drivers/infiniband/ulp/ipoib/ipoib_main.c

   1 /*
   2  * Copyright (c) 2004 Topspin Communications.  All rights reserved.
   3  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
   4  * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
   5  *
   6  * This software is available to you under a choice of one of two
   7  * licenses.  You may choose to be licensed under the terms of the GNU
   8  * General Public License (GPL) Version 2, available from the file
   9  * COPYING in the main directory of this source tree, or the
  10  * OpenIB.org BSD license below:
  11  *
  12  *     Redistribution and use in source and binary forms, with or
  13  *     without modification, are permitted provided that the following
  14  *     conditions are met:
  15  *
  16  *      - Redistributions of source code must retain the above
  17  *        copyright notice, this list of conditions and the following
  18  *        disclaimer.
  19  *
  20  *      - Redistributions in binary form must reproduce the above
  21  *        copyright notice, this list of conditions and the following
  22  *        disclaimer in the documentation and/or other materials
  23  *        provided with the distribution.
  24  *
  25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  32  * SOFTWARE.
  33  */
  34
  35 #include "ipoib.h"
  36
  37 #include <linux/module.h>
  38
  39 #include <linux/init.h>
  40 #include <linux/slab.h>
  41 #include <linux/kernel.h>
  42 #include <linux/vmalloc.h>
  43
  44 #include <linux/if_arp.h>       /* For ARPHRD_xxx */
  45
  46 #include <linux/ip.h>
  47 #include <linux/in.h>
  48
  49 #include <linux/jhash.h>
  50 #include <net/arp.h>
  51 #include <net/addrconf.h>
  52 #include <linux/inetdevice.h>
  53 #include <rdma/ib_cache.h>
  54 #include <linux/pci.h>
  55
  56 #define DRV_VERSION "1.0.0"
  57
  58 const char ipoib_driver_version[] = DRV_VERSION;
  59
  60 MODULE_AUTHOR("Roland Dreier");
  61 MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
  62 MODULE_LICENSE("Dual BSD/GPL");
  63 MODULE_VERSION(DRV_VERSION);
  64
  65 int ipoib_sendq_size __read_mostly = IPOIB_TX_RING_SIZE;
  66 int ipoib_recvq_size __read_mostly = IPOIB_RX_RING_SIZE;
  67
  68 module_param_named(send_queue_size, ipoib_sendq_size, int, 0444);
  69 MODULE_PARM_DESC(send_queue_size, "Number of descriptors in send queue");
  70 module_param_named(recv_queue_size, ipoib_recvq_size, int, 0444);
  71 MODULE_PARM_DESC(recv_queue_size, "Number of descriptors in receive queue");
  72
  73 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
  74 int ipoib_debug_level;
  75
  76 module_param_named(debug_level, ipoib_debug_level, int, 0644);
  77 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
  78 #endif
  79
  80 struct ipoib_path_iter {
  81         struct net_device *dev;
  82         struct ipoib_path  path;
  83 };
  84
  85 static const u8 ipv4_bcast_addr[] = {
  86         0x00, 0xff, 0xff, 0xff,
  87         0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
  88         0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff
  89 };
  90
  91 struct workqueue_struct *ipoib_workqueue;
  92
  93 struct ib_sa_client ipoib_sa_client;
  94
  95 static void ipoib_add_one(struct ib_device *device);
  96 static void ipoib_remove_one(struct ib_device *device, void *client_data);
  97 static void ipoib_neigh_reclaim(struct rcu_head *rp);
  98 static struct net_device *ipoib_get_net_dev_by_params(
  99                 struct ib_device *dev, u8 port, u16 pkey,
 100                 const union ib_gid *gid, const struct sockaddr *addr,
 101                 void *client_data);
 102 static int ipoib_set_mac(struct net_device *dev, void *addr);
 103
 104 static struct ib_client ipoib_client = {
 105         .name   = "ipoib",
 106         .add    = ipoib_add_one,
 107         .remove = ipoib_remove_one,
 108         .get_net_dev_by_params = ipoib_get_net_dev_by_params,
 109 };
 110
 111 int ipoib_open(struct net_device *dev)
 112 {
 113         struct ipoib_dev_priv *priv = netdev_priv(dev);
 114
 115         ipoib_dbg(priv, "bringing up interface\n");
 116
 117         netif_carrier_off(dev);
 118
 119         set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 120
 121         priv->sm_fullmember_sendonly_support = false;
 122
 123         if (ipoib_ib_dev_open(dev)) {
 124                 if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags))
 125                         return 0;
 126                 goto err_disable;
 127         }
 128
 129         if (ipoib_ib_dev_up(dev))
 130                 goto err_stop;
 131
 132         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 133                 struct ipoib_dev_priv *cpriv;
 134
 135                 /* Bring up any child interfaces too */
 136                 down_read(&priv->vlan_rwsem);
 137                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 138                         int flags;
 139
 140                         flags = cpriv->dev->flags;
 141                         if (flags & IFF_UP)
 142                                 continue;
 143
 144                         dev_change_flags(cpriv->dev, flags | IFF_UP);
 145                 }
 146                 up_read(&priv->vlan_rwsem);
 147         }
 148
 149         netif_start_queue(dev);
 150
 151         return 0;
 152
 153 err_stop:
 154         ipoib_ib_dev_stop(dev);
 155
 156 err_disable:
 157         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 158
 159         return -EINVAL;
 160 }
 161
 162 static int ipoib_stop(struct net_device *dev)
 163 {
 164         struct ipoib_dev_priv *priv = netdev_priv(dev);
 165
 166         ipoib_dbg(priv, "stopping interface\n");
 167
 168         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 169
 170         netif_stop_queue(dev);
 171
 172         ipoib_ib_dev_down(dev);
 173         ipoib_ib_dev_stop(dev);
 174
 175         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 176                 struct ipoib_dev_priv *cpriv;
 177
 178                 /* Bring down any child interfaces too */
 179                 down_read(&priv->vlan_rwsem);
 180                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 181                         int flags;
 182
 183                         flags = cpriv->dev->flags;
 184                         if (!(flags & IFF_UP))
 185                                 continue;
 186
 187                         dev_change_flags(cpriv->dev, flags & ~IFF_UP);
 188                 }
 189                 up_read(&priv->vlan_rwsem);
 190         }
 191
 192         return 0;
 193 }
 194
 195 static void ipoib_uninit(struct net_device *dev)
 196 {
 197         ipoib_dev_cleanup(dev);
 198 }
 199
 200 static netdev_features_t ipoib_fix_features(struct net_device *dev, netdev_features_t features)
 201 {
 202         struct ipoib_dev_priv *priv = netdev_priv(dev);
 203
 204         if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
 205                 features &= ~(NETIF_F_IP_CSUM | NETIF_F_TSO);
 206
 207         return features;
 208 }
 209
 210 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
 211 {
 212         struct ipoib_dev_priv *priv = netdev_priv(dev);
 213
 214         /* dev->mtu > 2K ==> connected mode */
 215         if (ipoib_cm_admin_enabled(dev)) {
 216                 if (new_mtu > ipoib_cm_max_mtu(dev))
 217                         return -EINVAL;
 218
 219                 if (new_mtu > priv->mcast_mtu)
 220                         ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
 221                                    priv->mcast_mtu);
 222
 223                 dev->mtu = new_mtu;
 224                 return 0;
 225         }
 226
 227         if (new_mtu > IPOIB_UD_MTU(priv->max_ib_mtu))
 228                 return -EINVAL;
 229
 230         priv->admin_mtu = new_mtu;
 231
 232         dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
 233
 234         return 0;
 235 }
 236
 237 /* Called with an RCU read lock taken */
 238 static bool ipoib_is_dev_match_addr_rcu(const struct sockaddr *addr,
 239                                         struct net_device *dev)
 240 {
 241         struct net *net = dev_net(dev);
 242         struct in_device *in_dev;
 243         struct sockaddr_in *addr_in = (struct sockaddr_in *)addr;
 244         struct sockaddr_in6 *addr_in6 = (struct sockaddr_in6 *)addr;
 245         __be32 ret_addr;
 246
 247         switch (addr->sa_family) {
 248         case AF_INET:
 249                 in_dev = in_dev_get(dev);
 250                 if (!in_dev)
 251                         return false;
 252
 253                 ret_addr = inet_confirm_addr(net, in_dev, 0,
 254                                              addr_in->sin_addr.s_addr,
 255                                              RT_SCOPE_HOST);
 256                 in_dev_put(in_dev);
 257                 if (ret_addr)
 258                         return true;
 259
 260                 break;
 261         case AF_INET6:
 262                 if (IS_ENABLED(CONFIG_IPV6) &&
 263                     ipv6_chk_addr(net, &addr_in6->sin6_addr, dev, 1))
 264                         return true;
 265
 266                 break;
 267         }
 268         return false;
 269 }
 270
 271 /**
 272  * Find the master net_device on top of the given net_device.
 273  * @dev: base IPoIB net_device
 274  *
 275  * Returns the master net_device with a reference held, or the same net_device
 276  * if no master exists.
 277  */
 278 static struct net_device *ipoib_get_master_net_dev(struct net_device *dev)
 279 {
 280         struct net_device *master;
 281
 282         rcu_read_lock();
 283         master = netdev_master_upper_dev_get_rcu(dev);
 284         if (master)
 285                 dev_hold(master);
 286         rcu_read_unlock();
 287
 288         if (master)
 289                 return master;
 290
 291         dev_hold(dev);
 292         return dev;
 293 }
 294
 295 /**
 296  * Find a net_device matching the given address, which is an upper device of
 297  * the given net_device.
 298  * @addr: IP address to look for.
 299  * @dev: base IPoIB net_device
 300  *
 301  * If found, returns the net_device with a reference held. Otherwise return
 302  * NULL.
