drivers/infiniband/ulp/ipoib/ipoib_main.c

   1 /*
   2  * Copyright (c) 2004 Topspin Communications.  All rights reserved.
   3  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
   4  * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
   5  *
   6  * This software is available to you under a choice of one of two
   7  * licenses.  You may choose to be licensed under the terms of the GNU
   8  * General Public License (GPL) Version 2, available from the file
   9  * COPYING in the main directory of this source tree, or the
  10  * OpenIB.org BSD license below:
  11  *
  12  *     Redistribution and use in source and binary forms, with or
  13  *     without modification, are permitted provided that the following
  14  *     conditions are met:
  15  *
  16  *      - Redistributions of source code must retain the above
  17  *        copyright notice, this list of conditions and the following
  18  *        disclaimer.
  19  *
  20  *      - Redistributions in binary form must reproduce the above
  21  *        copyright notice, this list of conditions and the following
  22  *        disclaimer in the documentation and/or other materials
  23  *        provided with the distribution.
  24  *
  25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  32  * SOFTWARE.
  33  */
  34
  35 #include "ipoib.h"
  36
  37 #include <linux/module.h>
  38
  39 #include <linux/init.h>
  40 #include <linux/slab.h>
  41 #include <linux/kernel.h>
  42 #include <linux/vmalloc.h>
  43
  44 #include <linux/if_arp.h>       /* For ARPHRD_xxx */
  45
  46 #include <linux/ip.h>
  47 #include <linux/in.h>
  48
  49 #include <linux/jhash.h>
  50 #include <net/arp.h>
  51 #include <net/addrconf.h>
  52 #include <linux/inetdevice.h>
  53 #include <rdma/ib_cache.h>
  54
  55 #define DRV_VERSION "1.0.0"
  56
  57 const char ipoib_driver_version[] = DRV_VERSION;
  58
  59 MODULE_AUTHOR("Roland Dreier");
  60 MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
  61 MODULE_LICENSE("Dual BSD/GPL");
  62
  63 int ipoib_sendq_size __read_mostly = IPOIB_TX_RING_SIZE;
  64 int ipoib_recvq_size __read_mostly = IPOIB_RX_RING_SIZE;
  65
  66 module_param_named(send_queue_size, ipoib_sendq_size, int, 0444);
  67 MODULE_PARM_DESC(send_queue_size, "Number of descriptors in send queue");
  68 module_param_named(recv_queue_size, ipoib_recvq_size, int, 0444);
  69 MODULE_PARM_DESC(recv_queue_size, "Number of descriptors in receive queue");
  70
  71 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
  72 int ipoib_debug_level;
  73
  74 module_param_named(debug_level, ipoib_debug_level, int, 0644);
  75 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
  76 #endif
  77
  78 struct ipoib_path_iter {
  79         struct net_device *dev;
  80         struct ipoib_path  path;
  81 };
  82
  83 static const u8 ipv4_bcast_addr[] = {
  84         0x00, 0xff, 0xff, 0xff,
  85         0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
  86         0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff
  87 };
  88
  89 struct workqueue_struct *ipoib_workqueue;
  90
  91 struct ib_sa_client ipoib_sa_client;
  92
  93 static void ipoib_add_one(struct ib_device *device);
  94 static void ipoib_remove_one(struct ib_device *device, void *client_data);
  95 static void ipoib_neigh_reclaim(struct rcu_head *rp);
  96 static struct net_device *ipoib_get_net_dev_by_params(
  97                 struct ib_device *dev, u8 port, u16 pkey,
  98                 const union ib_gid *gid, const struct sockaddr *addr,
  99                 void *client_data);
 100 static int ipoib_set_mac(struct net_device *dev, void *addr);
 101 static int ipoib_ioctl(struct net_device *dev, struct ifreq *ifr,
 102                        int cmd);
 103
 104 static struct ib_client ipoib_client = {
 105         .name   = "ipoib",
 106         .add    = ipoib_add_one,
 107         .remove = ipoib_remove_one,
 108         .get_net_dev_by_params = ipoib_get_net_dev_by_params,
 109 };
 110
 111 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 112 static int ipoib_netdev_event(struct notifier_block *this,
 113                               unsigned long event, void *ptr)
 114 {
 115         struct netdev_notifier_info *ni = ptr;
 116         struct net_device *dev = ni->dev;
 117
 118         if (dev->netdev_ops->ndo_open != ipoib_open)
 119                 return NOTIFY_DONE;
 120
 121         switch (event) {
 122         case NETDEV_REGISTER:
 123                 ipoib_create_debug_files(dev);
 124                 break;
 125         case NETDEV_CHANGENAME:
 126                 ipoib_delete_debug_files(dev);
 127                 ipoib_create_debug_files(dev);
 128                 break;
 129         case NETDEV_UNREGISTER:
 130                 ipoib_delete_debug_files(dev);
 131                 break;
 132         }
 133
 134         return NOTIFY_DONE;
 135 }
 136 #endif
 137
 138 int ipoib_open(struct net_device *dev)
 139 {
 140         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 141
 142         ipoib_dbg(priv, "bringing up interface\n");
 143
 144         netif_carrier_off(dev);
 145
 146         set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 147
 148         priv->sm_fullmember_sendonly_support = false;
 149
 150         if (ipoib_ib_dev_open(dev)) {
 151                 if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags))
 152                         return 0;
 153                 goto err_disable;
 154         }
 155
 156         ipoib_ib_dev_up(dev);
 157
 158         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 159                 struct ipoib_dev_priv *cpriv;
 160
 161                 /* Bring up any child interfaces too */
 162                 down_read(&priv->vlan_rwsem);
 163                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 164                         int flags;
 165
 166                         flags = cpriv->dev->flags;
 167                         if (flags & IFF_UP)
 168                                 continue;
 169
 170                         dev_change_flags(cpriv->dev, flags | IFF_UP);
 171                 }
 172                 up_read(&priv->vlan_rwsem);
 173         }
 174
 175         netif_start_queue(dev);
 176
 177         return 0;
 178
 179 err_disable:
 180         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 181
 182         return -EINVAL;
 183 }
 184
 185 static int ipoib_stop(struct net_device *dev)
 186 {
 187         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 188
 189         ipoib_dbg(priv, "stopping interface\n");
 190
 191         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 192
 193         netif_stop_queue(dev);
 194
 195         ipoib_ib_dev_down(dev);
 196         ipoib_ib_dev_stop(dev);
 197
 198         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 199                 struct ipoib_dev_priv *cpriv;
 200
 201                 /* Bring down any child interfaces too */
 202                 down_read(&priv->vlan_rwsem);
 203                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 204                         int flags;
 205
 206                         flags = cpriv->dev->flags;
 207                         if (!(flags & IFF_UP))
 208                                 continue;
 209
 210                         dev_change_flags(cpriv->dev, flags & ~IFF_UP);
 211                 }
 212                 up_read(&priv->vlan_rwsem);
 213         }
 214
 215         return 0;
 216 }
 217
 218 static netdev_features_t ipoib_fix_features(struct net_device *dev, netdev_features_t features)
 219 {
 220         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 221
 222         if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
 223                 features &= ~(NETIF_F_IP_CSUM | NETIF_F_TSO);
 224
 225         return features;
 226 }
 227
 228 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
 229 {
 230         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 231         int ret = 0;
 232
 233         /* dev->mtu > 2K ==> connected mode */
 234         if (ipoib_cm_admin_enabled(dev)) {
 235                 if (new_mtu > ipoib_cm_max_mtu(dev))
 236                         return -EINVAL;
 237
 238                 if (new_mtu > priv->mcast_mtu)
 239                         ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
 240                                    priv->mcast_mtu);
 241
 242                 dev->mtu = new_mtu;
 243                 return 0;
 244         }
 245
 246         if (new_mtu > IPOIB_UD_MTU(priv->max_ib_mtu))
 247                 return -EINVAL;
 248
 249         priv->admin_mtu = new_mtu;
 250
 251         if (priv->mcast_mtu < priv->admin_mtu)
 252                 ipoib_dbg(priv, "MTU must be smaller than the underlying "
 253                                 "link layer MTU - 4 (%u)\n", priv->mcast_mtu);
 254
 255         new_mtu = min(priv->mcast_mtu, priv->admin_mtu);
 256
 257         if (priv->rn_ops->ndo_change_mtu) {
 258                 bool carrier_status = netif_carrier_ok(dev);
 259
 260                 netif_carrier_off(dev);
 261
 262                 /* notify lower level on the real mtu */
 263                 ret = priv->rn_ops->ndo_change_mtu(dev, new_mtu);
 264
 265                 if (carrier_status)
 266                         netif_carrier_on(dev);
 267         } else {
 268                 dev->mtu = new_mtu;
 269         }
 270
 271         return ret;
 272 }
 273
 274 static void ipoib_get_stats(struct net_device *dev,
 275                             struct rtnl_link_stats64 *stats)
 276 {
 277         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 278
 279         if (priv->rn_ops->ndo_get_stats64)
 280                 priv->rn_ops->ndo_get_stats64(dev, stats);
 281         else
 282                 netdev_stats_to_stats64(stats, &dev->stats);
 283 }
 284
 285 /* Called with an RCU read lock taken */
 286 static bool ipoib_is_dev_match_addr_rcu(const struct sockaddr *addr,
 287                                         struct net_device *dev)
 288 {
 289         struct net *net = dev_net(dev);
 290         struct in_device *in_dev;
 291         struct sockaddr_in *addr_in = (struct sockaddr_in *)addr;
 292         struct sockaddr_in6 *addr_in6 = (struct sockaddr_in6 *)addr;
 293         __be32 ret_addr;
 294
 295         switch (addr->sa_family) {
 296         case AF_INET:
 297                 in_dev = in_dev_get(dev);
 298                 if (!in_dev)
 299                         return false;
 300
 301                 ret_addr = inet_confirm_addr(net, in_dev, 0,
 302                                              addr_in->sin_addr.s_addr,
 303                                              RT_SCOPE_HOST);
 304                 in_dev_put(in_dev);
 305                 if (ret_addr)
 306                         return true;
 307
 308                 break;
 309         case AF_INET6:
 310                 if (IS_ENABLED(CONFIG_IPV6) &&
 311                     ipv6_chk_addr(net, &addr_in6->sin6_addr, dev, 1))
 312                         return true;
 313
 314                 break;
 315         }
 316         return false;
 317 }
 318
 319 /**
 320  * Find the master net_device on top of the given net_device.
 321  * @dev: base IPoIB net_device
 322  *
 323  * Returns the master net_device with a reference held, or the same net_device
 324  * if no master exists.
