
2247         priv->dev->broadcast[8] = priv->pkey >> 8;
2248         priv->dev->broadcast[9] = priv->pkey & 0xff;
2249
2250         result = ib_query_gid(hca, port, 0, &priv->local_gid, NULL);
2251         if (result) {
2252                 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
2253                        hca->name, port, result);
2254                 goto device_init_failed;
2255         }
2256
2257         memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw,
2258                sizeof(union ib_gid));
2259         set_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags);
2260
2261         result = ipoib_dev_init(priv->dev, hca, port);
2262         if (result) {
2263                 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
2264                        hca->name, port, result);
2265                 goto device_init_failed;
2266         }
2267
2268         INIT_IB_EVENT_HANDLER(&priv->event_handler,
2269                               priv->ca, ipoib_event);
2270         ib_register_event_handler(&priv->event_handler);
2271
2272         result = register_netdev(priv->dev);
2273         if (result) {
2274                 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
2275                        hca->name, port, result);
2276                 goto register_failed;
2277         }
2278
2279         result = -ENOMEM;
2280         if (ipoib_cm_add_mode_attr(priv->dev))
2281                 goto sysfs_failed;
2282         if (ipoib_add_pkey_attr(priv->dev))
2283                 goto sysfs_failed;
2284         if (ipoib_add_umcast_attr(priv->dev))
2285                 goto sysfs_failed;
2286         if (device_create_file(&priv->dev->dev, &dev_attr_create_child))
2287                 goto sysfs_failed;
2288         if (device_create_file(&priv->dev->dev, &dev_attr_delete_child))
2289                 goto sysfs_failed;
2290
2291         return priv->dev;
2292
2293 sysfs_failed:
2294         unregister_netdev(priv->dev);
2295
2296 register_failed:
2297         ib_unregister_event_handler(&priv->event_handler);
2298         flush_workqueue(ipoib_workqueue);
2299         /* Stop GC if started before flush */
2300         set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
2301         cancel_delayed_work(&priv->neigh_reap_task);
2302         flush_workqueue(priv->wq);
2303         ipoib_dev_cleanup(priv->dev);
2304
2305 device_init_failed:
2306         rn = netdev_priv(priv->dev);
2307         rn->free_rdma_netdev(priv->dev);
2308         kfree(priv);
2309
2310 alloc_mem_failed:
2311         return ERR_PTR(result);
2312 }
2313
2314 static void ipoib_add_one(struct ib_device *device)
2315 {
2316         struct list_head *dev_list;
2317         struct net_device *dev;
2318         struct ipoib_dev_priv *priv;
2319         int p;
2320         int count = 0;
2321
2322         dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
2323         if (!dev_list)
2324                 return;
2325
2326         INIT_LIST_HEAD(dev_list);
2327
2328         for (p = rdma_start_port(device); p <= rdma_end_port(device); ++p) {
2329                 if (!rdma_protocol_ib(device, p))
2330                         continue;
2331                 dev = ipoib_add_port("ib%d", device, p);
2332                 if (!IS_ERR(dev)) {
2333                         priv = ipoib_priv(dev);
2334                         list_add_tail(&priv->list, dev_list);
2335                         count++;
2336                 }
2337         }
2338
2339         if (!count) {
2340                 pr_err("Failed to init port, removing it\n");
2341                 ipoib_remove_one(device, dev_list);
2342                 return;
2343         }
2344
2345         ib_set_client_data(device, &ipoib_client, dev_list);
2346 }
2347
2348 static void ipoib_remove_one(struct ib_device *device, void *client_data)
2349 {
2350         struct ipoib_dev_priv *priv, *tmp, *cpriv, *tcpriv;
2351         struct list_head *dev_list = client_data;
2352
2353         if (!dev_list)
2354                 return;
2355
2356         list_for_each_entry_safe(priv, tmp, dev_list, list) {
2357                 struct rdma_netdev *parent_rn = netdev_priv(priv->dev);
2358
2359                 ib_unregister_event_handler(&priv->event_handler);
2360                 flush_workqueue(ipoib_workqueue);
2361
2362                 /* mark interface in the middle of destruction */
2363                 set_bit(IPOIB_FLAG_GOING_DOWN, &priv->flags);
2364
2365                 rtnl_lock();
