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 MODULE_AUTHOR("Roland Dreier");
  56 MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
  57 MODULE_LICENSE("Dual BSD/GPL");
  58
  59 int ipoib_sendq_size __read_mostly = IPOIB_TX_RING_SIZE;
  60 int ipoib_recvq_size __read_mostly = IPOIB_RX_RING_SIZE;
  61
  62 module_param_named(send_queue_size, ipoib_sendq_size, int, 0444);
  63 MODULE_PARM_DESC(send_queue_size, "Number of descriptors in send queue");
  64 module_param_named(recv_queue_size, ipoib_recvq_size, int, 0444);
  65 MODULE_PARM_DESC(recv_queue_size, "Number of descriptors in receive queue");
  66
  67 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
  68 int ipoib_debug_level;
  69
  70 module_param_named(debug_level, ipoib_debug_level, int, 0644);
  71 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
  72 #endif
  73
  74 struct ipoib_path_iter {
  75         struct net_device *dev;
  76         struct ipoib_path  path;
  77 };
  78
  79 static const u8 ipv4_bcast_addr[] = {
  80         0x00, 0xff, 0xff, 0xff,
  81         0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
  82         0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff
  83 };
  84
  85 struct workqueue_struct *ipoib_workqueue;
  86
  87 struct ib_sa_client ipoib_sa_client;
  88
  89 static int ipoib_add_one(struct ib_device *device);
  90 static void ipoib_remove_one(struct ib_device *device, void *client_data);
  91 static void ipoib_neigh_reclaim(struct rcu_head *rp);
  92 static struct net_device *ipoib_get_net_dev_by_params(
  93                 struct ib_device *dev, u8 port, u16 pkey,
  94                 const union ib_gid *gid, const struct sockaddr *addr,
  95                 void *client_data);
  96 static int ipoib_set_mac(struct net_device *dev, void *addr);
  97 static int ipoib_ioctl(struct net_device *dev, struct ifreq *ifr,
  98                        int cmd);
  99
 100 static struct ib_client ipoib_client = {
 101         .name   = "ipoib",
 102         .add    = ipoib_add_one,
 103         .remove = ipoib_remove_one,
 104         .get_net_dev_by_params = ipoib_get_net_dev_by_params,
 105 };
 106
 107 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 108 static int ipoib_netdev_event(struct notifier_block *this,
 109                               unsigned long event, void *ptr)
 110 {
 111         struct netdev_notifier_info *ni = ptr;
 112         struct net_device *dev = ni->dev;
 113
 114         if (dev->netdev_ops->ndo_open != ipoib_open)
 115                 return NOTIFY_DONE;
 116
 117         switch (event) {
 118         case NETDEV_REGISTER:
 119                 ipoib_create_debug_files(dev);
 120                 break;
 121         case NETDEV_CHANGENAME:
 122                 ipoib_delete_debug_files(dev);
 123                 ipoib_create_debug_files(dev);
 124                 break;
 125         case NETDEV_UNREGISTER:
 126                 ipoib_delete_debug_files(dev);
 127                 break;
 128         }
 129
 130         return NOTIFY_DONE;
 131 }
 132 #endif
 133
 134 int ipoib_open(struct net_device *dev)
 135 {
 136         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 137
 138         ipoib_dbg(priv, "bringing up interface\n");
 139
 140         netif_carrier_off(dev);
 141
 142         set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 143
 144         priv->sm_fullmember_sendonly_support = false;
 145
 146         if (ipoib_ib_dev_open(dev)) {
 147                 if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags))
 148                         return 0;
 149                 goto err_disable;
 150         }
 151
 152         ipoib_ib_dev_up(dev);
 153
 154         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 155                 struct ipoib_dev_priv *cpriv;
 156
 157                 /* Bring up any child interfaces too */
 158                 down_read(&priv->vlan_rwsem);
 159                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 160                         int flags;
 161
 162                         flags = cpriv->dev->flags;
 163                         if (flags & IFF_UP)
 164                                 continue;
 165
 166                         dev_change_flags(cpriv->dev, flags | IFF_UP, NULL);
 167                 }
 168                 up_read(&priv->vlan_rwsem);
 169         }
 170
 171         netif_start_queue(dev);
 172
 173         return 0;
 174
 175 err_disable:
 176         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 177
 178         return -EINVAL;
 179 }
 180
 181 static int ipoib_stop(struct net_device *dev)
 182 {
 183         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 184
 185         ipoib_dbg(priv, "stopping interface\n");
 186
 187         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 188
 189         netif_stop_queue(dev);
 190
 191         ipoib_ib_dev_down(dev);
 192         ipoib_ib_dev_stop(dev);
 193
 194         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 195                 struct ipoib_dev_priv *cpriv;
 196
 197                 /* Bring down any child interfaces too */
 198                 down_read(&priv->vlan_rwsem);
 199                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 200                         int flags;
 201
 202                         flags = cpriv->dev->flags;
 203                         if (!(flags & IFF_UP))
 204                                 continue;
 205
 206                         dev_change_flags(cpriv->dev, flags & ~IFF_UP, NULL);
 207                 }
 208                 up_read(&priv->vlan_rwsem);
 209         }
 210
 211         return 0;
 212 }
 213
 214 static netdev_features_t ipoib_fix_features(struct net_device *dev, netdev_features_t features)
 215 {
 216         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 217
 218         if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
 219                 features &= ~(NETIF_F_IP_CSUM | NETIF_F_TSO);
 220
 221         return features;
 222 }
 223
 224 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
 225 {
 226         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 227         int ret = 0;
 228
 229         /* dev->mtu > 2K ==> connected mode */
 230         if (ipoib_cm_admin_enabled(dev)) {
 231                 if (new_mtu > ipoib_cm_max_mtu(dev))
 232                         return -EINVAL;
 233
 234                 if (new_mtu > priv->mcast_mtu)
 235                         ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
 236                                    priv->mcast_mtu);
 237
 238                 dev->mtu = new_mtu;
 239                 return 0;
 240         }
 241
 242         if (new_mtu < (ETH_MIN_MTU + IPOIB_ENCAP_LEN) ||
 243             new_mtu > IPOIB_UD_MTU(priv->max_ib_mtu))
 244                 return -EINVAL;
 245
 246         priv->admin_mtu = new_mtu;
 247
 248         if (priv->mcast_mtu < priv->admin_mtu)
 249                 ipoib_dbg(priv, "MTU must be smaller than the underlying "
 250                                 "link layer MTU - 4 (%u)\n", priv->mcast_mtu);
 251
 252         new_mtu = min(priv->mcast_mtu, priv->admin_mtu);
 253
 254         if (priv->rn_ops->ndo_change_mtu) {
 255                 bool carrier_status = netif_carrier_ok(dev);
 256
 257                 netif_carrier_off(dev);
 258
 259                 /* notify lower level on the real mtu */
 260                 ret = priv->rn_ops->ndo_change_mtu(dev, new_mtu);
 261
 262                 if (carrier_status)
 263                         netif_carrier_on(dev);
 264         } else {
 265                 dev->mtu = new_mtu;
 266         }
 267
 268         return ret;
 269 }
 270
 271 static void ipoib_get_stats(struct net_device *dev,
 272                             struct rtnl_link_stats64 *stats)
 273 {
 274         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 275
 276         if (priv->rn_ops->ndo_get_stats64)
 277                 priv->rn_ops->ndo_get_stats64(dev, stats);
 278         else
 279                 netdev_stats_to_stats64(stats, &dev->stats);
 280 }
 281
 282 /* Called with an RCU read lock taken */
 283 static bool ipoib_is_dev_match_addr_rcu(const struct sockaddr *addr,
 284                                         struct net_device *dev)
 285 {
 286         struct net *net = dev_net(dev);
 287         struct in_device *in_dev;
 288         struct sockaddr_in *addr_in = (struct sockaddr_in *)addr;
 289         struct sockaddr_in6 *addr_in6 = (struct sockaddr_in6 *)addr;
 290         __be32 ret_addr;
 291
 292         switch (addr->sa_family) {
 293         case AF_INET:
 294                 in_dev = in_dev_get(dev);
 295                 if (!in_dev)
 296                         return false;
 297
 298                 ret_addr = inet_confirm_addr(net, in_dev, 0,
 299                                              addr_in->sin_addr.s_addr,
 300                                              RT_SCOPE_HOST);
 301                 in_dev_put(in_dev);
 302                 if (ret_addr)
 303                         return true;
 304
 305                 break;
 306         case AF_INET6:
 307                 if (IS_ENABLED(CONFIG_IPV6) &&
 308                     ipv6_chk_addr(net, &addr_in6->sin6_addr, dev, 1))
 309                         return true;
 310
 311                 break;
 312         }
 313         return false;
 314 }
 315
 316 /**
 317  * Find the master net_device on top of the given net_device.
 318  * @dev: base IPoIB net_device
 319  *
 320  * Returns the master net_device with a reference held, or the same net_device
 321  * if no master exists.
 322  */
 323 static struct net_device *ipoib_get_master_net_dev(struct net_device *dev)
 324 {
 325         struct net_device *master;
 326
 327         rcu_read_lock();
 328         master = netdev_master_upper_dev_get_rcu(dev);
 329         if (master)
 330                 dev_hold(master);
 331         rcu_read_unlock();
 332
 333         if (master)
 334                 return master;
 335
 336         dev_hold(dev);
 337         return dev;
 338 }
 339
 340 struct ipoib_walk_data {
 341         const struct sockaddr *addr;
 342         struct net_device *result;
 343 };
 344
 345 static int ipoib_upper_walk(struct net_device *upper,
 346                             struct netdev_nested_priv *priv)
 347 {
 348         struct ipoib_walk_data *data = (struct ipoib_walk_data *)priv->data;
 349         int ret = 0;
 350
 351         if (ipoib_is_dev_match_addr_rcu(data->addr, upper)) {
 352                 dev_hold(upper);
 353                 data->result = upper;
 354                 ret = 1;
 355         }
 356
 357         return ret;
 358 }
 359
 360 /**
 361  * Find a net_device matching the given address, which is an upper device of
 362  * the given net_device.
 363  * @addr: IP address to look for.
 364  * @dev: base IPoIB net_device
 365  *
 366  * If found, returns the net_device with a reference held. Otherwise return
 367  * NULL.
 368  */
 369 static struct net_device *ipoib_get_net_dev_match_addr(
 370                 const struct sockaddr *addr, struct net_device *dev)
 371 {
 372         struct netdev_nested_priv priv;
 373         struct ipoib_walk_data data = {
 374                 .addr = addr,
 375         };
 376
 377         priv.data = (void *)&data;
 378         rcu_read_lock();
 379         if (ipoib_is_dev_match_addr_rcu(addr, dev)) {
 380                 dev_hold(dev);
 381                 data.result = dev;
 382                 goto out;
 383         }
 384
 385         netdev_walk_all_upper_dev_rcu(dev, ipoib_upper_walk, &priv);
 386 out:
 387         rcu_read_unlock();
 388         return data.result;
 389 }
 390
 391 /* returns the number of IPoIB netdevs on top a given ipoib device matching a
 392  * pkey_index and address, if one exists.