 303  */
 304 static struct net_device *ipoib_get_net_dev_match_addr(
 305                 const struct sockaddr *addr, struct net_device *dev)
 306 {
 307         struct net_device *upper,
 308                           *result = NULL;
 309         struct list_head *iter;
 310
 311         rcu_read_lock();
 312         if (ipoib_is_dev_match_addr_rcu(addr, dev)) {
 313                 dev_hold(dev);
 314                 result = dev;
 315                 goto out;
 316         }
 317
 318         netdev_for_each_all_upper_dev_rcu(dev, upper, iter) {
 319                 if (ipoib_is_dev_match_addr_rcu(addr, upper)) {
 320                         dev_hold(upper);
 321                         result = upper;
 322                         break;
 323                 }
 324         }
 325 out:
 326         rcu_read_unlock();
 327         return result;
 328 }
 329
 330 /* returns the number of IPoIB netdevs on top a given ipoib device matching a
 331  * pkey_index and address, if one exists.
 332  *
 333  * @found_net_dev: contains a matching net_device if the return value >= 1,
 334  * with a reference held. */
 335 static int ipoib_match_gid_pkey_addr(struct ipoib_dev_priv *priv,
 336                                      const union ib_gid *gid,
 337                                      u16 pkey_index,
 338                                      const struct sockaddr *addr,
 339                                      int nesting,
 340                                      struct net_device **found_net_dev)
 341 {
 342         struct ipoib_dev_priv *child_priv;
 343         struct net_device *net_dev = NULL;
 344         int matches = 0;
 345
 346         if (priv->pkey_index == pkey_index &&
 347             (!gid || !memcmp(gid, &priv->local_gid, sizeof(*gid)))) {
 348                 if (!addr) {
 349                         net_dev = ipoib_get_master_net_dev(priv->dev);
 350                 } else {
 351                         /* Verify the net_device matches the IP address, as
 352                          * IPoIB child devices currently share a GID. */
 353                         net_dev = ipoib_get_net_dev_match_addr(addr, priv->dev);
 354                 }
 355                 if (net_dev) {
 356                         if (!*found_net_dev)
 357                                 *found_net_dev = net_dev;
 358                         else
 359                                 dev_put(net_dev);
 360                         ++matches;
 361                 }
 362         }
 363
 364         /* Check child interfaces */
 365         down_read_nested(&priv->vlan_rwsem, nesting);
 366         list_for_each_entry(child_priv, &priv->child_intfs, list) {
 367                 matches += ipoib_match_gid_pkey_addr(child_priv, gid,
 368                                                     pkey_index, addr,
 369                                                     nesting + 1,
 370                                                     found_net_dev);
 371                 if (matches > 1)
 372                         break;
 373         }
 374         up_read(&priv->vlan_rwsem);
 375
 376         return matches;
 377 }
 378
 379 /* Returns the number of matching net_devs found (between 0 and 2). Also
 380  * return the matching net_device in the @net_dev parameter, holding a
 381  * reference to the net_device, if the number of matches >= 1 */
 382 static int __ipoib_get_net_dev_by_params(struct list_head *dev_list, u8 port,
 383                                          u16 pkey_index,
 384                                          const union ib_gid *gid,
 385                                          const struct sockaddr *addr,
 386                                          struct net_device **net_dev)
 387 {
 388         struct ipoib_dev_priv *priv;
 389         int matches = 0;
 390
 391         *net_dev = NULL;
 392
 393         list_for_each_entry(priv, dev_list, list) {
 394                 if (priv->port != port)
 395                         continue;
 396
 397                 matches += ipoib_match_gid_pkey_addr(priv, gid, pkey_index,
 398                                                      addr, 0, net_dev);
 399                 if (matches > 1)
 400                         break;
 401         }
 402
 403         return matches;
 404 }
 405
 406 static struct net_device *ipoib_get_net_dev_by_params(
 407                 struct ib_device *dev, u8 port, u16 pkey,
 408                 const union ib_gid *gid, const struct sockaddr *addr,
 409                 void *client_data)
 410 {
 411         struct net_device *net_dev;
 412         struct list_head *dev_list = client_data;
 413         u16 pkey_index;
 414         int matches;
 415         int ret;
 416
 417         if (!rdma_protocol_ib(dev, port))
 418                 return NULL;
 419
 420         ret = ib_find_cached_pkey(dev, port, pkey, &pkey_index);
 421         if (ret)
 422                 return NULL;
 423
 424         if (!dev_list)
 425                 return NULL;
 426
 427         /* See if we can find a unique device matching the L2 parameters */
 428         matches = __ipoib_get_net_dev_by_params(dev_list, port, pkey_index,
 429                                                 gid, NULL, &net_dev);
 430
 431         switch (matches) {
 432         case 0:
 433                 return NULL;
 434         case 1:
 435                 return net_dev;
 436         }
 437
 438         dev_put(net_dev);
 439
 440         /* Couldn't find a unique device with L2 parameters only. Use L3
 441          * address to uniquely match the net device */
 442         matches = __ipoib_get_net_dev_by_params(dev_list, port, pkey_index,
 443                                                 gid, addr, &net_dev);
 444         switch (matches) {
 445         case 0:
 446                 return NULL;
 447         default:
 448                 dev_warn_ratelimited(&dev->dev,
 449                                      "duplicate IP address detected\n");
 450                 /* Fall through */
 451         case 1:
 452                 return net_dev;
 453         }
 454 }
 455
 456 int ipoib_set_mode(struct net_device *dev, const char *buf)
 457 {
 458         struct ipoib_dev_priv *priv = netdev_priv(dev);
 459
 460         /* flush paths if we switch modes so that connections are restarted */
 461         if (IPOIB_CM_SUPPORTED(dev->dev_addr) && !strcmp(buf, "connected\n")) {
 462                 set_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
 463                 ipoib_warn(priv, "enabling connected mode "
 464                            "will cause multicast packet drops\n");
 465                 netdev_update_features(dev);
 466                 dev_set_mtu(dev, ipoib_cm_max_mtu(dev));
 467                 rtnl_unlock();
 468                 priv->tx_wr.wr.send_flags &= ~IB_SEND_IP_CSUM;
 469
 470                 ipoib_flush_paths(dev);
 471                 rtnl_lock();
 472                 return 0;
 473         }
 474
 475         if (!strcmp(buf, "datagram\n")) {
 476                 clear_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
 477                 netdev_update_features(dev);
 478                 dev_set_mtu(dev, min(priv->mcast_mtu, dev->mtu));
 479                 rtnl_unlock();
 480                 ipoib_flush_paths(dev);
 481                 rtnl_lock();
 482                 return 0;
 483         }
 484
 485         return -EINVAL;
 486 }
 487
 488 struct ipoib_path *__path_find(struct net_device *dev, void *gid)
 489 {
 490         struct ipoib_dev_priv *priv = netdev_priv(dev);
 491         struct rb_node *n = priv->path_tree.rb_node;
 492         struct ipoib_path *path;
 493         int ret;
 494
 495         while (n) {
 496                 path = rb_entry(n, struct ipoib_path, rb_node);
 497
 498                 ret = memcmp(gid, path->pathrec.dgid.raw,
 499                              sizeof (union ib_gid));
 500
 501                 if (ret < 0)
 502                         n = n->rb_left;
 503                 else if (ret > 0)
 504                         n = n->rb_right;
 505                 else
 506                         return path;
 507         }
 508
 509         return NULL;
 510 }
 511
 512 static int __path_add(struct net_device *dev, struct ipoib_path *path)
 513 {
 514         struct ipoib_dev_priv *priv = netdev_priv(dev);
 515         struct rb_node **n = &priv->path_tree.rb_node;
 516         struct rb_node *pn = NULL;
 517         struct ipoib_path *tpath;
 518         int ret;
 519
 520         while (*n) {
 521                 pn = *n;
 522                 tpath = rb_entry(pn, struct ipoib_path, rb_node);
 523
 524                 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
 525                              sizeof (union ib_gid));
 526                 if (ret < 0)
 527                         n = &pn->rb_left;
 528                 else if (ret > 0)
 529                         n = &pn->rb_right;
 530                 else
 531                         return -EEXIST;
 532         }
 533
 534         rb_link_node(&path->rb_node, pn, n);
 535         rb_insert_color(&path->rb_node, &priv->path_tree);
 536
 537         list_add_tail(&path->list, &priv->path_list);
 538
 539         return 0;
 540 }
 541
 542 static void path_free(struct net_device *dev, struct ipoib_path *path)
 543 {
 544         struct sk_buff *skb;
 545
 546         while ((skb = __skb_dequeue(&path->queue)))
 547                 dev_kfree_skb_irq(skb);
 548
 549         ipoib_dbg(netdev_priv(dev), "path_free\n");
 550
 551         /* remove all neigh connected to this path */
 552         ipoib_del_neighs_by_gid(dev, path->pathrec.dgid.raw);
 553
 554         if (path->ah)
 555                 ipoib_put_ah(path->ah);
 556
 557         kfree(path);
 558 }
 559
 560 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 561
 562 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
 563 {
 564         struct ipoib_path_iter *iter;
 565
 566         iter = kmalloc(sizeof *iter, GFP_KERNEL);
 567         if (!iter)
 568                 return NULL;
 569
 570         iter->dev = dev;
 571         memset(iter->path.pathrec.dgid.raw, 0, 16);
 572
 573         if (ipoib_path_iter_next(iter)) {
 574                 kfree(iter);
 575                 return NULL;
 576         }
 577
 578         return iter;
 579 }
 580
 581 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
 582 {
 583         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
 584         struct rb_node *n;
 585         struct ipoib_path *path;
 586         int ret = 1;
 587
 588         spin_lock_irq(&priv->lock);
 589
 590         n = rb_first(&priv->path_tree);
 591
 592         while (n) {
 593                 path = rb_entry(n, struct ipoib_path, rb_node);
 594
 595                 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
