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  * $Id: ipoib_main.c 1377 2004-12-23 19:57:12Z roland $
  35  */
  36
  37 #include "ipoib.h"
  38
  39 #include <linux/module.h>
  40
  41 #include <linux/init.h>
  42 #include <linux/slab.h>
  43 #include <linux/kernel.h>
  44
  45 #include <linux/if_arp.h>       /* For ARPHRD_xxx */
  46
  47 #include <linux/ip.h>
  48 #include <linux/in.h>
  49
  50 #include <net/dst.h>
  51
  52 MODULE_AUTHOR("Roland Dreier");
  53 MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
  54 MODULE_LICENSE("Dual BSD/GPL");
  55
  56 int ipoib_sendq_size __read_mostly = IPOIB_TX_RING_SIZE;
  57 int ipoib_recvq_size __read_mostly = IPOIB_RX_RING_SIZE;
  58
  59 module_param_named(send_queue_size, ipoib_sendq_size, int, 0444);
  60 MODULE_PARM_DESC(send_queue_size, "Number of descriptors in send queue");
  61 module_param_named(recv_queue_size, ipoib_recvq_size, int, 0444);
  62 MODULE_PARM_DESC(recv_queue_size, "Number of descriptors in receive queue");
  63
  64 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
  65 int ipoib_debug_level;
  66
  67 module_param_named(debug_level, ipoib_debug_level, int, 0644);
  68 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
  69 #endif
  70
  71 struct ipoib_path_iter {
  72         struct net_device *dev;
  73         struct ipoib_path  path;
  74 };
  75
  76 static const u8 ipv4_bcast_addr[] = {
  77         0x00, 0xff, 0xff, 0xff,
  78         0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
  79         0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff
  80 };
  81
  82 struct workqueue_struct *ipoib_workqueue;
  83
  84 struct ib_sa_client ipoib_sa_client;
  85
  86 static void ipoib_add_one(struct ib_device *device);
  87 static void ipoib_remove_one(struct ib_device *device);
  88
  89 static struct ib_client ipoib_client = {
  90         .name   = "ipoib",
  91         .add    = ipoib_add_one,
  92         .remove = ipoib_remove_one
  93 };
  94
  95 int ipoib_open(struct net_device *dev)
  96 {
  97         struct ipoib_dev_priv *priv = netdev_priv(dev);
  98
  99         ipoib_dbg(priv, "bringing up interface\n");
 100
 101         napi_enable(&priv->napi);
 102         set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 103
 104         if (ipoib_pkey_dev_delay_open(dev))
 105                 return 0;
 106
 107         if (ipoib_ib_dev_open(dev)) {
 108                 napi_disable(&priv->napi);
 109                 return -EINVAL;
 110         }
 111
 112         if (ipoib_ib_dev_up(dev)) {
 113                 ipoib_ib_dev_stop(dev, 1);
 114                 napi_disable(&priv->napi);
 115                 return -EINVAL;
 116         }
 117
 118         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 119                 struct ipoib_dev_priv *cpriv;
 120
 121                 /* Bring up any child interfaces too */
 122                 mutex_lock(&priv->vlan_mutex);
 123                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 124                         int flags;
 125
 126                         flags = cpriv->dev->flags;
 127                         if (flags & IFF_UP)
 128                                 continue;
 129
 130                         dev_change_flags(cpriv->dev, flags | IFF_UP);
 131                 }
 132                 mutex_unlock(&priv->vlan_mutex);
 133         }
 134
 135         netif_start_queue(dev);
 136
 137         return 0;
 138 }
 139
 140 static int ipoib_stop(struct net_device *dev)
 141 {
 142         struct ipoib_dev_priv *priv = netdev_priv(dev);
 143
 144         ipoib_dbg(priv, "stopping interface\n");
 145
 146         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 147         napi_disable(&priv->napi);
 148
 149         netif_stop_queue(dev);
 150
 151         /*
 152          * Now flush workqueue to make sure a scheduled task doesn't
 153          * bring our internal state back up.