 393  *
 394  * @found_net_dev: contains a matching net_device if the return value >= 1,
 395  * with a reference held. */
 396 static int ipoib_match_gid_pkey_addr(struct ipoib_dev_priv *priv,
 397                                      const union ib_gid *gid,
 398                                      u16 pkey_index,
 399                                      const struct sockaddr *addr,
 400                                      int nesting,
 401                                      struct net_device **found_net_dev)
 402 {
 403         struct ipoib_dev_priv *child_priv;
 404         struct net_device *net_dev = NULL;
 405         int matches = 0;
 406
 407         if (priv->pkey_index == pkey_index &&
 408             (!gid || !memcmp(gid, &priv->local_gid, sizeof(*gid)))) {
 409                 if (!addr) {
 410                         net_dev = ipoib_get_master_net_dev(priv->dev);
 411                 } else {
 412                         /* Verify the net_device matches the IP address, as
 413                          * IPoIB child devices currently share a GID. */
 414                         net_dev = ipoib_get_net_dev_match_addr(addr, priv->dev);
 415                 }
 416                 if (net_dev) {
 417                         if (!*found_net_dev)
 418                                 *found_net_dev = net_dev;
 419                         else
 420                                 dev_put(net_dev);
 421                         ++matches;
 422                 }
 423         }
 424
 425         /* Check child interfaces */
 426         down_read_nested(&priv->vlan_rwsem, nesting);
 427         list_for_each_entry(child_priv, &priv->child_intfs, list) {
 428                 matches += ipoib_match_gid_pkey_addr(child_priv, gid,
 429                                                     pkey_index, addr,
 430                                                     nesting + 1,
 431                                                     found_net_dev);
 432                 if (matches > 1)
 433                         break;
 434         }
 435         up_read(&priv->vlan_rwsem);
 436
 437         return matches;
 438 }
 439
 440 /* Returns the number of matching net_devs found (between 0 and 2). Also
 441  * return the matching net_device in the @net_dev parameter, holding a
 442  * reference to the net_device, if the number of matches >= 1 */
 443 static int __ipoib_get_net_dev_by_params(struct list_head *dev_list, u8 port,
 444                                          u16 pkey_index,
 445                                          const union ib_gid *gid,
 446                                          const struct sockaddr *addr,
 447                                          struct net_device **net_dev)
 448 {
 449         struct ipoib_dev_priv *priv;
 450         int matches = 0;
 451
 452         *net_dev = NULL;
 453
 454         list_for_each_entry(priv, dev_list, list) {
 455                 if (priv->port != port)
 456                         continue;
 457
 458                 matches += ipoib_match_gid_pkey_addr(priv, gid, pkey_index,
 459                                                      addr, 0, net_dev);
 460                 if (matches > 1)
 461                         break;
 462         }
 463
 464         return matches;
 465 }
 466
 467 static struct net_device *ipoib_get_net_dev_by_params(
 468                 struct ib_device *dev, u8 port, u16 pkey,
 469                 const union ib_gid *gid, const struct sockaddr *addr,
 470                 void *client_data)
 471 {
 472         struct net_device *net_dev;
 473         struct list_head *dev_list = client_data;
 474         u16 pkey_index;
 475         int matches;
 476         int ret;
 477
 478         if (!rdma_protocol_ib(dev, port))
 479                 return NULL;
 480
 481         ret = ib_find_cached_pkey(dev, port, pkey, &pkey_index);
 482         if (ret)
 483                 return NULL;
 484
 485         /* See if we can find a unique device matching the L2 parameters */
 486         matches = __ipoib_get_net_dev_by_params(dev_list, port, pkey_index,
 487                                                 gid, NULL, &net_dev);
 488
 489         switch (matches) {
 490         case 0:
 491                 return NULL;
 492         case 1:
 493                 return net_dev;
 494         }
 495
 496         dev_put(net_dev);
 497
 498         /* Couldn't find a unique device with L2 parameters only. Use L3
 499          * address to uniquely match the net device */
 500         matches = __ipoib_get_net_dev_by_params(dev_list, port, pkey_index,
 501                                                 gid, addr, &net_dev);
 502         switch (matches) {
 503         case 0:
 504                 return NULL;
 505         default:
 506                 dev_warn_ratelimited(&dev->dev,
 507                                      "duplicate IP address detected\n");
 508                 fallthrough;
 509         case 1:
 510                 return net_dev;
 511         }
 512 }
 513
 514 int ipoib_set_mode(struct net_device *dev, const char *buf)
 515 {
 516         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 517
 518         if ((test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags) &&
 519              !strcmp(buf, "connected\n")) ||
 520              (!test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags) &&
 521              !strcmp(buf, "datagram\n"))) {
 522                 return 0;
 523         }
 524
 525         /* flush paths if we switch modes so that connections are restarted */
 526         if (IPOIB_CM_SUPPORTED(dev->dev_addr) && !strcmp(buf, "connected\n")) {
 527                 set_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
 528                 ipoib_warn(priv, "enabling connected mode "
 529                            "will cause multicast packet drops\n");
 530                 netdev_update_features(dev);
 531                 dev_set_mtu(dev, ipoib_cm_max_mtu(dev));
 532                 netif_set_real_num_tx_queues(dev, 1);
 533                 rtnl_unlock();
 534                 priv->tx_wr.wr.send_flags &= ~IB_SEND_IP_CSUM;
 535
 536                 ipoib_flush_paths(dev);
 537                 return (!rtnl_trylock()) ? -EBUSY : 0;
 538         }
 539
 540         if (!strcmp(buf, "datagram\n")) {
 541                 clear_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
 542                 netdev_update_features(dev);
 543                 dev_set_mtu(dev, min(priv->mcast_mtu, dev->mtu));
 544                 netif_set_real_num_tx_queues(dev, dev->num_tx_queues);
 545                 rtnl_unlock();
 546                 ipoib_flush_paths(dev);
 547                 return (!rtnl_trylock()) ? -EBUSY : 0;
 548         }
 549
 550         return -EINVAL;
 551 }
 552
 553 struct ipoib_path *__path_find(struct net_device *dev, void *gid)
 554 {
 555         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 556         struct rb_node *n = priv->path_tree.rb_node;
 557         struct ipoib_path *path;
 558         int ret;
 559
 560         while (n) {
 561                 path = rb_entry(n, struct ipoib_path, rb_node);
 562
 563                 ret = memcmp(gid, path->pathrec.dgid.raw,
 564                              sizeof (union ib_gid));
 565
 566                 if (ret < 0)
 567                         n = n->rb_left;
 568                 else if (ret > 0)
 569                         n = n->rb_right;
 570                 else
 571                         return path;
 572         }
 573
 574         return NULL;
 575 }
 576
 577 static int __path_add(struct net_device *dev, struct ipoib_path *path)
 578 {
 579         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 580         struct rb_node **n = &priv->path_tree.rb_node;
 581         struct rb_node *pn = NULL;
 582         struct ipoib_path *tpath;
 583         int ret;
 584
 585         while (*n) {
 586                 pn = *n;
 587                 tpath = rb_entry(pn, struct ipoib_path, rb_node);
 588
 589                 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
 590                              sizeof (union ib_gid));
 591                 if (ret < 0)
 592                         n = &pn->rb_left;
 593                 else if (ret > 0)
 594                         n = &pn->rb_right;
 595                 else
 596                         return -EEXIST;
 597         }
 598
 599         rb_link_node(&path->rb_node, pn, n);
 600         rb_insert_color(&path->rb_node, &priv->path_tree);
 601
 602         list_add_tail(&path->list, &priv->path_list);
 603
 604         return 0;
 605 }
 606
 607 static void path_free(struct net_device *dev, struct ipoib_path *path)
 608 {
 609         struct sk_buff *skb;
 610
 611         while ((skb = __skb_dequeue(&path->queue)))
 612                 dev_kfree_skb_irq(skb);
 613
 614         ipoib_dbg(ipoib_priv(dev), "%s\n", __func__);
 615
 616         /* remove all neigh connected to this path */
 617         ipoib_del_neighs_by_gid(dev, path->pathrec.dgid.raw);
 618
 619         if (path->ah)
 620                 ipoib_put_ah(path->ah);
 621
 622         kfree(path);
 623 }
 624
 625 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 626
 627 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
 628 {
 629         struct ipoib_path_iter *iter;
 630
 631         iter = kmalloc(sizeof(*iter), GFP_KERNEL);
 632         if (!iter)
 633                 return NULL;
 634
 635         iter->dev = dev;
 636         memset(iter->path.pathrec.dgid.raw, 0, 16);
 637
 638         if (ipoib_path_iter_next(iter)) {
 639                 kfree(iter);
 640                 return NULL;
 641         }
 642
 643         return iter;
 644 }
 645
 646 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
 647 {
 648         struct ipoib_dev_priv *priv = ipoib_priv(iter->dev);
 649         struct rb_node *n;
 650         struct ipoib_path *path;
 651         int ret = 1;
 652
 653         spin_lock_irq(&priv->lock);
 654
 655         n = rb_first(&priv->path_tree);
 656
 657         while (n) {
 658                 path = rb_entry(n, struct ipoib_path, rb_node);
 659
 660                 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
 661                            sizeof (union ib_gid)) < 0) {
 662                         iter->path = *path;
 663                         ret = 0;
 664                         break;
 665                 }
 666
 667                 n = rb_next(n);
 668         }
 669
 670         spin_unlock_irq(&priv->lock);
 671
 672         return ret;
 673 }
 674
 675 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
 676                           struct ipoib_path *path)
 677 {
 678         *path = iter->path;
 679 }
 680
 681 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
 682
 683 void ipoib_mark_paths_invalid(struct net_device *dev)
 684 {
 685         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 686         struct ipoib_path *path, *tp;
 687
 688         spin_lock_irq(&priv->lock);
 689
 690         list_for_each_entry_safe(path, tp, &priv->path_list, list) {
 691                 ipoib_dbg(priv, "mark path LID 0x%08x GID %pI6 invalid\n",
 692                           be32_to_cpu(sa_path_get_dlid(&path->pathrec)),
 693                           path->pathrec.dgid.raw);
 694                 if (path->ah)
 695                         path->ah->valid = 0;
 696         }
 697
 698         spin_unlock_irq(&priv->lock);
 699 }
 700
 701 static void push_pseudo_header(struct sk_buff *skb, const char *daddr)
 702 {
 703         struct ipoib_pseudo_header *phdr;
 704
 705         phdr = skb_push(skb, sizeof(*phdr));
 706         memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
 707 }
 708
 709 void ipoib_flush_paths(struct net_device *dev)
 710 {
 711         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 712         struct ipoib_path *path, *tp;
 713         LIST_HEAD(remove_list);
 714         unsigned long flags;
 715
 716         netif_tx_lock_bh(dev);
 717         spin_lock_irqsave(&priv->lock, flags);
 718
 719         list_splice_init(&priv->path_list, &remove_list);
 720
 721         list_for_each_entry(path, &remove_list, list)
 722                 rb_erase(&path->rb_node, &priv->path_tree);
 723
 724         list_for_each_entry_safe(path, tp, &remove_list, list) {
 725                 if (path->query)
 726                         ib_sa_cancel_query(path->query_id, path->query);
 727                 spin_unlock_irqrestore(&priv->lock, flags);
 728                 netif_tx_unlock_bh(dev);
 729                 wait_for_completion(&path->done);
 730                 path_free(dev, path);
 731                 netif_tx_lock_bh(dev);
 732                 spin_lock_irqsave(&priv->lock, flags);
 733         }
 734
 735         spin_unlock_irqrestore(&priv->lock, flags);
 736         netif_tx_unlock_bh(dev);
 737 }
 738
 739 static void path_rec_completion(int status,
 740                                 struct sa_path_rec *pathrec,
 741                                 void *path_ptr)
 742 {
 743         struct ipoib_path *path = path_ptr;
 744         struct net_device *dev = path->dev;
 745         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 746         struct ipoib_ah *ah = NULL;
 747         struct ipoib_ah *old_ah = NULL;
 748         struct ipoib_neigh *neigh, *tn;
 749         struct sk_buff_head skqueue;
 750         struct sk_buff *skb;
 751         unsigned long flags;
 752
 753         if (!status)
 754                 ipoib_dbg(priv, "PathRec LID 0x%04x for GID %pI6\n",
 755                           be32_to_cpu(sa_path_get_dlid(pathrec)),
 756                           pathrec->dgid.raw);
 757         else
 758                 ipoib_dbg(priv, "PathRec status %d for GID %pI6\n",
 759                           status, path->pathrec.dgid.raw);
 760
 761         skb_queue_head_init(&skqueue);
 762
 763         if (!status) {
 764                 struct rdma_ah_attr av;
 765
 766                 if (!ib_init_ah_attr_from_path(priv->ca, priv->port,
 767                                                pathrec, &av, NULL)) {
 768                         ah = ipoib_create_ah(dev, priv->pd, &av);
 769                         rdma_destroy_ah_attr(&av);
 770                 }
 771         }
 772
 773         spin_lock_irqsave(&priv->lock, flags);
 774
 775         if (!IS_ERR_OR_NULL(ah)) {
 776                 /*
 777                  * pathrec.dgid is used as the database key from the LLADDR,
 778                  * it must remain unchanged even if the SA returns a different
 779                  * GID to use in the AH.