 596                            sizeof (union ib_gid)) < 0) {
 597                         iter->path = *path;
 598                         ret = 0;
 599                         break;
 600                 }
 601
 602                 n = rb_next(n);
 603         }
 604
 605         spin_unlock_irq(&priv->lock);
 606
 607         return ret;
 608 }
 609
 610 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
 611                           struct ipoib_path *path)
 612 {
 613         *path = iter->path;
 614 }
 615
 616 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
 617
 618 void ipoib_mark_paths_invalid(struct net_device *dev)
 619 {
 620         struct ipoib_dev_priv *priv = netdev_priv(dev);
 621         struct ipoib_path *path, *tp;
 622
 623         spin_lock_irq(&priv->lock);
 624
 625         list_for_each_entry_safe(path, tp, &priv->path_list, list) {
 626                 ipoib_dbg(priv, "mark path LID 0x%04x GID %pI6 invalid\n",
 627                         be16_to_cpu(path->pathrec.dlid),
 628                         path->pathrec.dgid.raw);
 629                 path->valid =  0;
 630         }
 631
 632         spin_unlock_irq(&priv->lock);
 633 }
 634
 635 struct classport_info_context {
 636         struct ipoib_dev_priv   *priv;
 637         struct completion       done;
 638         struct ib_sa_query      *sa_query;
 639 };
 640
 641 static void classport_info_query_cb(int status, struct ib_class_port_info *rec,
 642                                     void *context)
 643 {
 644         struct classport_info_context *cb_ctx = context;
 645         struct ipoib_dev_priv *priv;
 646
 647         WARN_ON(!context);
 648
 649         priv = cb_ctx->priv;
 650
 651         if (status || !rec) {
 652                 pr_debug("device: %s failed query classport_info status: %d\n",
 653                          priv->dev->name, status);
 654                 /* keeps the default, will try next mcast_restart */
 655                 priv->sm_fullmember_sendonly_support = false;
 656                 goto out;
 657         }
 658
 659         if (ib_get_cpi_capmask2(rec) &
 660             IB_SA_CAP_MASK2_SENDONLY_FULL_MEM_SUPPORT) {
 661                 pr_debug("device: %s enabled fullmember-sendonly for sendonly MCG\n",
 662                          priv->dev->name);
 663                 priv->sm_fullmember_sendonly_support = true;
 664         } else {
 665                 pr_debug("device: %s disabled fullmember-sendonly for sendonly MCG\n",
 666                          priv->dev->name);
 667                 priv->sm_fullmember_sendonly_support = false;
 668         }
 669
 670 out:
 671         complete(&cb_ctx->done);
 672 }
 673
 674 int ipoib_check_sm_sendonly_fullmember_support(struct ipoib_dev_priv *priv)
 675 {
 676         struct classport_info_context *callback_context;
 677         int ret;
 678
 679         callback_context = kmalloc(sizeof(*callback_context), GFP_KERNEL);
 680         if (!callback_context)
 681                 return -ENOMEM;
 682
 683         callback_context->priv = priv;
 684         init_completion(&callback_context->done);
 685
 686         ret = ib_sa_classport_info_rec_query(&ipoib_sa_client,
 687                                              priv->ca, priv->port, 3000,
 688                                              GFP_KERNEL,
 689                                              classport_info_query_cb,
 690                                              callback_context,
 691                                              &callback_context->sa_query);
 692         if (ret < 0) {
 693                 pr_info("%s failed to send ib_sa_classport_info query, ret: %d\n",
 694                         priv->dev->name, ret);
 695                 kfree(callback_context);
 696                 return ret;
 697         }
 698
 699         /* waiting for the callback to finish before returnning */
 700         wait_for_completion(&callback_context->done);
 701         kfree(callback_context);
 702
 703         return ret;
 704 }
 705
 706 void ipoib_flush_paths(struct net_device *dev)
 707 {
 708         struct ipoib_dev_priv *priv = netdev_priv(dev);
 709         struct ipoib_path *path, *tp;
 710         LIST_HEAD(remove_list);
 711         unsigned long flags;
 712
 713         netif_tx_lock_bh(dev);
 714         spin_lock_irqsave(&priv->lock, flags);
 715
 716         list_splice_init(&priv->path_list, &remove_list);
 717
 718         list_for_each_entry(path, &remove_list, list)
 719                 rb_erase(&path->rb_node, &priv->path_tree);
 720
 721         list_for_each_entry_safe(path, tp, &remove_list, list) {
 722                 if (path->query)
 723                         ib_sa_cancel_query(path->query_id, path->query);
 724                 spin_unlock_irqrestore(&priv->lock, flags);
 725                 netif_tx_unlock_bh(dev);
 726                 wait_for_completion(&path->done);
 727                 path_free(dev, path);
 728                 netif_tx_lock_bh(dev);
 729                 spin_lock_irqsave(&priv->lock, flags);
 730         }
 731
 732         spin_unlock_irqrestore(&priv->lock, flags);
 733         netif_tx_unlock_bh(dev);
 734 }
 735
 736 static void path_rec_completion(int status,
 737                                 struct ib_sa_path_rec *pathrec,
 738                                 void *path_ptr)
 739 {
 740         struct ipoib_path *path = path_ptr;
 741         struct net_device *dev = path->dev;
 742         struct ipoib_dev_priv *priv = netdev_priv(dev);
 743         struct ipoib_ah *ah = NULL;
 744         struct ipoib_ah *old_ah = NULL;
 745         struct ipoib_neigh *neigh, *tn;
 746         struct sk_buff_head skqueue;
 747         struct sk_buff *skb;
 748         unsigned long flags;
 749
 750         if (!status)
 751                 ipoib_dbg(priv, "PathRec LID 0x%04x for GID %pI6\n",
 752                           be16_to_cpu(pathrec->dlid), pathrec->dgid.raw);
 753         else
 754                 ipoib_dbg(priv, "PathRec status %d for GID %pI6\n",
 755                           status, path->pathrec.dgid.raw);
 756
 757         skb_queue_head_init(&skqueue);
 758
 759         if (!status) {
 760                 struct ib_ah_attr av;
 761
 762                 if (!ib_init_ah_from_path(priv->ca, priv->port, pathrec, &av))
 763                         ah = ipoib_create_ah(dev, priv->pd, &av);
 764         }
 765
 766         spin_lock_irqsave(&priv->lock, flags);
 767
 768         if (!IS_ERR_OR_NULL(ah)) {
 769                 path->pathrec = *pathrec;
 770
 771                 old_ah   = path->ah;
 772                 path->ah = ah;
 773
 774                 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
 775                           ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
 776
 777                 while ((skb = __skb_dequeue(&path->queue)))
 778                         __skb_queue_tail(&skqueue, skb);
 779
 780                 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
 781                         if (neigh->ah) {
 782                                 WARN_ON(neigh->ah != old_ah);
 783                                 /*
 784                                  * Dropping the ah reference inside
 785                                  * priv->lock is safe here, because we
 786                                  * will hold one more reference from
 787                                  * the original value of path->ah (ie
 788                                  * old_ah).
 789                                  */
 790                                 ipoib_put_ah(neigh->ah);
 791                         }
 792                         kref_get(&path->ah->ref);
 793                         neigh->ah = path->ah;
 794
 795                         if (ipoib_cm_enabled(dev, neigh->daddr)) {
 796                                 if (!ipoib_cm_get(neigh))
 797                                         ipoib_cm_set(neigh, ipoib_cm_create_tx(dev,
 798                                                                                path,
 799                                                                                neigh));
 800                                 if (!ipoib_cm_get(neigh)) {
 801                                         ipoib_neigh_free(neigh);
 802                                         continue;
 803                                 }
 804                         }
 805
 806                         while ((skb = __skb_dequeue(&neigh->queue)))
 807                                 __skb_queue_tail(&skqueue, skb);
 808                 }
 809                 path->valid = 1;
 810         }
 811
 812         path->query = NULL;
 813         complete(&path->done);
 814
 815         spin_unlock_irqrestore(&priv->lock, flags);
 816
 817         if (IS_ERR_OR_NULL(ah))
 818                 ipoib_del_neighs_by_gid(dev, path->pathrec.dgid.raw);
 819
 820         if (old_ah)
 821                 ipoib_put_ah(old_ah);
 822
 823         while ((skb = __skb_dequeue(&skqueue))) {
 824                 skb->dev = dev;
 825                 if (dev_queue_xmit(skb))
 826                         ipoib_warn(priv, "dev_queue_xmit failed "
 827                                    "to requeue packet\n");
 828         }
 829 }
 830
 831 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
 832 {
 833         struct ipoib_dev_priv *priv = netdev_priv(dev);
 834         struct ipoib_path *path;
 835
 836         if (!priv->broadcast)
 837                 return NULL;
 838
 839         path = kzalloc(sizeof *path, GFP_ATOMIC);
 840         if (!path)
 841                 return NULL;
 842
 843         path->dev = dev;
 844
 845         skb_queue_head_init(&path->queue);
 846
 847         INIT_LIST_HEAD(&path->neigh_list);
 848
 849         memcpy(path->pathrec.dgid.raw, gid, sizeof (union ib_gid));
 850         path->pathrec.sgid          = priv->local_gid;
 851         path->pathrec.pkey          = cpu_to_be16(priv->pkey);
 852         path->pathrec.numb_path     = 1;
 853         path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
 854
 855         return path;
 856 }
 857
 858 static int path_rec_start(struct net_device *dev,
 859                           struct ipoib_path *path)
 860 {
 861         struct ipoib_dev_priv *priv = netdev_priv(dev);
 862
 863         ipoib_dbg(priv, "Start path record lookup for %pI6\n",
 864                   path->pathrec.dgid.raw);
 865
 866         init_completion(&path->done);
 867
 868         path->query_id =
 869                 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
 870                                    &path->pathrec,
 871                                    IB_SA_PATH_REC_DGID          |
 872                                    IB_SA_PATH_REC_SGID          |
 873                                    IB_SA_PATH_REC_NUMB_PATH     |