 325  */
 326 static struct net_device *ipoib_get_master_net_dev(struct net_device *dev)
 327 {
 328         struct net_device *master;
 329
 330         rcu_read_lock();
 331         master = netdev_master_upper_dev_get_rcu(dev);
 332         if (master)
 333                 dev_hold(master);
 334         rcu_read_unlock();
 335
 336         if (master)
 337                 return master;
 338
 339         dev_hold(dev);
 340         return dev;
 341 }
 342
 343 struct ipoib_walk_data {
 344         const struct sockaddr *addr;
 345         struct net_device *result;
 346 };
 347
 348 static int ipoib_upper_walk(struct net_device *upper, void *_data)
 349 {
 350         struct ipoib_walk_data *data = _data;
 351         int ret = 0;
 352
 353         if (ipoib_is_dev_match_addr_rcu(data->addr, upper)) {
 354                 dev_hold(upper);
 355                 data->result = upper;
 356                 ret = 1;
 357         }
 358
 359         return ret;
 360 }
 361
 362 /**
 363  * Find a net_device matching the given address, which is an upper device of
 364  * the given net_device.
 365  * @addr: IP address to look for.
 366  * @dev: base IPoIB net_device
 367  *
 368  * If found, returns the net_device with a reference held. Otherwise return
 369  * NULL.
 370  */
 371 static struct net_device *ipoib_get_net_dev_match_addr(
 372                 const struct sockaddr *addr, struct net_device *dev)
 373 {
 374         struct ipoib_walk_data data = {
 375                 .addr = addr,
 376         };
 377
 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, &data);
 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         if (!dev_list)
 486                 return NULL;
 487
 488         /* See if we can find a unique device matching the L2 parameters */
 489         matches = __ipoib_get_net_dev_by_params(dev_list, port, pkey_index,
 490                                                 gid, NULL, &net_dev);
 491
 492         switch (matches) {
 493         case 0:
 494                 return NULL;
 495         case 1:
 496                 return net_dev;
 497         }
 498
 499         dev_put(net_dev);
 500
 501         /* Couldn't find a unique device with L2 parameters only. Use L3
 502          * address to uniquely match the net device */
 503         matches = __ipoib_get_net_dev_by_params(dev_list, port, pkey_index,
 504                                                 gid, addr, &net_dev);
 505         switch (matches) {
 506         case 0:
 507                 return NULL;
 508         default:
 509                 dev_warn_ratelimited(&dev->dev,
 510                                      "duplicate IP address detected\n");
 511                 /* Fall through */
 512         case 1:
 513                 return net_dev;
 514         }
 515 }
 516
 517 int ipoib_set_mode(struct net_device *dev, const char *buf)
 518 {
 519         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 520
 521         if ((test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags) &&
 522              !strcmp(buf, "connected\n")) ||
 523              (!test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags) &&
 524              !strcmp(buf, "datagram\n"))) {
 525                 return 0;
 526         }
 527
 528         /* flush paths if we switch modes so that connections are restarted */
 529         if (IPOIB_CM_SUPPORTED(dev->dev_addr) && !strcmp(buf, "connected\n")) {
 530                 set_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
 531                 ipoib_warn(priv, "enabling connected mode "
 532                            "will cause multicast packet drops\n");
 533                 netdev_update_features(dev);
 534                 dev_set_mtu(dev, ipoib_cm_max_mtu(dev));
 535                 rtnl_unlock();
 536                 priv->tx_wr.wr.send_flags &= ~IB_SEND_IP_CSUM;
 537
 538                 ipoib_flush_paths(dev);
 539                 return (!rtnl_trylock()) ? -EBUSY : 0;
 540         }
 541
 542         if (!strcmp(buf, "datagram\n")) {
 543                 clear_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
 544                 netdev_update_features(dev);
 545                 dev_set_mtu(dev, min(priv->mcast_mtu, dev->mtu));
 546                 rtnl_unlock();
 547                 ipoib_flush_paths(dev);
 548                 return (!rtnl_trylock()) ? -EBUSY : 0;
 549         }
 550
 551         return -EINVAL;
 552 }
 553
 554 struct ipoib_path *__path_find(struct net_device *dev, void *gid)
 555 {
 556         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 557         struct rb_node *n = priv->path_tree.rb_node;
 558         struct ipoib_path *path;
 559         int ret;
 560
 561         while (n) {
 562                 path = rb_entry(n, struct ipoib_path, rb_node);
 563
 564                 ret = memcmp(gid, path->pathrec.dgid.raw,
 565                              sizeof (union ib_gid));
 566
 567                 if (ret < 0)
 568                         n = n->rb_left;
 569                 else if (ret > 0)
 570                         n = n->rb_right;
 571                 else
 572                         return path;
 573         }
 574
 575         return NULL;
 576 }
 577
 578 static int __path_add(struct net_device *dev, struct ipoib_path *path)
 579 {
 580         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 581         struct rb_node **n = &priv->path_tree.rb_node;
 582         struct rb_node *pn = NULL;
 583         struct ipoib_path *tpath;
 584         int ret;
 585
 586         while (*n) {
 587                 pn = *n;
 588                 tpath = rb_entry(pn, struct ipoib_path, rb_node);
 589
 590                 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
 591                              sizeof (union ib_gid));
 592                 if (ret < 0)
 593                         n = &pn->rb_left;
 594                 else if (ret > 0)
 595                         n = &pn->rb_right;
 596                 else
 597                         return -EEXIST;
 598         }
 599
 600         rb_link_node(&path->rb_node, pn, n);
 601         rb_insert_color(&path->rb_node, &priv->path_tree);
 602
 603         list_add_tail(&path->list, &priv->path_list);
 604
 605         return 0;
 606 }
 607
 608 static void path_free(struct net_device *dev, struct ipoib_path *path)
 609 {
 610         struct sk_buff *skb;
 611
 612         while ((skb = __skb_dequeue(&path->queue)))
 613                 dev_kfree_skb_irq(skb);
 614
 615         ipoib_dbg(ipoib_priv(dev), "path_free\n");
 616
 617         /* remove all neigh connected to this path */
 618         ipoib_del_neighs_by_gid(dev, path->pathrec.dgid.raw);
 619
 620         if (path->ah)
 621                 ipoib_put_ah(path->ah);
 622
 623         kfree(path);
 624 }
 625
 626 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 627
 628 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
 629 {
 630         struct ipoib_path_iter *iter;
 631
 632         iter = kmalloc(sizeof(*iter), GFP_KERNEL);
 633         if (!iter)
 634                 return NULL;
 635
 636         iter->dev = dev;
 637         memset(iter->path.pathrec.dgid.raw, 0, 16);
 638
 639         if (ipoib_path_iter_next(iter)) {
 640                 kfree(iter);
 641                 return NULL;
 642         }
 643
 644         return iter;
 645 }
 646
 647 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
 648 {
 649         struct ipoib_dev_priv *priv = ipoib_priv(iter->dev);
 650         struct rb_node *n;
 651         struct ipoib_path *path;
 652         int ret = 1;
 653
 654         spin_lock_irq(&priv->lock);
 655
 656         n = rb_first(&priv->path_tree);
 657
 658         while (n) {
 659                 path = rb_entry(n, struct ipoib_path, rb_node);
 660
 661                 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
 662                            sizeof (union ib_gid)) < 0) {
 663                         iter->path = *path;
 664                         ret = 0;
 665                         break;
 666                 }
 667
 668                 n = rb_next(n);
 669         }
 670
 671         spin_unlock_irq(&priv->lock);
 672
 673         return ret;
 674 }
 675
 676 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
 677                           struct ipoib_path *path)
 678 {
 679         *path = iter->path;
 680 }
 681
 682 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
 683
 684 void ipoib_mark_paths_invalid(struct net_device *dev)
 685 {
 686         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 687         struct ipoib_path *path, *tp;
 688
 689         spin_lock_irq(&priv->lock);
 690
 691         list_for_each_entry_safe(path, tp, &priv->path_list, list) {
 692                 ipoib_dbg(priv, "mark path LID 0x%08x GID %pI6 invalid\n",
 693                           be32_to_cpu(sa_path_get_dlid(&path->pathrec)),
 694                           path->pathrec.dgid.raw);
 695                 if (path->ah)
 696                         path->ah->valid = 0;
 697         }
 698
 699         spin_unlock_irq(&priv->lock);
 700 }
 701
 702 static void push_pseudo_header(struct sk_buff *skb, const char *daddr)
 703 {
 704         struct ipoib_pseudo_header *phdr;
 705
 706         phdr = skb_push(skb, sizeof(*phdr));
 707         memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
 708 }
 709
 710 void ipoib_flush_paths(struct net_device *dev)
 711 {
 712         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 713         struct ipoib_path *path, *tp;
 714         LIST_HEAD(remove_list);
 715         unsigned long flags;
 716
 717         netif_tx_lock_bh(dev);
 718         spin_lock_irqsave(&priv->lock, flags);
 719
 720         list_splice_init(&priv->path_list, &remove_list);
 721
 722         list_for_each_entry(path, &remove_list, list)
 723                 rb_erase(&path->rb_node, &priv->path_tree);
 724
 725         list_for_each_entry_safe(path, tp, &remove_list, list) {
 726                 if (path->query)
 727                         ib_sa_cancel_query(path->query_id, path->query);
 728                 spin_unlock_irqrestore(&priv->lock, flags);
 729                 netif_tx_unlock_bh(dev);
 730                 wait_for_completion(&path->done);
 731                 path_free(dev, path);
 732                 netif_tx_lock_bh(dev);
 733                 spin_lock_irqsave(&priv->lock, flags);
 734         }
 735
 736         spin_unlock_irqrestore(&priv->lock, flags);
 737         netif_tx_unlock_bh(dev);
 738 }
 739
 740 static void path_rec_completion(int status,
 741                                 struct sa_path_rec *pathrec,
 742                                 void *path_ptr)
 743 {
 744         struct ipoib_path *path = path_ptr;
 745         struct net_device *dev = path->dev;
 746         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 747         struct ipoib_ah *ah = NULL;
 748         struct ipoib_ah *old_ah = NULL;
 749         struct ipoib_neigh *neigh, *tn;
 750         struct sk_buff_head skqueue;
 751         struct sk_buff *skb;
 752         unsigned long flags;
 753
 754         if (!status)
 755                 ipoib_dbg(priv, "PathRec LID 0x%04x for GID %pI6\n",
 756                           be32_to_cpu(sa_path_get_dlid(pathrec)),
 757                           pathrec->dgid.raw);
 758         else
 759                 ipoib_dbg(priv, "PathRec status %d for GID %pI6\n",
 760                           status, path->pathrec.dgid.raw);
 761
 762         skb_queue_head_init(&skqueue);
 763
 764         if (!status) {
 765                 struct rdma_ah_attr av;
 766
 767                 if (!ib_init_ah_attr_from_path(priv->ca, priv->port,
 768                                                pathrec, &av, NULL)) {
 769                         ah = ipoib_create_ah(dev, priv->pd, &av);
 770                         rdma_destroy_ah_attr(&av);
 771                 }
 772         }
 773
 774         spin_lock_irqsave(&priv->lock, flags);
 775
 776         if (!IS_ERR_OR_NULL(ah)) {
 777                 /*
 778                  * pathrec.dgid is used as the database key from the LLADDR,
 779                  * it must remain unchanged even if the SA returns a different
 780                  * GID to use in the AH.
