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 <net/netdev_lock.h>
  53 #include <net/pkt_sched.h>
  54 #include <linux/inetdevice.h>
  55 #include <rdma/ib_cache.h>
  56
  57 MODULE_AUTHOR("Roland Dreier");
  58 MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
  59 MODULE_LICENSE("Dual BSD/GPL");
  60
  61 int ipoib_sendq_size __read_mostly = IPOIB_TX_RING_SIZE;
  62 int ipoib_recvq_size __read_mostly = IPOIB_RX_RING_SIZE;
  63
  64 module_param_named(send_queue_size, ipoib_sendq_size, int, 0444);
  65 MODULE_PARM_DESC(send_queue_size, "Number of descriptors in send queue");
  66 module_param_named(recv_queue_size, ipoib_recvq_size, int, 0444);
  67 MODULE_PARM_DESC(recv_queue_size, "Number of descriptors in receive queue");
  68
  69 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
  70 int ipoib_debug_level;
  71
  72 module_param_named(debug_level, ipoib_debug_level, int, 0644);
  73 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
  74 #endif
  75
  76 struct ipoib_path_iter {
  77         struct net_device *dev;
  78         struct ipoib_path  path;
  79 };
  80
  81 static const u8 ipv4_bcast_addr[] = {
  82         0x00, 0xff, 0xff, 0xff,
  83         0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
  84         0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff
  85 };
  86
  87 struct workqueue_struct *ipoib_workqueue;
  88
  89 struct ib_sa_client ipoib_sa_client;
  90
  91 static int ipoib_add_one(struct ib_device *device);
  92 static void ipoib_remove_one(struct ib_device *device, void *client_data);
  93 static void ipoib_neigh_reclaim(struct rcu_head *rp);
  94 static struct net_device *ipoib_get_net_dev_by_params(
  95                 struct ib_device *dev, u32 port, u16 pkey,
  96                 const union ib_gid *gid, const struct sockaddr *addr,
  97                 void *client_data);
  98 static int ipoib_set_mac(struct net_device *dev, void *addr);
  99 static int ipoib_ioctl(struct net_device *dev, struct ifreq *ifr,
 100                        int cmd);
 101
 102 static struct ib_client ipoib_client = {
 103         .name   = "ipoib",
 104         .add    = ipoib_add_one,
 105         .remove = ipoib_remove_one,
 106         .get_net_dev_by_params = ipoib_get_net_dev_by_params,
 107 };
 108
 109 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 110 static int ipoib_netdev_event(struct notifier_block *this,
 111                               unsigned long event, void *ptr)
 112 {
 113         struct netdev_notifier_info *ni = ptr;
 114         struct net_device *dev = ni->dev;
 115
 116         if (dev->netdev_ops->ndo_open != ipoib_open)
 117                 return NOTIFY_DONE;
 118
 119         switch (event) {
 120         case NETDEV_REGISTER:
 121                 ipoib_create_debug_files(dev);
 122                 break;
 123         case NETDEV_CHANGENAME:
 124                 ipoib_delete_debug_files(dev);
 125                 ipoib_create_debug_files(dev);
 126                 break;
 127         case NETDEV_UNREGISTER:
 128                 ipoib_delete_debug_files(dev);
 129                 break;
 130         }
 131
 132         return NOTIFY_DONE;
 133 }
 134 #endif
 135
 136 struct ipoib_ifupdown_work {
 137         struct work_struct work;
 138         struct net_device *dev;
 139         netdevice_tracker dev_tracker;
 140         bool up;
 141 };
 142
 143 static void ipoib_ifupdown_task(struct work_struct *work)
 144 {
 145         struct ipoib_ifupdown_work *pwork =
 146                 container_of(work, struct ipoib_ifupdown_work, work);
 147         struct net_device *dev = pwork->dev;
 148         unsigned int flags;
 149
 150         rtnl_lock();
 151         flags = dev->flags;
 152         if (pwork->up)
 153                 flags |= IFF_UP;
 154         else
 155                 flags &= ~IFF_UP;
 156
 157         if (dev->flags != flags)
 158                 dev_change_flags(dev, flags, NULL);
 159         rtnl_unlock();
 160         netdev_put(dev, &pwork->dev_tracker);
 161         kfree(pwork);
 162 }
 163
 164 static void ipoib_schedule_ifupdown_task(struct net_device *dev, bool up)
 165 {
 166         struct ipoib_ifupdown_work *work;
 167
 168         if ((up && (dev->flags & IFF_UP)) ||
 169             (!up && !(dev->flags & IFF_UP)))
 170                 return;
 171
 172         work = kmalloc(sizeof(*work), GFP_KERNEL);
 173         if (!work)
 174                 return;
 175         work->dev = dev;
 176         netdev_hold(dev, &work->dev_tracker, GFP_KERNEL);
 177         work->up = up;
 178         INIT_WORK(&work->work, ipoib_ifupdown_task);
 179         queue_work(ipoib_workqueue, &work->work);
 180 }
 181
 182 int ipoib_open(struct net_device *dev)
 183 {
 184         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 185
 186         ipoib_dbg(priv, "bringing up interface\n");
 187
 188         netif_carrier_off(dev);
 189
 190         set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 191
 192         if (ipoib_ib_dev_open(dev)) {
 193                 if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags))
 194                         return 0;
 195                 goto err_disable;
 196         }
 197
 198         ipoib_ib_dev_up(dev);
 199
 200         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 201                 struct ipoib_dev_priv *cpriv;
 202
 203                 /* Bring up any child interfaces too */
 204                 netdev_lock_ops_to_full(dev);
 205                 list_for_each_entry(cpriv, &priv->child_intfs, list)
 206                         ipoib_schedule_ifupdown_task(cpriv->dev, true);
 207                 netdev_unlock_full_to_ops(dev);
 208         } else if (priv->parent) {
 209                 struct ipoib_dev_priv *ppriv = ipoib_priv(priv->parent);
 210
 211                 if (!test_bit(IPOIB_FLAG_ADMIN_UP, &ppriv->flags))
 212                         ipoib_dbg(priv, "parent device %s is not up, so child device may be not functioning.\n",
 213                                   ppriv->dev->name);
 214         }
 215         netif_start_queue(dev);
 216
 217         return 0;
 218
 219 err_disable:
 220         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 221
 222         return -EINVAL;
 223 }
 224
 225 static int ipoib_stop(struct net_device *dev)
 226 {
 227         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 228
 229         ipoib_dbg(priv, "stopping interface\n");
 230
 231         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 232
 233         netif_stop_queue(dev);
 234
 235         ipoib_ib_dev_down(dev);
 236         ipoib_ib_dev_stop(dev);
 237
 238         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 239                 struct ipoib_dev_priv *cpriv;
 240
 241                 /* Bring down any child interfaces too */
 242                 netdev_lock_ops_to_full(dev);
 243                 list_for_each_entry(cpriv, &priv->child_intfs, list)
 244                         ipoib_schedule_ifupdown_task(cpriv->dev, false);
 245                 netdev_unlock_full_to_ops(dev);
 246         }
 247
 248         return 0;
 249 }
 250
 251 static netdev_features_t ipoib_fix_features(struct net_device *dev, netdev_features_t features)
 252 {
 253         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 254
 255         if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
 256                 features &= ~(NETIF_F_IP_CSUM | NETIF_F_TSO);
 257
 258         return features;
 259 }
 260
 261 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
 262 {
 263         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 264         int ret = 0;
 265
 266         /* dev->mtu > 2K ==> connected mode */
 267         if (ipoib_cm_admin_enabled(dev)) {
 268                 if (new_mtu > ipoib_cm_max_mtu(dev))
 269                         return -EINVAL;
 270
 271                 if (new_mtu > priv->mcast_mtu)
 272                         ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
 273                                    priv->mcast_mtu);
 274
 275                 WRITE_ONCE(dev->mtu, new_mtu);
 276                 return 0;
 277         }
 278
 279         if (new_mtu < (ETH_MIN_MTU + IPOIB_ENCAP_LEN) ||
 280             new_mtu > IPOIB_UD_MTU(priv->max_ib_mtu))
 281                 return -EINVAL;
 282
 283         priv->admin_mtu = new_mtu;
 284
 285         if (priv->mcast_mtu < priv->admin_mtu)
 286                 ipoib_dbg(priv, "MTU must be smaller than the underlying "
 287                                 "link layer MTU - 4 (%u)\n", priv->mcast_mtu);
 288
 289         new_mtu = min(priv->mcast_mtu, priv->admin_mtu);
 290
 291         if (priv->rn_ops->ndo_change_mtu) {
 292                 bool carrier_status = netif_carrier_ok(dev);
 293
 294                 netif_carrier_off(dev);
 295
 296                 /* notify lower level on the real mtu */
 297                 ret = priv->rn_ops->ndo_change_mtu(dev, new_mtu);
 298
 299                 if (carrier_status)
 300                         netif_carrier_on(dev);
 301         } else {
 302                 WRITE_ONCE(dev->mtu, new_mtu);
 303         }
 304
 305         return ret;
 306 }
 307
 308 static void ipoib_get_stats(struct net_device *dev,
 309                             struct rtnl_link_stats64 *stats)
 310 {
 311         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 312
 313         if (priv->rn_ops->ndo_get_stats64)
 314                 priv->rn_ops->ndo_get_stats64(dev, stats);
 315         else
 316                 netdev_stats_to_stats64(stats, &dev->stats);
 317 }
 318
 319 /* Called with an RCU read lock taken */
 320 static bool ipoib_is_dev_match_addr_rcu(const struct sockaddr *addr,
 321                                         struct net_device *dev)
 322 {
 323         struct net *net = dev_net(dev);
 324         struct in_device *in_dev;
 325         struct sockaddr_in *addr_in = (struct sockaddr_in *)addr;
 326         struct sockaddr_in6 *addr_in6 = (struct sockaddr_in6 *)addr;
 327         __be32 ret_addr;
 328
 329         switch (addr->sa_family) {
 330         case AF_INET:
 331                 in_dev = in_dev_get(dev);
 332                 if (!in_dev)
 333                         return false;
 334
 335                 ret_addr = inet_confirm_addr(net, in_dev, 0,
 336                                              addr_in->sin_addr.s_addr,
 337                                              RT_SCOPE_HOST);
 338                 in_dev_put(in_dev);
 339                 if (ret_addr)
 340                         return true;
 341
 342                 break;
 343         case AF_INET6:
 344                 if (IS_ENABLED(CONFIG_IPV6) &&
 345                     ipv6_chk_addr(net, &addr_in6->sin6_addr, dev, 1))
 346                         return true;
 347
 348                 break;
 349         }
 350         return false;
 351 }
 352
 353 /*
 354  * Find the master net_device on top of the given net_device.
 355  * @dev: base IPoIB net_device
 356  *
 357  * Returns the master net_device with a reference held, or the same net_device
 358  * if no master exists.
 359  */
 360 static struct net_device *ipoib_get_master_net_dev(struct net_device *dev)
 361 {
 362         struct net_device *master;
 363
 364         rcu_read_lock();
 365         master = netdev_master_upper_dev_get_rcu(dev);
 366         dev_hold(master);
 367         rcu_read_unlock();
 368
 369         if (master)
 370                 return master;
 371
 372         dev_hold(dev);
 373         return dev;
 374 }
 375
 376 struct ipoib_walk_data {
 377         const struct sockaddr *addr;
 378         struct net_device *result;
 379 };
 380
 381 static int ipoib_upper_walk(struct net_device *upper,
 382                             struct netdev_nested_priv *priv)
 383 {
 384         struct ipoib_walk_data *data = (struct ipoib_walk_data *)priv->data;
 385         int ret = 0;
 386
 387         if (ipoib_is_dev_match_addr_rcu(data->addr, upper)) {
 388                 dev_hold(upper);
 389                 data->result = upper;
 390                 ret = 1;
 391         }
 392
 393         return ret;
 394 }
 395
 396 /**
 397  * ipoib_get_net_dev_match_addr - Find a net_device matching
 398  * the given address, which is an upper device of the given net_device.
 399  *
 400  * @addr: IP address to look for.
 401  * @dev: base IPoIB net_device
 402  *
 403  * If found, returns the net_device with a reference held. Otherwise return
 404  * NULL.