 780                  */
 781                 if (memcmp(pathrec->dgid.raw, path->pathrec.dgid.raw,
 782                            sizeof(union ib_gid))) {
 783                         ipoib_dbg(
 784                                 priv,
 785                                 "%s got PathRec for gid %pI6 while asked for %pI6\n",
 786                                 dev->name, pathrec->dgid.raw,
 787                                 path->pathrec.dgid.raw);
 788                         memcpy(pathrec->dgid.raw, path->pathrec.dgid.raw,
 789                                sizeof(union ib_gid));
 790                 }
 791
 792                 path->pathrec = *pathrec;
 793
 794                 old_ah   = path->ah;
 795                 path->ah = ah;
 796
 797                 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
 798                           ah, be32_to_cpu(sa_path_get_dlid(pathrec)),
 799                           pathrec->sl);
 800
 801                 while ((skb = __skb_dequeue(&path->queue)))
 802                         __skb_queue_tail(&skqueue, skb);
 803
 804                 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
 805                         if (neigh->ah) {
 806                                 WARN_ON(neigh->ah != old_ah);
 807                                 /*
 808                                  * Dropping the ah reference inside
 809                                  * priv->lock is safe here, because we
 810                                  * will hold one more reference from
 811                                  * the original value of path->ah (ie
 812                                  * old_ah).
 813                                  */
 814                                 ipoib_put_ah(neigh->ah);
 815                         }
 816                         kref_get(&path->ah->ref);
 817                         neigh->ah = path->ah;
 818
 819                         if (ipoib_cm_enabled(dev, neigh->daddr)) {
 820                                 if (!ipoib_cm_get(neigh))
 821                                         ipoib_cm_set(neigh, ipoib_cm_create_tx(dev,
 822                                                                                path,
 823                                                                                neigh));
 824                                 if (!ipoib_cm_get(neigh)) {
 825                                         ipoib_neigh_free(neigh);
 826                                         continue;
 827                                 }
 828                         }
 829
 830                         while ((skb = __skb_dequeue(&neigh->queue)))
 831                                 __skb_queue_tail(&skqueue, skb);
 832                 }
 833                 path->ah->valid = 1;
 834         }
 835
 836         path->query = NULL;
 837         complete(&path->done);
 838
 839         spin_unlock_irqrestore(&priv->lock, flags);
 840
 841         if (IS_ERR_OR_NULL(ah))
 842                 ipoib_del_neighs_by_gid(dev, path->pathrec.dgid.raw);
 843
 844         if (old_ah)
 845                 ipoib_put_ah(old_ah);
 846
 847         while ((skb = __skb_dequeue(&skqueue))) {
 848                 int ret;
 849                 skb->dev = dev;
 850                 ret = dev_queue_xmit(skb);
 851                 if (ret)
 852                         ipoib_warn(priv, "%s: dev_queue_xmit failed to re-queue packet, ret:%d\n",
 853                                    __func__, ret);
 854         }
 855 }
 856
 857 static void init_path_rec(struct ipoib_dev_priv *priv, struct ipoib_path *path,
 858                           void *gid)
 859 {
 860         path->dev = priv->dev;
 861
 862         if (rdma_cap_opa_ah(priv->ca, priv->port))
 863                 path->pathrec.rec_type = SA_PATH_REC_TYPE_OPA;
 864         else
 865                 path->pathrec.rec_type = SA_PATH_REC_TYPE_IB;
 866
 867         memcpy(path->pathrec.dgid.raw, gid, sizeof(union ib_gid));
 868         path->pathrec.sgid          = priv->local_gid;
 869         path->pathrec.pkey          = cpu_to_be16(priv->pkey);
 870         path->pathrec.numb_path     = 1;
 871         path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
 872 }
 873
 874 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
 875 {
 876         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 877         struct ipoib_path *path;
 878
 879         if (!priv->broadcast)
 880                 return NULL;
 881
 882         path = kzalloc(sizeof(*path), GFP_ATOMIC);
 883         if (!path)
 884                 return NULL;
 885
 886         skb_queue_head_init(&path->queue);
 887
 888         INIT_LIST_HEAD(&path->neigh_list);
 889
 890         init_path_rec(priv, path, gid);
 891
 892         return path;
 893 }
 894
 895 static int path_rec_start(struct net_device *dev,
 896                           struct ipoib_path *path)
 897 {
 898         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 899
 900         ipoib_dbg(priv, "Start path record lookup for %pI6\n",
 901                   path->pathrec.dgid.raw);
 902
 903         init_completion(&path->done);
 904
 905         path->query_id =
 906                 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
 907                                    &path->pathrec,
 908                                    IB_SA_PATH_REC_DGID          |
 909                                    IB_SA_PATH_REC_SGID          |
 910                                    IB_SA_PATH_REC_NUMB_PATH     |
 911                                    IB_SA_PATH_REC_TRAFFIC_CLASS |
 912                                    IB_SA_PATH_REC_PKEY,
 913                                    1000, GFP_ATOMIC,
 914                                    path_rec_completion,
 915                                    path, &path->query);
 916         if (path->query_id < 0) {
 917                 ipoib_warn(priv, "ib_sa_path_rec_get failed: %d\n", path->query_id);
 918                 path->query = NULL;
 919                 complete(&path->done);
 920                 return path->query_id;
 921         }
 922
 923         return 0;
 924 }
 925
 926 static void neigh_refresh_path(struct ipoib_neigh *neigh, u8 *daddr,
 927                                struct net_device *dev)
 928 {
 929         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 930         struct ipoib_path *path;
 931         unsigned long flags;
 932
 933         spin_lock_irqsave(&priv->lock, flags);
 934
 935         path = __path_find(dev, daddr + 4);
 936         if (!path)
 937                 goto out;
 938         if (!path->query)
 939                 path_rec_start(dev, path);
 940 out:
 941         spin_unlock_irqrestore(&priv->lock, flags);
 942 }
 943
 944 static struct ipoib_neigh *neigh_add_path(struct sk_buff *skb, u8 *daddr,
 945                                           struct net_device *dev)
 946 {
 947         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 948         struct rdma_netdev *rn = netdev_priv(dev);
 949         struct ipoib_path *path;
 950         struct ipoib_neigh *neigh;
 951         unsigned long flags;
 952
 953         spin_lock_irqsave(&priv->lock, flags);
 954         neigh = ipoib_neigh_alloc(daddr, dev);
 955         if (!neigh) {
 956                 spin_unlock_irqrestore(&priv->lock, flags);
 957                 ++dev->stats.tx_dropped;
 958                 dev_kfree_skb_any(skb);
 959                 return NULL;
 960         }
 961
 962         /* To avoid race condition, make sure that the
 963          * neigh will be added only once.
 964          */
 965         if (unlikely(!list_empty(&neigh->list))) {
 966                 spin_unlock_irqrestore(&priv->lock, flags);
 967                 return neigh;
 968         }
 969
 970         path = __path_find(dev, daddr + 4);
 971         if (!path) {
 972                 path = path_rec_create(dev, daddr + 4);
 973                 if (!path)
 974                         goto err_path;
 975
 976                 __path_add(dev, path);
 977         }
 978
 979         list_add_tail(&neigh->list, &path->neigh_list);
 980
 981         if (path->ah && path->ah->valid) {
 982                 kref_get(&path->ah->ref);
 983                 neigh->ah = path->ah;
 984
 985                 if (ipoib_cm_enabled(dev, neigh->daddr)) {
 986                         if (!ipoib_cm_get(neigh))
 987                                 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
 988                         if (!ipoib_cm_get(neigh)) {
 989                                 ipoib_neigh_free(neigh);
 990                                 goto err_drop;
 991                         }
 992                         if (skb_queue_len(&neigh->queue) <
 993                             IPOIB_MAX_PATH_REC_QUEUE) {
 994                                 push_pseudo_header(skb, neigh->daddr);
 995                                 __skb_queue_tail(&neigh->queue, skb);
 996                         } else {
 997                                 ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
 998                                            skb_queue_len(&neigh->queue));
 999                                 goto err_drop;
1000                         }
1001                 } else {
1002                         spin_unlock_irqrestore(&priv->lock, flags);
1003                         path->ah->last_send = rn->send(dev, skb, path->ah->ah,
1004                                                        IPOIB_QPN(daddr));
1005                         ipoib_neigh_put(neigh);
1006                         return NULL;
1007                 }
1008         } else {
1009                 neigh->ah  = NULL;
1010
1011                 if (!path->query && path_rec_start(dev, path))
1012                         goto err_path;
1013                 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
1014                         push_pseudo_header(skb, neigh->daddr);
1015                         __skb_queue_tail(&neigh->queue, skb);
1016                 } else {
1017                         goto err_drop;
1018                 }
1019         }
1020
1021         spin_unlock_irqrestore(&priv->lock, flags);
1022         ipoib_neigh_put(neigh);
1023         return NULL;
1024
1025 err_path:
1026         ipoib_neigh_free(neigh);
1027 err_drop:
1028         ++dev->stats.tx_dropped;
1029         dev_kfree_skb_any(skb);
1030
1031         spin_unlock_irqrestore(&priv->lock, flags);
1032         ipoib_neigh_put(neigh);
1033
1034         return NULL;
1035 }
1036
1037 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
1038                              struct ipoib_pseudo_header *phdr)
1039 {
1040         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1041         struct rdma_netdev *rn = netdev_priv(dev);
1042         struct ipoib_path *path;
1043         unsigned long flags;
1044
1045         spin_lock_irqsave(&priv->lock, flags);
1046
1047         /* no broadcast means that all paths are (going to be) not valid */
1048         if (!priv->broadcast)
1049                 goto drop_and_unlock;
1050
1051         path = __path_find(dev, phdr->hwaddr + 4);
1052         if (!path || !path->ah || !path->ah->valid) {
1053                 if (!path) {
1054                         path = path_rec_create(dev, phdr->hwaddr + 4);
1055                         if (!path)
1056                                 goto drop_and_unlock;
1057                         __path_add(dev, path);
1058                 } else {
1059                         /*
1060                          * make sure there are no changes in the existing
1061                          * path record
1062                          */
1063                         init_path_rec(priv, path, phdr->hwaddr + 4);
1064                 }
1065                 if (!path->query && path_rec_start(dev, path)) {
1066                         goto drop_and_unlock;
1067                 }
1068
1069                 if (skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
1070                         push_pseudo_header(skb, phdr->hwaddr);
1071                         __skb_queue_tail(&path->queue, skb);
1072                         goto unlock;
1073                 } else {
1074                         goto drop_and_unlock;
1075                 }
1076         }
1077
1078         spin_unlock_irqrestore(&priv->lock, flags);
1079         ipoib_dbg(priv, "Send unicast ARP to %08x\n",
1080                   be32_to_cpu(sa_path_get_dlid(&path->pathrec)));
1081         path->ah->last_send = rn->send(dev, skb, path->ah->ah,