 781                  */
 782                 if (memcmp(pathrec->dgid.raw, path->pathrec.dgid.raw,
 783                            sizeof(union ib_gid))) {
 784                         ipoib_dbg(
 785                                 priv,
 786                                 "%s got PathRec for gid %pI6 while asked for %pI6\n",
 787                                 dev->name, pathrec->dgid.raw,
 788                                 path->pathrec.dgid.raw);
 789                         memcpy(pathrec->dgid.raw, path->pathrec.dgid.raw,
 790                                sizeof(union ib_gid));
 791                 }
 792
 793                 path->pathrec = *pathrec;
 794
 795                 old_ah   = path->ah;
 796                 path->ah = ah;
 797
 798                 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
 799                           ah, be32_to_cpu(sa_path_get_dlid(pathrec)),
 800                           pathrec->sl);
 801
 802                 while ((skb = __skb_dequeue(&path->queue)))
 803                         __skb_queue_tail(&skqueue, skb);
 804
 805                 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
 806                         if (neigh->ah) {
 807                                 WARN_ON(neigh->ah != old_ah);
 808                                 /*
 809                                  * Dropping the ah reference inside
 810                                  * priv->lock is safe here, because we
 811                                  * will hold one more reference from
 812                                  * the original value of path->ah (ie
 813                                  * old_ah).
 814                                  */
 815                                 ipoib_put_ah(neigh->ah);
 816                         }
 817                         kref_get(&path->ah->ref);
 818                         neigh->ah = path->ah;
 819
 820                         if (ipoib_cm_enabled(dev, neigh->daddr)) {
 821                                 if (!ipoib_cm_get(neigh))
 822                                         ipoib_cm_set(neigh, ipoib_cm_create_tx(dev,
 823                                                                                path,
 824                                                                                neigh));
 825                                 if (!ipoib_cm_get(neigh)) {
 826                                         ipoib_neigh_free(neigh);
 827                                         continue;
 828                                 }
 829                         }
 830
 831                         while ((skb = __skb_dequeue(&neigh->queue)))
 832                                 __skb_queue_tail(&skqueue, skb);
 833                 }
 834                 path->ah->valid = 1;
 835         }
 836
 837         path->query = NULL;
 838         complete(&path->done);
 839
 840         spin_unlock_irqrestore(&priv->lock, flags);
 841
 842         if (IS_ERR_OR_NULL(ah))
 843                 ipoib_del_neighs_by_gid(dev, path->pathrec.dgid.raw);
 844
 845         if (old_ah)
 846                 ipoib_put_ah(old_ah);
 847
 848         while ((skb = __skb_dequeue(&skqueue))) {
 849                 int ret;
 850                 skb->dev = dev;
 851                 ret = dev_queue_xmit(skb);
 852                 if (ret)
 853                         ipoib_warn(priv, "%s: dev_queue_xmit failed to re-queue packet, ret:%d\n",
 854                                    __func__, ret);
 855         }
 856 }
 857
 858 static void init_path_rec(struct ipoib_dev_priv *priv, struct ipoib_path *path,
 859                           void *gid)
 860 {
 861         path->dev = priv->dev;
 862
 863         if (rdma_cap_opa_ah(priv->ca, priv->port))
 864                 path->pathrec.rec_type = SA_PATH_REC_TYPE_OPA;
 865         else
 866                 path->pathrec.rec_type = SA_PATH_REC_TYPE_IB;
 867
 868         memcpy(path->pathrec.dgid.raw, gid, sizeof(union ib_gid));
 869         path->pathrec.sgid          = priv->local_gid;
 870         path->pathrec.pkey          = cpu_to_be16(priv->pkey);
 871         path->pathrec.numb_path     = 1;
 872         path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
 873 }
 874
 875 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
 876 {
 877         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 878         struct ipoib_path *path;
 879
 880         if (!priv->broadcast)
 881                 return NULL;
 882
 883         path = kzalloc(sizeof(*path), GFP_ATOMIC);
 884         if (!path)
 885                 return NULL;
 886
 887         skb_queue_head_init(&path->queue);
 888
 889         INIT_LIST_HEAD(&path->neigh_list);
 890
 891         init_path_rec(priv, path, gid);
 892
 893         return path;
 894 }
 895
 896 static int path_rec_start(struct net_device *dev,
 897                           struct ipoib_path *path)
 898 {
 899         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 900
 901         ipoib_dbg(priv, "Start path record lookup for %pI6\n",
 902                   path->pathrec.dgid.raw);
 903
 904         init_completion(&path->done);
 905
 906         path->query_id =
 907                 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
 908                                    &path->pathrec,
 909                                    IB_SA_PATH_REC_DGID          |
 910                                    IB_SA_PATH_REC_SGID          |
 911                                    IB_SA_PATH_REC_NUMB_PATH     |
 912                                    IB_SA_PATH_REC_TRAFFIC_CLASS |
 913                                    IB_SA_PATH_REC_PKEY,
 914                                    1000, GFP_ATOMIC,
 915                                    path_rec_completion,
 916                                    path, &path->query);
 917         if (path->query_id < 0) {
 918                 ipoib_warn(priv, "ib_sa_path_rec_get failed: %d\n", path->query_id);
 919                 path->query = NULL;
 920                 complete(&path->done);
 921                 return path->query_id;
 922         }
 923
 924         return 0;
 925 }
 926
 927 static void neigh_refresh_path(struct ipoib_neigh *neigh, u8 *daddr,
 928                                struct net_device *dev)
 929 {
 930         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 931         struct ipoib_path *path;
 932         unsigned long flags;
 933
 934         spin_lock_irqsave(&priv->lock, flags);
 935
 936         path = __path_find(dev, daddr + 4);
 937         if (!path)
 938                 goto out;
 939         if (!path->query)
 940                 path_rec_start(dev, path);
 941 out:
 942         spin_unlock_irqrestore(&priv->lock, flags);
 943 }
 944
 945 static struct ipoib_neigh *neigh_add_path(struct sk_buff *skb, u8 *daddr,
 946                                           struct net_device *dev)
 947 {
 948         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 949         struct rdma_netdev *rn = netdev_priv(dev);
 950         struct ipoib_path *path;
 951         struct ipoib_neigh *neigh;
 952         unsigned long flags;
 953
 954         spin_lock_irqsave(&priv->lock, flags);
 955         neigh = ipoib_neigh_alloc(daddr, dev);
 956         if (!neigh) {
 957                 spin_unlock_irqrestore(&priv->lock, flags);
 958                 ++dev->stats.tx_dropped;
 959                 dev_kfree_skb_any(skb);
 960                 return NULL;
 961         }
 962
 963         /* To avoid race condition, make sure that the
 964          * neigh will be added only once.
 965          */
 966         if (unlikely(!list_empty(&neigh->list))) {
 967                 spin_unlock_irqrestore(&priv->lock, flags);
 968                 return neigh;
 969         }
 970
 971         path = __path_find(dev, daddr + 4);
 972         if (!path) {
 973                 path = path_rec_create(dev, daddr + 4);
 974                 if (!path)
 975                         goto err_path;
 976
 977                 __path_add(dev, path);
 978         }
 979
 980         list_add_tail(&neigh->list, &path->neigh_list);
 981
 982         if (path->ah && path->ah->valid) {
 983                 kref_get(&path->ah->ref);
 984                 neigh->ah = path->ah;
 985
 986                 if (ipoib_cm_enabled(dev, neigh->daddr)) {
 987                         if (!ipoib_cm_get(neigh))
 988                                 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
 989                         if (!ipoib_cm_get(neigh)) {
 990                                 ipoib_neigh_free(neigh);
 991                                 goto err_drop;
 992                         }
 993                         if (skb_queue_len(&neigh->queue) <
 994                             IPOIB_MAX_PATH_REC_QUEUE) {
 995                                 push_pseudo_header(skb, neigh->daddr);
 996                                 __skb_queue_tail(&neigh->queue, skb);
 997                         } else {
 998                                 ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
 999                                            skb_queue_len(&neigh->queue));
1000                                 goto err_drop;
1001                         }
1002                 } else {
1003                         spin_unlock_irqrestore(&priv->lock, flags);
1004                         path->ah->last_send = rn->send(dev, skb, path->ah->ah,
1005                                                        IPOIB_QPN(daddr));
1006                         ipoib_neigh_put(neigh);
1007                         return NULL;
1008                 }
1009         } else {
1010                 neigh->ah  = NULL;
1011
1012                 if (!path->query && path_rec_start(dev, path))
1013                         goto err_path;
1014                 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
1015                         push_pseudo_header(skb, neigh->daddr);
1016                         __skb_queue_tail(&neigh->queue, skb);
1017                 } else {
1018                         goto err_drop;
1019                 }
1020         }
1021
1022         spin_unlock_irqrestore(&priv->lock, flags);
1023         ipoib_neigh_put(neigh);
1024         return NULL;
1025
1026 err_path:
1027         ipoib_neigh_free(neigh);
1028 err_drop:
1029         ++dev->stats.tx_dropped;
1030         dev_kfree_skb_any(skb);
1031
1032         spin_unlock_irqrestore(&priv->lock, flags);
1033         ipoib_neigh_put(neigh);
1034
1035         return NULL;
1036 }
1037
1038 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
1039                              struct ipoib_pseudo_header *phdr)
1040 {
1041         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1042         struct rdma_netdev *rn = netdev_priv(dev);
1043         struct ipoib_path *path;
1044         unsigned long flags;
1045
1046         spin_lock_irqsave(&priv->lock, flags);
1047
1048         /* no broadcast means that all paths are (going to be) not valid */
1049         if (!priv->broadcast)
1050                 goto drop_and_unlock;
1051
1052         path = __path_find(dev, phdr->hwaddr + 4);
1053         if (!path || !path->ah || !path->ah->valid) {
1054                 if (!path) {
1055                         path = path_rec_create(dev, phdr->hwaddr + 4);
1056                         if (!path)
1057                                 goto drop_and_unlock;
1058                         __path_add(dev, path);
1059                 } else {
1060                         /*
1061                          * make sure there are no changes in the existing
1062                          * path record
1063                          */
1064                         init_path_rec(priv, path, phdr->hwaddr + 4);
1065                 }
1066                 if (!path->query && path_rec_start(dev, path)) {
1067                         goto drop_and_unlock;
1068                 }
1069
1070                 if (skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
1071                         push_pseudo_header(skb, phdr->hwaddr);
1072                         __skb_queue_tail(&path->queue, skb);
1073                         goto unlock;
1074                 } else {
1075                         goto drop_and_unlock;
1076                 }
1077         }
1078
1079         spin_unlock_irqrestore(&priv->lock, flags);
1080         ipoib_dbg(priv, "Send unicast ARP to %08x\n",
1081                   be32_to_cpu(sa_path_get_dlid(&path->pathrec)));