 405  */
 406 static struct net_device *ipoib_get_net_dev_match_addr(
 407                 const struct sockaddr *addr, struct net_device *dev)
 408 {
 409         struct netdev_nested_priv priv;
 410         struct ipoib_walk_data data = {
 411                 .addr = addr,
 412         };
 413
 414         priv.data = (void *)&data;
 415         rcu_read_lock();
 416         if (ipoib_is_dev_match_addr_rcu(addr, dev)) {
 417                 dev_hold(dev);
 418                 data.result = dev;
 419                 goto out;
 420         }
 421
 422         netdev_walk_all_upper_dev_rcu(dev, ipoib_upper_walk, &priv);
 423 out:
 424         rcu_read_unlock();
 425         return data.result;
 426 }
 427
 428 /* returns the number of IPoIB netdevs on top a given ipoib device matching a
 429  * pkey_index and address, if one exists.
 430  *
 431  * @found_net_dev: contains a matching net_device if the return value >= 1,
 432  * with a reference held. */
 433 static int ipoib_match_gid_pkey_addr(struct ipoib_dev_priv *priv,
 434                                      const union ib_gid *gid,
 435                                      u16 pkey_index,
 436                                      const struct sockaddr *addr,
 437                                      int nesting,
 438                                      struct net_device **found_net_dev)
 439 {
 440         struct ipoib_dev_priv *child_priv;
 441         struct net_device *net_dev = NULL;
 442         int matches = 0;
 443
 444         if (priv->pkey_index == pkey_index &&
 445             (!gid || !memcmp(gid, &priv->local_gid, sizeof(*gid)))) {
 446                 if (!addr) {
 447                         net_dev = ipoib_get_master_net_dev(priv->dev);
 448                 } else {
 449                         /* Verify the net_device matches the IP address, as
 450                          * IPoIB child devices currently share a GID. */
 451                         net_dev = ipoib_get_net_dev_match_addr(addr, priv->dev);
 452                 }
 453                 if (net_dev) {
 454                         if (!*found_net_dev)
 455                                 *found_net_dev = net_dev;
 456                         else
 457                                 dev_put(net_dev);
 458                         ++matches;
 459                 }
 460         }
 461
 462         if (test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags))
 463                 return matches;
 464
 465         /* Check child interfaces */
 466         netdev_lock(priv->dev);
 467         list_for_each_entry(child_priv, &priv->child_intfs, list) {
 468                 matches += ipoib_match_gid_pkey_addr(child_priv, gid,
 469                                                      pkey_index, addr,
 470                                                      nesting + 1,
 471                                                      found_net_dev);
 472                 if (matches > 1)
 473                         break;
 474         }
 475         netdev_unlock(priv->dev);
 476
 477         return matches;
 478 }
 479
 480 /* Returns the number of matching net_devs found (between 0 and 2). Also
 481  * return the matching net_device in the @net_dev parameter, holding a
 482  * reference to the net_device, if the number of matches >= 1 */
 483 static int __ipoib_get_net_dev_by_params(struct list_head *dev_list, u32 port,
 484                                          u16 pkey_index,
 485                                          const union ib_gid *gid,
 486                                          const struct sockaddr *addr,
 487                                          struct net_device **net_dev)
 488 {
 489         struct ipoib_dev_priv *priv;
 490         int matches = 0;
 491
 492         *net_dev = NULL;
 493
 494         list_for_each_entry(priv, dev_list, list) {
 495                 if (priv->port != port)
 496                         continue;
 497
 498                 matches += ipoib_match_gid_pkey_addr(priv, gid, pkey_index,
 499                                                      addr, 0, net_dev);
 500                 if (matches > 1)
 501                         break;
 502         }
 503
 504         return matches;
 505 }
 506
 507 static struct net_device *ipoib_get_net_dev_by_params(
 508                 struct ib_device *dev, u32 port, u16 pkey,
 509                 const union ib_gid *gid, const struct sockaddr *addr,
 510                 void *client_data)
 511 {
 512         struct net_device *net_dev;
 513         struct list_head *dev_list = client_data;
 514         u16 pkey_index;
 515         int matches;
 516         int ret;
 517
 518         if (!rdma_protocol_ib(dev, port))
 519                 return NULL;
 520
 521         ret = ib_find_cached_pkey(dev, port, pkey, &pkey_index);
 522         if (ret)
 523                 return NULL;
 524
 525         /* See if we can find a unique device matching the L2 parameters */
 526         matches = __ipoib_get_net_dev_by_params(dev_list, port, pkey_index,
 527                                                 gid, NULL, &net_dev);
 528
 529         switch (matches) {
 530         case 0:
 531                 return NULL;
 532         case 1:
 533                 return net_dev;
 534         }
 535
 536         dev_put(net_dev);
 537
 538         /* Couldn't find a unique device with L2 parameters only. Use L3
 539          * address to uniquely match the net device */
 540         matches = __ipoib_get_net_dev_by_params(dev_list, port, pkey_index,
 541                                                 gid, addr, &net_dev);
 542         switch (matches) {
 543         case 0:
 544                 return NULL;
 545         default:
 546                 dev_warn_ratelimited(&dev->dev,
 547                                      "duplicate IP address detected\n");
 548                 fallthrough;
 549         case 1:
 550                 return net_dev;
 551         }
 552 }
 553
 554 int ipoib_set_mode(struct net_device *dev, const char *buf)
 555 {
 556         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 557
 558         if ((test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags) &&
 559              !strcmp(buf, "connected\n")) ||
 560              (!test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags) &&
 561              !strcmp(buf, "datagram\n"))) {
 562                 return 0;
 563         }
 564
 565         /* flush paths if we switch modes so that connections are restarted */
 566         if (IPOIB_CM_SUPPORTED(dev->dev_addr) && !strcmp(buf, "connected\n")) {
 567                 set_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
 568                 ipoib_warn(priv, "enabling connected mode "
 569                            "will cause multicast packet drops\n");
 570                 netdev_lock_ops(dev);
 571                 netdev_update_features(dev);
 572                 netif_set_mtu(dev, ipoib_cm_max_mtu(dev));
 573                 netif_set_real_num_tx_queues(dev, 1);
 574                 netdev_unlock_ops(dev);
 575                 rtnl_unlock();
 576                 priv->tx_wr.wr.send_flags &= ~IB_SEND_IP_CSUM;
 577
 578                 ipoib_flush_paths(dev);
 579                 return (!rtnl_trylock()) ? -EBUSY : 0;
 580         }
 581
 582         if (!strcmp(buf, "datagram\n")) {
 583                 clear_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
 584                 netdev_lock_ops(dev);
 585                 netdev_update_features(dev);
 586                 netif_set_mtu(dev, min(priv->mcast_mtu, dev->mtu));
 587                 netif_set_real_num_tx_queues(dev, dev->num_tx_queues);
 588                 netdev_unlock_ops(dev);
 589                 rtnl_unlock();
 590                 ipoib_flush_paths(dev);
 591                 return (!rtnl_trylock()) ? -EBUSY : 0;
 592         }
 593
 594         return -EINVAL;
 595 }
 596
 597 struct ipoib_path *__path_find(struct net_device *dev, void *gid)
 598 {
 599         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 600         struct rb_node *n = priv->path_tree.rb_node;
 601         struct ipoib_path *path;
 602         int ret;
 603
 604         while (n) {
 605                 path = rb_entry(n, struct ipoib_path, rb_node);
 606
 607                 ret = memcmp(gid, path->pathrec.dgid.raw,
 608                              sizeof (union ib_gid));
 609
 610                 if (ret < 0)
 611                         n = n->rb_left;
 612                 else if (ret > 0)
 613                         n = n->rb_right;
 614                 else
 615                         return path;
 616         }
 617
 618         return NULL;
 619 }
 620
 621 static int __path_add(struct net_device *dev, struct ipoib_path *path)
 622 {
 623         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 624         struct rb_node **n = &priv->path_tree.rb_node;
 625         struct rb_node *pn = NULL;
 626         struct ipoib_path *tpath;
 627         int ret;
 628
 629         while (*n) {
 630                 pn = *n;
 631                 tpath = rb_entry(pn, struct ipoib_path, rb_node);
 632
 633                 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
 634                              sizeof (union ib_gid));
 635                 if (ret < 0)
 636                         n = &pn->rb_left;
 637                 else if (ret > 0)
 638                         n = &pn->rb_right;
 639                 else
 640                         return -EEXIST;
 641         }
 642
 643         rb_link_node(&path->rb_node, pn, n);
 644         rb_insert_color(&path->rb_node, &priv->path_tree);
 645
 646         list_add_tail(&path->list, &priv->path_list);
 647
 648         return 0;
 649 }
 650
 651 static void path_free(struct net_device *dev, struct ipoib_path *path)
 652 {
 653         struct sk_buff *skb;
 654
 655         while ((skb = __skb_dequeue(&path->queue)))
 656                 dev_kfree_skb_irq(skb);
 657
 658         ipoib_dbg(ipoib_priv(dev), "%s\n", __func__);
 659
 660         /* remove all neigh connected to this path */
 661         ipoib_del_neighs_by_gid(dev, path->pathrec.dgid.raw);
 662
 663         if (path->ah)
 664                 ipoib_put_ah(path->ah);
 665
 666         kfree(path);
 667 }
 668
 669 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 670
 671 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
 672 {
 673         struct ipoib_path_iter *iter;
 674
 675         iter = kmalloc(sizeof(*iter), GFP_KERNEL);
 676         if (!iter)
 677                 return NULL;
 678
 679         iter->dev = dev;
 680         memset(iter->path.pathrec.dgid.raw, 0, 16);
 681
 682         if (ipoib_path_iter_next(iter)) {
 683                 kfree(iter);
 684                 return NULL;
 685         }
 686
 687         return iter;
 688 }
 689
 690 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
 691 {
 692         struct ipoib_dev_priv *priv = ipoib_priv(iter->dev);
 693         struct rb_node *n;
 694         struct ipoib_path *path;
 695         int ret = 1;
 696
 697         spin_lock_irq(&priv->lock);
 698
 699         n = rb_first(&priv->path_tree);
 700
 701         while (n) {
 702                 path = rb_entry(n, struct ipoib_path, rb_node);
 703
 704                 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
 705                            sizeof (union ib_gid)) < 0) {
 706                         iter->path = *path;
 707                         ret = 0;
 708                         break;
 709                 }
 710
 711                 n = rb_next(n);
 712         }
 713
 714         spin_unlock_irq(&priv->lock);
 715
 716         return ret;
 717 }
 718
 719 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
 720                           struct ipoib_path *path)
 721 {
 722         *path = iter->path;
 723 }
 724
 725 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
 726
 727 void ipoib_mark_paths_invalid(struct net_device *dev)
 728 {
 729         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 730         struct ipoib_path *path, *tp;
 731
 732         spin_lock_irq(&priv->lock);
 733
 734         list_for_each_entry_safe(path, tp, &priv->path_list, list) {
 735                 ipoib_dbg(priv, "mark path LID 0x%08x GID %pI6 invalid\n",
 736                           be32_to_cpu(sa_path_get_dlid(&path->pathrec)),
 737                           path->pathrec.dgid.raw);
 738                 if (path->ah)
 739                         path->ah->valid = 0;
 740         }
 741
 742         spin_unlock_irq(&priv->lock);
 743 }
 744
 745 static void push_pseudo_header(struct sk_buff *skb, const char *daddr)
 746 {
 747         struct ipoib_pseudo_header *phdr;
 748
 749         phdr = skb_push(skb, sizeof(*phdr));
 750         memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
 751 }
 752
 753 void ipoib_flush_paths(struct net_device *dev)
 754 {
 755         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 756         struct ipoib_path *path, *tp;
 757         LIST_HEAD(remove_list);
 758         unsigned long flags;
 759
 760         netif_tx_lock_bh(dev);
 761         spin_lock_irqsave(&priv->lock, flags);
 762
 763         list_splice_init(&priv->path_list, &remove_list);
 764
 765         list_for_each_entry(path, &remove_list, list)
 766                 rb_erase(&path->rb_node, &priv->path_tree);
 767
 768         list_for_each_entry_safe(path, tp, &remove_list, list) {
 769                 if (path->query)
 770                         ib_sa_cancel_query(path->query_id, path->query);
 771                 spin_unlock_irqrestore(&priv->lock, flags);
 772                 netif_tx_unlock_bh(dev);
 773                 wait_for_completion(&path->done);
 774                 path_free(dev, path);
 775                 netif_tx_lock_bh(dev);
 776                 spin_lock_irqsave(&priv->lock, flags);
 777         }
 778
 779         spin_unlock_irqrestore(&priv->lock, flags);
 780         netif_tx_unlock_bh(dev);
 781 }
 782
 783 static void path_rec_completion(int status,
 784                                 struct sa_path_rec *pathrec,
 785                                 unsigned int num_prs, void *path_ptr)
 786 {
 787         struct ipoib_path *path = path_ptr;
 788         struct net_device *dev = path->dev;
 789         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 790         struct ipoib_ah *ah = NULL;
 791         struct ipoib_ah *old_ah = NULL;
 792         struct ipoib_neigh *neigh, *tn;
 793         struct sk_buff_head skqueue;
 794         struct sk_buff *skb;
 795         unsigned long flags;
 796
 797         if (!status)
 798                 ipoib_dbg(priv, "PathRec LID 0x%04x for GID %pI6\n",
 799                           be32_to_cpu(sa_path_get_dlid(pathrec)),
 800                           pathrec->dgid.raw);
 801         else
 802                 ipoib_dbg(priv, "PathRec status %d for GID %pI6\n",
 803                           status, path->pathrec.dgid.raw);
 804
 805         skb_queue_head_init(&skqueue);
 806
 807         if (!status) {
 808                 struct rdma_ah_attr av;
 809
 810                 if (!ib_init_ah_attr_from_path(priv->ca, priv->port,
 811                                                pathrec, &av, NULL)) {
 812                         ah = ipoib_create_ah(dev, priv->pd, &av);
 813                         rdma_destroy_ah_attr(&av);
 814                 }
 815         }
 816
 817         spin_lock_irqsave(&priv->lock, flags);
 818
 819         if (!IS_ERR_OR_NULL(ah)) {
 820                 /*
 821                  * pathrec.dgid is used as the database key from the LLADDR,
 822                  * it must remain unchanged even if the SA returns a different
 823                  * GID to use in the AH.