 874                                    IB_SA_PATH_REC_TRAFFIC_CLASS |
 875                                    IB_SA_PATH_REC_PKEY,
 876                                    1000, GFP_ATOMIC,
 877                                    path_rec_completion,
 878                                    path, &path->query);
 879         if (path->query_id < 0) {
 880                 ipoib_warn(priv, "ib_sa_path_rec_get failed: %d\n", path->query_id);
 881                 path->query = NULL;
 882                 complete(&path->done);
 883                 return path->query_id;
 884         }
 885
 886         return 0;
 887 }
 888
 889 static void neigh_add_path(struct sk_buff *skb, u8 *daddr,
 890                            struct net_device *dev)
 891 {
 892         struct ipoib_dev_priv *priv = netdev_priv(dev);
 893         struct ipoib_path *path;
 894         struct ipoib_neigh *neigh;
 895         unsigned long flags;
 896
 897         spin_lock_irqsave(&priv->lock, flags);
 898         neigh = ipoib_neigh_alloc(daddr, dev);
 899         if (!neigh) {
 900                 spin_unlock_irqrestore(&priv->lock, flags);
 901                 ++dev->stats.tx_dropped;
 902                 dev_kfree_skb_any(skb);
 903                 return;
 904         }
 905
 906         path = __path_find(dev, daddr + 4);
 907         if (!path) {
 908                 path = path_rec_create(dev, daddr + 4);
 909                 if (!path)
 910                         goto err_path;
 911
 912                 __path_add(dev, path);
 913         }
 914
 915         list_add_tail(&neigh->list, &path->neigh_list);
 916
 917         if (path->ah) {
 918                 kref_get(&path->ah->ref);
 919                 neigh->ah = path->ah;
 920
 921                 if (ipoib_cm_enabled(dev, neigh->daddr)) {
 922                         if (!ipoib_cm_get(neigh))
 923                                 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
 924                         if (!ipoib_cm_get(neigh)) {
 925                                 ipoib_neigh_free(neigh);
 926                                 goto err_drop;
 927                         }
 928                         if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE)
 929                                 __skb_queue_tail(&neigh->queue, skb);
 930                         else {
 931                                 ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
 932                                            skb_queue_len(&neigh->queue));
 933                                 goto err_drop;
 934                         }
 935                 } else {
 936                         spin_unlock_irqrestore(&priv->lock, flags);
 937                         ipoib_send(dev, skb, path->ah, IPOIB_QPN(daddr));
 938                         ipoib_neigh_put(neigh);
 939                         return;
 940                 }
 941         } else {
 942                 neigh->ah  = NULL;
 943
 944                 if (!path->query && path_rec_start(dev, path))
 945                         goto err_path;
 946                 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE)
 947                         __skb_queue_tail(&neigh->queue, skb);
 948                 else
 949                         goto err_drop;
 950         }
 951
 952         spin_unlock_irqrestore(&priv->lock, flags);
 953         ipoib_neigh_put(neigh);
 954         return;
 955
 956 err_path:
 957         ipoib_neigh_free(neigh);
 958 err_drop:
 959         ++dev->stats.tx_dropped;
 960         dev_kfree_skb_any(skb);
 961
 962         spin_unlock_irqrestore(&priv->lock, flags);
 963         ipoib_neigh_put(neigh);
 964 }
 965
 966 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
 967                              struct ipoib_cb *cb)
 968 {
 969         struct ipoib_dev_priv *priv = netdev_priv(dev);
 970         struct ipoib_path *path;
 971         unsigned long flags;
 972
 973         spin_lock_irqsave(&priv->lock, flags);
 974
 975         path = __path_find(dev, cb->hwaddr + 4);
 976         if (!path || !path->valid) {
 977                 int new_path = 0;
 978
 979                 if (!path) {
 980                         path = path_rec_create(dev, cb->hwaddr + 4);
 981                         new_path = 1;
 982                 }
 983                 if (path) {
 984                         if (skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 985                                 __skb_queue_tail(&path->queue, skb);
 986                         } else {
 987                                 ++dev->stats.tx_dropped;
 988                                 dev_kfree_skb_any(skb);
 989                         }
 990
 991                         if (!path->query && path_rec_start(dev, path)) {
 992                                 spin_unlock_irqrestore(&priv->lock, flags);
 993                                 if (new_path)
 994                                         path_free(dev, path);
 995                                 return;
 996                         } else
 997                                 __path_add(dev, path);
 998                 } else {
 999                         ++dev->stats.tx_dropped;
1000                         dev_kfree_skb_any(skb);
1001                 }
1002
1003                 spin_unlock_irqrestore(&priv->lock, flags);
1004                 return;
1005         }
1006
1007         if (path->ah) {
1008                 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
1009                           be16_to_cpu(path->pathrec.dlid));
1010
1011                 spin_unlock_irqrestore(&priv->lock, flags);
1012                 ipoib_send(dev, skb, path->ah, IPOIB_QPN(cb->hwaddr));
1013                 return;
1014         } else if ((path->query || !path_rec_start(dev, path)) &&
1015                    skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
1016                 __skb_queue_tail(&path->queue, skb);
1017         } else {
1018                 ++dev->stats.tx_dropped;
1019                 dev_kfree_skb_any(skb);
1020         }
1021
1022         spin_unlock_irqrestore(&priv->lock, flags);
1023 }
1024
1025 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
1026 {
1027         struct ipoib_dev_priv *priv = netdev_priv(dev);
1028         struct ipoib_neigh *neigh;
1029         struct ipoib_cb *cb = ipoib_skb_cb(skb);
1030         struct ipoib_header *header;
1031         unsigned long flags;
1032
1033         header = (struct ipoib_header *) skb->data;
1034
1035         if (unlikely(cb->hwaddr[4] == 0xff)) {
1036                 /* multicast, arrange "if" according to probability */
1037                 if ((header->proto != htons(ETH_P_IP)) &&
1038                     (header->proto != htons(ETH_P_IPV6)) &&
1039                     (header->proto != htons(ETH_P_ARP)) &&
1040                     (header->proto != htons(ETH_P_RARP)) &&
1041                     (header->proto != htons(ETH_P_TIPC))) {
1042                         /* ethertype not supported by IPoIB */
1043                         ++dev->stats.tx_dropped;
1044                         dev_kfree_skb_any(skb);
1045                         return NETDEV_TX_OK;
1046                 }
1047                 /* Add in the P_Key for multicast*/
1048                 cb->hwaddr[8] = (priv->pkey >> 8) & 0xff;
1049                 cb->hwaddr[9] = priv->pkey & 0xff;
1050
1051                 neigh = ipoib_neigh_get(dev, cb->hwaddr);
1052                 if (likely(neigh))
1053                         goto send_using_neigh;
1054                 ipoib_mcast_send(dev, cb->hwaddr, skb);
1055                 return NETDEV_TX_OK;
1056         }
1057
1058         /* unicast, arrange "switch" according to probability */
1059         switch (header->proto) {
1060         case htons(ETH_P_IP):
1061         case htons(ETH_P_IPV6):
1062         case htons(ETH_P_TIPC):
1063                 neigh = ipoib_neigh_get(dev, cb->hwaddr);
1064                 if (unlikely(!neigh)) {
1065                         neigh_add_path(skb, cb->hwaddr, dev);
1066                         return NETDEV_TX_OK;
1067                 }
1068                 break;
1069         case htons(ETH_P_ARP):
1070         case htons(ETH_P_RARP):
1071                 /* for unicast ARP and RARP should always perform path find */
1072                 unicast_arp_send(skb, dev, cb);
1073                 return NETDEV_TX_OK;
1074         default:
1075                 /* ethertype not supported by IPoIB */
1076                 ++dev->stats.tx_dropped;
1077                 dev_kfree_skb_any(skb);
1078                 return NETDEV_TX_OK;
1079         }
1080
1081 send_using_neigh:
1082         /* note we now hold a ref to neigh */
1083         if (ipoib_cm_get(neigh)) {
1084                 if (ipoib_cm_up(neigh)) {
1085                         ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
1086                         goto unref;
1087                 }
1088         } else if (neigh->ah) {
1089                 ipoib_send(dev, skb, neigh->ah, IPOIB_QPN(cb->hwaddr));
1090                 goto unref;
1091         }
1092
1093         if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
1094                 spin_lock_irqsave(&priv->lock, flags);
1095                 __skb_queue_tail(&neigh->queue, skb);
1096                 spin_unlock_irqrestore(&priv->lock, flags);
1097         } else {
1098                 ++dev->stats.tx_dropped;
1099                 dev_kfree_skb_any(skb);
1100         }
1101
1102 unref:
1103         ipoib_neigh_put(neigh);
1104
1105         return NETDEV_TX_OK;
1106 }
1107
1108 static void ipoib_timeout(struct net_device *dev)
1109 {
1110         struct ipoib_dev_priv *priv = netdev_priv(dev);
1111
1112         ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
1113                    jiffies_to_msecs(jiffies - dev_trans_start(dev)));
1114         ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
1115                    netif_queue_stopped(dev),
1116                    priv->tx_head, priv->tx_tail);
1117         /* XXX reset QP, etc. */
1118 }
1119
1120 static int ipoib_hard_header(struct sk_buff *skb,
1121                              struct net_device *dev,
1122                              unsigned short type,
1123                              const void *daddr, const void *saddr, unsigned len)
1124 {
1125         struct ipoib_header *header;
1126         struct ipoib_cb *cb = ipoib_skb_cb(skb);
1127
1128         header = (struct ipoib_header *) skb_push(skb, sizeof *header);
1129
1130         header->proto = htons(type);
1131         header->reserved = 0;
1132
1133         /*
1134          * we don't rely on dst_entry structure,  always stuff the
1135          * destination address into skb->cb so we can figure out where
1136          * to send the packet later.