1082         path->ah->last_send = rn->send(dev, skb, path->ah->ah,
1083                                        IPOIB_QPN(phdr->hwaddr));
1084         return;
1085
1086 drop_and_unlock:
1087         ++dev->stats.tx_dropped;
1088         dev_kfree_skb_any(skb);
1089 unlock:
1090         spin_unlock_irqrestore(&priv->lock, flags);
1091 }
1092
1093 static netdev_tx_t ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
1094 {
1095         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1096         struct rdma_netdev *rn = netdev_priv(dev);
1097         struct ipoib_neigh *neigh;
1098         struct ipoib_pseudo_header *phdr;
1099         struct ipoib_header *header;
1100         unsigned long flags;
1101
1102         phdr = (struct ipoib_pseudo_header *) skb->data;
1103         skb_pull(skb, sizeof(*phdr));
1104         header = (struct ipoib_header *) skb->data;
1105
1106         if (unlikely(phdr->hwaddr[4] == 0xff)) {
1107                 /* multicast, arrange "if" according to probability */
1108                 if ((header->proto != htons(ETH_P_IP)) &&
1109                     (header->proto != htons(ETH_P_IPV6)) &&
1110                     (header->proto != htons(ETH_P_ARP)) &&
1111                     (header->proto != htons(ETH_P_RARP)) &&
1112                     (header->proto != htons(ETH_P_TIPC))) {
1113                         /* ethertype not supported by IPoIB */
1114                         ++dev->stats.tx_dropped;
1115                         dev_kfree_skb_any(skb);
1116                         return NETDEV_TX_OK;
1117                 }
1118                 /* Add in the P_Key for multicast*/
1119                 phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
1120                 phdr->hwaddr[9] = priv->pkey & 0xff;
1121
1122                 neigh = ipoib_neigh_get(dev, phdr->hwaddr);
1123                 if (likely(neigh))
1124                         goto send_using_neigh;
1125                 ipoib_mcast_send(dev, phdr->hwaddr, skb);
1126                 return NETDEV_TX_OK;
1127         }
1128
1129         /* unicast, arrange "switch" according to probability */
1130         switch (header->proto) {
1131         case htons(ETH_P_IP):
1132         case htons(ETH_P_IPV6):
1133         case htons(ETH_P_TIPC):
1134                 neigh = ipoib_neigh_get(dev, phdr->hwaddr);
1135                 if (unlikely(!neigh)) {
1136                         neigh = neigh_add_path(skb, phdr->hwaddr, dev);
1137                         if (likely(!neigh))
1138                                 return NETDEV_TX_OK;
1139                 }
1140                 break;
1141         case htons(ETH_P_ARP):
1142         case htons(ETH_P_RARP):
1143                 /* for unicast ARP and RARP should always perform path find */
1144                 unicast_arp_send(skb, dev, phdr);
1145                 return NETDEV_TX_OK;
1146         default:
1147                 /* ethertype not supported by IPoIB */
1148                 ++dev->stats.tx_dropped;
1149                 dev_kfree_skb_any(skb);
1150                 return NETDEV_TX_OK;
1151         }
1152
1153 send_using_neigh:
1154         /* note we now hold a ref to neigh */
1155         if (ipoib_cm_get(neigh)) {
1156                 if (ipoib_cm_up(neigh)) {
1157                         ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
1158                         goto unref;
1159                 }
1160         } else if (neigh->ah && neigh->ah->valid) {
1161                 neigh->ah->last_send = rn->send(dev, skb, neigh->ah->ah,
1162                                                 IPOIB_QPN(phdr->hwaddr));
1163                 goto unref;
1164         } else if (neigh->ah) {
1165                 neigh_refresh_path(neigh, phdr->hwaddr, dev);
1166         }
1167
1168         if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
1169                 push_pseudo_header(skb, phdr->hwaddr);
1170                 spin_lock_irqsave(&priv->lock, flags);
1171                 __skb_queue_tail(&neigh->queue, skb);
1172                 spin_unlock_irqrestore(&priv->lock, flags);
1173         } else {
1174                 ++dev->stats.tx_dropped;
1175                 dev_kfree_skb_any(skb);
1176         }
1177
1178 unref:
1179         ipoib_neigh_put(neigh);
1180
1181         return NETDEV_TX_OK;
1182 }
1183
1184 static void ipoib_timeout(struct net_device *dev)
1185 {
1186         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1187
1188         ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
1189                    jiffies_to_msecs(jiffies - dev_trans_start(dev)));
1190         ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
1191                    netif_queue_stopped(dev),
1192                    priv->tx_head, priv->tx_tail);
1193         /* XXX reset QP, etc. */
1194 }
1195
1196 static int ipoib_hard_header(struct sk_buff *skb,
1197                              struct net_device *dev,
1198                              unsigned short type,
1199                              const void *daddr,
1200                              const void *saddr,
1201                              unsigned int len)
1202 {
1203         struct ipoib_header *header;
1204
1205         header = skb_push(skb, sizeof(*header));
1206
1207         header->proto = htons(type);
1208         header->reserved = 0;
1209
1210         /*
1211          * we don't rely on dst_entry structure,  always stuff the
1212          * destination address into skb hard header so we can figure out where
1213          * to send the packet later.
1214          */
1215         push_pseudo_header(skb, daddr);
1216
1217         return IPOIB_HARD_LEN;
1218 }
1219
1220 static void ipoib_set_mcast_list(struct net_device *dev)
1221 {
1222         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1223
1224         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
1225                 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
1226                 return;
1227         }
1228
1229         queue_work(priv->wq, &priv->restart_task);
1230 }
1231
1232 static int ipoib_get_iflink(const struct net_device *dev)
1233 {
1234         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1235
1236         /* parent interface */
1237         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags))
1238                 return dev->ifindex;
1239
1240         /* child/vlan interface */
1241         return priv->parent->ifindex;
1242 }
1243
1244 static u32 ipoib_addr_hash(struct ipoib_neigh_hash *htbl, u8 *daddr)
1245 {
1246         /*
1247          * Use only the address parts that contributes to spreading
1248          * The subnet prefix is not used as one can not connect to
1249          * same remote port (GUID) using the same remote QPN via two
1250          * different subnets.
1251          */
1252          /* qpn octets[1:4) & port GUID octets[12:20) */
1253         u32 *d32 = (u32 *) daddr;
1254         u32 hv;
1255
1256         hv = jhash_3words(d32[3], d32[4], IPOIB_QPN_MASK & d32[0], 0);
1257         return hv & htbl->mask;
1258 }
1259
1260 struct ipoib_neigh *ipoib_neigh_get(struct net_device *dev, u8 *daddr)
1261 {
1262         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1263         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1264         struct ipoib_neigh_hash *htbl;
1265         struct ipoib_neigh *neigh = NULL;
1266         u32 hash_val;
1267
1268         rcu_read_lock_bh();
1269
1270         htbl = rcu_dereference_bh(ntbl->htbl);
1271
1272         if (!htbl)
1273                 goto out_unlock;
1274
1275         hash_val = ipoib_addr_hash(htbl, daddr);
1276         for (neigh = rcu_dereference_bh(htbl->buckets[hash_val]);
1277              neigh != NULL;
1278              neigh = rcu_dereference_bh(neigh->hnext)) {
1279                 if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) {
1280                         /* found, take one ref on behalf of the caller */
1281                         if (!atomic_inc_not_zero(&neigh->refcnt)) {
1282                                 /* deleted */
1283                                 neigh = NULL;
1284                                 goto out_unlock;
1285                         }
1286
1287                         if (likely(skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE))
1288                                 neigh->alive = jiffies;
1289                         goto out_unlock;
1290                 }
1291         }
1292
1293 out_unlock:
1294         rcu_read_unlock_bh();
1295         return neigh;
1296 }
1297
1298 static void __ipoib_reap_neigh(struct ipoib_dev_priv *priv)
1299 {
1300         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1301         struct ipoib_neigh_hash *htbl;
1302         unsigned long neigh_obsolete;
1303         unsigned long dt;
1304         unsigned long flags;
1305         int i;
1306         LIST_HEAD(remove_list);
1307
1308         spin_lock_irqsave(&priv->lock, flags);
1309
1310         htbl = rcu_dereference_protected(ntbl->htbl,
1311                                          lockdep_is_held(&priv->lock));
1312
1313         if (!htbl)
1314                 goto out_unlock;
1315
1316         /* neigh is obsolete if it was idle for two GC periods */
1317         dt = 2 * arp_tbl.gc_interval;
1318         neigh_obsolete = jiffies - dt;
1319
1320         for (i = 0; i < htbl->size; i++) {
1321                 struct ipoib_neigh *neigh;
1322                 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1323
1324                 while ((neigh = rcu_dereference_protected(*np,
1325                                                           lockdep_is_held(&priv->lock))) != NULL) {
1326                         /* was the neigh idle for two GC periods */
1327                         if (time_after(neigh_obsolete, neigh->alive)) {
1328
1329                                 ipoib_check_and_add_mcast_sendonly(priv, neigh->daddr + 4, &remove_list);
1330
1331                                 rcu_assign_pointer(*np,
1332                                                    rcu_dereference_protected(neigh->hnext,
1333                                                                              lockdep_is_held(&priv->lock)));
1334                                 /* remove from path/mc list */
1335                                 list_del_init(&neigh->list);
1336                                 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1337                         } else {
1338                                 np = &neigh->hnext;
1339                         }
1340
1341                 }
1342         }
1343
1344 out_unlock:
1345         spin_unlock_irqrestore(&priv->lock, flags);
1346         ipoib_mcast_remove_list(&remove_list);
1347 }
1348
1349 static void ipoib_reap_neigh(struct work_struct *work)
1350 {
1351         struct ipoib_dev_priv *priv =
1352                 container_of(work, struct ipoib_dev_priv, neigh_reap_task.work);
1353
1354         __ipoib_reap_neigh(priv);
1355
1356         queue_delayed_work(priv->wq, &priv->neigh_reap_task,
1357                            arp_tbl.gc_interval);
1358 }
1359
1360
1361 static struct ipoib_neigh *ipoib_neigh_ctor(u8 *daddr,
1362                                       struct net_device *dev)
1363 {
1364         struct ipoib_neigh *neigh;
1365
1366         neigh = kzalloc(sizeof(*neigh), GFP_ATOMIC);
1367         if (!neigh)
1368                 return NULL;
1369
1370         neigh->dev = dev;
1371         memcpy(&neigh->daddr, daddr, sizeof(neigh->daddr));
1372         skb_queue_head_init(&neigh->queue);
1373         INIT_LIST_HEAD(&neigh->list);
1374         ipoib_cm_set(neigh, NULL);
1375         /* one ref on behalf of the caller */
1376         atomic_set(&neigh->refcnt, 1);
1377
1378         return neigh;
1379 }
1380
1381 struct ipoib_neigh *ipoib_neigh_alloc(u8 *daddr,
1382                                       struct net_device *dev)
1383 {
1384         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1385         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1386         struct ipoib_neigh_hash *htbl;
1387         struct ipoib_neigh *neigh;
1388         u32 hash_val;
1389
1390         htbl = rcu_dereference_protected(ntbl->htbl,
1391                                          lockdep_is_held(&priv->lock));
1392         if (!htbl) {
1393                 neigh = NULL;
1394                 goto out_unlock;
1395         }
1396
1397         /* need to add a new neigh, but maybe some other thread succeeded?