 824                  */
 825                 if (memcmp(pathrec->dgid.raw, path->pathrec.dgid.raw,
 826                            sizeof(union ib_gid))) {
 827                         ipoib_dbg(
 828                                 priv,
 829                                 "%s got PathRec for gid %pI6 while asked for %pI6\n",
 830                                 dev->name, pathrec->dgid.raw,
 831                                 path->pathrec.dgid.raw);
 832                         memcpy(pathrec->dgid.raw, path->pathrec.dgid.raw,
 833                                sizeof(union ib_gid));
 834                 }
 835
 836                 path->pathrec = *pathrec;
 837
 838                 old_ah   = path->ah;
 839                 path->ah = ah;
 840
 841                 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
 842                           ah, be32_to_cpu(sa_path_get_dlid(pathrec)),
 843                           pathrec->sl);
 844
 845                 while ((skb = __skb_dequeue(&path->queue)))
 846                         __skb_queue_tail(&skqueue, skb);
 847
 848                 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
 849                         if (neigh->ah) {
 850                                 WARN_ON(neigh->ah != old_ah);
 851                                 /*
 852                                  * Dropping the ah reference inside
 853                                  * priv->lock is safe here, because we
 854                                  * will hold one more reference from
 855                                  * the original value of path->ah (ie
 856                                  * old_ah).
 857                                  */
 858                                 ipoib_put_ah(neigh->ah);
 859                         }
 860                         kref_get(&path->ah->ref);
 861                         neigh->ah = path->ah;
 862
 863                         if (ipoib_cm_enabled(dev, neigh->daddr)) {
 864                                 if (!ipoib_cm_get(neigh))
 865                                         ipoib_cm_set(neigh, ipoib_cm_create_tx(dev,
 866                                                                                path,
 867                                                                                neigh));
 868                                 if (!ipoib_cm_get(neigh)) {
 869                                         ipoib_neigh_free(neigh);
 870                                         continue;
 871                                 }
 872                         }
 873
 874                         while ((skb = __skb_dequeue(&neigh->queue)))
 875                                 __skb_queue_tail(&skqueue, skb);
 876                 }
 877                 path->ah->valid = 1;
 878         }
 879
 880         path->query = NULL;
 881         complete(&path->done);
 882
 883         spin_unlock_irqrestore(&priv->lock, flags);
 884
 885         if (IS_ERR_OR_NULL(ah))
 886                 ipoib_del_neighs_by_gid(dev, path->pathrec.dgid.raw);
 887
 888         if (old_ah)
 889                 ipoib_put_ah(old_ah);
 890
 891         while ((skb = __skb_dequeue(&skqueue))) {
 892                 int ret;
 893                 skb->dev = dev;
 894                 ret = dev_queue_xmit(skb);
 895                 if (ret)
 896                         ipoib_warn(priv, "%s: dev_queue_xmit failed to re-queue packet, ret:%d\n",
 897                                    __func__, ret);
 898         }
 899 }
 900
 901 static void init_path_rec(struct ipoib_dev_priv *priv, struct ipoib_path *path,
 902                           void *gid)
 903 {
 904         path->dev = priv->dev;
 905
 906         if (rdma_cap_opa_ah(priv->ca, priv->port))
 907                 path->pathrec.rec_type = SA_PATH_REC_TYPE_OPA;
 908         else
 909                 path->pathrec.rec_type = SA_PATH_REC_TYPE_IB;
 910
 911         memcpy(path->pathrec.dgid.raw, gid, sizeof(union ib_gid));
 912         path->pathrec.sgid          = priv->local_gid;
 913         path->pathrec.pkey          = cpu_to_be16(priv->pkey);
 914         path->pathrec.numb_path     = 1;
 915         path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
 916 }
 917
 918 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
 919 {
 920         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 921         struct ipoib_path *path;
 922
 923         if (!priv->broadcast)
 924                 return NULL;
 925
 926         path = kzalloc(sizeof(*path), GFP_ATOMIC);
 927         if (!path)
 928                 return NULL;
 929
 930         skb_queue_head_init(&path->queue);
 931
 932         INIT_LIST_HEAD(&path->neigh_list);
 933
 934         init_path_rec(priv, path, gid);
 935
 936         return path;
 937 }
 938
 939 static int path_rec_start(struct net_device *dev,
 940                           struct ipoib_path *path)
 941 {
 942         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 943
 944         ipoib_dbg(priv, "Start path record lookup for %pI6\n",
 945                   path->pathrec.dgid.raw);
 946
 947         init_completion(&path->done);
 948
 949         path->query_id =
 950                 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
 951                                    &path->pathrec,
 952                                    IB_SA_PATH_REC_DGID          |
 953                                    IB_SA_PATH_REC_SGID          |
 954                                    IB_SA_PATH_REC_NUMB_PATH     |
 955                                    IB_SA_PATH_REC_TRAFFIC_CLASS |
 956                                    IB_SA_PATH_REC_PKEY,
 957                                    1000, GFP_ATOMIC,
 958                                    path_rec_completion,
 959                                    path, &path->query);
 960         if (path->query_id < 0) {
 961                 ipoib_warn(priv, "ib_sa_path_rec_get failed: %d\n", path->query_id);
 962                 path->query = NULL;
 963                 complete(&path->done);
 964                 return path->query_id;
 965         }
 966
 967         return 0;
 968 }
 969
 970 static void neigh_refresh_path(struct ipoib_neigh *neigh, u8 *daddr,
 971                                struct net_device *dev)
 972 {
 973         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 974         struct ipoib_path *path;
 975         unsigned long flags;
 976
 977         spin_lock_irqsave(&priv->lock, flags);
 978
 979         path = __path_find(dev, daddr + 4);
 980         if (!path)
 981                 goto out;
 982         if (!path->query)
 983                 path_rec_start(dev, path);
 984 out:
 985         spin_unlock_irqrestore(&priv->lock, flags);
 986 }
 987
 988 static struct ipoib_neigh *neigh_add_path(struct sk_buff *skb, u8 *daddr,
 989                                           struct net_device *dev)
 990 {
 991         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 992         struct rdma_netdev *rn = netdev_priv(dev);
 993         struct ipoib_path *path;
 994         struct ipoib_neigh *neigh;
 995         unsigned long flags;
 996
 997         spin_lock_irqsave(&priv->lock, flags);
 998         neigh = ipoib_neigh_alloc(daddr, dev);
 999         if (!neigh) {
1000                 spin_unlock_irqrestore(&priv->lock, flags);
1001                 ++dev->stats.tx_dropped;
1002                 dev_kfree_skb_any(skb);
1003                 return NULL;
1004         }
1005
1006         /* To avoid race condition, make sure that the
1007          * neigh will be added only once.
1008          */
1009         if (unlikely(!list_empty(&neigh->list))) {
1010                 spin_unlock_irqrestore(&priv->lock, flags);
1011                 return neigh;
1012         }
1013
1014         path = __path_find(dev, daddr + 4);
1015         if (!path) {
1016                 path = path_rec_create(dev, daddr + 4);
1017                 if (!path)
1018                         goto err_path;
1019
1020                 __path_add(dev, path);
1021         }
1022
1023         list_add_tail(&neigh->list, &path->neigh_list);
1024
1025         if (path->ah && path->ah->valid) {
1026                 kref_get(&path->ah->ref);
1027                 neigh->ah = path->ah;
1028
1029                 if (ipoib_cm_enabled(dev, neigh->daddr)) {
1030                         if (!ipoib_cm_get(neigh))
1031                                 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
1032                         if (!ipoib_cm_get(neigh)) {
1033                                 ipoib_neigh_free(neigh);
1034                                 goto err_drop;
1035                         }
1036                         if (skb_queue_len(&neigh->queue) <
1037                             IPOIB_MAX_PATH_REC_QUEUE) {
1038                                 push_pseudo_header(skb, neigh->daddr);
1039                                 __skb_queue_tail(&neigh->queue, skb);
1040                         } else {
1041                                 ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
1042                                            skb_queue_len(&neigh->queue));
1043                                 goto err_drop;
1044                         }
1045                 } else {
1046                         spin_unlock_irqrestore(&priv->lock, flags);
1047                         path->ah->last_send = rn->send(dev, skb, path->ah->ah,
1048                                                        IPOIB_QPN(daddr));
1049                         ipoib_neigh_put(neigh);
1050                         return NULL;
1051                 }
1052         } else {
1053                 neigh->ah  = NULL;
1054
1055                 if (!path->query && path_rec_start(dev, path))
1056                         goto err_path;
1057                 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
1058                         push_pseudo_header(skb, neigh->daddr);
1059                         __skb_queue_tail(&neigh->queue, skb);
1060                 } else {
1061                         goto err_drop;
1062                 }
1063         }
1064
1065         spin_unlock_irqrestore(&priv->lock, flags);
1066         ipoib_neigh_put(neigh);
1067         return NULL;
1068
1069 err_path:
1070         ipoib_neigh_free(neigh);
1071 err_drop:
1072         ++dev->stats.tx_dropped;
1073         dev_kfree_skb_any(skb);
1074
1075         spin_unlock_irqrestore(&priv->lock, flags);
1076         ipoib_neigh_put(neigh);
1077
1078         return NULL;
1079 }
1080
1081 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
1082                              struct ipoib_pseudo_header *phdr)
1083 {
1084         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1085         struct rdma_netdev *rn = netdev_priv(dev);
1086         struct ipoib_path *path;
1087         unsigned long flags;
1088
1089         spin_lock_irqsave(&priv->lock, flags);
1090
1091         /* no broadcast means that all paths are (going to be) not valid */
1092         if (!priv->broadcast)
1093                 goto drop_and_unlock;
1094
1095         path = __path_find(dev, phdr->hwaddr + 4);
1096         if (!path || !path->ah || !path->ah->valid) {
1097                 if (!path) {
1098                         path = path_rec_create(dev, phdr->hwaddr + 4);
1099                         if (!path)
1100                                 goto drop_and_unlock;
1101                         __path_add(dev, path);
1102                 } else {
1103                         /*
1104                          * make sure there are no changes in the existing
1105                          * path record
1106                          */
1107                         init_path_rec(priv, path, phdr->hwaddr + 4);
1108                 }
1109                 if (!path->query && path_rec_start(dev, path)) {
1110                         goto drop_and_unlock;
1111                 }
1112
1113                 if (skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
1114                         push_pseudo_header(skb, phdr->hwaddr);
1115                         __skb_queue_tail(&path->queue, skb);
1116                         goto unlock;
1117                 } else {
1118                         goto drop_and_unlock;
1119                 }
1120         }
1121
1122         spin_unlock_irqrestore(&priv->lock, flags);
1123         ipoib_dbg(priv, "Send unicast ARP to %08x\n",
1124                   be32_to_cpu(sa_path_get_dlid(&path->pathrec)));
1125         path->ah->last_send = rn->send(dev, skb, path->ah->ah,
1126                                        IPOIB_QPN(phdr->hwaddr));
1127         return;
1128
1129 drop_and_unlock:
1130         ++dev->stats.tx_dropped;
1131         dev_kfree_skb_any(skb);
1132 unlock:
1133         spin_unlock_irqrestore(&priv->lock, flags);
1134 }
1135
1136 static netdev_tx_t ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
1137 {
1138         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1139         struct rdma_netdev *rn = netdev_priv(dev);
1140         struct ipoib_neigh *neigh;
1141         struct ipoib_pseudo_header *phdr;
1142         struct ipoib_header *header;