1137          */
1138         memcpy(cb->hwaddr, daddr, INFINIBAND_ALEN);
1139
1140         return sizeof *header;
1141 }
1142
1143 static void ipoib_set_mcast_list(struct net_device *dev)
1144 {
1145         struct ipoib_dev_priv *priv = netdev_priv(dev);
1146
1147         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
1148                 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
1149                 return;
1150         }
1151
1152         queue_work(priv->wq, &priv->restart_task);
1153 }
1154
1155 static int ipoib_get_iflink(const struct net_device *dev)
1156 {
1157         struct ipoib_dev_priv *priv = netdev_priv(dev);
1158
1159         /* parent interface */
1160         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags))
1161                 return dev->ifindex;
1162
1163         /* child/vlan interface */
1164         return priv->parent->ifindex;
1165 }
1166
1167 static u32 ipoib_addr_hash(struct ipoib_neigh_hash *htbl, u8 *daddr)
1168 {
1169         /*
1170          * Use only the address parts that contributes to spreading
1171          * The subnet prefix is not used as one can not connect to
1172          * same remote port (GUID) using the same remote QPN via two
1173          * different subnets.
1174          */
1175          /* qpn octets[1:4) & port GUID octets[12:20) */
1176         u32 *d32 = (u32 *) daddr;
1177         u32 hv;
1178
1179         hv = jhash_3words(d32[3], d32[4], IPOIB_QPN_MASK & d32[0], 0);
1180         return hv & htbl->mask;
1181 }
1182
1183 struct ipoib_neigh *ipoib_neigh_get(struct net_device *dev, u8 *daddr)
1184 {
1185         struct ipoib_dev_priv *priv = netdev_priv(dev);
1186         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1187         struct ipoib_neigh_hash *htbl;
1188         struct ipoib_neigh *neigh = NULL;
1189         u32 hash_val;
1190
1191         rcu_read_lock_bh();
1192
1193         htbl = rcu_dereference_bh(ntbl->htbl);
1194
1195         if (!htbl)
1196                 goto out_unlock;
1197
1198         hash_val = ipoib_addr_hash(htbl, daddr);
1199         for (neigh = rcu_dereference_bh(htbl->buckets[hash_val]);
1200              neigh != NULL;
1201              neigh = rcu_dereference_bh(neigh->hnext)) {
1202                 if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) {
1203                         /* found, take one ref on behalf of the caller */
1204                         if (!atomic_inc_not_zero(&neigh->refcnt)) {
1205                                 /* deleted */
1206                                 neigh = NULL;
1207                                 goto out_unlock;
1208                         }
1209
1210                         if (likely(skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE))
1211                                 neigh->alive = jiffies;
1212                         goto out_unlock;
1213                 }
1214         }
1215
1216 out_unlock:
1217         rcu_read_unlock_bh();
1218         return neigh;
1219 }
1220
1221 static void __ipoib_reap_neigh(struct ipoib_dev_priv *priv)
1222 {
1223         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1224         struct ipoib_neigh_hash *htbl;
1225         unsigned long neigh_obsolete;
1226         unsigned long dt;
1227         unsigned long flags;
1228         int i;
1229         LIST_HEAD(remove_list);
1230
1231         if (test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
1232                 return;
1233
1234         spin_lock_irqsave(&priv->lock, flags);
1235
1236         htbl = rcu_dereference_protected(ntbl->htbl,
1237                                          lockdep_is_held(&priv->lock));
1238
1239         if (!htbl)
1240                 goto out_unlock;
1241
1242         /* neigh is obsolete if it was idle for two GC periods */
1243         dt = 2 * arp_tbl.gc_interval;
1244         neigh_obsolete = jiffies - dt;
1245         /* handle possible race condition */
1246         if (test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
1247                 goto out_unlock;
1248
1249         for (i = 0; i < htbl->size; i++) {
1250                 struct ipoib_neigh *neigh;
1251                 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1252
1253                 while ((neigh = rcu_dereference_protected(*np,
1254                                                           lockdep_is_held(&priv->lock))) != NULL) {
1255                         /* was the neigh idle for two GC periods */
1256                         if (time_after(neigh_obsolete, neigh->alive)) {
1257
1258                                 ipoib_check_and_add_mcast_sendonly(priv, neigh->daddr + 4, &remove_list);
1259
1260                                 rcu_assign_pointer(*np,
1261                                                    rcu_dereference_protected(neigh->hnext,
1262                                                                              lockdep_is_held(&priv->lock)));
1263                                 /* remove from path/mc list */
1264                                 list_del(&neigh->list);
1265                                 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1266                         } else {
1267                                 np = &neigh->hnext;
1268                         }
1269
1270                 }
1271         }
1272
1273 out_unlock:
1274         spin_unlock_irqrestore(&priv->lock, flags);
1275         ipoib_mcast_remove_list(&remove_list);
1276 }
1277
1278 static void ipoib_reap_neigh(struct work_struct *work)
1279 {
1280         struct ipoib_dev_priv *priv =
1281                 container_of(work, struct ipoib_dev_priv, neigh_reap_task.work);
1282
1283         __ipoib_reap_neigh(priv);
1284
1285         if (!test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
1286                 queue_delayed_work(priv->wq, &priv->neigh_reap_task,
1287                                    arp_tbl.gc_interval);
1288 }
1289
1290
1291 static struct ipoib_neigh *ipoib_neigh_ctor(u8 *daddr,
1292                                       struct net_device *dev)
1293 {
1294         struct ipoib_neigh *neigh;
1295
1296         neigh = kzalloc(sizeof *neigh, GFP_ATOMIC);
1297         if (!neigh)
1298                 return NULL;
1299
1300         neigh->dev = dev;
1301         memcpy(&neigh->daddr, daddr, sizeof(neigh->daddr));
1302         skb_queue_head_init(&neigh->queue);
1303         INIT_LIST_HEAD(&neigh->list);
1304         ipoib_cm_set(neigh, NULL);
1305         /* one ref on behalf of the caller */
1306         atomic_set(&neigh->refcnt, 1);
1307
1308         return neigh;
1309 }
1310
1311 struct ipoib_neigh *ipoib_neigh_alloc(u8 *daddr,
1312                                       struct net_device *dev)
1313 {
1314         struct ipoib_dev_priv *priv = netdev_priv(dev);
1315         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1316         struct ipoib_neigh_hash *htbl;
1317         struct ipoib_neigh *neigh;
1318         u32 hash_val;
1319
1320         htbl = rcu_dereference_protected(ntbl->htbl,
1321                                          lockdep_is_held(&priv->lock));
1322         if (!htbl) {
1323                 neigh = NULL;
1324                 goto out_unlock;
1325         }
1326
1327         /* need to add a new neigh, but maybe some other thread succeeded?
1328          * recalc hash, maybe hash resize took place so we do a search
1329          */
1330         hash_val = ipoib_addr_hash(htbl, daddr);
1331         for (neigh = rcu_dereference_protected(htbl->buckets[hash_val],
1332                                                lockdep_is_held(&priv->lock));
1333              neigh != NULL;
1334              neigh = rcu_dereference_protected(neigh->hnext,
1335                                                lockdep_is_held(&priv->lock))) {
1336                 if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) {
1337                         /* found, take one ref on behalf of the caller */
1338                         if (!atomic_inc_not_zero(&neigh->refcnt)) {
1339                                 /* deleted */
1340                                 neigh = NULL;
1341                                 break;
1342                         }
1343                         neigh->alive = jiffies;
1344                         goto out_unlock;
1345                 }
1346         }
1347
1348         neigh = ipoib_neigh_ctor(daddr, dev);
1349         if (!neigh)
1350                 goto out_unlock;
1351
1352         /* one ref on behalf of the hash table */
1353         atomic_inc(&neigh->refcnt);
1354         neigh->alive = jiffies;
1355         /* put in hash */
1356         rcu_assign_pointer(neigh->hnext,
1357                            rcu_dereference_protected(htbl->buckets[hash_val],
1358                                                      lockdep_is_held(&priv->lock)));
1359         rcu_assign_pointer(htbl->buckets[hash_val], neigh);
1360         atomic_inc(&ntbl->entries);
1361
1362 out_unlock:
1363
1364         return neigh;
1365 }
1366
1367 void ipoib_neigh_dtor(struct ipoib_neigh *neigh)
1368 {
1369         /* neigh reference count was dropprd to zero */
1370         struct net_device *dev = neigh->dev;
1371         struct ipoib_dev_priv *priv = netdev_priv(dev);
1372         struct sk_buff *skb;
1373         if (neigh->ah)
1374                 ipoib_put_ah(neigh->ah);
1375         while ((skb = __skb_dequeue(&neigh->queue))) {
1376                 ++dev->stats.tx_dropped;
1377                 dev_kfree_skb_any(skb);
1378         }
1379         if (ipoib_cm_get(neigh))
1380                 ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
1381         ipoib_dbg(netdev_priv(dev),
1382                   "neigh free for %06x %pI6\n",
1383                   IPOIB_QPN(neigh->daddr),
1384                   neigh->daddr + 4);
1385         kfree(neigh);
1386         if (atomic_dec_and_test(&priv->ntbl.entries)) {
1387                 if (test_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags))
1388                         complete(&priv->ntbl.flushed);
1389         }
1390 }
1391
1392 static void ipoib_neigh_reclaim(struct rcu_head *rp)
1393 {
1394         /* Called as a result of removal from hash table */
1395         struct ipoib_neigh *neigh = container_of(rp, struct ipoib_neigh, rcu);