1398          * recalc hash, maybe hash resize took place so we do a search
1399          */
1400         hash_val = ipoib_addr_hash(htbl, daddr);
1401         for (neigh = rcu_dereference_protected(htbl->buckets[hash_val],
1402                                                lockdep_is_held(&priv->lock));
1403              neigh != NULL;
1404              neigh = rcu_dereference_protected(neigh->hnext,
1405                                                lockdep_is_held(&priv->lock))) {
1406                 if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) {
1407                         /* found, take one ref on behalf of the caller */
1408                         if (!atomic_inc_not_zero(&neigh->refcnt)) {
1409                                 /* deleted */
1410                                 neigh = NULL;
1411                                 break;
1412                         }
1413                         neigh->alive = jiffies;
1414                         goto out_unlock;
1415                 }
1416         }
1417
1418         neigh = ipoib_neigh_ctor(daddr, dev);
1419         if (!neigh)
1420                 goto out_unlock;
1421
1422         /* one ref on behalf of the hash table */
1423         atomic_inc(&neigh->refcnt);
1424         neigh->alive = jiffies;
1425         /* put in hash */
1426         rcu_assign_pointer(neigh->hnext,
1427                            rcu_dereference_protected(htbl->buckets[hash_val],
1428                                                      lockdep_is_held(&priv->lock)));
1429         rcu_assign_pointer(htbl->buckets[hash_val], neigh);
1430         atomic_inc(&ntbl->entries);
1431
1432 out_unlock:
1433
1434         return neigh;
1435 }
1436
1437 void ipoib_neigh_dtor(struct ipoib_neigh *neigh)
1438 {
1439         /* neigh reference count was dropprd to zero */
1440         struct net_device *dev = neigh->dev;
1441         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1442         struct sk_buff *skb;
1443         if (neigh->ah)
1444                 ipoib_put_ah(neigh->ah);
1445         while ((skb = __skb_dequeue(&neigh->queue))) {
1446                 ++dev->stats.tx_dropped;
1447                 dev_kfree_skb_any(skb);
1448         }
1449         if (ipoib_cm_get(neigh))
1450                 ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
1451         ipoib_dbg(ipoib_priv(dev),
1452                   "neigh free for %06x %pI6\n",
1453                   IPOIB_QPN(neigh->daddr),
1454                   neigh->daddr + 4);
1455         kfree(neigh);
1456         if (atomic_dec_and_test(&priv->ntbl.entries)) {
1457                 if (test_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags))
1458                         complete(&priv->ntbl.flushed);
1459         }
1460 }
1461
1462 static void ipoib_neigh_reclaim(struct rcu_head *rp)
1463 {
1464         /* Called as a result of removal from hash table */
1465         struct ipoib_neigh *neigh = container_of(rp, struct ipoib_neigh, rcu);
1466         /* note TX context may hold another ref */
1467         ipoib_neigh_put(neigh);
1468 }
1469
1470 void ipoib_neigh_free(struct ipoib_neigh *neigh)
1471 {
1472         struct net_device *dev = neigh->dev;
1473         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1474         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1475         struct ipoib_neigh_hash *htbl;
1476         struct ipoib_neigh __rcu **np;
1477         struct ipoib_neigh *n;
1478         u32 hash_val;
1479
1480         htbl = rcu_dereference_protected(ntbl->htbl,
1481                                         lockdep_is_held(&priv->lock));
1482         if (!htbl)
1483                 return;
1484
1485         hash_val = ipoib_addr_hash(htbl, neigh->daddr);
1486         np = &htbl->buckets[hash_val];
1487         for (n = rcu_dereference_protected(*np,
1488                                             lockdep_is_held(&priv->lock));
1489              n != NULL;
1490              n = rcu_dereference_protected(*np,
1491                                         lockdep_is_held(&priv->lock))) {
1492                 if (n == neigh) {
1493                         /* found */
1494                         rcu_assign_pointer(*np,
1495                                            rcu_dereference_protected(neigh->hnext,
1496                                                                      lockdep_is_held(&priv->lock)));
1497                         /* remove from parent list */
1498                         list_del_init(&neigh->list);
1499                         call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1500                         return;
1501                 } else {
1502                         np = &n->hnext;
1503                 }
1504         }
1505 }
1506
1507 static int ipoib_neigh_hash_init(struct ipoib_dev_priv *priv)
1508 {
1509         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1510         struct ipoib_neigh_hash *htbl;
1511         struct ipoib_neigh __rcu **buckets;
1512         u32 size;
1513
1514         clear_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags);
1515         ntbl->htbl = NULL;
1516         htbl = kzalloc(sizeof(*htbl), GFP_KERNEL);
1517         if (!htbl)
1518                 return -ENOMEM;
1519         size = roundup_pow_of_two(arp_tbl.gc_thresh3);
1520         buckets = kvcalloc(size, sizeof(*buckets), GFP_KERNEL);
1521         if (!buckets) {
1522                 kfree(htbl);
1523                 return -ENOMEM;
1524         }
1525         htbl->size = size;
1526         htbl->mask = (size - 1);
1527         htbl->buckets = buckets;
1528         RCU_INIT_POINTER(ntbl->htbl, htbl);
1529         htbl->ntbl = ntbl;
1530         atomic_set(&ntbl->entries, 0);
1531
1532         /* start garbage collection */
1533         queue_delayed_work(priv->wq, &priv->neigh_reap_task,
1534                            arp_tbl.gc_interval);
1535
1536         return 0;
1537 }
1538
1539 static void neigh_hash_free_rcu(struct rcu_head *head)
1540 {
1541         struct ipoib_neigh_hash *htbl = container_of(head,
1542                                                     struct ipoib_neigh_hash,
1543                                                     rcu);
1544         struct ipoib_neigh __rcu **buckets = htbl->buckets;
1545         struct ipoib_neigh_table *ntbl = htbl->ntbl;
1546
1547         kvfree(buckets);
1548         kfree(htbl);
1549         complete(&ntbl->deleted);
1550 }
1551
1552 void ipoib_del_neighs_by_gid(struct net_device *dev, u8 *gid)
1553 {
1554         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1555         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1556         struct ipoib_neigh_hash *htbl;
1557         unsigned long flags;
1558         int i;
1559
1560         /* remove all neigh connected to a given path or mcast */
1561         spin_lock_irqsave(&priv->lock, flags);
1562
1563         htbl = rcu_dereference_protected(ntbl->htbl,
1564                                          lockdep_is_held(&priv->lock));
1565
1566         if (!htbl)
1567                 goto out_unlock;
1568
1569         for (i = 0; i < htbl->size; i++) {
1570                 struct ipoib_neigh *neigh;
1571                 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1572
1573                 while ((neigh = rcu_dereference_protected(*np,
1574                                                           lockdep_is_held(&priv->lock))) != NULL) {
1575                         /* delete neighs belong to this parent */
1576                         if (!memcmp(gid, neigh->daddr + 4, sizeof (union ib_gid))) {
1577                                 rcu_assign_pointer(*np,
1578                                                    rcu_dereference_protected(neigh->hnext,
1579                                                                              lockdep_is_held(&priv->lock)));
1580                                 /* remove from parent list */
1581                                 list_del_init(&neigh->list);
1582                                 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1583                         } else {
1584                                 np = &neigh->hnext;
1585                         }
1586
1587                 }
1588         }
1589 out_unlock:
1590         spin_unlock_irqrestore(&priv->lock, flags);
1591 }
1592
1593 static void ipoib_flush_neighs(struct ipoib_dev_priv *priv)
1594 {
1595         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1596         struct ipoib_neigh_hash *htbl;
1597         unsigned long flags;
1598         int i, wait_flushed = 0;
1599
1600         init_completion(&priv->ntbl.flushed);
1601         set_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags);
1602
1603         spin_lock_irqsave(&priv->lock, flags);
1604
1605         htbl = rcu_dereference_protected(ntbl->htbl,
1606                                         lockdep_is_held(&priv->lock));
1607         if (!htbl)
1608                 goto out_unlock;
1609
1610         wait_flushed = atomic_read(&priv->ntbl.entries);
1611         if (!wait_flushed)
1612                 goto free_htbl;
1613
1614         for (i = 0; i < htbl->size; i++) {
1615                 struct ipoib_neigh *neigh;
1616                 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1617
1618                 while ((neigh = rcu_dereference_protected(*np,
1619                                        lockdep_is_held(&priv->lock))) != NULL) {
1620                         rcu_assign_pointer(*np,
1621                                            rcu_dereference_protected(neigh->hnext,
1622                                                                      lockdep_is_held(&priv->lock)));
1623                         /* remove from path/mc list */
1624                         list_del_init(&neigh->list);
1625                         call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1626                 }
1627         }
1628
1629 free_htbl:
1630         rcu_assign_pointer(ntbl->htbl, NULL);
1631         call_rcu(&htbl->rcu, neigh_hash_free_rcu);
1632
1633 out_unlock:
1634         spin_unlock_irqrestore(&priv->lock, flags);
1635         if (wait_flushed)
1636                 wait_for_completion(&priv->ntbl.flushed);