1143         unsigned long flags;
1144
1145         phdr = (struct ipoib_pseudo_header *) skb->data;
1146         skb_pull(skb, sizeof(*phdr));
1147         header = (struct ipoib_header *) skb->data;
1148
1149         if (unlikely(phdr->hwaddr[4] == 0xff)) {
1150                 /* multicast, arrange "if" according to probability */
1151                 if ((header->proto != htons(ETH_P_IP)) &&
1152                     (header->proto != htons(ETH_P_IPV6)) &&
1153                     (header->proto != htons(ETH_P_ARP)) &&
1154                     (header->proto != htons(ETH_P_RARP)) &&
1155                     (header->proto != htons(ETH_P_TIPC))) {
1156                         /* ethertype not supported by IPoIB */
1157                         ++dev->stats.tx_dropped;
1158                         dev_kfree_skb_any(skb);
1159                         return NETDEV_TX_OK;
1160                 }
1161                 /* Add in the P_Key for multicast*/
1162                 phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
1163                 phdr->hwaddr[9] = priv->pkey & 0xff;
1164
1165                 neigh = ipoib_neigh_get(dev, phdr->hwaddr);
1166                 if (likely(neigh))
1167                         goto send_using_neigh;
1168                 ipoib_mcast_send(dev, phdr->hwaddr, skb);
1169                 return NETDEV_TX_OK;
1170         }
1171
1172         /* unicast, arrange "switch" according to probability */
1173         switch (header->proto) {
1174         case htons(ETH_P_IP):
1175         case htons(ETH_P_IPV6):
1176         case htons(ETH_P_TIPC):
1177                 neigh = ipoib_neigh_get(dev, phdr->hwaddr);
1178                 if (unlikely(!neigh)) {
1179                         neigh = neigh_add_path(skb, phdr->hwaddr, dev);
1180                         if (likely(!neigh))
1181                                 return NETDEV_TX_OK;
1182                 }
1183                 break;
1184         case htons(ETH_P_ARP):
1185         case htons(ETH_P_RARP):
1186                 /* for unicast ARP and RARP should always perform path find */
1187                 unicast_arp_send(skb, dev, phdr);
1188                 return NETDEV_TX_OK;
1189         default:
1190                 /* ethertype not supported by IPoIB */
1191                 ++dev->stats.tx_dropped;
1192                 dev_kfree_skb_any(skb);
1193                 return NETDEV_TX_OK;
1194         }
1195
1196 send_using_neigh:
1197         /* note we now hold a ref to neigh */
1198         if (ipoib_cm_get(neigh)) {
1199                 if (ipoib_cm_up(neigh)) {
1200                         ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
1201                         goto unref;
1202                 }
1203         } else if (neigh->ah && neigh->ah->valid) {
1204                 neigh->ah->last_send = rn->send(dev, skb, neigh->ah->ah,
1205                                                 IPOIB_QPN(phdr->hwaddr));
1206                 goto unref;
1207         } else if (neigh->ah) {
1208                 neigh_refresh_path(neigh, phdr->hwaddr, dev);
1209         }
1210
1211         if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
1212                 push_pseudo_header(skb, phdr->hwaddr);
1213                 spin_lock_irqsave(&priv->lock, flags);
1214                 __skb_queue_tail(&neigh->queue, skb);
1215                 spin_unlock_irqrestore(&priv->lock, flags);
1216         } else {
1217                 ++dev->stats.tx_dropped;
1218                 dev_kfree_skb_any(skb);
1219         }
1220
1221 unref:
1222         ipoib_neigh_put(neigh);
1223
1224         return NETDEV_TX_OK;
1225 }
1226
1227 static void ipoib_timeout(struct net_device *dev, unsigned int txqueue)
1228 {
1229         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1230         struct rdma_netdev *rn = netdev_priv(dev);
1231
1232         if (rn->tx_timeout) {
1233                 rn->tx_timeout(dev, txqueue);
1234                 return;
1235         }
1236         ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
1237                    jiffies_to_msecs(jiffies - dev_trans_start(dev)));
1238         ipoib_warn(priv,
1239                    "queue stopped %d, tx_head %u, tx_tail %u, global_tx_head %u, global_tx_tail %u\n",
1240                    netif_queue_stopped(dev), priv->tx_head, priv->tx_tail,
1241                    priv->global_tx_head, priv->global_tx_tail);
1242
1243
1244         schedule_work(&priv->tx_timeout_work);
1245 }
1246
1247 void ipoib_ib_tx_timeout_work(struct work_struct *work)
1248 {
1249         struct ipoib_dev_priv *priv = container_of(work,
1250                                                    struct ipoib_dev_priv,
1251                                                    tx_timeout_work);
1252         int err;
1253
1254         rtnl_lock();
1255         netdev_lock_ops(priv->dev);
1256
1257         if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
1258                 goto unlock;
1259
1260         ipoib_stop(priv->dev);
1261         err = ipoib_open(priv->dev);
1262         if (err) {
1263                 ipoib_warn(priv, "ipoib_open failed recovering from a tx_timeout, err(%d).\n",
1264                                 err);
1265                 goto unlock;
1266         }
1267
1268         netif_tx_wake_all_queues(priv->dev);
1269 unlock:
1270         netdev_unlock_ops(priv->dev);
1271         rtnl_unlock();
1272
1273 }
1274
1275 static int ipoib_hard_header(struct sk_buff *skb,
1276                              struct net_device *dev,
1277                              unsigned short type,
1278                              const void *daddr,
1279                              const void *saddr,
1280                              unsigned int len)
1281 {
1282         struct ipoib_header *header;
1283
1284         header = skb_push(skb, sizeof(*header));
1285
1286         header->proto = htons(type);
1287         header->reserved = 0;
1288
1289         /*
1290          * we don't rely on dst_entry structure,  always stuff the
1291          * destination address into skb hard header so we can figure out where
1292          * to send the packet later.
1293          */
1294         push_pseudo_header(skb, daddr);
1295
1296         return IPOIB_HARD_LEN;
1297 }
1298
1299 static void ipoib_set_mcast_list(struct net_device *dev)
1300 {
1301         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1302
1303         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
1304                 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
1305                 return;
1306         }
1307
1308         queue_work(priv->wq, &priv->restart_task);
1309 }
1310
1311 static int ipoib_get_iflink(const struct net_device *dev)
1312 {
1313         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1314
1315         /* parent interface */
1316         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags))
1317                 return READ_ONCE(dev->ifindex);
1318
1319         /* child/vlan interface */
1320         return READ_ONCE(priv->parent->ifindex);
1321 }
1322
1323 static u32 ipoib_addr_hash(struct ipoib_neigh_hash *htbl, u8 *daddr)
1324 {
1325         /*
1326          * Use only the address parts that contributes to spreading
1327          * The subnet prefix is not used as one can not connect to
1328          * same remote port (GUID) using the same remote QPN via two
1329          * different subnets.
1330          */
1331          /* qpn octets[1:4) & port GUID octets[12:20) */
1332         u32 *d32 = (u32 *) daddr;
1333         u32 hv;
1334
1335         hv = jhash_3words(d32[3], d32[4], IPOIB_QPN_MASK & d32[0], 0);
1336         return hv & htbl->mask;
1337 }
1338
1339 struct ipoib_neigh *ipoib_neigh_get(struct net_device *dev, u8 *daddr)
1340 {
1341         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1342         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1343         struct ipoib_neigh_hash *htbl;
1344         struct ipoib_neigh *neigh = NULL;
1345         u32 hash_val;
1346
1347         rcu_read_lock_bh();
1348
1349         htbl = rcu_dereference_bh(ntbl->htbl);
1350
1351         if (!htbl)
1352                 goto out_unlock;
1353
1354         hash_val = ipoib_addr_hash(htbl, daddr);
1355         for (neigh = rcu_dereference_bh(htbl->buckets[hash_val]);
1356              neigh != NULL;
1357              neigh = rcu_dereference_bh(neigh->hnext)) {
1358                 if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) {
1359                         /* found, take one ref on behalf of the caller */
1360                         if (!refcount_inc_not_zero(&neigh->refcnt)) {
1361                                 /* deleted */
1362                                 neigh = NULL;
1363                                 goto out_unlock;
1364                         }
1365
1366                         if (likely(skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE))
1367                                 neigh->alive = jiffies;
1368                         goto out_unlock;
1369                 }
1370         }
1371
1372 out_unlock:
1373         rcu_read_unlock_bh();
1374         return neigh;
1375 }
1376
1377 static void __ipoib_reap_neigh(struct ipoib_dev_priv *priv)
1378 {
1379         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1380         struct ipoib_neigh_hash *htbl;
1381         unsigned long neigh_obsolete;
1382         unsigned long dt;
1383         unsigned long flags;
1384         int i;
1385         LIST_HEAD(remove_list);
1386
1387         spin_lock_irqsave(&priv->lock, flags);
1388
1389         htbl = rcu_dereference_protected(ntbl->htbl,
1390                                          lockdep_is_held(&priv->lock));
1391
1392         if (!htbl)
1393                 goto out_unlock;
1394
1395         /* neigh is obsolete if it was idle for two GC periods */
1396         dt = 2 * arp_tbl.gc_interval;
1397         neigh_obsolete = jiffies - dt;
1398
1399         for (i = 0; i < htbl->size; i++) {
1400                 struct ipoib_neigh *neigh;
1401                 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1402
1403                 while ((neigh = rcu_dereference_protected(*np,
1404                                                           lockdep_is_held(&priv->lock))) != NULL) {
1405                         /* was the neigh idle for two GC periods */
1406                         if (time_after(neigh_obsolete, neigh->alive)) {
1407
1408                                 ipoib_check_and_add_mcast_sendonly(priv, neigh->daddr + 4, &remove_list);
1409
1410                                 rcu_assign_pointer(*np,
1411                                                    rcu_dereference_protected(neigh->hnext,
1412                                                                              lockdep_is_held(&priv->lock)));
1413                                 /* remove from path/mc list */
1414                                 list_del_init(&neigh->list);
1415                                 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1416                         } else {
1417                                 np = &neigh->hnext;
1418                         }
1419
1420                 }
1421         }
1422
1423 out_unlock:
1424         spin_unlock_irqrestore(&priv->lock, flags);
1425         ipoib_mcast_remove_list(&remove_list);
1426 }
1427
1428 static void ipoib_reap_neigh(struct work_struct *work)
1429 {
1430         struct ipoib_dev_priv *priv =
1431                 container_of(work, struct ipoib_dev_priv, neigh_reap_task.work);
1432
1433         __ipoib_reap_neigh(priv);
1434
1435         queue_delayed_work(priv->wq, &priv->neigh_reap_task,
1436                            arp_tbl.gc_interval);
1437 }
1438
1439
1440 static struct ipoib_neigh *ipoib_neigh_ctor(u8 *daddr,
1441                                       struct net_device *dev)
1442 {
1443         struct ipoib_neigh *neigh;
1444
1445         neigh = kzalloc(sizeof(*neigh), GFP_ATOMIC);
1446         if (!neigh)
1447                 return NULL;
1448
1449         neigh->dev = dev;
1450         memcpy(&neigh->daddr, daddr, sizeof(neigh->daddr));
1451         skb_queue_head_init(&neigh->queue);
1452         INIT_LIST_HEAD(&neigh->list);
1453         ipoib_cm_set(neigh, NULL);
1454         /* one ref on behalf of the caller */
1455         refcount_set(&neigh->refcnt, 1);
1456
1457         return neigh;
1458 }
1459
1460 struct ipoib_neigh *ipoib_neigh_alloc(u8 *daddr,
1461                                       struct net_device *dev)
1462 {
1463         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1464         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1465         struct ipoib_neigh_hash *htbl;
1466         struct ipoib_neigh *neigh;
1467         u32 hash_val;
1468
1469         htbl = rcu_dereference_protected(ntbl->htbl,
1470                                          lockdep_is_held(&priv->lock));
1471         if (!htbl) {
1472                 neigh = NULL;
1473                 goto out_unlock;
1474         }
1475
1476         /* need to add a new neigh, but maybe some other thread succeeded?