1396         /* note TX context may hold another ref */
1397         ipoib_neigh_put(neigh);
1398 }
1399
1400 void ipoib_neigh_free(struct ipoib_neigh *neigh)
1401 {
1402         struct net_device *dev = neigh->dev;
1403         struct ipoib_dev_priv *priv = netdev_priv(dev);
1404         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1405         struct ipoib_neigh_hash *htbl;
1406         struct ipoib_neigh __rcu **np;
1407         struct ipoib_neigh *n;
1408         u32 hash_val;
1409
1410         htbl = rcu_dereference_protected(ntbl->htbl,
1411                                         lockdep_is_held(&priv->lock));
1412         if (!htbl)
1413                 return;
1414
1415         hash_val = ipoib_addr_hash(htbl, neigh->daddr);
1416         np = &htbl->buckets[hash_val];
1417         for (n = rcu_dereference_protected(*np,
1418                                             lockdep_is_held(&priv->lock));
1419              n != NULL;
1420              n = rcu_dereference_protected(*np,
1421                                         lockdep_is_held(&priv->lock))) {
1422                 if (n == neigh) {
1423                         /* found */
1424                         rcu_assign_pointer(*np,
1425                                            rcu_dereference_protected(neigh->hnext,
1426                                                                      lockdep_is_held(&priv->lock)));
1427                         /* remove from parent list */
1428                         list_del(&neigh->list);
1429                         call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1430                         return;
1431                 } else {
1432                         np = &n->hnext;
1433                 }
1434         }
1435 }
1436
1437 static int ipoib_neigh_hash_init(struct ipoib_dev_priv *priv)
1438 {
1439         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1440         struct ipoib_neigh_hash *htbl;
1441         struct ipoib_neigh __rcu **buckets;
1442         u32 size;
1443
1444         clear_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags);
1445         ntbl->htbl = NULL;
1446         htbl = kzalloc(sizeof(*htbl), GFP_KERNEL);
1447         if (!htbl)
1448                 return -ENOMEM;
1449         set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1450         size = roundup_pow_of_two(arp_tbl.gc_thresh3);
1451         buckets = kzalloc(size * sizeof(*buckets), GFP_KERNEL);
1452         if (!buckets) {
1453                 kfree(htbl);
1454                 return -ENOMEM;
1455         }
1456         htbl->size = size;
1457         htbl->mask = (size - 1);
1458         htbl->buckets = buckets;
1459         RCU_INIT_POINTER(ntbl->htbl, htbl);
1460         htbl->ntbl = ntbl;
1461         atomic_set(&ntbl->entries, 0);
1462
1463         /* start garbage collection */
1464         clear_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1465         queue_delayed_work(priv->wq, &priv->neigh_reap_task,
1466                            arp_tbl.gc_interval);
1467
1468         return 0;
1469 }
1470
1471 static void neigh_hash_free_rcu(struct rcu_head *head)
1472 {
1473         struct ipoib_neigh_hash *htbl = container_of(head,
1474                                                     struct ipoib_neigh_hash,
1475                                                     rcu);
1476         struct ipoib_neigh __rcu **buckets = htbl->buckets;
1477         struct ipoib_neigh_table *ntbl = htbl->ntbl;
1478
1479         kfree(buckets);
1480         kfree(htbl);
1481         complete(&ntbl->deleted);
1482 }
1483
1484 void ipoib_del_neighs_by_gid(struct net_device *dev, u8 *gid)
1485 {
1486         struct ipoib_dev_priv *priv = netdev_priv(dev);
1487         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1488         struct ipoib_neigh_hash *htbl;
1489         unsigned long flags;
1490         int i;
1491
1492         /* remove all neigh connected to a given path or mcast */
1493         spin_lock_irqsave(&priv->lock, flags);
1494
1495         htbl = rcu_dereference_protected(ntbl->htbl,
1496                                          lockdep_is_held(&priv->lock));
1497
1498         if (!htbl)
1499                 goto out_unlock;
1500
1501         for (i = 0; i < htbl->size; i++) {
1502                 struct ipoib_neigh *neigh;
1503                 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1504
1505                 while ((neigh = rcu_dereference_protected(*np,
1506                                                           lockdep_is_held(&priv->lock))) != NULL) {
1507                         /* delete neighs belong to this parent */
1508                         if (!memcmp(gid, neigh->daddr + 4, sizeof (union ib_gid))) {
1509                                 rcu_assign_pointer(*np,
1510                                                    rcu_dereference_protected(neigh->hnext,
1511                                                                              lockdep_is_held(&priv->lock)));
1512                                 /* remove from parent list */
1513                                 list_del(&neigh->list);
1514                                 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1515                         } else {
1516                                 np = &neigh->hnext;
1517                         }
1518
1519                 }
1520         }
1521 out_unlock:
1522         spin_unlock_irqrestore(&priv->lock, flags);
1523 }
1524
1525 static void ipoib_flush_neighs(struct ipoib_dev_priv *priv)
1526 {
1527         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1528         struct ipoib_neigh_hash *htbl;
1529         unsigned long flags;
1530         int i, wait_flushed = 0;
1531
1532         init_completion(&priv->ntbl.flushed);
1533
1534         spin_lock_irqsave(&priv->lock, flags);
1535
1536         htbl = rcu_dereference_protected(ntbl->htbl,
1537                                         lockdep_is_held(&priv->lock));
1538         if (!htbl)
1539                 goto out_unlock;
1540
1541         wait_flushed = atomic_read(&priv->ntbl.entries);
1542         if (!wait_flushed)
1543                 goto free_htbl;
1544
1545         for (i = 0; i < htbl->size; i++) {
1546                 struct ipoib_neigh *neigh;
1547                 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1548
1549                 while ((neigh = rcu_dereference_protected(*np,
1550                                        lockdep_is_held(&priv->lock))) != NULL) {
1551                         rcu_assign_pointer(*np,
1552                                            rcu_dereference_protected(neigh->hnext,
1553                                                                      lockdep_is_held(&priv->lock)));
1554                         /* remove from path/mc list */
1555                         list_del(&neigh->list);
1556                         call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1557                 }
1558         }
1559
1560 free_htbl:
1561         rcu_assign_pointer(ntbl->htbl, NULL);
1562         call_rcu(&htbl->rcu, neigh_hash_free_rcu);
1563
1564 out_unlock:
1565         spin_unlock_irqrestore(&priv->lock, flags);
1566         if (wait_flushed)
1567                 wait_for_completion(&priv->ntbl.flushed);
1568 }
1569
1570 static void ipoib_neigh_hash_uninit(struct net_device *dev)
1571 {
1572         struct ipoib_dev_priv *priv = netdev_priv(dev);
1573         int stopped;
1574
1575         ipoib_dbg(priv, "ipoib_neigh_hash_uninit\n");
1576         init_completion(&priv->ntbl.deleted);
1577         set_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags);
1578
1579         /* Stop GC if called at init fail need to cancel work */
1580         stopped = test_and_set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1581         if (!stopped)
1582                 cancel_delayed_work(&priv->neigh_reap_task);
1583
1584         ipoib_flush_neighs(priv);
1585
1586         wait_for_completion(&priv->ntbl.deleted);
1587 }
1588
1589
1590 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
1591 {
1592         struct ipoib_dev_priv *priv = netdev_priv(dev);
1593
1594         /* Allocate RX/TX "rings" to hold queued skbs */
1595         priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
1596                                 GFP_KERNEL);
1597         if (!priv->rx_ring) {
1598                 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
1599                        ca->name, ipoib_recvq_size);
1600                 goto out;
1601         }
1602
1603         priv->tx_ring = vzalloc(ipoib_sendq_size * sizeof *priv->tx_ring);
1604         if (!priv->tx_ring) {
1605                 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
1606                        ca->name, ipoib_sendq_size);
1607                 goto out_rx_ring_cleanup;
1608         }
1609
1610         /* priv->tx_head, tx_tail & tx_outstanding are already 0 */
1611
1612         if (ipoib_ib_dev_init(dev, ca, port))
1613                 goto out_tx_ring_cleanup;
1614
1615         /*
1616          * Must be after ipoib_ib_dev_init so we can allocate a per
1617          * device wq there and use it here
1618          */
1619         if (ipoib_neigh_hash_init(priv) < 0)
1620                 goto out_dev_uninit;
1621
1622         return 0;
1623
1624 out_dev_uninit:
1625         ipoib_ib_dev_cleanup(dev);
1626
1627 out_tx_ring_cleanup:
1628         vfree(priv->tx_ring);
1629
1630 out_rx_ring_cleanup:
1631         kfree(priv->rx_ring);
1632
1633 out:
1634         return -ENOMEM;
1635 }
1636
1637 void ipoib_dev_cleanup(struct net_device *dev)
1638 {
1639         struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
1640         LIST_HEAD(head);
1641
1642         ASSERT_RTNL();
1643
1644         ipoib_delete_debug_files(dev);
1645
1646         /* Delete any child interfaces first */
1647         list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
1648                 /* Stop GC on child */
1649                 set_bit(IPOIB_STOP_NEIGH_GC, &cpriv->flags);
1650                 cancel_delayed_work(&cpriv->neigh_reap_task);
1651                 unregister_netdevice_queue(cpriv->dev, &head);
1652         }