1637 }
1638
1639 static void ipoib_neigh_hash_uninit(struct net_device *dev)
1640 {
1641         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1642
1643         ipoib_dbg(priv, "ipoib_neigh_hash_uninit\n");
1644         init_completion(&priv->ntbl.deleted);
1645
1646         cancel_delayed_work_sync(&priv->neigh_reap_task);
1647
1648         ipoib_flush_neighs(priv);
1649
1650         wait_for_completion(&priv->ntbl.deleted);
1651 }
1652
1653 static void ipoib_napi_add(struct net_device *dev)
1654 {
1655         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1656
1657         netif_napi_add(dev, &priv->recv_napi, ipoib_rx_poll, IPOIB_NUM_WC);
1658         netif_napi_add(dev, &priv->send_napi, ipoib_tx_poll, MAX_SEND_CQE);
1659 }
1660
1661 static void ipoib_napi_del(struct net_device *dev)
1662 {
1663         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1664
1665         netif_napi_del(&priv->recv_napi);
1666         netif_napi_del(&priv->send_napi);
1667 }
1668
1669 static void ipoib_dev_uninit_default(struct net_device *dev)
1670 {
1671         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1672
1673         ipoib_transport_dev_cleanup(dev);
1674
1675         ipoib_napi_del(dev);
1676
1677         ipoib_cm_dev_cleanup(dev);
1678
1679         kfree(priv->rx_ring);
1680         vfree(priv->tx_ring);
1681
1682         priv->rx_ring = NULL;
1683         priv->tx_ring = NULL;
1684 }
1685
1686 static int ipoib_dev_init_default(struct net_device *dev)
1687 {
1688         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1689
1690         ipoib_napi_add(dev);
1691
1692         /* Allocate RX/TX "rings" to hold queued skbs */
1693         priv->rx_ring = kcalloc(ipoib_recvq_size,
1694                                        sizeof(*priv->rx_ring),
1695                                        GFP_KERNEL);
1696         if (!priv->rx_ring)
1697                 goto out;
1698
1699         priv->tx_ring = vzalloc(array_size(ipoib_sendq_size,
1700                                            sizeof(*priv->tx_ring)));
1701         if (!priv->tx_ring) {
1702                 pr_warn("%s: failed to allocate TX ring (%d entries)\n",
1703                         priv->ca->name, ipoib_sendq_size);
1704                 goto out_rx_ring_cleanup;
1705         }
1706
1707         /* priv->tx_head, tx_tail & tx_outstanding are already 0 */
1708
1709         if (ipoib_transport_dev_init(dev, priv->ca)) {
1710                 pr_warn("%s: ipoib_transport_dev_init failed\n",
1711                         priv->ca->name);
1712                 goto out_tx_ring_cleanup;
1713         }
1714
1715         /* after qp created set dev address */
1716         priv->dev->dev_addr[1] = (priv->qp->qp_num >> 16) & 0xff;
1717         priv->dev->dev_addr[2] = (priv->qp->qp_num >>  8) & 0xff;
1718         priv->dev->dev_addr[3] = (priv->qp->qp_num) & 0xff;
1719
1720         return 0;
1721
1722 out_tx_ring_cleanup:
1723         vfree(priv->tx_ring);
1724
1725 out_rx_ring_cleanup:
1726         kfree(priv->rx_ring);
1727
1728 out:
1729         ipoib_napi_del(dev);
1730         return -ENOMEM;
1731 }
1732
1733 static int ipoib_ioctl(struct net_device *dev, struct ifreq *ifr,
1734                        int cmd)
1735 {
1736         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1737
1738         if (!priv->rn_ops->ndo_do_ioctl)
1739                 return -EOPNOTSUPP;
1740
1741         return priv->rn_ops->ndo_do_ioctl(dev, ifr, cmd);
1742 }
1743
1744 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
1745 {
1746         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1747         int ret = -ENOMEM;
1748
1749         priv->ca = ca;
1750         priv->port = port;
1751         priv->qp = NULL;
1752
1753         /*
1754          * the various IPoIB tasks assume they will never race against
1755          * themselves, so always use a single thread workqueue
1756          */
1757         priv->wq = alloc_ordered_workqueue("ipoib_wq", WQ_MEM_RECLAIM);
1758         if (!priv->wq) {
1759                 pr_warn("%s: failed to allocate device WQ\n", dev->name);
1760                 goto out;
1761         }
1762
1763         /* create pd, which used both for control and datapath*/
1764         priv->pd = ib_alloc_pd(priv->ca, 0);
1765         if (IS_ERR(priv->pd)) {
1766                 pr_warn("%s: failed to allocate PD\n", ca->name);
1767                 goto clean_wq;
1768         }
1769
1770         ret = priv->rn_ops->ndo_init(dev);
1771         if (ret) {
1772                 pr_warn("%s failed to init HW resource\n", dev->name);
1773                 goto out_free_pd;
1774         }
1775
1776         ret = ipoib_neigh_hash_init(priv);
1777         if (ret) {
1778                 pr_warn("%s failed to init neigh hash\n", dev->name);
1779                 goto out_dev_uninit;
1780         }
1781
1782         if (dev->flags & IFF_UP) {
1783                 if (ipoib_ib_dev_open(dev)) {
1784                         pr_warn("%s failed to open device\n", dev->name);
1785                         ret = -ENODEV;
1786                         goto out_hash_uninit;
1787                 }
1788         }
1789
1790         return 0;
1791
1792 out_hash_uninit:
1793         ipoib_neigh_hash_uninit(dev);
1794
1795 out_dev_uninit:
1796         ipoib_ib_dev_cleanup(dev);
1797
1798 out_free_pd:
1799         if (priv->pd) {
1800                 ib_dealloc_pd(priv->pd);
1801                 priv->pd = NULL;
1802         }
1803
1804 clean_wq:
1805         if (priv->wq) {
1806                 destroy_workqueue(priv->wq);
1807                 priv->wq = NULL;
1808         }
1809
1810 out:
1811         return ret;
1812 }
1813
1814 /*
1815  * This must be called before doing an unregister_netdev on a parent device to
1816  * shutdown the IB event handler.
1817  */
1818 static void ipoib_parent_unregister_pre(struct net_device *ndev)
1819 {
1820         struct ipoib_dev_priv *priv = ipoib_priv(ndev);
1821
1822         /*
1823          * ipoib_set_mac checks netif_running before pushing work, clearing
1824          * running ensures the it will not add more work.
1825          */
1826         rtnl_lock();
1827         dev_change_flags(priv->dev, priv->dev->flags & ~IFF_UP);
1828         rtnl_unlock();
1829
1830         /* ipoib_event() cannot be running once this returns */
1831         ib_unregister_event_handler(&priv->event_handler);
1832
1833         /*
1834          * Work on the queue grabs the rtnl lock, so this cannot be done while
1835          * also holding it.
1836          */
1837         flush_workqueue(ipoib_workqueue);
1838 }
1839
1840 static void ipoib_ndo_uninit(struct net_device *dev)
1841 {
1842         struct ipoib_dev_priv *priv = ipoib_priv(dev), *cpriv, *tcpriv;
1843         LIST_HEAD(head);
1844
1845         ASSERT_RTNL();
1846
1847         /* Delete any child interfaces first */
1848         list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
1849                 /* Stop GC on child */
1850                 cancel_delayed_work_sync(&cpriv->neigh_reap_task);
1851                 unregister_netdevice_queue(cpriv->dev, &head);
1852         }
1853         unregister_netdevice_many(&head);
1854
1855         ipoib_neigh_hash_uninit(dev);
1856
1857         ipoib_ib_dev_cleanup(dev);
1858
1859         /* no more works over the priv->wq */
1860         if (priv->wq) {
1861                 flush_workqueue(priv->wq);
1862                 destroy_workqueue(priv->wq);
1863                 priv->wq = NULL;
1864         }
1865 }
1866
1867 static int ipoib_set_vf_link_state(struct net_device *dev, int vf, int link_state)
1868 {
1869         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1870
1871         return ib_set_vf_link_state(priv->ca, vf, priv->port, link_state);
1872 }
1873
1874 static int ipoib_get_vf_config(struct net_device *dev, int vf,
1875                                struct ifla_vf_info *ivf)
1876 {
1877         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1878         int err;
1879
1880         err = ib_get_vf_config(priv->ca, vf, priv->port, ivf);
1881         if (err)
1882                 return err;
1883
1884         ivf->vf = vf;
1885
1886         return 0;
1887 }
1888
1889 static int ipoib_set_vf_guid(struct net_device *dev, int vf, u64 guid, int type)
1890 {
1891         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1892
1893         if (type != IFLA_VF_IB_NODE_GUID && type != IFLA_VF_IB_PORT_GUID)
1894                 return -EINVAL;
1895
1896         return ib_set_vf_guid(priv->ca, vf, priv->port, guid, type);
1897 }
1898
1899 static int ipoib_get_vf_stats(struct net_device *dev, int vf,
1900                               struct ifla_vf_stats *vf_stats)
1901 {
1902         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1903
1904         return ib_get_vf_stats(priv->ca, vf, priv->port, vf_stats);
1905 }
1906
1907 static const struct header_ops ipoib_header_ops = {
1908         .create = ipoib_hard_header,
1909 };
1910
1911 static const struct net_device_ops ipoib_netdev_ops_pf = {
1912         .ndo_uninit              = ipoib_ndo_uninit,
1913         .ndo_open                = ipoib_open,
1914         .ndo_stop                = ipoib_stop,
1915         .ndo_change_mtu          = ipoib_change_mtu,
1916         .ndo_fix_features        = ipoib_fix_features,
1917         .ndo_start_xmit          = ipoib_start_xmit,
1918         .ndo_tx_timeout          = ipoib_timeout,
1919         .ndo_set_rx_mode         = ipoib_set_mcast_list,
1920         .ndo_get_iflink          = ipoib_get_iflink,
1921         .ndo_set_vf_link_state   = ipoib_set_vf_link_state,
1922         .ndo_get_vf_config       = ipoib_get_vf_config,
1923         .ndo_get_vf_stats        = ipoib_get_vf_stats,
1924         .ndo_set_vf_guid         = ipoib_set_vf_guid,
1925         .ndo_set_mac_address     = ipoib_set_mac,
1926         .ndo_get_stats64         = ipoib_get_stats,
1927         .ndo_do_ioctl            = ipoib_ioctl,
1928 };
1929
1930 static const struct net_device_ops ipoib_netdev_ops_vf = {
1931         .ndo_uninit              = ipoib_ndo_uninit,
1932         .ndo_open                = ipoib_open,
1933         .ndo_stop                = ipoib_stop,