1477          * recalc hash, maybe hash resize took place so we do a search
1478          */
1479         hash_val = ipoib_addr_hash(htbl, daddr);
1480         for (neigh = rcu_dereference_protected(htbl->buckets[hash_val],
1481                                                lockdep_is_held(&priv->lock));
1482              neigh != NULL;
1483              neigh = rcu_dereference_protected(neigh->hnext,
1484                                                lockdep_is_held(&priv->lock))) {
1485                 if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) {
1486                         /* found, take one ref on behalf of the caller */
1487                         if (!refcount_inc_not_zero(&neigh->refcnt)) {
1488                                 /* deleted */
1489                                 neigh = NULL;
1490                                 break;
1491                         }
1492                         neigh->alive = jiffies;
1493                         goto out_unlock;
1494                 }
1495         }
1496
1497         neigh = ipoib_neigh_ctor(daddr, dev);
1498         if (!neigh)
1499                 goto out_unlock;
1500
1501         /* one ref on behalf of the hash table */
1502         refcount_inc(&neigh->refcnt);
1503         neigh->alive = jiffies;
1504         /* put in hash */
1505         rcu_assign_pointer(neigh->hnext,
1506                            rcu_dereference_protected(htbl->buckets[hash_val],
1507                                                      lockdep_is_held(&priv->lock)));
1508         rcu_assign_pointer(htbl->buckets[hash_val], neigh);
1509         atomic_inc(&ntbl->entries);
1510
1511 out_unlock:
1512
1513         return neigh;
1514 }
1515
1516 void ipoib_neigh_dtor(struct ipoib_neigh *neigh)
1517 {
1518         /* neigh reference count was dropprd to zero */
1519         struct net_device *dev = neigh->dev;
1520         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1521         struct sk_buff *skb;
1522         if (neigh->ah)
1523                 ipoib_put_ah(neigh->ah);
1524         while ((skb = __skb_dequeue(&neigh->queue))) {
1525                 ++dev->stats.tx_dropped;
1526                 dev_kfree_skb_any(skb);
1527         }
1528         if (ipoib_cm_get(neigh))
1529                 ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
1530         ipoib_dbg(ipoib_priv(dev),
1531                   "neigh free for %06x %pI6\n",
1532                   IPOIB_QPN(neigh->daddr),
1533                   neigh->daddr + 4);
1534         kfree(neigh);
1535         if (atomic_dec_and_test(&priv->ntbl.entries)) {
1536                 if (test_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags))
1537                         complete(&priv->ntbl.flushed);
1538         }
1539 }
1540
1541 static void ipoib_neigh_reclaim(struct rcu_head *rp)
1542 {
1543         /* Called as a result of removal from hash table */
1544         struct ipoib_neigh *neigh = container_of(rp, struct ipoib_neigh, rcu);
1545         /* note TX context may hold another ref */
1546         ipoib_neigh_put(neigh);
1547 }
1548
1549 void ipoib_neigh_free(struct ipoib_neigh *neigh)
1550 {
1551         struct net_device *dev = neigh->dev;
1552         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1553         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1554         struct ipoib_neigh_hash *htbl;
1555         struct ipoib_neigh __rcu **np;
1556         struct ipoib_neigh *n;
1557         u32 hash_val;
1558
1559         htbl = rcu_dereference_protected(ntbl->htbl,
1560                                         lockdep_is_held(&priv->lock));
1561         if (!htbl)
1562                 return;
1563
1564         hash_val = ipoib_addr_hash(htbl, neigh->daddr);
1565         np = &htbl->buckets[hash_val];
1566         for (n = rcu_dereference_protected(*np,
1567                                             lockdep_is_held(&priv->lock));
1568              n != NULL;
1569              n = rcu_dereference_protected(*np,
1570                                         lockdep_is_held(&priv->lock))) {
1571                 if (n == neigh) {
1572                         /* found */
1573                         rcu_assign_pointer(*np,
1574                                            rcu_dereference_protected(neigh->hnext,
1575                                                                      lockdep_is_held(&priv->lock)));
1576                         /* remove from parent list */
1577                         list_del_init(&neigh->list);
1578                         call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1579                         return;
1580                 } else {
1581                         np = &n->hnext;
1582                 }
1583         }
1584 }
1585
1586 static int ipoib_neigh_hash_init(struct ipoib_dev_priv *priv)
1587 {
1588         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1589         struct ipoib_neigh_hash *htbl;
1590         struct ipoib_neigh __rcu **buckets;
1591         u32 size;
1592
1593         clear_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags);
1594         ntbl->htbl = NULL;
1595         htbl = kzalloc(sizeof(*htbl), GFP_KERNEL);
1596         if (!htbl)
1597                 return -ENOMEM;
1598         size = roundup_pow_of_two(arp_tbl.gc_thresh3);
1599         buckets = kvcalloc(size, sizeof(*buckets), GFP_KERNEL);
1600         if (!buckets) {
1601                 kfree(htbl);
1602                 return -ENOMEM;
1603         }
1604         htbl->size = size;
1605         htbl->mask = (size - 1);
1606         htbl->buckets = buckets;
1607         RCU_INIT_POINTER(ntbl->htbl, htbl);
1608         htbl->ntbl = ntbl;
1609         atomic_set(&ntbl->entries, 0);
1610
1611         /* start garbage collection */
1612         queue_delayed_work(priv->wq, &priv->neigh_reap_task,
1613                            arp_tbl.gc_interval);
1614
1615         return 0;
1616 }
1617
1618 static void neigh_hash_free_rcu(struct rcu_head *head)
1619 {
1620         struct ipoib_neigh_hash *htbl = container_of(head,
1621                                                     struct ipoib_neigh_hash,
1622                                                     rcu);
1623         struct ipoib_neigh __rcu **buckets = htbl->buckets;
1624         struct ipoib_neigh_table *ntbl = htbl->ntbl;
1625
1626         kvfree(buckets);
1627         kfree(htbl);
1628         complete(&ntbl->deleted);
1629 }
1630
1631 void ipoib_del_neighs_by_gid(struct net_device *dev, u8 *gid)
1632 {
1633         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1634         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1635         struct ipoib_neigh_hash *htbl;
1636         unsigned long flags;
1637         int i;
1638
1639         /* remove all neigh connected to a given path or mcast */
1640         spin_lock_irqsave(&priv->lock, flags);
1641
1642         htbl = rcu_dereference_protected(ntbl->htbl,
1643                                          lockdep_is_held(&priv->lock));
1644
1645         if (!htbl)
1646                 goto out_unlock;
1647
1648         for (i = 0; i < htbl->size; i++) {
1649                 struct ipoib_neigh *neigh;
1650                 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1651
1652                 while ((neigh = rcu_dereference_protected(*np,
1653                                                           lockdep_is_held(&priv->lock))) != NULL) {
1654                         /* delete neighs belong to this parent */
1655                         if (!memcmp(gid, neigh->daddr + 4, sizeof (union ib_gid))) {
1656                                 rcu_assign_pointer(*np,
1657                                                    rcu_dereference_protected(neigh->hnext,
1658                                                                              lockdep_is_held(&priv->lock)));
1659                                 /* remove from parent list */
1660                                 list_del_init(&neigh->list);
1661                                 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1662                         } else {
1663                                 np = &neigh->hnext;
1664                         }
1665
1666                 }
1667         }
1668 out_unlock:
1669         spin_unlock_irqrestore(&priv->lock, flags);
1670 }
1671
1672 static void ipoib_flush_neighs(struct ipoib_dev_priv *priv)
1673 {
1674         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1675         struct ipoib_neigh_hash *htbl;
1676         unsigned long flags;
1677         int i, wait_flushed = 0;
1678
1679         init_completion(&priv->ntbl.flushed);
1680         set_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags);
1681
1682         spin_lock_irqsave(&priv->lock, flags);
1683
1684         htbl = rcu_dereference_protected(ntbl->htbl,
1685                                         lockdep_is_held(&priv->lock));
1686         if (!htbl)
1687                 goto out_unlock;
1688
1689         wait_flushed = atomic_read(&priv->ntbl.entries);
1690         if (!wait_flushed)
1691                 goto free_htbl;
1692
1693         for (i = 0; i < htbl->size; i++) {
1694                 struct ipoib_neigh *neigh;
1695                 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1696
1697                 while ((neigh = rcu_dereference_protected(*np,
1698                                        lockdep_is_held(&priv->lock))) != NULL) {
1699                         rcu_assign_pointer(*np,
1700                                            rcu_dereference_protected(neigh->hnext,
1701                                                                      lockdep_is_held(&priv->lock)));
1702                         /* remove from path/mc list */
1703                         list_del_init(&neigh->list);
1704                         call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1705                 }
1706         }
1707
1708 free_htbl:
1709         rcu_assign_pointer(ntbl->htbl, NULL);
1710         call_rcu(&htbl->rcu, neigh_hash_free_rcu);
1711
1712 out_unlock:
1713         spin_unlock_irqrestore(&priv->lock, flags);
1714         if (wait_flushed)
1715                 wait_for_completion(&priv->ntbl.flushed);
1716 }
1717
1718 static void ipoib_neigh_hash_uninit(struct net_device *dev)
1719 {
1720         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1721
1722         ipoib_dbg(priv, "%s\n", __func__);
1723         init_completion(&priv->ntbl.deleted);
1724
1725         cancel_delayed_work_sync(&priv->neigh_reap_task);
1726
1727         ipoib_flush_neighs(priv);
1728
1729         wait_for_completion(&priv->ntbl.deleted);
1730 }
1731
1732 static void ipoib_napi_add(struct net_device *dev)
1733 {
1734         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1735
1736         netif_napi_add_weight(dev, &priv->recv_napi, ipoib_rx_poll,
1737                               IPOIB_NUM_WC);
1738         netif_napi_add_weight(dev, &priv->send_napi, ipoib_tx_poll,
1739                               MAX_SEND_CQE);
1740 }
1741
1742 static void ipoib_napi_del(struct net_device *dev)
1743 {
1744         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1745
1746         netif_napi_del(&priv->recv_napi);
1747         netif_napi_del(&priv->send_napi);
1748 }
1749
1750 static void ipoib_dev_uninit_default(struct net_device *dev)
1751 {
1752         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1753
1754         ipoib_transport_dev_cleanup(dev);
1755
1756         ipoib_napi_del(dev);
1757
1758         ipoib_cm_dev_cleanup(dev);
1759
1760         kfree(priv->rx_ring);
1761         vfree(priv->tx_ring);
1762
1763         priv->rx_ring = NULL;
1764         priv->tx_ring = NULL;
1765 }
1766
1767 static int ipoib_dev_init_default(struct net_device *dev)
1768 {
1769         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1770         u8 addr_mod[3];
1771
1772         ipoib_napi_add(dev);
1773
1774         /* Allocate RX/TX "rings" to hold queued skbs */
1775         priv->rx_ring = kcalloc(ipoib_recvq_size,
1776                                        sizeof(*priv->rx_ring),
1777                                        GFP_KERNEL);
1778         if (!priv->rx_ring)
1779                 goto out;
1780
1781         priv->tx_ring = vzalloc(array_size(ipoib_sendq_size,
1782                                            sizeof(*priv->tx_ring)));
1783         if (!priv->tx_ring) {
1784                 pr_warn("%s: failed to allocate TX ring (%d entries)\n",
1785                         priv->ca->name, ipoib_sendq_size);
1786                 goto out_rx_ring_cleanup;
1787         }
1788
1789         /* priv->tx_head, tx_tail and global_tx_tail/head are already 0 */
1790
1791         if (ipoib_transport_dev_init(dev, priv->ca)) {
1792                 pr_warn("%s: ipoib_transport_dev_init failed\n",
1793                         priv->ca->name);
1794                 goto out_tx_ring_cleanup;
1795         }
1796
1797         /* after qp created set dev address */
1798         addr_mod[0] = (priv->qp->qp_num >> 16) & 0xff;
1799         addr_mod[1] = (priv->qp->qp_num >>  8) & 0xff;
1800         addr_mod[2] = (priv->qp->qp_num) & 0xff;
1801         dev_addr_mod(priv->dev, 1, addr_mod, sizeof(addr_mod));
1802
1803         return 0;
1804
1805 out_tx_ring_cleanup:
1806         vfree(priv->tx_ring);
1807
1808 out_rx_ring_cleanup:
1809         kfree(priv->rx_ring);
1810
1811 out:
1812         ipoib_napi_del(dev);
1813         return -ENOMEM;
1814 }
1815
1816 static int ipoib_ioctl(struct net_device *dev, struct ifreq *ifr,
1817                        int cmd)
1818 {
1819         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1820
1821         if (!priv->rn_ops->ndo_eth_ioctl)
1822                 return -EOPNOTSUPP;
1823
1824         return priv->rn_ops->ndo_eth_ioctl(dev, ifr, cmd);
1825 }
1826
1827 static int ipoib_dev_init(struct net_device *dev)
1828 {
1829         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1830         int ret = -ENOMEM;
1831
1832         priv->qp = NULL;
1833
1834         /*
1835          * the various IPoIB tasks assume they will never race against
1836          * themselves, so always use a single thread workqueue
1837          */
1838         priv->wq = alloc_ordered_workqueue("ipoib_wq", WQ_MEM_RECLAIM);
1839         if (!priv->wq) {
1840                 pr_warn("%s: failed to allocate device WQ\n", dev->name);
1841                 goto out;
1842         }
1843
1844         /* create pd, which used both for control and datapath*/
1845         priv->pd = ib_alloc_pd(priv->ca, 0);
1846         if (IS_ERR(priv->pd)) {
1847                 pr_warn("%s: failed to allocate PD\n", priv->ca->name);
1848                 goto clean_wq;
1849         }
1850
1851         ret = priv->rn_ops->ndo_init(dev);
1852         if (ret) {
1853                 pr_warn("%s failed to init HW resource\n", dev->name);
1854                 goto out_free_pd;
1855         }
1856
1857         ret = ipoib_neigh_hash_init(priv);
1858         if (ret) {
1859                 pr_warn("%s failed to init neigh hash\n", dev->name);
1860                 goto out_dev_uninit;
1861         }
1862
1863         if (dev->flags & IFF_UP) {
1864                 if (ipoib_ib_dev_open(dev)) {
1865                         pr_warn("%s failed to open device\n", dev->name);
1866                         ret = -ENODEV;
1867                         goto out_hash_uninit;
1868                 }
1869         }
1870
1871         return 0;
1872
1873 out_hash_uninit:
1874         ipoib_neigh_hash_uninit(dev);
1875
1876 out_dev_uninit:
1877         ipoib_ib_dev_cleanup(dev);
1878
1879 out_free_pd:
1880         if (priv->pd) {
1881                 ib_dealloc_pd(priv->pd);
1882                 priv->pd = NULL;
1883         }
1884
1885 clean_wq:
1886         if (priv->wq) {
1887                 destroy_workqueue(priv->wq);
1888                 priv->wq = NULL;
1889         }
1890
1891 out:
1892         return ret;
1893 }
1894
1895 /*
1896  * This must be called before doing an unregister_netdev on a parent device to
1897  * shutdown the IB event handler.