1082                                        IPOIB_QPN(phdr->hwaddr));
1083         return;
1084
1085 drop_and_unlock:
1086         ++dev->stats.tx_dropped;
1087         dev_kfree_skb_any(skb);
1088 unlock:
1089         spin_unlock_irqrestore(&priv->lock, flags);
1090 }
1091
1092 static netdev_tx_t ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
1093 {
1094         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1095         struct rdma_netdev *rn = netdev_priv(dev);
1096         struct ipoib_neigh *neigh;
1097         struct ipoib_pseudo_header *phdr;
1098         struct ipoib_header *header;
1099         unsigned long flags;
1100
1101         phdr = (struct ipoib_pseudo_header *) skb->data;
1102         skb_pull(skb, sizeof(*phdr));
1103         header = (struct ipoib_header *) skb->data;
1104
1105         if (unlikely(phdr->hwaddr[4] == 0xff)) {
1106                 /* multicast, arrange "if" according to probability */
1107                 if ((header->proto != htons(ETH_P_IP)) &&
1108                     (header->proto != htons(ETH_P_IPV6)) &&
1109                     (header->proto != htons(ETH_P_ARP)) &&
1110                     (header->proto != htons(ETH_P_RARP)) &&
1111                     (header->proto != htons(ETH_P_TIPC))) {
1112                         /* ethertype not supported by IPoIB */
1113                         ++dev->stats.tx_dropped;
1114                         dev_kfree_skb_any(skb);
1115                         return NETDEV_TX_OK;
1116                 }
1117                 /* Add in the P_Key for multicast*/
1118                 phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
1119                 phdr->hwaddr[9] = priv->pkey & 0xff;
1120
1121                 neigh = ipoib_neigh_get(dev, phdr->hwaddr);
1122                 if (likely(neigh))
1123                         goto send_using_neigh;
1124                 ipoib_mcast_send(dev, phdr->hwaddr, skb);
1125                 return NETDEV_TX_OK;
1126         }
1127
1128         /* unicast, arrange "switch" according to probability */
1129         switch (header->proto) {
1130         case htons(ETH_P_IP):
1131         case htons(ETH_P_IPV6):
1132         case htons(ETH_P_TIPC):
1133                 neigh = ipoib_neigh_get(dev, phdr->hwaddr);
1134                 if (unlikely(!neigh)) {
1135                         neigh = neigh_add_path(skb, phdr->hwaddr, dev);
1136                         if (likely(!neigh))
1137                                 return NETDEV_TX_OK;
1138                 }
1139                 break;
1140         case htons(ETH_P_ARP):
1141         case htons(ETH_P_RARP):
1142                 /* for unicast ARP and RARP should always perform path find */
1143                 unicast_arp_send(skb, dev, phdr);
1144                 return NETDEV_TX_OK;
1145         default:
1146                 /* ethertype not supported by IPoIB */
1147                 ++dev->stats.tx_dropped;
1148                 dev_kfree_skb_any(skb);
1149                 return NETDEV_TX_OK;
1150         }
1151
1152 send_using_neigh:
1153         /* note we now hold a ref to neigh */
1154         if (ipoib_cm_get(neigh)) {
1155                 if (ipoib_cm_up(neigh)) {
1156                         ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
1157                         goto unref;
1158                 }
1159         } else if (neigh->ah && neigh->ah->valid) {
1160                 neigh->ah->last_send = rn->send(dev, skb, neigh->ah->ah,
1161                                                 IPOIB_QPN(phdr->hwaddr));
1162                 goto unref;
1163         } else if (neigh->ah) {
1164                 neigh_refresh_path(neigh, phdr->hwaddr, dev);
1165         }
1166
1167         if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
1168                 push_pseudo_header(skb, phdr->hwaddr);
1169                 spin_lock_irqsave(&priv->lock, flags);
1170                 __skb_queue_tail(&neigh->queue, skb);
1171                 spin_unlock_irqrestore(&priv->lock, flags);
1172         } else {
1173                 ++dev->stats.tx_dropped;
1174                 dev_kfree_skb_any(skb);
1175         }
1176
1177 unref:
1178         ipoib_neigh_put(neigh);
1179
1180         return NETDEV_TX_OK;
1181 }
1182
1183 static void ipoib_timeout(struct net_device *dev, unsigned int txqueue)
1184 {
1185         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1186
1187         ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
1188                    jiffies_to_msecs(jiffies - dev_trans_start(dev)));
1189         ipoib_warn(priv,
1190                    "queue stopped %d, tx_head %u, tx_tail %u, global_tx_head %u, global_tx_tail %u\n",
1191                    netif_queue_stopped(dev), priv->tx_head, priv->tx_tail,
1192                    priv->global_tx_head, priv->global_tx_tail);
1193
1194         /* XXX reset QP, etc. */
1195 }
1196
1197 static int ipoib_hard_header(struct sk_buff *skb,
1198                              struct net_device *dev,
1199                              unsigned short type,
1200                              const void *daddr,
1201                              const void *saddr,
1202                              unsigned int len)
1203 {
1204         struct ipoib_header *header;
1205
1206         header = skb_push(skb, sizeof(*header));
1207
1208         header->proto = htons(type);
1209         header->reserved = 0;
1210
1211         /*
1212          * we don't rely on dst_entry structure,  always stuff the
1213          * destination address into skb hard header so we can figure out where
1214          * to send the packet later.
1215          */
1216         push_pseudo_header(skb, daddr);
1217
1218         return IPOIB_HARD_LEN;
1219 }
1220
1221 static void ipoib_set_mcast_list(struct net_device *dev)
1222 {
1223         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1224
1225         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
1226                 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
1227                 return;
1228         }
1229
1230         queue_work(priv->wq, &priv->restart_task);
1231 }
1232
1233 static int ipoib_get_iflink(const struct net_device *dev)
1234 {
1235         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1236
1237         /* parent interface */
1238         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags))
1239                 return dev->ifindex;
1240
1241         /* child/vlan interface */
1242         return priv->parent->ifindex;
1243 }
1244
1245 static u32 ipoib_addr_hash(struct ipoib_neigh_hash *htbl, u8 *daddr)
1246 {
1247         /*
1248          * Use only the address parts that contributes to spreading
1249          * The subnet prefix is not used as one can not connect to
1250          * same remote port (GUID) using the same remote QPN via two
1251          * different subnets.
1252          */
1253          /* qpn octets[1:4) & port GUID octets[12:20) */
1254         u32 *d32 = (u32 *) daddr;
1255         u32 hv;
1256
1257         hv = jhash_3words(d32[3], d32[4], IPOIB_QPN_MASK & d32[0], 0);
1258         return hv & htbl->mask;
1259 }
1260
1261 struct ipoib_neigh *ipoib_neigh_get(struct net_device *dev, u8 *daddr)
1262 {
1263         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1264         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1265         struct ipoib_neigh_hash *htbl;
1266         struct ipoib_neigh *neigh = NULL;
1267         u32 hash_val;
1268
1269         rcu_read_lock_bh();
1270
1271         htbl = rcu_dereference_bh(ntbl->htbl);
1272
1273         if (!htbl)
1274                 goto out_unlock;
1275
1276         hash_val = ipoib_addr_hash(htbl, daddr);
1277         for (neigh = rcu_dereference_bh(htbl->buckets[hash_val]);
1278              neigh != NULL;
1279              neigh = rcu_dereference_bh(neigh->hnext)) {
1280                 if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) {
1281                         /* found, take one ref on behalf of the caller */
1282                         if (!atomic_inc_not_zero(&neigh->refcnt)) {
1283                                 /* deleted */
1284                                 neigh = NULL;
1285                                 goto out_unlock;
1286                         }
1287
1288                         if (likely(skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE))
1289                                 neigh->alive = jiffies;
1290                         goto out_unlock;
1291                 }
1292         }
1293
1294 out_unlock:
1295         rcu_read_unlock_bh();
1296         return neigh;
1297 }
1298
1299 static void __ipoib_reap_neigh(struct ipoib_dev_priv *priv)
1300 {
1301         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1302         struct ipoib_neigh_hash *htbl;
1303         unsigned long neigh_obsolete;
1304         unsigned long dt;
1305         unsigned long flags;
1306         int i;
1307         LIST_HEAD(remove_list);
1308
1309         spin_lock_irqsave(&priv->lock, flags);
1310
1311         htbl = rcu_dereference_protected(ntbl->htbl,
1312                                          lockdep_is_held(&priv->lock));
1313
1314         if (!htbl)
1315                 goto out_unlock;
1316
1317         /* neigh is obsolete if it was idle for two GC periods */
1318         dt = 2 * arp_tbl.gc_interval;
1319         neigh_obsolete = jiffies - dt;
1320
1321         for (i = 0; i < htbl->size; i++) {
1322                 struct ipoib_neigh *neigh;
1323                 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1324
1325                 while ((neigh = rcu_dereference_protected(*np,
1326                                                           lockdep_is_held(&priv->lock))) != NULL) {
1327                         /* was the neigh idle for two GC periods */
1328                         if (time_after(neigh_obsolete, neigh->alive)) {
1329
1330                                 ipoib_check_and_add_mcast_sendonly(priv, neigh->daddr + 4, &remove_list);
1331
1332                                 rcu_assign_pointer(*np,
1333                                                    rcu_dereference_protected(neigh->hnext,
1334                                                                              lockdep_is_held(&priv->lock)));
1335                                 /* remove from path/mc list */
1336                                 list_del_init(&neigh->list);
1337                                 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1338                         } else {
1339                                 np = &neigh->hnext;
1340                         }
1341
1342                 }
1343         }
1344
1345 out_unlock:
1346         spin_unlock_irqrestore(&priv->lock, flags);
1347         ipoib_mcast_remove_list(&remove_list);
1348 }
1349
1350 static void ipoib_reap_neigh(struct work_struct *work)
1351 {
1352         struct ipoib_dev_priv *priv =
1353                 container_of(work, struct ipoib_dev_priv, neigh_reap_task.work);
1354
1355         __ipoib_reap_neigh(priv);
1356
1357         queue_delayed_work(priv->wq, &priv->neigh_reap_task,
1358                            arp_tbl.gc_interval);
1359 }
1360
1361
1362 static struct ipoib_neigh *ipoib_neigh_ctor(u8 *daddr,
1363                                       struct net_device *dev)
1364 {
1365         struct ipoib_neigh *neigh;
1366
1367         neigh = kzalloc(sizeof(*neigh), GFP_ATOMIC);
1368         if (!neigh)
1369                 return NULL;
1370
1371         neigh->dev = dev;
1372         memcpy(&neigh->daddr, daddr, sizeof(neigh->daddr));
1373         skb_queue_head_init(&neigh->queue);
1374         INIT_LIST_HEAD(&neigh->list);
1375         ipoib_cm_set(neigh, NULL);
1376         /* one ref on behalf of the caller */
1377         atomic_set(&neigh->refcnt, 1);
1378
1379         return neigh;
1380 }
1381
1382 struct ipoib_neigh *ipoib_neigh_alloc(u8 *daddr,
1383                                       struct net_device *dev)
1384 {
1385         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1386         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1387         struct ipoib_neigh_hash *htbl;
1388         struct ipoib_neigh *neigh;
1389         u32 hash_val;
1390
1391         htbl = rcu_dereference_protected(ntbl->htbl,
1392                                          lockdep_is_held(&priv->lock));
1393         if (!htbl) {
1394                 neigh = NULL;
1395                 goto out_unlock;
1396         }
1397
1398         /* need to add a new neigh, but maybe some other thread succeeded?