1934         .ndo_change_mtu          = ipoib_change_mtu,
1935         .ndo_fix_features        = ipoib_fix_features,
1936         .ndo_start_xmit          = ipoib_start_xmit,
1937         .ndo_tx_timeout          = ipoib_timeout,
1938         .ndo_set_rx_mode         = ipoib_set_mcast_list,
1939         .ndo_get_iflink          = ipoib_get_iflink,
1940         .ndo_get_stats64         = ipoib_get_stats,
1941         .ndo_do_ioctl            = ipoib_ioctl,
1942 };
1943
1944 void ipoib_setup_common(struct net_device *dev)
1945 {
1946         dev->header_ops          = &ipoib_header_ops;
1947
1948         ipoib_set_ethtool_ops(dev);
1949
1950         dev->watchdog_timeo      = HZ;
1951
1952         dev->flags              |= IFF_BROADCAST | IFF_MULTICAST;
1953
1954         dev->hard_header_len     = IPOIB_HARD_LEN;
1955         dev->addr_len            = INFINIBAND_ALEN;
1956         dev->type                = ARPHRD_INFINIBAND;
1957         dev->tx_queue_len        = ipoib_sendq_size * 2;
1958         dev->features            = (NETIF_F_VLAN_CHALLENGED     |
1959                                     NETIF_F_HIGHDMA);
1960         netif_keep_dst(dev);
1961
1962         memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
1963 }
1964
1965 static void ipoib_build_priv(struct net_device *dev)
1966 {
1967         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1968
1969         priv->dev = dev;
1970         spin_lock_init(&priv->lock);
1971         init_rwsem(&priv->vlan_rwsem);
1972         mutex_init(&priv->mcast_mutex);
1973         mutex_init(&priv->sysfs_mutex);
1974
1975         INIT_LIST_HEAD(&priv->path_list);
1976         INIT_LIST_HEAD(&priv->child_intfs);
1977         INIT_LIST_HEAD(&priv->dead_ahs);
1978         INIT_LIST_HEAD(&priv->multicast_list);
1979
1980         INIT_DELAYED_WORK(&priv->mcast_task,   ipoib_mcast_join_task);
1981         INIT_WORK(&priv->carrier_on_task, ipoib_mcast_carrier_on_task);
1982         INIT_WORK(&priv->flush_light,   ipoib_ib_dev_flush_light);
1983         INIT_WORK(&priv->flush_normal,   ipoib_ib_dev_flush_normal);
1984         INIT_WORK(&priv->flush_heavy,   ipoib_ib_dev_flush_heavy);
1985         INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
1986         INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
1987         INIT_DELAYED_WORK(&priv->neigh_reap_task, ipoib_reap_neigh);
1988 }
1989
1990 static const struct net_device_ops ipoib_netdev_default_pf = {
1991         .ndo_init                = ipoib_dev_init_default,
1992         .ndo_uninit              = ipoib_dev_uninit_default,
1993         .ndo_open                = ipoib_ib_dev_open_default,
1994         .ndo_stop                = ipoib_ib_dev_stop_default,
1995 };
1996
1997 static struct net_device
1998 *ipoib_create_netdev_default(struct ib_device *hca,
1999                              const char *name,
2000                              unsigned char name_assign_type,
2001                              void (*setup)(struct net_device *))
2002 {
2003         struct net_device *dev;
2004         struct rdma_netdev *rn;
2005
2006         dev = alloc_netdev((int)sizeof(struct rdma_netdev),
2007                            name,
2008                            name_assign_type, setup);
2009         if (!dev)
2010                 return NULL;
2011
2012         rn = netdev_priv(dev);
2013
2014         rn->send = ipoib_send;
2015         rn->attach_mcast = ipoib_mcast_attach;
2016         rn->detach_mcast = ipoib_mcast_detach;
2017         rn->free_rdma_netdev = free_netdev;
2018         rn->hca = hca;
2019
2020         dev->netdev_ops = &ipoib_netdev_default_pf;
2021
2022         return dev;
2023 }
2024
2025 static struct net_device *ipoib_get_netdev(struct ib_device *hca, u8 port,
2026                                            const char *name)
2027 {
2028         struct net_device *dev;
2029
2030         if (hca->alloc_rdma_netdev) {
2031                 dev = hca->alloc_rdma_netdev(hca, port,
2032                                              RDMA_NETDEV_IPOIB, name,
2033                                              NET_NAME_UNKNOWN,
2034                                              ipoib_setup_common);
2035                 if (IS_ERR_OR_NULL(dev) && PTR_ERR(dev) != -EOPNOTSUPP)
2036                         return NULL;
2037         }
2038
2039         if (!hca->alloc_rdma_netdev || PTR_ERR(dev) == -EOPNOTSUPP)
2040                 dev = ipoib_create_netdev_default(hca, name, NET_NAME_UNKNOWN,
2041                                                   ipoib_setup_common);
2042
2043         return dev;
2044 }
2045
2046 struct ipoib_dev_priv *ipoib_intf_alloc(struct ib_device *hca, u8 port,
2047                                         const char *name)
2048 {
2049         struct net_device *dev;
2050         struct ipoib_dev_priv *priv;
2051         struct rdma_netdev *rn;
2052
2053         priv = kzalloc(sizeof(*priv), GFP_KERNEL);
2054         if (!priv)
2055                 return NULL;
2056
2057         dev = ipoib_get_netdev(hca, port, name);
2058         if (!dev)
2059                 goto free_priv;
2060
2061         priv->rn_ops = dev->netdev_ops;
2062
2063         /* fixme : should be after the query_cap */
2064         if (priv->hca_caps & IB_DEVICE_VIRTUAL_FUNCTION)
2065                 dev->netdev_ops = &ipoib_netdev_ops_vf;
2066         else
2067                 dev->netdev_ops = &ipoib_netdev_ops_pf;
2068
2069         rn = netdev_priv(dev);
2070         rn->clnt_priv = priv;
2071         ipoib_build_priv(dev);
2072
2073         return priv;
2074 free_priv:
2075         kfree(priv);
2076         return NULL;
2077 }
2078
2079 static ssize_t show_pkey(struct device *dev,
2080                          struct device_attribute *attr, char *buf)
2081 {
2082         struct net_device *ndev = to_net_dev(dev);
2083         struct ipoib_dev_priv *priv = ipoib_priv(ndev);
2084
2085         return sprintf(buf, "0x%04x\n", priv->pkey);
2086 }
2087 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
2088
2089 static ssize_t show_umcast(struct device *dev,
2090                            struct device_attribute *attr, char *buf)
2091 {
2092         struct net_device *ndev = to_net_dev(dev);
2093         struct ipoib_dev_priv *priv = ipoib_priv(ndev);
2094
2095         return sprintf(buf, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
2096 }
2097
2098 void ipoib_set_umcast(struct net_device *ndev, int umcast_val)
2099 {
2100         struct ipoib_dev_priv *priv = ipoib_priv(ndev);
2101
2102         if (umcast_val > 0) {
2103                 set_bit(IPOIB_FLAG_UMCAST, &priv->flags);
2104                 ipoib_warn(priv, "ignoring multicast groups joined directly "
2105                                 "by userspace\n");
2106         } else
2107                 clear_bit(IPOIB_FLAG_UMCAST, &priv->flags);
2108 }
2109
2110 static ssize_t set_umcast(struct device *dev,
2111                           struct device_attribute *attr,
2112                           const char *buf, size_t count)
2113 {
2114         unsigned long umcast_val = simple_strtoul(buf, NULL, 0);
2115
2116         ipoib_set_umcast(to_net_dev(dev), umcast_val);
2117
2118         return count;
2119 }
2120 static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast);
2121
2122 int ipoib_add_umcast_attr(struct net_device *dev)
2123 {
2124         return device_create_file(&dev->dev, &dev_attr_umcast);
2125 }
2126
2127 static void set_base_guid(struct ipoib_dev_priv *priv, union ib_gid *gid)
2128 {
2129         struct ipoib_dev_priv *child_priv;
2130         struct net_device *netdev = priv->dev;
2131
2132         netif_addr_lock_bh(netdev);
2133
2134         memcpy(&priv->local_gid.global.interface_id,
2135                &gid->global.interface_id,
2136                sizeof(gid->global.interface_id));
2137         memcpy(netdev->dev_addr + 4, &priv->local_gid, sizeof(priv->local_gid));
2138         clear_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags);
2139
2140         netif_addr_unlock_bh(netdev);
2141
2142         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
2143                 down_read(&priv->vlan_rwsem);
2144                 list_for_each_entry(child_priv, &priv->child_intfs, list)
2145                         set_base_guid(child_priv, gid);
2146                 up_read(&priv->vlan_rwsem);
2147         }
2148 }
2149
2150 static int ipoib_check_lladdr(struct net_device *dev,
2151                               struct sockaddr_storage *ss)
2152 {
2153         union ib_gid *gid = (union ib_gid *)(ss->__data + 4);
2154         int ret = 0;
2155
2156         netif_addr_lock_bh(dev);
2157
2158         /* Make sure the QPN, reserved and subnet prefix match the current
2159          * lladdr, it also makes sure the lladdr is unicast.
2160          */
2161         if (memcmp(dev->dev_addr, ss->__data,
2162                    4 + sizeof(gid->global.subnet_prefix)) ||
2163             gid->global.interface_id == 0)
2164                 ret = -EINVAL;
2165
2166         netif_addr_unlock_bh(dev);
2167
2168         return ret;
2169 }
2170
2171 static int ipoib_set_mac(struct net_device *dev, void *addr)
2172 {
2173         struct ipoib_dev_priv *priv = ipoib_priv(dev);
2174         struct sockaddr_storage *ss = addr;
2175         int ret;
2176
2177         if (!(dev->priv_flags & IFF_LIVE_ADDR_CHANGE) && netif_running(dev))
2178                 return -EBUSY;
2179
2180         ret = ipoib_check_lladdr(dev, ss);
2181         if (ret)
2182                 return ret;
2183
2184         set_base_guid(priv, (union ib_gid *)(ss->__data + 4));
2185
2186         queue_work(ipoib_workqueue, &priv->flush_light);
2187
2188         return 0;
2189 }
2190
2191 static ssize_t create_child(struct device *dev,
2192                             struct device_attribute *attr,
2193                             const char *buf, size_t count)
2194 {
2195         int pkey;
2196         int ret;
2197
2198         if (sscanf(buf, "%i", &pkey) != 1)
2199                 return -EINVAL;
2200
2201         if (pkey <= 0 || pkey > 0xffff || pkey == 0x8000)
2202                 return -EINVAL;
2203
2204         /*
2205          * Set the full membership bit, so that we join the right
2206          * broadcast group, etc.