1898  */
1899 static void ipoib_parent_unregister_pre(struct net_device *ndev)
1900 {
1901         struct ipoib_dev_priv *priv = ipoib_priv(ndev);
1902
1903         /*
1904          * ipoib_set_mac checks netif_running before pushing work, clearing
1905          * running ensures the it will not add more work.
1906          */
1907         rtnl_lock();
1908         dev_change_flags(priv->dev, priv->dev->flags & ~IFF_UP, NULL);
1909         rtnl_unlock();
1910
1911         /* ipoib_event() cannot be running once this returns */
1912         ib_unregister_event_handler(&priv->event_handler);
1913
1914         /*
1915          * Work on the queue grabs the rtnl lock, so this cannot be done while
1916          * also holding it.
1917          */
1918         flush_workqueue(ipoib_workqueue);
1919 }
1920
1921 static void ipoib_set_dev_features(struct ipoib_dev_priv *priv)
1922 {
1923         priv->hca_caps = priv->ca->attrs.device_cap_flags;
1924         priv->kernel_caps = priv->ca->attrs.kernel_cap_flags;
1925
1926         if (priv->hca_caps & IB_DEVICE_UD_IP_CSUM) {
1927                 priv->dev->hw_features |= NETIF_F_IP_CSUM | NETIF_F_RXCSUM;
1928
1929                 if (priv->kernel_caps & IBK_UD_TSO)
1930                         priv->dev->hw_features |= NETIF_F_TSO;
1931
1932                 priv->dev->features |= priv->dev->hw_features;
1933         }
1934 }
1935
1936 static int ipoib_parent_init(struct net_device *ndev)
1937 {
1938         struct ipoib_dev_priv *priv = ipoib_priv(ndev);
1939         struct ib_port_attr attr;
1940         int result;
1941
1942         result = ib_query_port(priv->ca, priv->port, &attr);
1943         if (result) {
1944                 pr_warn("%s: ib_query_port %d failed\n", priv->ca->name,
1945                         priv->port);
1946                 return result;
1947         }
1948         priv->max_ib_mtu = rdma_mtu_from_attr(priv->ca, priv->port, &attr);
1949
1950         result = ib_query_pkey(priv->ca, priv->port, 0, &priv->pkey);
1951         if (result) {
1952                 pr_warn("%s: ib_query_pkey port %d failed (ret = %d)\n",
1953                         priv->ca->name, priv->port, result);
1954                 return result;
1955         }
1956
1957         result = rdma_query_gid(priv->ca, priv->port, 0, &priv->local_gid);
1958         if (result) {
1959                 pr_warn("%s: rdma_query_gid port %d failed (ret = %d)\n",
1960                         priv->ca->name, priv->port, result);
1961                 return result;
1962         }
1963         dev_addr_mod(priv->dev, 4, priv->local_gid.raw, sizeof(union ib_gid));
1964
1965         SET_NETDEV_DEV(priv->dev, priv->ca->dev.parent);
1966         priv->dev->dev_port = priv->port - 1;
1967         /* Let's set this one too for backwards compatibility. */
1968         priv->dev->dev_id = priv->port - 1;
1969
1970         return 0;
1971 }
1972
1973 static void ipoib_child_init(struct net_device *ndev)
1974 {
1975         struct ipoib_dev_priv *priv = ipoib_priv(ndev);
1976         struct ipoib_dev_priv *ppriv = ipoib_priv(priv->parent);
1977
1978         priv->max_ib_mtu = ppriv->max_ib_mtu;
1979         set_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags);
1980         if (memchr_inv(priv->dev->dev_addr, 0, INFINIBAND_ALEN))
1981                 memcpy(&priv->local_gid, priv->dev->dev_addr + 4,
1982                        sizeof(priv->local_gid));
1983         else {
1984                 __dev_addr_set(priv->dev, ppriv->dev->dev_addr,
1985                                INFINIBAND_ALEN);
1986                 memcpy(&priv->local_gid, &ppriv->local_gid,
1987                        sizeof(priv->local_gid));
1988         }
1989 }
1990
1991 static int ipoib_ndo_init(struct net_device *ndev)
1992 {
1993         struct ipoib_dev_priv *priv = ipoib_priv(ndev);
1994         int rc;
1995         struct rdma_netdev *rn = netdev_priv(ndev);
1996
1997         if (priv->parent) {
1998                 ipoib_child_init(ndev);
1999         } else {
2000                 rc = ipoib_parent_init(ndev);
2001                 if (rc)
2002                         return rc;
2003         }
2004
2005         /* MTU will be reset when mcast join happens */
2006         ndev->mtu = IPOIB_UD_MTU(priv->max_ib_mtu);
2007         priv->mcast_mtu = priv->admin_mtu = ndev->mtu;
2008         rn->mtu = priv->mcast_mtu;
2009         ndev->max_mtu = IPOIB_CM_MTU;
2010
2011         ndev->neigh_priv_len = sizeof(struct ipoib_neigh);
2012
2013         /*
2014          * Set the full membership bit, so that we join the right
2015          * broadcast group, etc.
2016          */
2017         priv->pkey |= 0x8000;
2018
2019         ndev->broadcast[8] = priv->pkey >> 8;
2020         ndev->broadcast[9] = priv->pkey & 0xff;
2021         set_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags);
2022
2023         ipoib_set_dev_features(priv);
2024
2025         rc = ipoib_dev_init(ndev);
2026         if (rc) {
2027                 pr_warn("%s: failed to initialize device: %s port %d (ret = %d)\n",
2028                         priv->ca->name, priv->dev->name, priv->port, rc);
2029                 return rc;
2030         }
2031
2032         if (priv->parent) {
2033                 struct ipoib_dev_priv *ppriv = ipoib_priv(priv->parent);
2034
2035                 dev_hold(priv->parent);
2036
2037                 netdev_lock(priv->parent);
2038                 list_add_tail(&priv->list, &ppriv->child_intfs);
2039                 netdev_unlock(priv->parent);
2040         }
2041
2042         return 0;
2043 }
2044
2045 static void ipoib_ndo_uninit(struct net_device *dev)
2046 {
2047         struct ipoib_dev_priv *priv = ipoib_priv(dev);
2048
2049         /*
2050          * ipoib_remove_one guarantees the children are removed before the
2051          * parent, and that is the only place where a parent can be removed.
2052          */
2053         WARN_ON(!list_empty(&priv->child_intfs));
2054
2055         if (priv->parent) {
2056                 struct ipoib_dev_priv *ppriv = ipoib_priv(priv->parent);
2057
2058                 netdev_lock(ppriv->dev);
2059                 list_del(&priv->list);
2060                 netdev_unlock(ppriv->dev);
2061         }
2062
2063         ipoib_neigh_hash_uninit(dev);
2064
2065         ipoib_ib_dev_cleanup(dev);
2066
2067         /* no more works over the priv->wq */
2068         if (priv->wq) {
2069                 /* See ipoib_mcast_carrier_on_task() */
2070                 WARN_ON(test_bit(IPOIB_FLAG_OPER_UP, &priv->flags));
2071                 destroy_workqueue(priv->wq);
2072                 priv->wq = NULL;
2073         }
2074
2075         dev_put(priv->parent);
2076 }
2077
2078 static int ipoib_set_vf_link_state(struct net_device *dev, int vf, int link_state)
2079 {
2080         struct ipoib_dev_priv *priv = ipoib_priv(dev);
2081
2082         return ib_set_vf_link_state(priv->ca, vf, priv->port, link_state);
2083 }
2084
2085 static int ipoib_get_vf_config(struct net_device *dev, int vf,
2086                                struct ifla_vf_info *ivf)
2087 {
2088         struct ipoib_dev_priv *priv = ipoib_priv(dev);
2089         int err;
2090
2091         err = ib_get_vf_config(priv->ca, vf, priv->port, ivf);
2092         if (err)
2093                 return err;
2094
2095         ivf->vf = vf;
2096         memcpy(ivf->mac, dev->dev_addr, dev->addr_len);
2097
2098         return 0;
2099 }
2100
2101 static int ipoib_set_vf_guid(struct net_device *dev, int vf, u64 guid, int type)
2102 {
2103         struct ipoib_dev_priv *priv = ipoib_priv(dev);
2104
2105         if (type != IFLA_VF_IB_NODE_GUID && type != IFLA_VF_IB_PORT_GUID)
2106                 return -EINVAL;
2107
2108         return ib_set_vf_guid(priv->ca, vf, priv->port, guid, type);
2109 }
2110
2111 static int ipoib_get_vf_guid(struct net_device *dev, int vf,
2112                              struct ifla_vf_guid *node_guid,
2113                              struct ifla_vf_guid *port_guid)
2114 {
2115         struct ipoib_dev_priv *priv = ipoib_priv(dev);
2116
2117         return ib_get_vf_guid(priv->ca, vf, priv->port, node_guid, port_guid);
2118 }
2119
2120 static int ipoib_get_vf_stats(struct net_device *dev, int vf,
2121                               struct ifla_vf_stats *vf_stats)
2122 {
2123         struct ipoib_dev_priv *priv = ipoib_priv(dev);
2124
2125         return ib_get_vf_stats(priv->ca, vf, priv->port, vf_stats);
2126 }
2127
2128 static const struct header_ops ipoib_header_ops = {
2129         .create = ipoib_hard_header,
2130 };
2131
2132 static const struct net_device_ops ipoib_netdev_ops_pf = {
2133         .ndo_init                = ipoib_ndo_init,
2134         .ndo_uninit              = ipoib_ndo_uninit,
2135         .ndo_open                = ipoib_open,
2136         .ndo_stop                = ipoib_stop,
2137         .ndo_change_mtu          = ipoib_change_mtu,
2138         .ndo_fix_features        = ipoib_fix_features,
2139         .ndo_start_xmit          = ipoib_start_xmit,
2140         .ndo_tx_timeout          = ipoib_timeout,
2141         .ndo_set_rx_mode         = ipoib_set_mcast_list,
2142         .ndo_get_iflink          = ipoib_get_iflink,
2143         .ndo_set_vf_link_state   = ipoib_set_vf_link_state,
2144         .ndo_get_vf_config       = ipoib_get_vf_config,
2145         .ndo_get_vf_stats        = ipoib_get_vf_stats,
2146         .ndo_get_vf_guid         = ipoib_get_vf_guid,
2147         .ndo_set_vf_guid         = ipoib_set_vf_guid,
2148         .ndo_set_mac_address     = ipoib_set_mac,
2149         .ndo_get_stats64         = ipoib_get_stats,
2150         .ndo_eth_ioctl           = ipoib_ioctl,
2151 };
2152
2153 static const struct net_device_ops ipoib_netdev_ops_vf = {
2154         .ndo_init                = ipoib_ndo_init,
2155         .ndo_uninit              = ipoib_ndo_uninit,
2156         .ndo_open                = ipoib_open,
2157         .ndo_stop                = ipoib_stop,
2158         .ndo_change_mtu          = ipoib_change_mtu,
2159         .ndo_fix_features        = ipoib_fix_features,
2160         .ndo_start_xmit          = ipoib_start_xmit,
2161         .ndo_tx_timeout          = ipoib_timeout,
2162         .ndo_set_rx_mode         = ipoib_set_mcast_list,
2163         .ndo_get_iflink          = ipoib_get_iflink,
2164         .ndo_get_stats64         = ipoib_get_stats,
2165         .ndo_eth_ioctl           = ipoib_ioctl,
2166 };
2167
2168 static const struct net_device_ops ipoib_netdev_default_pf = {
2169         .ndo_init                = ipoib_dev_init_default,
2170         .ndo_uninit              = ipoib_dev_uninit_default,
2171         .ndo_open                = ipoib_ib_dev_open_default,
2172         .ndo_stop                = ipoib_ib_dev_stop_default,
2173 };
2174
2175 void ipoib_setup_common(struct net_device *dev)
2176 {
2177         dev->header_ops          = &ipoib_header_ops;
2178         dev->netdev_ops          = &ipoib_netdev_default_pf;
2179
2180         ipoib_set_ethtool_ops(dev);
2181
2182         dev->watchdog_timeo      = 10 * HZ;
2183
2184         dev->flags              |= IFF_BROADCAST | IFF_MULTICAST;
2185
2186         dev->hard_header_len     = IPOIB_HARD_LEN;
2187         dev->addr_len            = INFINIBAND_ALEN;
2188         dev->type                = ARPHRD_INFINIBAND;
2189         dev->tx_queue_len        = DEFAULT_TX_QUEUE_LEN;
2190         dev->features            = (NETIF_F_VLAN_CHALLENGED     |
2191                                     NETIF_F_HIGHDMA);
2192         netif_keep_dst(dev);
2193
2194         memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
2195
2196         /*
2197          * unregister_netdev always frees the netdev, we use this mode
2198          * consistently to unify all the various unregister paths, including
2199          * those connected to rtnl_link_ops which require it.