1399          * recalc hash, maybe hash resize took place so we do a search
1400          */
1401         hash_val = ipoib_addr_hash(htbl, daddr);
1402         for (neigh = rcu_dereference_protected(htbl->buckets[hash_val],
1403                                                lockdep_is_held(&priv->lock));
1404              neigh != NULL;
1405              neigh = rcu_dereference_protected(neigh->hnext,
1406                                                lockdep_is_held(&priv->lock))) {
1407                 if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) {
1408                         /* found, take one ref on behalf of the caller */
1409                         if (!atomic_inc_not_zero(&neigh->refcnt)) {
1410                                 /* deleted */
1411                                 neigh = NULL;
1412                                 break;
1413                         }
1414                         neigh->alive = jiffies;
1415                         goto out_unlock;
1416                 }
1417         }
1418
1419         neigh = ipoib_neigh_ctor(daddr, dev);
1420         if (!neigh)
1421                 goto out_unlock;
1422
1423         /* one ref on behalf of the hash table */
1424         atomic_inc(&neigh->refcnt);
1425         neigh->alive = jiffies;
1426         /* put in hash */
1427         rcu_assign_pointer(neigh->hnext,
1428                            rcu_dereference_protected(htbl->buckets[hash_val],
1429                                                      lockdep_is_held(&priv->lock)));
1430         rcu_assign_pointer(htbl->buckets[hash_val], neigh);
1431         atomic_inc(&ntbl->entries);
1432
1433 out_unlock:
1434
1435         return neigh;
1436 }
1437
1438 void ipoib_neigh_dtor(struct ipoib_neigh *neigh)
1439 {
1440         /* neigh reference count was dropprd to zero */
1441         struct net_device *dev = neigh->dev;
1442         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1443         struct sk_buff *skb;
1444         if (neigh->ah)
1445                 ipoib_put_ah(neigh->ah);
1446         while ((skb = __skb_dequeue(&neigh->queue))) {
1447                 ++dev->stats.tx_dropped;
1448                 dev_kfree_skb_any(skb);
1449         }
1450         if (ipoib_cm_get(neigh))
1451                 ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
1452         ipoib_dbg(ipoib_priv(dev),
1453                   "neigh free for %06x %pI6\n",
1454                   IPOIB_QPN(neigh->daddr),
1455                   neigh->daddr + 4);
1456         kfree(neigh);
1457         if (atomic_dec_and_test(&priv->ntbl.entries)) {
1458                 if (test_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags))
1459                         complete(&priv->ntbl.flushed);
1460         }
1461 }
1462
1463 static void ipoib_neigh_reclaim(struct rcu_head *rp)
1464 {
1465         /* Called as a result of removal from hash table */
1466         struct ipoib_neigh *neigh = container_of(rp, struct ipoib_neigh, rcu);
1467         /* note TX context may hold another ref */
1468         ipoib_neigh_put(neigh);
1469 }
1470
1471 void ipoib_neigh_free(struct ipoib_neigh *neigh)
1472 {
1473         struct net_device *dev = neigh->dev;
1474         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1475         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1476         struct ipoib_neigh_hash *htbl;
1477         struct ipoib_neigh __rcu **np;
1478         struct ipoib_neigh *n;
1479         u32 hash_val;
1480
1481         htbl = rcu_dereference_protected(ntbl->htbl,
1482                                         lockdep_is_held(&priv->lock));
1483         if (!htbl)
1484                 return;
1485
1486         hash_val = ipoib_addr_hash(htbl, neigh->daddr);
1487         np = &htbl->buckets[hash_val];
1488         for (n = rcu_dereference_protected(*np,
1489                                             lockdep_is_held(&priv->lock));
1490              n != NULL;
1491              n = rcu_dereference_protected(*np,
1492                                         lockdep_is_held(&priv->lock))) {
1493                 if (n == neigh) {
1494                         /* found */
1495                         rcu_assign_pointer(*np,
1496                                            rcu_dereference_protected(neigh->hnext,
1497                                                                      lockdep_is_held(&priv->lock)));
1498                         /* remove from parent list */
1499                         list_del_init(&neigh->list);
1500                         call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1501                         return;
1502                 } else {
1503                         np = &n->hnext;
1504                 }
1505         }
1506 }
1507
1508 static int ipoib_neigh_hash_init(struct ipoib_dev_priv *priv)
1509 {
1510         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1511         struct ipoib_neigh_hash *htbl;
1512         struct ipoib_neigh __rcu **buckets;
1513         u32 size;
1514
1515         clear_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags);
1516         ntbl->htbl = NULL;
1517         htbl = kzalloc(sizeof(*htbl), GFP_KERNEL);
1518         if (!htbl)
1519                 return -ENOMEM;
1520         size = roundup_pow_of_two(arp_tbl.gc_thresh3);
1521         buckets = kvcalloc(size, sizeof(*buckets), GFP_KERNEL);
1522         if (!buckets) {
1523                 kfree(htbl);
1524                 return -ENOMEM;
1525         }
1526         htbl->size = size;
1527         htbl->mask = (size - 1);
1528         htbl->buckets = buckets;
1529         RCU_INIT_POINTER(ntbl->htbl, htbl);
1530         htbl->ntbl = ntbl;
1531         atomic_set(&ntbl->entries, 0);
1532
1533         /* start garbage collection */
1534         queue_delayed_work(priv->wq, &priv->neigh_reap_task,
1535                            arp_tbl.gc_interval);
1536
1537         return 0;
1538 }
1539
1540 static void neigh_hash_free_rcu(struct rcu_head *head)
1541 {
1542         struct ipoib_neigh_hash *htbl = container_of(head,
1543                                                     struct ipoib_neigh_hash,
1544                                                     rcu);
1545         struct ipoib_neigh __rcu **buckets = htbl->buckets;
1546         struct ipoib_neigh_table *ntbl = htbl->ntbl;
1547
1548         kvfree(buckets);
1549         kfree(htbl);
1550         complete(&ntbl->deleted);
1551 }
1552
1553 void ipoib_del_neighs_by_gid(struct net_device *dev, u8 *gid)
1554 {
1555         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1556         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1557         struct ipoib_neigh_hash *htbl;
1558         unsigned long flags;
1559         int i;
1560
1561         /* remove all neigh connected to a given path or mcast */
1562         spin_lock_irqsave(&priv->lock, flags);
1563
1564         htbl = rcu_dereference_protected(ntbl->htbl,
1565                                          lockdep_is_held(&priv->lock));
1566
1567         if (!htbl)
1568                 goto out_unlock;
1569
1570         for (i = 0; i < htbl->size; i++) {
1571                 struct ipoib_neigh *neigh;
1572                 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1573
1574                 while ((neigh = rcu_dereference_protected(*np,
1575                                                           lockdep_is_held(&priv->lock))) != NULL) {
1576                         /* delete neighs belong to this parent */
1577                         if (!memcmp(gid, neigh->daddr + 4, sizeof (union ib_gid))) {
1578                                 rcu_assign_pointer(*np,
1579                                                    rcu_dereference_protected(neigh->hnext,
1580                                                                              lockdep_is_held(&priv->lock)));
1581                                 /* remove from parent list */
1582                                 list_del_init(&neigh->list);
1583                                 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1584                         } else {
1585                                 np = &neigh->hnext;
1586                         }
1587
1588                 }
1589         }
1590 out_unlock:
1591         spin_unlock_irqrestore(&priv->lock, flags);
1592 }
1593
1594 static void ipoib_flush_neighs(struct ipoib_dev_priv *priv)
1595 {
1596         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1597         struct ipoib_neigh_hash *htbl;
1598         unsigned long flags;
1599         int i, wait_flushed = 0;
1600
1601         init_completion(&priv->ntbl.flushed);
1602         set_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags);
1603
1604         spin_lock_irqsave(&priv->lock, flags);
1605
1606         htbl = rcu_dereference_protected(ntbl->htbl,
1607                                         lockdep_is_held(&priv->lock));
1608         if (!htbl)
1609                 goto out_unlock;
1610
1611         wait_flushed = atomic_read(&priv->ntbl.entries);
1612         if (!wait_flushed)
1613                 goto free_htbl;
1614
1615         for (i = 0; i < htbl->size; i++) {
1616                 struct ipoib_neigh *neigh;
1617                 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1618
1619                 while ((neigh = rcu_dereference_protected(*np,
1620                                        lockdep_is_held(&priv->lock))) != NULL) {
1621                         rcu_assign_pointer(*np,
1622                                            rcu_dereference_protected(neigh->hnext,
1623                                                                      lockdep_is_held(&priv->lock)));
1624                         /* remove from path/mc list */
1625                         list_del_init(&neigh->list);
1626                         call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1627                 }
1628         }
1629
1630 free_htbl:
1631         rcu_assign_pointer(ntbl->htbl, NULL);
1632         call_rcu(&htbl->rcu, neigh_hash_free_rcu);
1633
1634 out_unlock:
1635         spin_unlock_irqrestore(&priv->lock, flags);
1636         if (wait_flushed)
1637                 wait_for_completion(&priv->ntbl.flushed);
1638 }
1639
1640 static void ipoib_neigh_hash_uninit(struct net_device *dev)
1641 {
1642         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1643
1644         ipoib_dbg(priv, "%s\n", __func__);
1645         init_completion(&priv->ntbl.deleted);
1646
1647         cancel_delayed_work_sync(&priv->neigh_reap_task);
1648
1649         ipoib_flush_neighs(priv);
1650
1651         wait_for_completion(&priv->ntbl.deleted);
1652 }
1653
1654 static void ipoib_napi_add(struct net_device *dev)
1655 {
1656         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1657
1658         netif_napi_add(dev, &priv->recv_napi, ipoib_rx_poll, IPOIB_NUM_WC);
1659         netif_napi_add(dev, &priv->send_napi, ipoib_tx_poll, MAX_SEND_CQE);
1660 }
1661
1662 static void ipoib_napi_del(struct net_device *dev)
1663 {
1664         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1665
1666         netif_napi_del(&priv->recv_napi);
1667         netif_napi_del(&priv->send_napi);
1668 }
1669
1670 static void ipoib_dev_uninit_default(struct net_device *dev)
1671 {
1672         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1673
1674         ipoib_transport_dev_cleanup(dev);
1675
1676         ipoib_napi_del(dev);
1677
1678         ipoib_cm_dev_cleanup(dev);
1679
1680         kfree(priv->rx_ring);
1681         vfree(priv->tx_ring);
1682
1683         priv->rx_ring = NULL;
1684         priv->tx_ring = NULL;
1685 }
1686
1687 static int ipoib_dev_init_default(struct net_device *dev)
1688 {
1689         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1690
1691         ipoib_napi_add(dev);
1692
1693         /* Allocate RX/TX "rings" to hold queued skbs */
1694         priv->rx_ring = kcalloc(ipoib_recvq_size,
1695                                        sizeof(*priv->rx_ring),
1696                                        GFP_KERNEL);
1697         if (!priv->rx_ring)
1698                 goto out;
1699
1700         priv->tx_ring = vzalloc(array_size(ipoib_sendq_size,
1701                                            sizeof(*priv->tx_ring)));
1702         if (!priv->tx_ring) {
1703                 pr_warn("%s: failed to allocate TX ring (%d entries)\n",
1704                         priv->ca->name, ipoib_sendq_size);
1705                 goto out_rx_ring_cleanup;
1706         }
1707
1708         /* priv->tx_head, tx_tail and global_tx_tail/head are already 0 */
1709
1710         if (ipoib_transport_dev_init(dev, priv->ca)) {
1711                 pr_warn("%s: ipoib_transport_dev_init failed\n",
1712                         priv->ca->name);
1713                 goto out_tx_ring_cleanup;
1714         }
1715
1716         /* after qp created set dev address */
1717         priv->dev->dev_addr[1] = (priv->qp->qp_num >> 16) & 0xff;
1718         priv->dev->dev_addr[2] = (priv->qp->qp_num >>  8) & 0xff;
1719         priv->dev->dev_addr[3] = (priv->qp->qp_num) & 0xff;
1720
1721         return 0;
1722
1723 out_tx_ring_cleanup:
1724         vfree(priv->tx_ring);
1725
1726 out_rx_ring_cleanup:
1727         kfree(priv->rx_ring);
1728
1729 out:
1730         ipoib_napi_del(dev);
1731         return -ENOMEM;
1732 }
1733
1734 static int ipoib_ioctl(struct net_device *dev, struct ifreq *ifr,
1735                        int cmd)
1736 {
1737         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1738
1739         if (!priv->rn_ops->ndo_do_ioctl)
1740                 return -EOPNOTSUPP;
1741
1742         return priv->rn_ops->ndo_do_ioctl(dev, ifr, cmd);
1743 }
1744
1745 static int ipoib_dev_init(struct net_device *dev)
1746 {
1747         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1748         int ret = -ENOMEM;
1749
1750         priv->qp = NULL;
1751
1752         /*
1753          * the various IPoIB tasks assume they will never race against
1754          * themselves, so always use a single thread workqueue
1755          */
1756         priv->wq = alloc_ordered_workqueue("ipoib_wq", WQ_MEM_RECLAIM);
1757         if (!priv->wq) {
1758                 pr_warn("%s: failed to allocate device WQ\n", dev->name);
1759                 goto out;
1760         }
1761
1762         /* create pd, which used both for control and datapath*/
1763         priv->pd = ib_alloc_pd(priv->ca, 0);
1764         if (IS_ERR(priv->pd)) {
1765                 pr_warn("%s: failed to allocate PD\n", priv->ca->name);
1766                 goto clean_wq;
1767         }
1768
1769         ret = priv->rn_ops->ndo_init(dev);
1770         if (ret) {
1771                 pr_warn("%s failed to init HW resource\n", dev->name);
1772                 goto out_free_pd;
1773         }
1774
1775         ret = ipoib_neigh_hash_init(priv);
1776         if (ret) {
1777                 pr_warn("%s failed to init neigh hash\n", dev->name);
1778                 goto out_dev_uninit;
1779         }
1780
1781         if (dev->flags & IFF_UP) {
1782                 if (ipoib_ib_dev_open(dev)) {
1783                         pr_warn("%s failed to open device\n", dev->name);
1784                         ret = -ENODEV;
1785                         goto out_hash_uninit;
1786                 }
1787         }
1788
1789         return 0;
1790
1791 out_hash_uninit:
1792         ipoib_neigh_hash_uninit(dev);
1793
1794 out_dev_uninit:
1795         ipoib_ib_dev_cleanup(dev);
1796
1797 out_free_pd:
1798         if (priv->pd) {
1799                 ib_dealloc_pd(priv->pd);
1800                 priv->pd = NULL;
1801         }
1802
1803 clean_wq:
1804         if (priv->wq) {
1805                 destroy_workqueue(priv->wq);
1806                 priv->wq = NULL;
1807         }
1808
1809 out:
1810         return ret;
1811 }
1812
1813 /*
1814  * This must be called before doing an unregister_netdev on a parent device to
1815  * shutdown the IB event handler.