1653         unregister_netdevice_many(&head);
1654
1655         /*
1656          * Must be before ipoib_ib_dev_cleanup or we delete an in use
1657          * work queue
1658          */
1659         ipoib_neigh_hash_uninit(dev);
1660
1661         ipoib_ib_dev_cleanup(dev);
1662
1663         kfree(priv->rx_ring);
1664         vfree(priv->tx_ring);
1665
1666         priv->rx_ring = NULL;
1667         priv->tx_ring = NULL;
1668 }
1669
1670 static int ipoib_set_vf_link_state(struct net_device *dev, int vf, int link_state)
1671 {
1672         struct ipoib_dev_priv *priv = netdev_priv(dev);
1673
1674         return ib_set_vf_link_state(priv->ca, vf, priv->port, link_state);
1675 }
1676
1677 static int ipoib_get_vf_config(struct net_device *dev, int vf,
1678                                struct ifla_vf_info *ivf)
1679 {
1680         struct ipoib_dev_priv *priv = netdev_priv(dev);
1681         int err;
1682
1683         err = ib_get_vf_config(priv->ca, vf, priv->port, ivf);
1684         if (err)
1685                 return err;
1686
1687         ivf->vf = vf;
1688
1689         return 0;
1690 }
1691
1692 static int ipoib_set_vf_guid(struct net_device *dev, int vf, u64 guid, int type)
1693 {
1694         struct ipoib_dev_priv *priv = netdev_priv(dev);
1695
1696         if (type != IFLA_VF_IB_NODE_GUID && type != IFLA_VF_IB_PORT_GUID)
1697                 return -EINVAL;
1698
1699         return ib_set_vf_guid(priv->ca, vf, priv->port, guid, type);
1700 }
1701
1702 static int ipoib_get_vf_stats(struct net_device *dev, int vf,
1703                               struct ifla_vf_stats *vf_stats)
1704 {
1705         struct ipoib_dev_priv *priv = netdev_priv(dev);
1706
1707         return ib_get_vf_stats(priv->ca, vf, priv->port, vf_stats);
1708 }
1709
1710 static const struct header_ops ipoib_header_ops = {
1711         .create = ipoib_hard_header,
1712 };
1713
1714 static const struct net_device_ops ipoib_netdev_ops_pf = {
1715         .ndo_uninit              = ipoib_uninit,
1716         .ndo_open                = ipoib_open,
1717         .ndo_stop                = ipoib_stop,
1718         .ndo_change_mtu          = ipoib_change_mtu,
1719         .ndo_fix_features        = ipoib_fix_features,
1720         .ndo_start_xmit          = ipoib_start_xmit,
1721         .ndo_tx_timeout          = ipoib_timeout,
1722         .ndo_set_rx_mode         = ipoib_set_mcast_list,
1723         .ndo_get_iflink          = ipoib_get_iflink,
1724         .ndo_set_vf_link_state   = ipoib_set_vf_link_state,
1725         .ndo_get_vf_config       = ipoib_get_vf_config,
1726         .ndo_get_vf_stats        = ipoib_get_vf_stats,
1727         .ndo_set_vf_guid         = ipoib_set_vf_guid,
1728         .ndo_set_mac_address     = ipoib_set_mac,
1729 };
1730
1731 static const struct net_device_ops ipoib_netdev_ops_vf = {
1732         .ndo_uninit              = ipoib_uninit,
1733         .ndo_open                = ipoib_open,
1734         .ndo_stop                = ipoib_stop,
1735         .ndo_change_mtu          = ipoib_change_mtu,
1736         .ndo_fix_features        = ipoib_fix_features,
1737         .ndo_start_xmit          = ipoib_start_xmit,
1738         .ndo_tx_timeout          = ipoib_timeout,
1739         .ndo_set_rx_mode         = ipoib_set_mcast_list,
1740         .ndo_get_iflink          = ipoib_get_iflink,
1741 };
1742
1743 void ipoib_setup(struct net_device *dev)
1744 {
1745         struct ipoib_dev_priv *priv = netdev_priv(dev);
1746
1747         if (priv->hca_caps & IB_DEVICE_VIRTUAL_FUNCTION)
1748                 dev->netdev_ops = &ipoib_netdev_ops_vf;
1749         else
1750                 dev->netdev_ops = &ipoib_netdev_ops_pf;
1751
1752         dev->header_ops          = &ipoib_header_ops;
1753
1754         ipoib_set_ethtool_ops(dev);
1755
1756         netif_napi_add(dev, &priv->napi, ipoib_poll, NAPI_POLL_WEIGHT);
1757
1758         dev->watchdog_timeo      = HZ;
1759
1760         dev->flags              |= IFF_BROADCAST | IFF_MULTICAST;
1761
1762         dev->hard_header_len     = IPOIB_ENCAP_LEN;
1763         dev->addr_len            = INFINIBAND_ALEN;
1764         dev->type                = ARPHRD_INFINIBAND;
1765         dev->tx_queue_len        = ipoib_sendq_size * 2;
1766         dev->features            = (NETIF_F_VLAN_CHALLENGED     |
1767                                     NETIF_F_HIGHDMA);
1768         netif_keep_dst(dev);
1769
1770         memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
1771
1772         priv->dev = dev;
1773
1774         spin_lock_init(&priv->lock);
1775
1776         init_rwsem(&priv->vlan_rwsem);
1777
1778         INIT_LIST_HEAD(&priv->path_list);
1779         INIT_LIST_HEAD(&priv->child_intfs);
1780         INIT_LIST_HEAD(&priv->dead_ahs);
1781         INIT_LIST_HEAD(&priv->multicast_list);
1782
1783         INIT_DELAYED_WORK(&priv->mcast_task,   ipoib_mcast_join_task);
1784         INIT_WORK(&priv->carrier_on_task, ipoib_mcast_carrier_on_task);
1785         INIT_WORK(&priv->flush_light,   ipoib_ib_dev_flush_light);
1786         INIT_WORK(&priv->flush_normal,   ipoib_ib_dev_flush_normal);
1787         INIT_WORK(&priv->flush_heavy,   ipoib_ib_dev_flush_heavy);
1788         INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
1789         INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
1790         INIT_DELAYED_WORK(&priv->neigh_reap_task, ipoib_reap_neigh);
1791 }
1792
1793 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
1794 {
1795         struct net_device *dev;
1796
1797         dev = alloc_netdev((int)sizeof(struct ipoib_dev_priv), name,
1798                            NET_NAME_UNKNOWN, ipoib_setup);
1799         if (!dev)
1800                 return NULL;
1801
1802         return netdev_priv(dev);
1803 }
1804
1805 static ssize_t show_pkey(struct device *dev,
1806                          struct device_attribute *attr, char *buf)
1807 {
1808         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1809
1810         return sprintf(buf, "0x%04x\n", priv->pkey);
1811 }
1812 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1813
1814 static ssize_t show_umcast(struct device *dev,
1815                            struct device_attribute *attr, char *buf)
1816 {
1817         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1818
1819         return sprintf(buf, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
1820 }
1821
1822 void ipoib_set_umcast(struct net_device *ndev, int umcast_val)
1823 {
1824         struct ipoib_dev_priv *priv = netdev_priv(ndev);
1825
1826         if (umcast_val > 0) {
1827                 set_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1828                 ipoib_warn(priv, "ignoring multicast groups joined directly "
1829                                 "by userspace\n");
1830         } else
1831                 clear_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1832 }
1833
1834 static ssize_t set_umcast(struct device *dev,
1835                           struct device_attribute *attr,
1836                           const char *buf, size_t count)
1837 {
1838         unsigned long umcast_val = simple_strtoul(buf, NULL, 0);
1839
1840         ipoib_set_umcast(to_net_dev(dev), umcast_val);
1841
1842         return count;
1843 }
1844 static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast);
1845
1846 int ipoib_add_umcast_attr(struct net_device *dev)
1847 {
1848         return device_create_file(&dev->dev, &dev_attr_umcast);
1849 }
1850
1851 static void set_base_guid(struct ipoib_dev_priv *priv, union ib_gid *gid)
1852 {
1853         struct ipoib_dev_priv *child_priv;
1854         struct net_device *netdev = priv->dev;
1855
1856         netif_addr_lock_bh(netdev);
1857
1858         memcpy(&priv->local_gid.global.interface_id,
1859                &gid->global.interface_id,
1860                sizeof(gid->global.interface_id));
1861         memcpy(netdev->dev_addr + 4, &priv->local_gid, sizeof(priv->local_gid));
1862         clear_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags);
1863
1864         netif_addr_unlock_bh(netdev);
1865
1866         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
1867                 down_read(&priv->vlan_rwsem);
1868                 list_for_each_entry(child_priv, &priv->child_intfs, list)
1869                         set_base_guid(child_priv, gid);
1870                 up_read(&priv->vlan_rwsem);
1871         }
1872 }
1873
1874 static int ipoib_check_lladdr(struct net_device *dev,
1875                               struct sockaddr_storage *ss)
1876 {
1877         union ib_gid *gid = (union ib_gid *)(ss->__data + 4);
1878         int ret = 0;
1879
1880         netif_addr_lock_bh(dev);
1881
1882         /* Make sure the QPN, reserved and subnet prefix match the current
1883          * lladdr, it also makes sure the lladdr is unicast.
1884          */
1885         if (memcmp(dev->dev_addr, ss->__data,
1886                    4 + sizeof(gid->global.subnet_prefix)) ||
1887             gid->global.interface_id == 0)
1888                 ret = -EINVAL;
1889
1890         netif_addr_unlock_bh(dev);
1891
1892         return ret;
1893 }
1894
1895 static int ipoib_set_mac(struct net_device *dev, void *addr)
1896 {
1897         struct ipoib_dev_priv *priv = netdev_priv(dev);
1898         struct sockaddr_storage *ss = addr;
1899         int ret;
1900
1901         if (!(dev->priv_flags & IFF_LIVE_ADDR_CHANGE) && netif_running(dev))
1902                 return -EBUSY;
1903
1904         ret = ipoib_check_lladdr(dev, ss);
1905         if (ret)
1906                 return ret;
1907
1908         set_base_guid(priv, (union ib_gid *)(ss->__data + 4));
1909
1910         queue_work(ipoib_workqueue, &priv->flush_light);
1911
1912         return 0;
1913 }
1914
1915 static ssize_t create_child(struct device *dev,
1916                             struct device_attribute *attr,
1917                             const char *buf, size_t count)
1918 {
1919         int pkey;
1920         int ret;
1921
1922         if (sscanf(buf, "%i", &pkey) != 1)
1923                 return -EINVAL;
1924
1925         if (pkey <= 0 || pkey > 0xffff || pkey == 0x8000)
1926                 return -EINVAL;
1927
1928         /*
1929          * Set the full membership bit, so that we join the right
1930          * broadcast group, etc.