 154          */
 155         flush_workqueue(ipoib_workqueue);
 156
 157         ipoib_ib_dev_down(dev, 1);
 158         ipoib_ib_dev_stop(dev, 1);
 159
 160         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 161                 struct ipoib_dev_priv *cpriv;
 162
 163                 /* Bring down any child interfaces too */
 164                 mutex_lock(&priv->vlan_mutex);
 165                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 166                         int flags;
 167
 168                         flags = cpriv->dev->flags;
 169                         if (!(flags & IFF_UP))
 170                                 continue;
 171
 172                         dev_change_flags(cpriv->dev, flags & ~IFF_UP);
 173                 }
 174                 mutex_unlock(&priv->vlan_mutex);
 175         }
 176
 177         return 0;
 178 }
 179
 180 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
 181 {
 182         struct ipoib_dev_priv *priv = netdev_priv(dev);
 183
 184         /* dev->mtu > 2K ==> connected mode */
 185         if (ipoib_cm_admin_enabled(dev)) {
 186                 if (new_mtu > ipoib_cm_max_mtu(dev))
 187                         return -EINVAL;
 188
 189                 if (new_mtu > priv->mcast_mtu)
 190                         ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
 191                                    priv->mcast_mtu);
 192
 193                 dev->mtu = new_mtu;
 194                 return 0;
 195         }
 196
 197         if (new_mtu > IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN)
 198                 return -EINVAL;
 199
 200         priv->admin_mtu = new_mtu;
 201
 202         dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
 203
 204         return 0;
 205 }
 206
 207 static struct ipoib_path *__path_find(struct net_device *dev, void *gid)
 208 {
 209         struct ipoib_dev_priv *priv = netdev_priv(dev);
 210         struct rb_node *n = priv->path_tree.rb_node;
 211         struct ipoib_path *path;
 212         int ret;
 213
 214         while (n) {
 215                 path = rb_entry(n, struct ipoib_path, rb_node);
 216
 217                 ret = memcmp(gid, path->pathrec.dgid.raw,
 218                              sizeof (union ib_gid));
 219
 220                 if (ret < 0)
 221                         n = n->rb_left;
 222                 else if (ret > 0)
 223                         n = n->rb_right;
 224                 else
 225                         return path;
 226         }
 227
 228         return NULL;
 229 }
 230
 231 static int __path_add(struct net_device *dev, struct ipoib_path *path)
 232 {
 233         struct ipoib_dev_priv *priv = netdev_priv(dev);
 234         struct rb_node **n = &priv->path_tree.rb_node;
 235         struct rb_node *pn = NULL;
 236         struct ipoib_path *tpath;
 237         int ret;
 238
 239         while (*n) {
 240                 pn = *n;
 241                 tpath = rb_entry(pn, struct ipoib_path, rb_node);
 242
 243                 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
 244                              sizeof (union ib_gid));
 245                 if (ret < 0)
 246                         n = &pn->rb_left;
 247                 else if (ret > 0)
 248                         n = &pn->rb_right;
 249                 else
 250                         return -EEXIST;
 251         }
 252
 253         rb_link_node(&path->rb_node, pn, n);
 254         rb_insert_color(&path->rb_node, &priv->path_tree);
 255
 256         list_add_tail(&path->list, &priv->path_list);
 257
 258         return 0;
 259 }
 260
 261 static void path_free(struct net_device *dev, struct ipoib_path *path)
 262 {
 263         struct ipoib_dev_priv *priv = netdev_priv(dev);
 264         struct ipoib_neigh *neigh, *tn;
 265         struct sk_buff *skb;
 266         unsigned long flags;
 267
 268         while ((skb = __skb_dequeue(&path->queue)))
 269                 dev_kfree_skb_irq(skb);
 270
 271         spin_lock_irqsave(&priv->lock, flags);
 272
 273         list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
 274                 /*
 275                  * It's safe to call ipoib_put_ah() inside priv->lock
 276                  * here, because we know that path->ah will always
 277                  * hold one more reference, so ipoib_put_ah() will
 278                  * never do more than decrement the ref count.
 279                  */
 280                 if (neigh->ah)
 281                         ipoib_put_ah(neigh->ah);
 282
 283                 ipoib_neigh_free(dev, neigh);
 284         }
 285
 286         spin_unlock_irqrestore(&priv->lock, flags);
 287
 288         if (path->ah)
 289                 ipoib_put_ah(path->ah);
 290
 291         kfree(path);
 292 }
 293
 294 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 295
 296 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
 297 {
 298         struct ipoib_path_iter *iter;
 299
 300         iter = kmalloc(sizeof *iter, GFP_KERNEL);
 301         if (!iter)
 302                 return NULL;
 303
 304         iter->dev = dev;
 305         memset(iter->path.pathrec.dgid.raw, 0, 16);
 306
 307         if (ipoib_path_iter_next(iter)) {
 308                 kfree(iter);
 309                 return NULL;
 310         }
 311
 312         return iter;
 313 }
 314
 315 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
 316 {
 317         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
 318         struct rb_node *n;
 319         struct ipoib_path *path;
 320         int ret = 1;
 321
 322         spin_lock_irq(&priv->lock);
 323
 324         n = rb_first(&priv->path_tree);
 325
 326         while (n) {
 327                 path = rb_entry(n, struct ipoib_path, rb_node);
 328
 329                 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
 330                            sizeof (union ib_gid)) < 0) {
 331                         iter->path = *path;
 332                         ret = 0;
 333                         break;
 334                 }