1816  */
1817 static void ipoib_parent_unregister_pre(struct net_device *ndev)
1818 {
1819         struct ipoib_dev_priv *priv = ipoib_priv(ndev);
1820
1821         /*
1822          * ipoib_set_mac checks netif_running before pushing work, clearing
1823          * running ensures the it will not add more work.
1824          */
1825         rtnl_lock();
1826         dev_change_flags(priv->dev, priv->dev->flags & ~IFF_UP, NULL);
1827         rtnl_unlock();
1828
1829         /* ipoib_event() cannot be running once this returns */
1830         ib_unregister_event_handler(&priv->event_handler);
1831
1832         /*
1833          * Work on the queue grabs the rtnl lock, so this cannot be done while
1834          * also holding it.
1835          */
1836         flush_workqueue(ipoib_workqueue);
1837 }
1838
1839 static void ipoib_set_dev_features(struct ipoib_dev_priv *priv)
1840 {
1841         priv->hca_caps = priv->ca->attrs.device_cap_flags;
1842
1843         if (priv->hca_caps & IB_DEVICE_UD_IP_CSUM) {
1844                 priv->dev->hw_features |= NETIF_F_IP_CSUM | NETIF_F_RXCSUM;
1845
1846                 if (priv->hca_caps & IB_DEVICE_UD_TSO)
1847                         priv->dev->hw_features |= NETIF_F_TSO;
1848
1849                 priv->dev->features |= priv->dev->hw_features;
1850         }
1851 }
1852
1853 static int ipoib_parent_init(struct net_device *ndev)
1854 {
1855         struct ipoib_dev_priv *priv = ipoib_priv(ndev);
1856         struct ib_port_attr attr;
1857         int result;
1858
1859         result = ib_query_port(priv->ca, priv->port, &attr);
1860         if (result) {
1861                 pr_warn("%s: ib_query_port %d failed\n", priv->ca->name,
1862                         priv->port);
1863                 return result;
1864         }
1865         priv->max_ib_mtu = rdma_mtu_from_attr(priv->ca, priv->port, &attr);
1866
1867         result = ib_query_pkey(priv->ca, priv->port, 0, &priv->pkey);
1868         if (result) {
1869                 pr_warn("%s: ib_query_pkey port %d failed (ret = %d)\n",
1870                         priv->ca->name, priv->port, result);
1871                 return result;
1872         }
1873
1874         result = rdma_query_gid(priv->ca, priv->port, 0, &priv->local_gid);
1875         if (result) {
1876                 pr_warn("%s: rdma_query_gid port %d failed (ret = %d)\n",
1877                         priv->ca->name, priv->port, result);
1878                 return result;
1879         }
1880         memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw,
1881                sizeof(union ib_gid));
1882
1883         SET_NETDEV_DEV(priv->dev, priv->ca->dev.parent);
1884         priv->dev->dev_port = priv->port - 1;
1885         /* Let's set this one too for backwards compatibility. */
1886         priv->dev->dev_id = priv->port - 1;
1887
1888         return 0;
1889 }
1890
1891 static void ipoib_child_init(struct net_device *ndev)
1892 {
1893         struct ipoib_dev_priv *priv = ipoib_priv(ndev);
1894         struct ipoib_dev_priv *ppriv = ipoib_priv(priv->parent);
1895
1896         priv->max_ib_mtu = ppriv->max_ib_mtu;
1897         set_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags);
1898         if (memchr_inv(priv->dev->dev_addr, 0, INFINIBAND_ALEN))
1899                 memcpy(&priv->local_gid, priv->dev->dev_addr + 4,
1900                        sizeof(priv->local_gid));
1901         else {
1902                 memcpy(priv->dev->dev_addr, ppriv->dev->dev_addr,
1903                        INFINIBAND_ALEN);
1904                 memcpy(&priv->local_gid, &ppriv->local_gid,
1905                        sizeof(priv->local_gid));
1906         }
1907 }
1908
1909 static int ipoib_ndo_init(struct net_device *ndev)
1910 {
1911         struct ipoib_dev_priv *priv = ipoib_priv(ndev);
1912         int rc;
1913         struct rdma_netdev *rn = netdev_priv(ndev);
1914
1915         if (priv->parent) {
1916                 ipoib_child_init(ndev);
1917         } else {
1918                 rc = ipoib_parent_init(ndev);
1919                 if (rc)
1920                         return rc;
1921         }
1922
1923         /* MTU will be reset when mcast join happens */
1924         ndev->mtu = IPOIB_UD_MTU(priv->max_ib_mtu);
1925         priv->mcast_mtu = priv->admin_mtu = ndev->mtu;
1926         rn->mtu = priv->mcast_mtu;
1927         ndev->max_mtu = IPOIB_CM_MTU;
1928
1929         ndev->neigh_priv_len = sizeof(struct ipoib_neigh);
1930
1931         /*
1932          * Set the full membership bit, so that we join the right
1933          * broadcast group, etc.
1934          */
1935         priv->pkey |= 0x8000;
1936
1937         ndev->broadcast[8] = priv->pkey >> 8;
1938         ndev->broadcast[9] = priv->pkey & 0xff;
1939         set_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags);
1940
1941         ipoib_set_dev_features(priv);
1942
1943         rc = ipoib_dev_init(ndev);
1944         if (rc) {
1945                 pr_warn("%s: failed to initialize device: %s port %d (ret = %d)\n",
1946                         priv->ca->name, priv->dev->name, priv->port, rc);
1947                 return rc;
1948         }
1949
1950         if (priv->parent) {
1951                 struct ipoib_dev_priv *ppriv = ipoib_priv(priv->parent);
1952
1953                 dev_hold(priv->parent);
1954
1955                 down_write(&ppriv->vlan_rwsem);
1956                 list_add_tail(&priv->list, &ppriv->child_intfs);
1957                 up_write(&ppriv->vlan_rwsem);
1958         }
1959
1960         return 0;
1961 }
1962
1963 static void ipoib_ndo_uninit(struct net_device *dev)
1964 {
1965         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1966
1967         ASSERT_RTNL();
1968
1969         /*
1970          * ipoib_remove_one guarantees the children are removed before the
1971          * parent, and that is the only place where a parent can be removed.
1972          */
1973         WARN_ON(!list_empty(&priv->child_intfs));
1974
1975         if (priv->parent) {
1976                 struct ipoib_dev_priv *ppriv = ipoib_priv(priv->parent);
1977
1978                 down_write(&ppriv->vlan_rwsem);
1979                 list_del(&priv->list);
1980                 up_write(&ppriv->vlan_rwsem);
1981         }
1982
1983         ipoib_neigh_hash_uninit(dev);
1984
1985         ipoib_ib_dev_cleanup(dev);
1986
1987         /* no more works over the priv->wq */
1988         if (priv->wq) {
1989                 /* See ipoib_mcast_carrier_on_task() */
1990                 WARN_ON(test_bit(IPOIB_FLAG_OPER_UP, &priv->flags));
1991                 flush_workqueue(priv->wq);
1992                 destroy_workqueue(priv->wq);
1993                 priv->wq = NULL;
1994         }
1995
1996         if (priv->parent)
1997                 dev_put(priv->parent);
1998 }
1999
2000 static int ipoib_set_vf_link_state(struct net_device *dev, int vf, int link_state)
2001 {
2002         struct ipoib_dev_priv *priv = ipoib_priv(dev);
2003
2004         return ib_set_vf_link_state(priv->ca, vf, priv->port, link_state);
2005 }
2006
2007 static int ipoib_get_vf_config(struct net_device *dev, int vf,
2008                                struct ifla_vf_info *ivf)
2009 {
2010         struct ipoib_dev_priv *priv = ipoib_priv(dev);
2011         int err;
2012
2013         err = ib_get_vf_config(priv->ca, vf, priv->port, ivf);
2014         if (err)
2015                 return err;
2016
2017         ivf->vf = vf;
2018         memcpy(ivf->mac, dev->dev_addr, dev->addr_len);
2019
2020         return 0;
2021 }
2022
2023 static int ipoib_set_vf_guid(struct net_device *dev, int vf, u64 guid, int type)
2024 {
2025         struct ipoib_dev_priv *priv = ipoib_priv(dev);
2026
2027         if (type != IFLA_VF_IB_NODE_GUID && type != IFLA_VF_IB_PORT_GUID)
2028                 return -EINVAL;
2029
2030         return ib_set_vf_guid(priv->ca, vf, priv->port, guid, type);
2031 }
2032
2033 static int ipoib_get_vf_guid(struct net_device *dev, int vf,
2034                              struct ifla_vf_guid *node_guid,
2035                              struct ifla_vf_guid *port_guid)
2036 {
2037         struct ipoib_dev_priv *priv = ipoib_priv(dev);
2038
2039         return ib_get_vf_guid(priv->ca, vf, priv->port, node_guid, port_guid);
2040 }
2041
2042 static int ipoib_get_vf_stats(struct net_device *dev, int vf,
2043                               struct ifla_vf_stats *vf_stats)
2044 {
2045         struct ipoib_dev_priv *priv = ipoib_priv(dev);
2046
2047         return ib_get_vf_stats(priv->ca, vf, priv->port, vf_stats);
2048 }
2049
2050 static const struct header_ops ipoib_header_ops = {
2051         .create = ipoib_hard_header,
2052 };
2053
2054 static const struct net_device_ops ipoib_netdev_ops_pf = {
2055         .ndo_init                = ipoib_ndo_init,
2056         .ndo_uninit              = ipoib_ndo_uninit,
2057         .ndo_open                = ipoib_open,
2058         .ndo_stop                = ipoib_stop,
2059         .ndo_change_mtu          = ipoib_change_mtu,
2060         .ndo_fix_features        = ipoib_fix_features,
2061         .ndo_start_xmit          = ipoib_start_xmit,
2062         .ndo_tx_timeout          = ipoib_timeout,
2063         .ndo_set_rx_mode         = ipoib_set_mcast_list,
2064         .ndo_get_iflink          = ipoib_get_iflink,
2065         .ndo_set_vf_link_state   = ipoib_set_vf_link_state,
2066         .ndo_get_vf_config       = ipoib_get_vf_config,
2067         .ndo_get_vf_stats        = ipoib_get_vf_stats,
2068         .ndo_get_vf_guid         = ipoib_get_vf_guid,
2069         .ndo_set_vf_guid         = ipoib_set_vf_guid,
2070         .ndo_set_mac_address     = ipoib_set_mac,
2071         .ndo_get_stats64         = ipoib_get_stats,
2072         .ndo_do_ioctl            = ipoib_ioctl,
2073 };
2074
2075 static const struct net_device_ops ipoib_netdev_ops_vf = {
2076         .ndo_init                = ipoib_ndo_init,
2077         .ndo_uninit              = ipoib_ndo_uninit,
2078         .ndo_open                = ipoib_open,
2079         .ndo_stop                = ipoib_stop,
2080         .ndo_change_mtu          = ipoib_change_mtu,
2081         .ndo_fix_features        = ipoib_fix_features,
2082         .ndo_start_xmit          = ipoib_start_xmit,
2083         .ndo_tx_timeout          = ipoib_timeout,
2084         .ndo_set_rx_mode         = ipoib_set_mcast_list,
2085         .ndo_get_iflink          = ipoib_get_iflink,
2086         .ndo_get_stats64         = ipoib_get_stats,
2087         .ndo_do_ioctl            = ipoib_ioctl,
2088 };
2089
2090 static const struct net_device_ops ipoib_netdev_default_pf = {
2091         .ndo_init                = ipoib_dev_init_default,
2092         .ndo_uninit              = ipoib_dev_uninit_default,
2093         .ndo_open                = ipoib_ib_dev_open_default,
2094         .ndo_stop                = ipoib_ib_dev_stop_default,
2095 };
2096
2097 void ipoib_setup_common(struct net_device *dev)
2098 {
2099         dev->header_ops          = &ipoib_header_ops;
2100         dev->netdev_ops          = &ipoib_netdev_default_pf;
2101
2102         ipoib_set_ethtool_ops(dev);
2103
2104         dev->watchdog_timeo      = HZ;
2105
2106         dev->flags              |= IFF_BROADCAST | IFF_MULTICAST;
2107
2108         dev->hard_header_len     = IPOIB_HARD_LEN;
2109         dev->addr_len            = INFINIBAND_ALEN;
2110         dev->type                = ARPHRD_INFINIBAND;
2111         dev->tx_queue_len        = ipoib_sendq_size * 2;
2112         dev->features            = (NETIF_F_VLAN_CHALLENGED     |
2113                                     NETIF_F_HIGHDMA);
2114         netif_keep_dst(dev);
2115
2116         memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
2117
2118         /*
2119          * unregister_netdev always frees the netdev, we use this mode
2120          * consistently to unify all the various unregister paths, including
2121          * those connected to rtnl_link_ops which require it.