1931          */
1932         pkey |= 0x8000;
1933
1934         ret = ipoib_vlan_add(to_net_dev(dev), pkey);
1935
1936         return ret ? ret : count;
1937 }
1938 static DEVICE_ATTR(create_child, S_IWUSR, NULL, create_child);
1939
1940 static ssize_t delete_child(struct device *dev,
1941                             struct device_attribute *attr,
1942                             const char *buf, size_t count)
1943 {
1944         int pkey;
1945         int ret;
1946
1947         if (sscanf(buf, "%i", &pkey) != 1)
1948                 return -EINVAL;
1949
1950         if (pkey < 0 || pkey > 0xffff)
1951                 return -EINVAL;
1952
1953         ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
1954
1955         return ret ? ret : count;
1956
1957 }
1958 static DEVICE_ATTR(delete_child, S_IWUSR, NULL, delete_child);
1959
1960 int ipoib_add_pkey_attr(struct net_device *dev)
1961 {
1962         return device_create_file(&dev->dev, &dev_attr_pkey);
1963 }
1964
1965 int ipoib_set_dev_features(struct ipoib_dev_priv *priv, struct ib_device *hca)
1966 {
1967         priv->hca_caps = hca->attrs.device_cap_flags;
1968
1969         if (priv->hca_caps & IB_DEVICE_UD_IP_CSUM) {
1970                 priv->dev->hw_features = NETIF_F_IP_CSUM | NETIF_F_RXCSUM;
1971
1972                 if (priv->hca_caps & IB_DEVICE_UD_TSO)
1973                         priv->dev->hw_features |= NETIF_F_TSO;
1974
1975                 priv->dev->features |= priv->dev->hw_features;
1976         }
1977
1978         return 0;
1979 }
1980
1981 static struct net_device *ipoib_add_port(const char *format,
1982                                          struct ib_device *hca, u8 port)
1983 {
1984         struct ipoib_dev_priv *priv;
1985         struct ib_port_attr attr;
1986         int result = -ENOMEM;
1987
1988         priv = ipoib_intf_alloc(format);
1989         if (!priv)
1990                 goto alloc_mem_failed;
1991
1992         SET_NETDEV_DEV(priv->dev, hca->dma_device);
1993         priv->dev->dev_id = port - 1;
1994
1995         result = ib_query_port(hca, port, &attr);
1996         if (!result)
1997                 priv->max_ib_mtu = ib_mtu_enum_to_int(attr.max_mtu);
1998         else {
1999                 printk(KERN_WARNING "%s: ib_query_port %d failed\n",
2000                        hca->name, port);
2001                 goto device_init_failed;
2002         }
2003
2004         /* MTU will be reset when mcast join happens */
2005         priv->dev->mtu  = IPOIB_UD_MTU(priv->max_ib_mtu);
2006         priv->mcast_mtu  = priv->admin_mtu = priv->dev->mtu;
2007
2008         priv->dev->neigh_priv_len = sizeof(struct ipoib_neigh);
2009
2010         result = ib_query_pkey(hca, port, 0, &priv->pkey);
2011         if (result) {
2012                 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
2013                        hca->name, port, result);
2014                 goto device_init_failed;
2015         }
2016
2017         result = ipoib_set_dev_features(priv, hca);
2018         if (result)
2019                 goto device_init_failed;
2020
2021         /*
2022          * Set the full membership bit, so that we join the right
2023          * broadcast group, etc.
2024          */
2025         priv->pkey |= 0x8000;
2026
2027         priv->dev->broadcast[8] = priv->pkey >> 8;
2028         priv->dev->broadcast[9] = priv->pkey & 0xff;
2029
2030         result = ib_query_gid(hca, port, 0, &priv->local_gid, NULL);
2031         if (result) {
2032                 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
2033                        hca->name, port, result);
2034                 goto device_init_failed;
2035         } else
2036                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
2037         set_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags);
2038
2039         result = ipoib_dev_init(priv->dev, hca, port);
2040         if (result < 0) {
2041                 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
2042                        hca->name, port, result);
2043                 goto device_init_failed;
2044         }
2045
2046         INIT_IB_EVENT_HANDLER(&priv->event_handler,
2047                               priv->ca, ipoib_event);
2048         result = ib_register_event_handler(&priv->event_handler);
2049         if (result < 0) {
2050                 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
2051                        "port %d (ret = %d)\n",
2052                        hca->name, port, result);
2053                 goto event_failed;
2054         }
2055
2056         result = register_netdev(priv->dev);
2057         if (result) {
2058                 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
2059                        hca->name, port, result);
2060                 goto register_failed;
2061         }
2062
2063         ipoib_create_debug_files(priv->dev);
2064
2065         if (ipoib_cm_add_mode_attr(priv->dev))
2066                 goto sysfs_failed;
2067         if (ipoib_add_pkey_attr(priv->dev))
2068                 goto sysfs_failed;
2069         if (ipoib_add_umcast_attr(priv->dev))
2070                 goto sysfs_failed;
2071         if (device_create_file(&priv->dev->dev, &dev_attr_create_child))
2072                 goto sysfs_failed;
2073         if (device_create_file(&priv->dev->dev, &dev_attr_delete_child))
2074                 goto sysfs_failed;
2075
2076         return priv->dev;
2077
2078 sysfs_failed:
2079         ipoib_delete_debug_files(priv->dev);
2080         unregister_netdev(priv->dev);
2081
2082 register_failed:
2083         ib_unregister_event_handler(&priv->event_handler);
2084         flush_workqueue(ipoib_workqueue);
2085         /* Stop GC if started before flush */
2086         set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
2087         cancel_delayed_work(&priv->neigh_reap_task);
2088         flush_workqueue(priv->wq);
2089
2090 event_failed:
2091         ipoib_dev_cleanup(priv->dev);
2092
2093 device_init_failed:
2094         free_netdev(priv->dev);
2095
2096 alloc_mem_failed:
2097         return ERR_PTR(result);
2098 }
2099
2100 static void ipoib_add_one(struct ib_device *device)
2101 {
2102         struct list_head *dev_list;
2103         struct net_device *dev;
2104         struct ipoib_dev_priv *priv;
2105         int p;
2106         int count = 0;
2107
2108         dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
2109         if (!dev_list)
2110                 return;
2111
2112         INIT_LIST_HEAD(dev_list);
2113
2114         for (p = rdma_start_port(device); p <= rdma_end_port(device); ++p) {
2115                 if (!rdma_protocol_ib(device, p))
2116                         continue;
2117                 dev = ipoib_add_port("ib%d", device, p);
2118                 if (!IS_ERR(dev)) {
2119                         priv = netdev_priv(dev);
2120                         list_add_tail(&priv->list, dev_list);
2121                         count++;
2122                 }
2123         }
2124
2125         if (!count) {
2126                 kfree(dev_list);
2127                 return;
2128         }
2129
2130         ib_set_client_data(device, &ipoib_client, dev_list);
2131 }
2132
2133 static void ipoib_remove_one(struct ib_device *device, void *client_data)
2134 {
2135         struct ipoib_dev_priv *priv, *tmp;
2136         struct list_head *dev_list = client_data;
2137
2138         if (!dev_list)
2139                 return;
2140
2141         list_for_each_entry_safe(priv, tmp, dev_list, list) {
2142                 ib_unregister_event_handler(&priv->event_handler);
2143                 flush_workqueue(ipoib_workqueue);
2144
2145                 /* mark interface in the middle of destruction */
2146                 set_bit(IPOIB_FLAG_GOING_DOWN, &priv->flags);
2147
2148                 rtnl_lock();
2149                 dev_change_flags(priv->dev, priv->dev->flags & ~IFF_UP);
2150                 rtnl_unlock();
2151
2152                 /* Stop GC */
2153                 set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
2154                 cancel_delayed_work(&priv->neigh_reap_task);
2155                 flush_workqueue(priv->wq);
2156
2157                 unregister_netdev(priv->dev);
2158                 free_netdev(priv->dev);
2159         }
2160
2161         kfree(dev_list);
2162 }
2163
2164 static int __init ipoib_init_module(void)
2165 {
2166         int ret;
2167
2168         ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
2169         ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
2170         ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
2171
2172         ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
2173         ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
2174         ipoib_sendq_size = max3(ipoib_sendq_size, 2 * MAX_SEND_CQE, IPOIB_MIN_QUEUE_SIZE);
2175 #ifdef CONFIG_INFINIBAND_IPOIB_CM
2176         ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
2177 #endif
2178
2179         /*
2180          * When copying small received packets, we only copy from the
2181          * linear data part of the SKB, so we rely on this condition.
2182          */
2183         BUILD_BUG_ON(IPOIB_CM_COPYBREAK > IPOIB_CM_HEAD_SIZE);
2184
2185         ret = ipoib_register_debugfs();
2186         if (ret)
2187                 return ret;
2188
2189         /*
2190          * We create a global workqueue here that is used for all flush
2191          * operations.  However, if you attempt to flush a workqueue
2192          * from a task on that same workqueue, it deadlocks the system.
2193          * We want to be able to flush the tasks associated with a
2194          * specific net device, so we also create a workqueue for each
2195          * netdevice.  We queue up the tasks for that device only on
2196          * its private workqueue, and we only queue up flush events
2197          * on our global flush workqueue.  This avoids the deadlocks.
2198          */
2199         ipoib_workqueue = create_singlethread_workqueue("ipoib_flush");
2200         if (!ipoib_workqueue) {
2201                 ret = -ENOMEM;
2202                 goto err_fs;
2203         }
2204
2205         ib_sa_register_client(&ipoib_sa_client);
2206
2207         ret = ib_register_client(&ipoib_client);
2208         if (ret)
2209                 goto err_sa;
2210
2211         ret = ipoib_netlink_init();
2212         if (ret)
2213                 goto err_client;
2214
2215         return 0;
2216
2217 err_client:
2218         ib_unregister_client(&ipoib_client);
2219
2220 err_sa:
2221         ib_sa_unregister_client(&ipoib_sa_client);
2222         destroy_workqueue(ipoib_workqueue);
2223
2224 err_fs:
2225         ipoib_unregister_debugfs();
2226
2227         return ret;
2228 }
2229
2230 static void __exit ipoib_cleanup_module(void)
2231 {
2232         ipoib_netlink_fini();
2233         ib_unregister_client(&ipoib_client);
2234         ib_sa_unregister_client(&ipoib_sa_client);
2235         ipoib_unregister_debugfs();
2236         destroy_workqueue(ipoib_workqueue);
2237 }
2238
2239 module_init(ipoib_init_module);
2240 module_exit(ipoib_cleanup_module);