2366                 dev_change_flags(priv->dev, priv->dev->flags & ~IFF_UP);
2367                 rtnl_unlock();
2368
2369                 /* Stop GC */
2370                 set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
2371                 cancel_delayed_work(&priv->neigh_reap_task);
2372                 flush_workqueue(priv->wq);
2373
2374                 /* Wrap rtnl_lock/unlock with mutex to protect sysfs calls */
2375                 mutex_lock(&priv->sysfs_mutex);
2376                 unregister_netdev(priv->dev);
2377                 mutex_unlock(&priv->sysfs_mutex);
2378
2379                 parent_rn->free_rdma_netdev(priv->dev);
2380
2381                 list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
2382                         struct rdma_netdev *child_rn;
2383
2384                         child_rn = netdev_priv(cpriv->dev);
2385                         child_rn->free_rdma_netdev(cpriv->dev);
2386                         kfree(cpriv);
2387                 }
2388
2389                 kfree(priv);
2390         }
2391
2392         kfree(dev_list);
2393 }
2394
2395 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
2396 static struct notifier_block ipoib_netdev_notifier = {
2397         .notifier_call = ipoib_netdev_event,
2398 };
2399 #endif
2400
2401 static int __init ipoib_init_module(void)
2402 {
2403         int ret;
2404
2405         ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
2406         ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
2407         ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
2408
2409         ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
2410         ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
2411         ipoib_sendq_size = max3(ipoib_sendq_size, 2 * MAX_SEND_CQE, IPOIB_MIN_QUEUE_SIZE);
2412 #ifdef CONFIG_INFINIBAND_IPOIB_CM
2413         ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
2414         ipoib_max_conn_qp = max(ipoib_max_conn_qp, 0);
2415 #endif
2416
2417         /*
2418          * When copying small received packets, we only copy from the
2419          * linear data part of the SKB, so we rely on this condition.
2420          */
2421         BUILD_BUG_ON(IPOIB_CM_COPYBREAK > IPOIB_CM_HEAD_SIZE);
2422
2423         ret = ipoib_register_debugfs();
2424         if (ret)
2425                 return ret;
2426
2427         /*
2428          * We create a global workqueue here that is used for all flush
2429          * operations.  However, if you attempt to flush a workqueue
2430          * from a task on that same workqueue, it deadlocks the system.
2431          * We want to be able to flush the tasks associated with a
2432          * specific net device, so we also create a workqueue for each
2433          * netdevice.  We queue up the tasks for that device only on
2434          * its private workqueue, and we only queue up flush events
2435          * on our global flush workqueue.  This avoids the deadlocks.
2436          */
2437         ipoib_workqueue = alloc_ordered_workqueue("ipoib_flush",
2438                                                   WQ_MEM_RECLAIM);
2439         if (!ipoib_workqueue) {
2440                 ret = -ENOMEM;
2441                 goto err_fs;
2442         }
2443
2444         ib_sa_register_client(&ipoib_sa_client);
2445
2446         ret = ib_register_client(&ipoib_client);
2447         if (ret)
2448                 goto err_sa;
2449
2450         ret = ipoib_netlink_init();
2451         if (ret)
2452                 goto err_client;
2453
2454 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
2455         register_netdevice_notifier(&ipoib_netdev_notifier);
2456 #endif
2457         return 0;
2458
2459 err_client:
2460         ib_unregister_client(&ipoib_client);
2461
2462 err_sa:
2463         ib_sa_unregister_client(&ipoib_sa_client);
2464         destroy_workqueue(ipoib_workqueue);
2465
2466 err_fs:
2467         ipoib_unregister_debugfs();
2468
2469         return ret;
2470 }
2471
2472 static void __exit ipoib_cleanup_module(void)
2473 {
2474 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
2475         unregister_netdevice_notifier(&ipoib_netdev_notifier);
2476 #endif
2477         ipoib_netlink_fini();
2478         ib_unregister_client(&ipoib_client);
2479         ib_sa_unregister_client(&ipoib_sa_client);
2480         ipoib_unregister_debugfs();
2481         destroy_workqueue(ipoib_workqueue);
2482 }
2483
2484 module_init(ipoib_init_module);
2485 module_exit(ipoib_cleanup_module);