2122          */
2123         dev->needs_free_netdev = true;
2124 }
2125
2126 static void ipoib_build_priv(struct net_device *dev)
2127 {
2128         struct ipoib_dev_priv *priv = ipoib_priv(dev);
2129
2130         priv->dev = dev;
2131         spin_lock_init(&priv->lock);
2132         init_rwsem(&priv->vlan_rwsem);
2133         mutex_init(&priv->mcast_mutex);
2134
2135         INIT_LIST_HEAD(&priv->path_list);
2136         INIT_LIST_HEAD(&priv->child_intfs);
2137         INIT_LIST_HEAD(&priv->dead_ahs);
2138         INIT_LIST_HEAD(&priv->multicast_list);
2139
2140         INIT_DELAYED_WORK(&priv->mcast_task,   ipoib_mcast_join_task);
2141         INIT_WORK(&priv->carrier_on_task, ipoib_mcast_carrier_on_task);
2142         INIT_WORK(&priv->flush_light,   ipoib_ib_dev_flush_light);
2143         INIT_WORK(&priv->flush_normal,   ipoib_ib_dev_flush_normal);
2144         INIT_WORK(&priv->flush_heavy,   ipoib_ib_dev_flush_heavy);
2145         INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
2146         INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
2147         INIT_DELAYED_WORK(&priv->neigh_reap_task, ipoib_reap_neigh);
2148 }
2149
2150 static struct net_device *ipoib_alloc_netdev(struct ib_device *hca, u8 port,
2151                                              const char *name)
2152 {
2153         struct net_device *dev;
2154
2155         dev = rdma_alloc_netdev(hca, port, RDMA_NETDEV_IPOIB, name,
2156                                 NET_NAME_UNKNOWN, ipoib_setup_common);
2157         if (!IS_ERR(dev) || PTR_ERR(dev) != -EOPNOTSUPP)
2158                 return dev;
2159
2160         dev = alloc_netdev(sizeof(struct rdma_netdev), name, NET_NAME_UNKNOWN,
2161                            ipoib_setup_common);
2162         if (!dev)
2163                 return ERR_PTR(-ENOMEM);
2164         return dev;
2165 }
2166
2167 int ipoib_intf_init(struct ib_device *hca, u8 port, const char *name,
2168                     struct net_device *dev)
2169 {
2170         struct rdma_netdev *rn = netdev_priv(dev);
2171         struct ipoib_dev_priv *priv;
2172         int rc;
2173
2174         priv = kzalloc(sizeof(*priv), GFP_KERNEL);
2175         if (!priv)
2176                 return -ENOMEM;
2177
2178         priv->ca = hca;
2179         priv->port = port;
2180
2181         rc = rdma_init_netdev(hca, port, RDMA_NETDEV_IPOIB, name,
2182                               NET_NAME_UNKNOWN, ipoib_setup_common, dev);
2183         if (rc) {
2184                 if (rc != -EOPNOTSUPP)
2185                         goto out;
2186
2187                 rn->send = ipoib_send;
2188                 rn->attach_mcast = ipoib_mcast_attach;
2189                 rn->detach_mcast = ipoib_mcast_detach;
2190                 rn->hca = hca;
2191         }
2192
2193         priv->rn_ops = dev->netdev_ops;
2194
2195         if (hca->attrs.device_cap_flags & IB_DEVICE_VIRTUAL_FUNCTION)
2196                 dev->netdev_ops = &ipoib_netdev_ops_vf;
2197         else
2198                 dev->netdev_ops = &ipoib_netdev_ops_pf;
2199
2200         rn->clnt_priv = priv;
2201         /*
2202          * Only the child register_netdev flows can handle priv_destructor
2203          * being set, so we force it to NULL here and handle manually until it
2204          * is safe to turn on.
2205          */
2206         priv->next_priv_destructor = dev->priv_destructor;
2207         dev->priv_destructor = NULL;
2208
2209         ipoib_build_priv(dev);
2210
2211         return 0;
2212
2213 out:
2214         kfree(priv);
2215         return rc;
2216 }
2217
2218 struct net_device *ipoib_intf_alloc(struct ib_device *hca, u8 port,
2219                                     const char *name)
2220 {
2221         struct net_device *dev;
2222         int rc;
2223
2224         dev = ipoib_alloc_netdev(hca, port, name);
2225         if (IS_ERR(dev))
2226                 return dev;
2227
2228         rc = ipoib_intf_init(hca, port, name, dev);
2229         if (rc) {
2230                 free_netdev(dev);
2231                 return ERR_PTR(rc);
2232         }
2233
2234         /*
2235          * Upon success the caller must ensure ipoib_intf_free is called or
2236          * register_netdevice succeed'd and priv_destructor is set to
2237          * ipoib_intf_free.
2238          */
2239         return dev;
2240 }
2241
2242 void ipoib_intf_free(struct net_device *dev)
2243 {
2244         struct ipoib_dev_priv *priv = ipoib_priv(dev);
2245         struct rdma_netdev *rn = netdev_priv(dev);
2246
2247         dev->priv_destructor = priv->next_priv_destructor;
2248         if (dev->priv_destructor)
2249                 dev->priv_destructor(dev);
2250
2251         /*
2252          * There are some error flows around register_netdev failing that may
2253          * attempt to call priv_destructor twice, prevent that from happening.
2254          */
2255         dev->priv_destructor = NULL;
2256
2257         /* unregister/destroy is very complicated. Make bugs more obvious. */
2258         rn->clnt_priv = NULL;
2259
2260         kfree(priv);
2261 }
2262
2263 static ssize_t show_pkey(struct device *dev,
2264                          struct device_attribute *attr, char *buf)
2265 {
2266         struct net_device *ndev = to_net_dev(dev);
2267         struct ipoib_dev_priv *priv = ipoib_priv(ndev);
2268
2269         return sprintf(buf, "0x%04x\n", priv->pkey);
2270 }
2271 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
2272
2273 static ssize_t show_umcast(struct device *dev,
2274                            struct device_attribute *attr, char *buf)
2275 {
2276         struct net_device *ndev = to_net_dev(dev);
2277         struct ipoib_dev_priv *priv = ipoib_priv(ndev);
2278
2279         return sprintf(buf, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
2280 }
2281
2282 void ipoib_set_umcast(struct net_device *ndev, int umcast_val)
2283 {
2284         struct ipoib_dev_priv *priv = ipoib_priv(ndev);
2285
2286         if (umcast_val > 0) {
2287                 set_bit(IPOIB_FLAG_UMCAST, &priv->flags);
2288                 ipoib_warn(priv, "ignoring multicast groups joined directly "
2289                                 "by userspace\n");
2290         } else
2291                 clear_bit(IPOIB_FLAG_UMCAST, &priv->flags);
2292 }
2293
2294 static ssize_t set_umcast(struct device *dev,
2295                           struct device_attribute *attr,
2296                           const char *buf, size_t count)
2297 {
2298         unsigned long umcast_val = simple_strtoul(buf, NULL, 0);
2299
2300         ipoib_set_umcast(to_net_dev(dev), umcast_val);
2301
2302         return count;
2303 }
2304 static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast);
2305
2306 int ipoib_add_umcast_attr(struct net_device *dev)
2307 {
2308         return device_create_file(&dev->dev, &dev_attr_umcast);
2309 }
2310
2311 static void set_base_guid(struct ipoib_dev_priv *priv, union ib_gid *gid)
2312 {
2313         struct ipoib_dev_priv *child_priv;
2314         struct net_device *netdev = priv->dev;
2315
2316         netif_addr_lock_bh(netdev);
2317
2318         memcpy(&priv->local_gid.global.interface_id,
2319                &gid->global.interface_id,
2320                sizeof(gid->global.interface_id));
2321         memcpy(netdev->dev_addr + 4, &priv->local_gid, sizeof(priv->local_gid));
2322         clear_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags);
2323
2324         netif_addr_unlock_bh(netdev);
2325
2326         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
2327                 down_read(&priv->vlan_rwsem);
2328                 list_for_each_entry(child_priv, &priv->child_intfs, list)
2329                         set_base_guid(child_priv, gid);
2330                 up_read(&priv->vlan_rwsem);
2331         }
2332 }
2333
2334 static int ipoib_check_lladdr(struct net_device *dev,
2335                               struct sockaddr_storage *ss)
2336 {
2337         union ib_gid *gid = (union ib_gid *)(ss->__data + 4);
2338         int ret = 0;
2339
2340         netif_addr_lock_bh(dev);
2341
2342         /* Make sure the QPN, reserved and subnet prefix match the current
2343          * lladdr, it also makes sure the lladdr is unicast.