2200          */
2201         dev->needs_free_netdev = true;
2202 }
2203
2204 static void ipoib_build_priv(struct net_device *dev)
2205 {
2206         struct ipoib_dev_priv *priv = ipoib_priv(dev);
2207
2208         priv->dev = dev;
2209         spin_lock_init(&priv->lock);
2210         mutex_init(&priv->mcast_mutex);
2211
2212         INIT_LIST_HEAD(&priv->path_list);
2213         INIT_LIST_HEAD(&priv->child_intfs);
2214         INIT_LIST_HEAD(&priv->dead_ahs);
2215         INIT_LIST_HEAD(&priv->multicast_list);
2216
2217         INIT_DELAYED_WORK(&priv->mcast_task,   ipoib_mcast_join_task);
2218         INIT_WORK(&priv->carrier_on_task, ipoib_mcast_carrier_on_task);
2219         INIT_WORK(&priv->reschedule_napi_work, ipoib_napi_schedule_work);
2220         INIT_WORK(&priv->flush_light,   ipoib_ib_dev_flush_light);
2221         INIT_WORK(&priv->flush_normal,   ipoib_ib_dev_flush_normal);
2222         INIT_WORK(&priv->flush_heavy,   ipoib_ib_dev_flush_heavy);
2223         INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
2224         INIT_WORK(&priv->tx_timeout_work, ipoib_ib_tx_timeout_work);
2225         INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
2226         INIT_DELAYED_WORK(&priv->neigh_reap_task, ipoib_reap_neigh);
2227 }
2228
2229 static struct net_device *ipoib_alloc_netdev(struct ib_device *hca, u32 port,
2230                                              const char *name)
2231 {
2232         struct net_device *dev;
2233
2234         dev = rdma_alloc_netdev(hca, port, RDMA_NETDEV_IPOIB, name,
2235                                 NET_NAME_UNKNOWN, ipoib_setup_common);
2236         if (!IS_ERR(dev) || PTR_ERR(dev) != -EOPNOTSUPP)
2237                 return dev;
2238
2239         dev = alloc_netdev(sizeof(struct rdma_netdev), name, NET_NAME_UNKNOWN,
2240                            ipoib_setup_common);
2241         if (!dev)
2242                 return ERR_PTR(-ENOMEM);
2243         return dev;
2244 }
2245
2246 int ipoib_intf_init(struct ib_device *hca, u32 port, const char *name,
2247                     struct net_device *dev)
2248 {
2249         struct rdma_netdev *rn = netdev_priv(dev);
2250         struct ipoib_dev_priv *priv;
2251         int rc;
2252
2253         priv = kzalloc(sizeof(*priv), GFP_KERNEL);
2254         if (!priv)
2255                 return -ENOMEM;
2256
2257         priv->ca = hca;
2258         priv->port = port;
2259
2260         rc = rdma_init_netdev(hca, port, RDMA_NETDEV_IPOIB, name,
2261                               NET_NAME_UNKNOWN, ipoib_setup_common, dev);
2262         if (rc) {
2263                 if (rc != -EOPNOTSUPP)
2264                         goto out;
2265
2266                 rn->send = ipoib_send;
2267                 rn->attach_mcast = ipoib_mcast_attach;
2268                 rn->detach_mcast = ipoib_mcast_detach;
2269                 rn->hca = hca;
2270
2271                 rc = netif_set_real_num_tx_queues(dev, 1);
2272                 if (rc)
2273                         goto out;
2274
2275                 rc = netif_set_real_num_rx_queues(dev, 1);
2276                 if (rc)
2277                         goto out;
2278         }
2279
2280         priv->rn_ops = dev->netdev_ops;
2281
2282         if (hca->attrs.kernel_cap_flags & IBK_VIRTUAL_FUNCTION)
2283                 dev->netdev_ops = &ipoib_netdev_ops_vf;
2284         else
2285                 dev->netdev_ops = &ipoib_netdev_ops_pf;
2286
2287         rn->clnt_priv = priv;
2288         /*
2289          * Only the child register_netdev flows can handle priv_destructor
2290          * being set, so we force it to NULL here and handle manually until it
2291          * is safe to turn on.
2292          */
2293         priv->next_priv_destructor = dev->priv_destructor;
2294         dev->priv_destructor = NULL;
2295
2296         ipoib_build_priv(dev);
2297
2298         return 0;
2299
2300 out:
2301         kfree(priv);
2302         return rc;
2303 }
2304
2305 struct net_device *ipoib_intf_alloc(struct ib_device *hca, u32 port,
2306                                     const char *name)
2307 {
2308         struct net_device *dev;
2309         int rc;
2310
2311         dev = ipoib_alloc_netdev(hca, port, name);
2312         if (IS_ERR(dev))
2313                 return dev;
2314
2315         rc = ipoib_intf_init(hca, port, name, dev);
2316         if (rc) {
2317                 free_netdev(dev);
2318                 return ERR_PTR(rc);
2319         }
2320
2321         /*
2322          * Upon success the caller must ensure ipoib_intf_free is called or
2323          * register_netdevice succeed'd and priv_destructor is set to
2324          * ipoib_intf_free.
2325          */
2326         return dev;
2327 }
2328
2329 void ipoib_intf_free(struct net_device *dev)
2330 {
2331         struct ipoib_dev_priv *priv = ipoib_priv(dev);
2332         struct rdma_netdev *rn = netdev_priv(dev);
2333
2334         dev->priv_destructor = priv->next_priv_destructor;
2335         if (dev->priv_destructor)
2336                 dev->priv_destructor(dev);
2337
2338         /*
2339          * There are some error flows around register_netdev failing that may
2340          * attempt to call priv_destructor twice, prevent that from happening.
2341          */
2342         dev->priv_destructor = NULL;
2343
2344         /* unregister/destroy is very complicated. Make bugs more obvious. */
2345         rn->clnt_priv = NULL;
2346
2347         kfree(priv);
2348 }
2349
2350 static ssize_t pkey_show(struct device *dev, struct device_attribute *attr,
2351                          char *buf)
2352 {
2353         struct net_device *ndev = to_net_dev(dev);
2354         struct ipoib_dev_priv *priv = ipoib_priv(ndev);
2355
2356         return sysfs_emit(buf, "0x%04x\n", priv->pkey);
2357 }
2358 static DEVICE_ATTR_RO(pkey);
2359
2360 static ssize_t umcast_show(struct device *dev, struct device_attribute *attr,
2361                            char *buf)
2362 {
2363         struct net_device *ndev = to_net_dev(dev);
2364         struct ipoib_dev_priv *priv = ipoib_priv(ndev);
2365
2366         return sysfs_emit(buf, "%d\n",
2367                           test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
2368 }
2369
2370 void ipoib_set_umcast(struct net_device *ndev, int umcast_val)
2371 {
2372         struct ipoib_dev_priv *priv = ipoib_priv(ndev);
2373
2374         if (umcast_val > 0) {
2375                 set_bit(IPOIB_FLAG_UMCAST, &priv->flags);
2376                 ipoib_warn(priv, "ignoring multicast groups joined directly "
2377                                 "by userspace\n");
2378         } else
2379                 clear_bit(IPOIB_FLAG_UMCAST, &priv->flags);
2380 }
2381
2382 static ssize_t umcast_store(struct device *dev, struct device_attribute *attr,
2383                             const char *buf, size_t count)
2384 {
2385         unsigned long umcast_val = simple_strtoul(buf, NULL, 0);
2386
2387         ipoib_set_umcast(to_net_dev(dev), umcast_val);
2388
2389         return count;
2390 }
2391 static DEVICE_ATTR_RW(umcast);
2392
2393 int ipoib_add_umcast_attr(struct net_device *dev)
2394 {
2395         return device_create_file(&dev->dev, &dev_attr_umcast);
2396 }
2397
2398 static void set_base_guid(struct ipoib_dev_priv *priv, union ib_gid *gid)
2399 {
2400         struct ipoib_dev_priv *child_priv;
2401         struct net_device *netdev = priv->dev;
2402
2403         netif_addr_lock_bh(netdev);
2404
2405         memcpy(&priv->local_gid.global.interface_id,
2406                &gid->global.interface_id,
2407                sizeof(gid->global.interface_id));
2408         dev_addr_mod(netdev, 4, (u8 *)&priv->local_gid, sizeof(priv->local_gid));
2409         clear_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags);
2410
2411         netif_addr_unlock_bh(netdev);
2412
2413         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
2414                 netdev_lock_ops_to_full(priv->dev);
2415                 list_for_each_entry(child_priv, &priv->child_intfs, list)
2416                         set_base_guid(child_priv, gid);
2417                 netdev_unlock_full_to_ops(priv->dev);
2418         }
2419 }
2420
2421 static int ipoib_check_lladdr(struct net_device *dev,
2422                               struct sockaddr_storage *ss)
2423 {
2424         union ib_gid *gid = (union ib_gid *)(ss->__data + 4);
2425         int ret = 0;
2426
2427         netif_addr_lock_bh(dev);
2428
2429         /* Make sure the QPN, reserved and subnet prefix match the current
2430          * lladdr, it also makes sure the lladdr is unicast.