2207          */
2208         pkey |= 0x8000;
2209
2210         ret = ipoib_vlan_add(to_net_dev(dev), pkey);
2211
2212         return ret ? ret : count;
2213 }
2214 static DEVICE_ATTR(create_child, S_IWUSR, NULL, create_child);
2215
2216 static ssize_t delete_child(struct device *dev,
2217                             struct device_attribute *attr,
2218                             const char *buf, size_t count)
2219 {
2220         int pkey;
2221         int ret;
2222
2223         if (sscanf(buf, "%i", &pkey) != 1)
2224                 return -EINVAL;
2225
2226         if (pkey < 0 || pkey > 0xffff)
2227                 return -EINVAL;
2228
2229         ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
2230
2231         return ret ? ret : count;
2232
2233 }
2234 static DEVICE_ATTR(delete_child, S_IWUSR, NULL, delete_child);
2235
2236 int ipoib_add_pkey_attr(struct net_device *dev)
2237 {
2238         return device_create_file(&dev->dev, &dev_attr_pkey);
2239 }
2240
2241 void ipoib_set_dev_features(struct ipoib_dev_priv *priv, struct ib_device *hca)
2242 {
2243         priv->hca_caps = hca->attrs.device_cap_flags;
2244
2245         if (priv->hca_caps & IB_DEVICE_UD_IP_CSUM) {
2246                 priv->dev->hw_features |= NETIF_F_IP_CSUM | NETIF_F_RXCSUM;
2247
2248                 if (priv->hca_caps & IB_DEVICE_UD_TSO)
2249                         priv->dev->hw_features |= NETIF_F_TSO;
2250
2251                 priv->dev->features |= priv->dev->hw_features;
2252         }
2253 }
2254
2255 static struct net_device *ipoib_add_port(const char *format,
2256                                          struct ib_device *hca, u8 port)
2257 {
2258         struct ipoib_dev_priv *priv;
2259         struct ib_port_attr attr;
2260         struct rdma_netdev *rn;
2261         int result = -ENOMEM;
2262
2263         priv = ipoib_intf_alloc(hca, port, format);
2264         if (!priv) {
2265                 pr_warn("%s, %d: ipoib_intf_alloc failed\n", hca->name, port);
2266                 goto alloc_mem_failed;
2267         }
2268
2269         SET_NETDEV_DEV(priv->dev, hca->dev.parent);
2270         priv->dev->dev_id = port - 1;
2271
2272         result = ib_query_port(hca, port, &attr);
2273         if (result) {
2274                 pr_warn("%s: ib_query_port %d failed\n", hca->name, port);
2275                 goto device_init_failed;
2276         }
2277
2278         priv->max_ib_mtu = ib_mtu_enum_to_int(attr.max_mtu);
2279
2280         /* MTU will be reset when mcast join happens */
2281         priv->dev->mtu  = IPOIB_UD_MTU(priv->max_ib_mtu);
2282         priv->mcast_mtu  = priv->admin_mtu = priv->dev->mtu;
2283         priv->dev->max_mtu = IPOIB_CM_MTU;
2284
2285         priv->dev->neigh_priv_len = sizeof(struct ipoib_neigh);
2286
2287         result = ib_query_pkey(hca, port, 0, &priv->pkey);
2288         if (result) {
2289                 pr_warn("%s: ib_query_pkey port %d failed (ret = %d)\n",
2290                         hca->name, port, result);
2291                 goto device_init_failed;
2292         }
2293
2294         ipoib_set_dev_features(priv, hca);
2295
2296         /*
2297          * Set the full membership bit, so that we join the right
2298          * broadcast group, etc.
2299          */
2300         priv->pkey |= 0x8000;
2301
2302         priv->dev->broadcast[8] = priv->pkey >> 8;
2303         priv->dev->broadcast[9] = priv->pkey & 0xff;
2304
2305         result = rdma_query_gid(hca, port, 0, &priv->local_gid);
2306         if (result) {
2307                 pr_warn("%s: rdma_query_gid port %d failed (ret = %d)\n",
2308                         hca->name, port, result);
2309                 goto device_init_failed;
2310         }
2311
2312         memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw,
2313                sizeof(union ib_gid));
2314         set_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags);
2315
2316         result = ipoib_dev_init(priv->dev, hca, port);
2317         if (result) {
2318                 pr_warn("%s: failed to initialize port %d (ret = %d)\n",
2319                         hca->name, port, result);
2320                 goto device_init_failed;
2321         }
2322
2323         INIT_IB_EVENT_HANDLER(&priv->event_handler,
2324                               priv->ca, ipoib_event);
2325         ib_register_event_handler(&priv->event_handler);
2326
2327         /* call event handler to ensure pkey in sync */
2328         queue_work(ipoib_workqueue, &priv->flush_heavy);
2329
2330         result = register_netdev(priv->dev);
2331         if (result) {
2332                 pr_warn("%s: couldn't register ipoib port %d; error %d\n",
2333                         hca->name, port, result);
2334                 ipoib_parent_unregister_pre(priv->dev);
2335                 goto device_init_failed;
2336         }
2337
2338         result = -ENOMEM;
2339         if (ipoib_cm_add_mode_attr(priv->dev))
2340                 goto sysfs_failed;
2341         if (ipoib_add_pkey_attr(priv->dev))
2342                 goto sysfs_failed;
2343         if (ipoib_add_umcast_attr(priv->dev))
2344                 goto sysfs_failed;
2345         if (device_create_file(&priv->dev->dev, &dev_attr_create_child))
2346                 goto sysfs_failed;
2347         if (device_create_file(&priv->dev->dev, &dev_attr_delete_child))
2348                 goto sysfs_failed;
2349
2350         return priv->dev;
2351
2352 sysfs_failed:
2353         ipoib_parent_unregister_pre(priv->dev);
2354         unregister_netdev(priv->dev);
2355
2356 device_init_failed:
2357         rn = netdev_priv(priv->dev);
2358         rn->free_rdma_netdev(priv->dev);
2359         kfree(priv);
2360
2361 alloc_mem_failed:
2362         return ERR_PTR(result);
2363 }
2364
2365 static void ipoib_add_one(struct ib_device *device)
2366 {
2367         struct list_head *dev_list;
2368         struct net_device *dev;
2369         struct ipoib_dev_priv *priv;
2370         int p;
2371         int count = 0;
2372
2373         dev_list = kmalloc(sizeof(*dev_list), GFP_KERNEL);
2374         if (!dev_list)
2375                 return;
2376
2377         INIT_LIST_HEAD(dev_list);
2378
2379         for (p = rdma_start_port(device); p <= rdma_end_port(device); ++p) {
2380                 if (!rdma_protocol_ib(device, p))
2381                         continue;
2382                 dev = ipoib_add_port("ib%d", device, p);
2383                 if (!IS_ERR(dev)) {
2384                         priv = ipoib_priv(dev);
2385                         list_add_tail(&priv->list, dev_list);
2386                         count++;
2387                 }
2388         }
2389
2390         if (!count) {
2391                 kfree(dev_list);
2392                 return;
2393         }
2394
2395         ib_set_client_data(device, &ipoib_client, dev_list);
2396 }
2397
2398 static void ipoib_remove_one(struct ib_device *device, void *client_data)
2399 {
2400         struct ipoib_dev_priv *priv, *tmp, *cpriv, *tcpriv;
2401         struct list_head *dev_list = client_data;
2402
2403         if (!dev_list)
2404                 return;
2405
2406         list_for_each_entry_safe(priv, tmp, dev_list, list) {
2407                 struct rdma_netdev *parent_rn = netdev_priv(priv->dev);
2408
2409                 ipoib_parent_unregister_pre(priv->dev);
2410
2411                 /* Wrap rtnl_lock/unlock with mutex to protect sysfs calls */
2412                 mutex_lock(&priv->sysfs_mutex);
2413                 unregister_netdev(priv->dev);
2414                 mutex_unlock(&priv->sysfs_mutex);
2415
2416                 parent_rn->free_rdma_netdev(priv->dev);
2417
2418                 list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
2419                         struct rdma_netdev *child_rn;
2420
2421                         child_rn = netdev_priv(cpriv->dev);
2422                         child_rn->free_rdma_netdev(cpriv->dev);
2423                         kfree(cpriv);
2424                 }
2425
2426                 kfree(priv);
2427         }
2428
2429         kfree(dev_list);
2430 }
2431
2432 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
2433 static struct notifier_block ipoib_netdev_notifier = {
2434         .notifier_call = ipoib_netdev_event,
2435 };
2436 #endif
2437
2438 static int __init ipoib_init_module(void)
2439 {
2440         int ret;
2441
2442         ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
2443         ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
2444         ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
2445
2446         ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
2447         ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
2448         ipoib_sendq_size = max3(ipoib_sendq_size, 2 * MAX_SEND_CQE, IPOIB_MIN_QUEUE_SIZE);
2449 #ifdef CONFIG_INFINIBAND_IPOIB_CM
2450         ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
2451         ipoib_max_conn_qp = max(ipoib_max_conn_qp, 0);
2452 #endif
2453
2454         /*
2455          * When copying small received packets, we only copy from the
2456          * linear data part of the SKB, so we rely on this condition.
2457          */
2458         BUILD_BUG_ON(IPOIB_CM_COPYBREAK > IPOIB_CM_HEAD_SIZE);
2459
2460         ret = ipoib_register_debugfs();
2461         if (ret)
2462                 return ret;
2463
2464         /*
2465          * We create a global workqueue here that is used for all flush
2466          * operations.  However, if you attempt to flush a workqueue
2467          * from a task on that same workqueue, it deadlocks the system.
2468          * We want to be able to flush the tasks associated with a
2469          * specific net device, so we also create a workqueue for each
2470          * netdevice.  We queue up the tasks for that device only on
2471          * its private workqueue, and we only queue up flush events
2472          * on our global flush workqueue.  This avoids the deadlocks.
2473          */
2474         ipoib_workqueue = alloc_ordered_workqueue("ipoib_flush",
2475                                                   WQ_MEM_RECLAIM);
2476         if (!ipoib_workqueue) {
2477                 ret = -ENOMEM;
2478                 goto err_fs;
2479         }
2480
2481         ib_sa_register_client(&ipoib_sa_client);
2482
2483         ret = ib_register_client(&ipoib_client);
2484         if (ret)
2485                 goto err_sa;
2486
2487         ret = ipoib_netlink_init();
2488         if (ret)
2489                 goto err_client;
2490
2491 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
2492         register_netdevice_notifier(&ipoib_netdev_notifier);
2493 #endif
2494         return 0;
2495
2496 err_client:
2497         ib_unregister_client(&ipoib_client);
2498
2499 err_sa:
2500         ib_sa_unregister_client(&ipoib_sa_client);
2501         destroy_workqueue(ipoib_workqueue);
2502
2503 err_fs:
2504         ipoib_unregister_debugfs();
2505
2506         return ret;
2507 }
2508
2509 static void __exit ipoib_cleanup_module(void)
2510 {
2511 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
2512         unregister_netdevice_notifier(&ipoib_netdev_notifier);
2513 #endif
2514         ipoib_netlink_fini();
2515         ib_unregister_client(&ipoib_client);
2516         ib_sa_unregister_client(&ipoib_sa_client);
2517         ipoib_unregister_debugfs();
2518         destroy_workqueue(ipoib_workqueue);
2519 }
2520
2521 module_init(ipoib_init_module);
2522 module_exit(ipoib_cleanup_module);