2344          */
2345         if (memcmp(dev->dev_addr, ss->__data,
2346                    4 + sizeof(gid->global.subnet_prefix)) ||
2347             gid->global.interface_id == 0)
2348                 ret = -EINVAL;
2349
2350         netif_addr_unlock_bh(dev);
2351
2352         return ret;
2353 }
2354
2355 static int ipoib_set_mac(struct net_device *dev, void *addr)
2356 {
2357         struct ipoib_dev_priv *priv = ipoib_priv(dev);
2358         struct sockaddr_storage *ss = addr;
2359         int ret;
2360
2361         if (!(dev->priv_flags & IFF_LIVE_ADDR_CHANGE) && netif_running(dev))
2362                 return -EBUSY;
2363
2364         ret = ipoib_check_lladdr(dev, ss);
2365         if (ret)
2366                 return ret;
2367
2368         set_base_guid(priv, (union ib_gid *)(ss->__data + 4));
2369
2370         queue_work(ipoib_workqueue, &priv->flush_light);
2371
2372         return 0;
2373 }
2374
2375 static ssize_t create_child(struct device *dev,
2376                             struct device_attribute *attr,
2377                             const char *buf, size_t count)
2378 {
2379         int pkey;
2380         int ret;
2381
2382         if (sscanf(buf, "%i", &pkey) != 1)
2383                 return -EINVAL;
2384
2385         if (pkey <= 0 || pkey > 0xffff || pkey == 0x8000)
2386                 return -EINVAL;
2387
2388         ret = ipoib_vlan_add(to_net_dev(dev), pkey);
2389
2390         return ret ? ret : count;
2391 }
2392 static DEVICE_ATTR(create_child, S_IWUSR, NULL, create_child);
2393
2394 static ssize_t delete_child(struct device *dev,
2395                             struct device_attribute *attr,
2396                             const char *buf, size_t count)
2397 {
2398         int pkey;
2399         int ret;
2400
2401         if (sscanf(buf, "%i", &pkey) != 1)
2402                 return -EINVAL;
2403
2404         if (pkey < 0 || pkey > 0xffff)
2405                 return -EINVAL;
2406
2407         ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
2408
2409         return ret ? ret : count;
2410
2411 }
2412 static DEVICE_ATTR(delete_child, S_IWUSR, NULL, delete_child);
2413
2414 int ipoib_add_pkey_attr(struct net_device *dev)
2415 {
2416         return device_create_file(&dev->dev, &dev_attr_pkey);
2417 }
2418
2419 /*
2420  * We erroneously exposed the iface's port number in the dev_id
2421  * sysfs field long after dev_port was introduced for that purpose[1],
2422  * and we need to stop everyone from relying on that.
2423  * Let's overload the shower routine for the dev_id file here
2424  * to gently bring the issue up.
2425  *
2426  * [1] https://www.spinics.net/lists/netdev/msg272123.html
2427  */
2428 static ssize_t dev_id_show(struct device *dev,
2429                            struct device_attribute *attr, char *buf)
2430 {
2431         struct net_device *ndev = to_net_dev(dev);
2432
2433         /*
2434          * ndev->dev_port will be equal to 0 in old kernel prior to commit
2435          * 9b8b2a323008 ("IB/ipoib: Use dev_port to expose network interface
2436          * port numbers") Zero was chosen as special case for user space
2437          * applications to fallback and query dev_id to check if it has
2438          * different value or not.
2439          *
2440          * Don't print warning in such scenario.
2441          *
2442          * https://github.com/systemd/systemd/blob/master/src/udev/udev-builtin-net_id.c#L358
2443          */
2444         if (ndev->dev_port && ndev->dev_id == ndev->dev_port)
2445                 netdev_info_once(ndev,
2446                         "\"%s\" wants to know my dev_id. Should it look at dev_port instead? See Documentation/ABI/testing/sysfs-class-net for more info.\n",
2447                         current->comm);
2448
2449         return sprintf(buf, "%#x\n", ndev->dev_id);
2450 }
2451 static DEVICE_ATTR_RO(dev_id);
2452
2453 static int ipoib_intercept_dev_id_attr(struct net_device *dev)
2454 {
2455         device_remove_file(&dev->dev, &dev_attr_dev_id);
2456         return device_create_file(&dev->dev, &dev_attr_dev_id);
2457 }
2458
2459 static struct net_device *ipoib_add_port(const char *format,
2460                                          struct ib_device *hca, u8 port)
2461 {
2462         struct rtnl_link_ops *ops = ipoib_get_link_ops();
2463         struct rdma_netdev_alloc_params params;
2464         struct ipoib_dev_priv *priv;
2465         struct net_device *ndev;
2466         int result;
2467
2468         ndev = ipoib_intf_alloc(hca, port, format);
2469         if (IS_ERR(ndev)) {
2470                 pr_warn("%s, %d: ipoib_intf_alloc failed %ld\n", hca->name, port,
2471                         PTR_ERR(ndev));
2472                 return ndev;
2473         }
2474         priv = ipoib_priv(ndev);
2475
2476         INIT_IB_EVENT_HANDLER(&priv->event_handler,
2477                               priv->ca, ipoib_event);
2478         ib_register_event_handler(&priv->event_handler);
2479
2480         /* call event handler to ensure pkey in sync */
2481         queue_work(ipoib_workqueue, &priv->flush_heavy);
2482
2483         ndev->rtnl_link_ops = ipoib_get_link_ops();
2484
2485         result = register_netdev(ndev);
2486         if (result) {
2487                 pr_warn("%s: couldn't register ipoib port %d; error %d\n",
2488                         hca->name, port, result);
2489
2490                 ipoib_parent_unregister_pre(ndev);
2491                 ipoib_intf_free(ndev);
2492                 free_netdev(ndev);
2493
2494                 return ERR_PTR(result);
2495         }
2496
2497         if (hca->ops.rdma_netdev_get_params) {
2498                 int rc = hca->ops.rdma_netdev_get_params(hca, port,
2499                                                      RDMA_NETDEV_IPOIB,
2500                                                      &params);
2501
2502                 if (!rc && ops->priv_size < params.sizeof_priv)
2503                         ops->priv_size = params.sizeof_priv;
2504         }
2505         /*
2506          * We cannot set priv_destructor before register_netdev because we
2507          * need priv to be always valid during the error flow to execute
2508          * ipoib_parent_unregister_pre(). Instead handle it manually and only
2509          * enter priv_destructor mode once we are completely registered.
2510          */
2511         ndev->priv_destructor = ipoib_intf_free;
2512
2513         if (ipoib_intercept_dev_id_attr(ndev))
2514                 goto sysfs_failed;
2515         if (ipoib_cm_add_mode_attr(ndev))
2516                 goto sysfs_failed;
2517         if (ipoib_add_pkey_attr(ndev))
2518                 goto sysfs_failed;
2519         if (ipoib_add_umcast_attr(ndev))
2520                 goto sysfs_failed;
2521         if (device_create_file(&ndev->dev, &dev_attr_create_child))
2522                 goto sysfs_failed;
2523         if (device_create_file(&ndev->dev, &dev_attr_delete_child))
2524                 goto sysfs_failed;
2525
2526         return ndev;
2527
2528 sysfs_failed:
2529         ipoib_parent_unregister_pre(ndev);
2530         unregister_netdev(ndev);
2531         return ERR_PTR(-ENOMEM);
2532 }
2533
2534 static int ipoib_add_one(struct ib_device *device)
2535 {
2536         struct list_head *dev_list;
2537         struct net_device *dev;
2538         struct ipoib_dev_priv *priv;
2539         unsigned int p;
2540         int count = 0;
2541
2542         dev_list = kmalloc(sizeof(*dev_list), GFP_KERNEL);
2543         if (!dev_list)
2544                 return -ENOMEM;
2545
2546         INIT_LIST_HEAD(dev_list);
2547
2548         rdma_for_each_port (device, p) {
2549                 if (!rdma_protocol_ib(device, p))
2550                         continue;
2551                 dev = ipoib_add_port("ib%d", device, p);
2552                 if (!IS_ERR(dev)) {
2553                         priv = ipoib_priv(dev);
2554                         list_add_tail(&priv->list, dev_list);
2555                         count++;
2556                 }
2557         }
2558
2559         if (!count) {
2560                 kfree(dev_list);
2561                 return -EOPNOTSUPP;
2562         }
2563
2564         ib_set_client_data(device, &ipoib_client, dev_list);
2565         return 0;
2566 }
2567
2568 static void ipoib_remove_one(struct ib_device *device, void *client_data)
2569 {
2570         struct ipoib_dev_priv *priv, *tmp, *cpriv, *tcpriv;
2571         struct list_head *dev_list = client_data;
2572
2573         list_for_each_entry_safe(priv, tmp, dev_list, list) {
2574                 LIST_HEAD(head);
2575                 ipoib_parent_unregister_pre(priv->dev);
2576
2577                 rtnl_lock();
2578
2579                 list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs,
2580                                          list)
2581                         unregister_netdevice_queue(cpriv->dev, &head);
2582                 unregister_netdevice_queue(priv->dev, &head);
2583                 unregister_netdevice_many(&head);
2584
2585                 rtnl_unlock();
2586         }
2587
2588         kfree(dev_list);
2589 }
2590
2591 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
2592 static struct notifier_block ipoib_netdev_notifier = {
2593         .notifier_call = ipoib_netdev_event,
2594 };
2595 #endif
2596
2597 static int __init ipoib_init_module(void)
2598 {
2599         int ret;
2600
2601         ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
2602         ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
2603         ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
2604
2605         ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
2606         ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
2607         ipoib_sendq_size = max3(ipoib_sendq_size, 2 * MAX_SEND_CQE, IPOIB_MIN_QUEUE_SIZE);
2608 #ifdef CONFIG_INFINIBAND_IPOIB_CM
2609         ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
2610         ipoib_max_conn_qp = max(ipoib_max_conn_qp, 0);
2611 #endif
2612
2613         /*
2614          * When copying small received packets, we only copy from the
2615          * linear data part of the SKB, so we rely on this condition.
2616          */
2617         BUILD_BUG_ON(IPOIB_CM_COPYBREAK > IPOIB_CM_HEAD_SIZE);
2618
2619         ipoib_register_debugfs();
2620
2621         /*
2622          * We create a global workqueue here that is used for all flush
2623          * operations.  However, if you attempt to flush a workqueue
2624          * from a task on that same workqueue, it deadlocks the system.
2625          * We want to be able to flush the tasks associated with a
2626          * specific net device, so we also create a workqueue for each
2627          * netdevice.  We queue up the tasks for that device only on
2628          * its private workqueue, and we only queue up flush events
2629          * on our global flush workqueue.  This avoids the deadlocks.
2630          */
2631         ipoib_workqueue = alloc_ordered_workqueue("ipoib_flush", 0);
2632         if (!ipoib_workqueue) {
2633                 ret = -ENOMEM;
2634                 goto err_fs;
2635         }
2636
2637         ib_sa_register_client(&ipoib_sa_client);
2638
2639         ret = ib_register_client(&ipoib_client);
2640         if (ret)
2641                 goto err_sa;
2642
2643         ret = ipoib_netlink_init();
2644         if (ret)
2645                 goto err_client;
2646
2647 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
2648         register_netdevice_notifier(&ipoib_netdev_notifier);
2649 #endif
2650         return 0;
2651
2652 err_client:
2653         ib_unregister_client(&ipoib_client);
2654
2655 err_sa:
2656         ib_sa_unregister_client(&ipoib_sa_client);
2657         destroy_workqueue(ipoib_workqueue);
2658
2659 err_fs:
2660         ipoib_unregister_debugfs();
2661
2662         return ret;
2663 }
2664
2665 static void __exit ipoib_cleanup_module(void)
2666 {
2667 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
2668         unregister_netdevice_notifier(&ipoib_netdev_notifier);
2669 #endif
2670         ipoib_netlink_fini();
2671         ib_unregister_client(&ipoib_client);
2672         ib_sa_unregister_client(&ipoib_sa_client);
2673         ipoib_unregister_debugfs();
2674         destroy_workqueue(ipoib_workqueue);
2675 }
2676
2677 module_init(ipoib_init_module);
2678 module_exit(ipoib_cleanup_module);