2431          */
2432         if (memcmp(dev->dev_addr, ss->__data,
2433                    4 + sizeof(gid->global.subnet_prefix)) ||
2434             gid->global.interface_id == 0)
2435                 ret = -EINVAL;
2436
2437         netif_addr_unlock_bh(dev);
2438
2439         return ret;
2440 }
2441
2442 static int ipoib_set_mac(struct net_device *dev, void *addr)
2443 {
2444         struct ipoib_dev_priv *priv = ipoib_priv(dev);
2445         struct sockaddr_storage *ss = addr;
2446         int ret;
2447
2448         if (!(dev->priv_flags & IFF_LIVE_ADDR_CHANGE) && netif_running(dev))
2449                 return -EBUSY;
2450
2451         ret = ipoib_check_lladdr(dev, ss);
2452         if (ret)
2453                 return ret;
2454
2455         set_base_guid(priv, (union ib_gid *)(ss->__data + 4));
2456
2457         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
2458                 struct ipoib_dev_priv *cpriv;
2459
2460                 netdev_lock_ops_to_full(dev);
2461                 list_for_each_entry(cpriv, &priv->child_intfs, list)
2462                         queue_work(ipoib_workqueue, &cpriv->flush_light);
2463                 netdev_unlock_full_to_ops(dev);
2464         }
2465         queue_work(ipoib_workqueue, &priv->flush_light);
2466
2467         return 0;
2468 }
2469
2470 static ssize_t create_child_store(struct device *dev,
2471                                   struct device_attribute *attr,
2472                                   const char *buf, size_t count)
2473 {
2474         int pkey;
2475         int ret;
2476
2477         if (sscanf(buf, "%i", &pkey) != 1)
2478                 return -EINVAL;
2479
2480         if (pkey <= 0 || pkey > 0xffff || pkey == 0x8000)
2481                 return -EINVAL;
2482
2483         ret = ipoib_vlan_add(to_net_dev(dev), pkey);
2484
2485         return ret ? ret : count;
2486 }
2487 static DEVICE_ATTR_WO(create_child);
2488
2489 static ssize_t delete_child_store(struct device *dev,
2490                                   struct device_attribute *attr,
2491                                   const char *buf, size_t count)
2492 {
2493         int pkey;
2494         int ret;
2495
2496         if (sscanf(buf, "%i", &pkey) != 1)
2497                 return -EINVAL;
2498
2499         if (pkey < 0 || pkey > 0xffff)
2500                 return -EINVAL;
2501
2502         ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
2503
2504         return ret ? ret : count;
2505
2506 }
2507 static DEVICE_ATTR_WO(delete_child);
2508
2509 int ipoib_add_pkey_attr(struct net_device *dev)
2510 {
2511         return device_create_file(&dev->dev, &dev_attr_pkey);
2512 }
2513
2514 /*
2515  * We erroneously exposed the iface's port number in the dev_id
2516  * sysfs field long after dev_port was introduced for that purpose[1],
2517  * and we need to stop everyone from relying on that.
2518  * Let's overload the shower routine for the dev_id file here
2519  * to gently bring the issue up.
2520  *
2521  * [1] https://www.spinics.net/lists/netdev/msg272123.html
2522  */
2523 static ssize_t dev_id_show(struct device *dev,
2524                            struct device_attribute *attr, char *buf)
2525 {
2526         struct net_device *ndev = to_net_dev(dev);
2527
2528         /*
2529          * ndev->dev_port will be equal to 0 in old kernel prior to commit
2530          * 9b8b2a323008 ("IB/ipoib: Use dev_port to expose network interface
2531          * port numbers") Zero was chosen as special case for user space
2532          * applications to fallback and query dev_id to check if it has
2533          * different value or not.
2534          *
2535          * Don't print warning in such scenario.
2536          *
2537          * https://github.com/systemd/systemd/blob/master/src/udev/udev-builtin-net_id.c#L358
2538          */
2539         if (ndev->dev_port && ndev->dev_id == ndev->dev_port)
2540                 netdev_info_once(ndev,
2541                         "\"%s\" wants to know my dev_id. Should it look at dev_port instead? See Documentation/ABI/testing/sysfs-class-net for more info.\n",
2542                         current->comm);
2543
2544         return sysfs_emit(buf, "%#x\n", ndev->dev_id);
2545 }
2546 static DEVICE_ATTR_RO(dev_id);
2547
2548 static int ipoib_intercept_dev_id_attr(struct net_device *dev)
2549 {
2550         device_remove_file(&dev->dev, &dev_attr_dev_id);
2551         return device_create_file(&dev->dev, &dev_attr_dev_id);
2552 }
2553
2554 static struct net_device *ipoib_add_port(const char *format,
2555                                          struct ib_device *hca, u32 port)
2556 {
2557         struct rtnl_link_ops *ops = ipoib_get_link_ops();
2558         struct rdma_netdev_alloc_params params;
2559         struct ipoib_dev_priv *priv;
2560         struct net_device *ndev;
2561         int result;
2562
2563         ndev = ipoib_intf_alloc(hca, port, format);
2564         if (IS_ERR(ndev)) {
2565                 pr_warn("%s, %d: ipoib_intf_alloc failed %ld\n", hca->name, port,
2566                         PTR_ERR(ndev));
2567                 return ndev;
2568         }
2569         priv = ipoib_priv(ndev);
2570
2571         INIT_IB_EVENT_HANDLER(&priv->event_handler,
2572                               priv->ca, ipoib_event);
2573         ib_register_event_handler(&priv->event_handler);
2574
2575         /* call event handler to ensure pkey in sync */
2576         ipoib_queue_work(priv, IPOIB_FLUSH_HEAVY);
2577
2578         ndev->rtnl_link_ops = ipoib_get_link_ops();
2579
2580         result = register_netdev(ndev);
2581         if (result) {
2582                 pr_warn("%s: couldn't register ipoib port %d; error %d\n",
2583                         hca->name, port, result);
2584
2585                 ipoib_parent_unregister_pre(ndev);
2586                 ipoib_intf_free(ndev);
2587                 free_netdev(ndev);
2588
2589                 return ERR_PTR(result);
2590         }
2591
2592         if (hca->ops.rdma_netdev_get_params) {
2593                 int rc = hca->ops.rdma_netdev_get_params(hca, port,
2594                                                      RDMA_NETDEV_IPOIB,
2595                                                      &params);
2596
2597                 if (!rc && ops->priv_size < params.sizeof_priv)
2598                         ops->priv_size = params.sizeof_priv;
2599         }
2600         /*
2601          * We cannot set priv_destructor before register_netdev because we
2602          * need priv to be always valid during the error flow to execute
2603          * ipoib_parent_unregister_pre(). Instead handle it manually and only
2604          * enter priv_destructor mode once we are completely registered.
2605          */
2606         ndev->priv_destructor = ipoib_intf_free;
2607
2608         if (ipoib_intercept_dev_id_attr(ndev))
2609                 goto sysfs_failed;
2610         if (ipoib_cm_add_mode_attr(ndev))
2611                 goto sysfs_failed;
2612         if (ipoib_add_pkey_attr(ndev))
2613                 goto sysfs_failed;
2614         if (ipoib_add_umcast_attr(ndev))
2615                 goto sysfs_failed;
2616         if (device_create_file(&ndev->dev, &dev_attr_create_child))
2617                 goto sysfs_failed;
2618         if (device_create_file(&ndev->dev, &dev_attr_delete_child))
2619                 goto sysfs_failed;
2620
2621         return ndev;
2622
2623 sysfs_failed:
2624         ipoib_parent_unregister_pre(ndev);
2625         unregister_netdev(ndev);
2626         return ERR_PTR(-ENOMEM);
2627 }
2628
2629 static int ipoib_add_one(struct ib_device *device)
2630 {
2631         struct list_head *dev_list;
2632         struct net_device *dev;
2633         struct ipoib_dev_priv *priv;
2634         unsigned int p;
2635         int count = 0;
2636
2637         dev_list = kmalloc(sizeof(*dev_list), GFP_KERNEL);
2638         if (!dev_list)
2639                 return -ENOMEM;
2640
2641         INIT_LIST_HEAD(dev_list);
2642
2643         rdma_for_each_port (device, p) {
2644                 if (!rdma_protocol_ib(device, p))
2645                         continue;
2646                 dev = ipoib_add_port("ib%d", device, p);
2647                 if (!IS_ERR(dev)) {
2648                         priv = ipoib_priv(dev);
2649                         list_add_tail(&priv->list, dev_list);
2650                         count++;
2651                 }
2652         }
2653
2654         if (!count) {
2655                 kfree(dev_list);
2656                 return -EOPNOTSUPP;
2657         }
2658
2659         ib_set_client_data(device, &ipoib_client, dev_list);
2660         return 0;
2661 }
2662
2663 static void ipoib_remove_one(struct ib_device *device, void *client_data)
2664 {
2665         struct ipoib_dev_priv *priv, *tmp, *cpriv, *tcpriv;
2666         struct list_head *dev_list = client_data;
2667
2668         list_for_each_entry_safe(priv, tmp, dev_list, list) {
2669                 LIST_HEAD(head);
2670                 ipoib_parent_unregister_pre(priv->dev);
2671
2672                 rtnl_lock();
2673
2674                 netdev_lock(priv->dev);
2675                 list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs,
2676                                          list)
2677                         unregister_netdevice_queue(cpriv->dev, &head);
2678                 netdev_unlock(priv->dev);
2679                 unregister_netdevice_queue(priv->dev, &head);
2680                 unregister_netdevice_many(&head);
2681
2682                 rtnl_unlock();
2683         }
2684
2685         kfree(dev_list);
2686 }
2687
2688 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
2689 static struct notifier_block ipoib_netdev_notifier = {
2690         .notifier_call = ipoib_netdev_event,
2691 };
2692 #endif
2693
2694 static int __init ipoib_init_module(void)
2695 {
2696         int ret;
2697
2698         ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
2699         ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
2700         ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
2701
2702         ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
2703         ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
2704         ipoib_sendq_size = max3(ipoib_sendq_size, 2 * MAX_SEND_CQE, IPOIB_MIN_QUEUE_SIZE);
2705 #ifdef CONFIG_INFINIBAND_IPOIB_CM
2706         ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
2707         ipoib_max_conn_qp = max(ipoib_max_conn_qp, 0);
2708 #endif
2709
2710         /*
2711          * When copying small received packets, we only copy from the
2712          * linear data part of the SKB, so we rely on this condition.
2713          */
2714         BUILD_BUG_ON(IPOIB_CM_COPYBREAK > IPOIB_CM_HEAD_SIZE);
2715
2716         ipoib_register_debugfs();
2717
2718         /*
2719          * We create a global workqueue here that is used for all flush
2720          * operations.  However, if you attempt to flush a workqueue
2721          * from a task on that same workqueue, it deadlocks the system.
2722          * We want to be able to flush the tasks associated with a
2723          * specific net device, so we also create a workqueue for each
2724          * netdevice.  We queue up the tasks for that device only on
2725          * its private workqueue, and we only queue up flush events
2726          * on our global flush workqueue.  This avoids the deadlocks.
2727          */
2728         ipoib_workqueue = alloc_ordered_workqueue("ipoib_flush", 0);
2729         if (!ipoib_workqueue) {
2730                 ret = -ENOMEM;
2731                 goto err_fs;
2732         }
2733
2734         ib_sa_register_client(&ipoib_sa_client);
2735
2736         ret = ib_register_client(&ipoib_client);
2737         if (ret)
2738                 goto err_sa;
2739
2740         ret = ipoib_netlink_init();
2741         if (ret)
2742                 goto err_client;
2743
2744 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
2745         register_netdevice_notifier(&ipoib_netdev_notifier);
2746 #endif
2747         return 0;
2748
2749 err_client:
2750         ib_unregister_client(&ipoib_client);
2751
2752 err_sa:
2753         ib_sa_unregister_client(&ipoib_sa_client);
2754         destroy_workqueue(ipoib_workqueue);
2755
2756 err_fs:
2757         ipoib_unregister_debugfs();
2758
2759         return ret;
2760 }
2761
2762 static void __exit ipoib_cleanup_module(void)
2763 {
2764 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
2765         unregister_netdevice_notifier(&ipoib_netdev_notifier);
2766 #endif
2767         ipoib_netlink_fini();
2768         ib_unregister_client(&ipoib_client);
2769         ib_sa_unregister_client(&ipoib_sa_client);
2770         ipoib_unregister_debugfs();
2771         destroy_workqueue(ipoib_workqueue);
2772 }
2773
2774 module_init(ipoib_init_module);
2775 module_exit(ipoib_cleanup_module);