 335
 336                 n = rb_next(n);
 337         }
 338
 339         spin_unlock_irq(&priv->lock);
 340
 341         return ret;
 342 }
 343
 344 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
 345                           struct ipoib_path *path)
 346 {
 347         *path = iter->path;
 348 }
 349
 350 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
 351
 352 void ipoib_flush_paths(struct net_device *dev)
 353 {
 354         struct ipoib_dev_priv *priv = netdev_priv(dev);
 355         struct ipoib_path *path, *tp;
 356         LIST_HEAD(remove_list);
 357
 358         spin_lock_irq(&priv->tx_lock);
 359         spin_lock(&priv->lock);
 360
 361         list_splice(&priv->path_list, &remove_list);
 362         INIT_LIST_HEAD(&priv->path_list);
 363
 364         list_for_each_entry(path, &remove_list, list)
 365                 rb_erase(&path->rb_node, &priv->path_tree);
 366
 367         list_for_each_entry_safe(path, tp, &remove_list, list) {
 368                 if (path->query)
 369                         ib_sa_cancel_query(path->query_id, path->query);
 370                 spin_unlock(&priv->lock);
 371                 spin_unlock_irq(&priv->tx_lock);
 372                 wait_for_completion(&path->done);
 373                 path_free(dev, path);
 374                 spin_lock_irq(&priv->tx_lock);
 375                 spin_lock(&priv->lock);
 376         }
 377         spin_unlock(&priv->lock);
 378         spin_unlock_irq(&priv->tx_lock);
 379 }
 380
 381 static void path_rec_completion(int status,
 382                                 struct ib_sa_path_rec *pathrec,
 383                                 void *path_ptr)
 384 {
 385         struct ipoib_path *path = path_ptr;
 386         struct net_device *dev = path->dev;
 387         struct ipoib_dev_priv *priv = netdev_priv(dev);
 388         struct ipoib_ah *ah = NULL;
 389         struct ipoib_neigh *neigh, *tn;
 390         struct sk_buff_head skqueue;
 391         struct sk_buff *skb;
 392         unsigned long flags;
 393
 394         if (!status)
 395                 ipoib_dbg(priv, "PathRec LID 0x%04x for GID " IPOIB_GID_FMT "\n",
 396                           be16_to_cpu(pathrec->dlid), IPOIB_GID_ARG(pathrec->dgid));
 397         else
 398                 ipoib_dbg(priv, "PathRec status %d for GID " IPOIB_GID_FMT "\n",
 399                           status, IPOIB_GID_ARG(path->pathrec.dgid));
 400
 401         skb_queue_head_init(&skqueue);
 402
 403         if (!status) {
 404                 struct ib_ah_attr av;
 405
 406                 if (!ib_init_ah_from_path(priv->ca, priv->port, pathrec, &av))
 407                         ah = ipoib_create_ah(dev, priv->pd, &av);
 408         }
 409
 410         spin_lock_irqsave(&priv->lock, flags);
 411
 412         path->ah = ah;
 413
 414         if (ah) {
 415                 path->pathrec = *pathrec;
 416
 417                 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
 418                           ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
 419
 420                 while ((skb = __skb_dequeue(&path->queue)))
 421                         __skb_queue_tail(&skqueue, skb);
 422
 423                 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
 424                         kref_get(&path->ah->ref);
 425                         neigh->ah = path->ah;
 426                         memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
 427                                sizeof(union ib_gid));
 428
 429                         if (ipoib_cm_enabled(dev, neigh->neighbour)) {
 430                                 if (!ipoib_cm_get(neigh))
 431                                         ipoib_cm_set(neigh, ipoib_cm_create_tx(dev,
 432                                                                                path,
 433                                                                                neigh));
 434                                 if (!ipoib_cm_get(neigh)) {
 435                                         list_del(&neigh->list);
 436                                         if (neigh->ah)
 437                                                 ipoib_put_ah(neigh->ah);
 438                                         ipoib_neigh_free(dev, neigh);
 439                                         continue;
 440                                 }
 441                         }
 442
 443                         while ((skb = __skb_dequeue(&neigh->queue)))
 444                                 __skb_queue_tail(&skqueue, skb);
 445                 }
 446         }
 447
 448         path->query = NULL;
 449         complete(&path->done);
 450
 451         spin_unlock_irqrestore(&priv->lock, flags);
 452
 453         while ((skb = __skb_dequeue(&skqueue))) {
 454                 skb->dev = dev;
 455                 if (dev_queue_xmit(skb))
 456                         ipoib_warn(priv, "dev_queue_xmit failed "
 457                                    "to requeue packet\n");
 458         }
 459 }
 460
 461 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
 462 {
 463         struct ipoib_dev_priv *priv = netdev_priv(dev);
 464         struct ipoib_path *path;
 465
 466         if (!priv->broadcast)
 467                 return NULL;
 468
 469         path = kzalloc(sizeof *path, GFP_ATOMIC);
 470         if (!path)
 471                 return NULL;
 472
 473         path->dev = dev;
 474
 475         skb_queue_head_init(&path->queue);
 476
 477         INIT_LIST_HEAD(&path->neigh_list);
 478
 479         memcpy(path->pathrec.dgid.raw, gid, sizeof (union ib_gid));
 480         path->pathrec.sgid          = priv->local_gid;
 481         path->pathrec.pkey          = cpu_to_be16(priv->pkey);
 482         path->pathrec.numb_path     = 1;
 483         path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
 484
 485         return path;
 486 }
 487
 488 static int path_rec_start(struct net_device *dev,
 489                           struct ipoib_path *path)
 490 {
 491         struct ipoib_dev_priv *priv = netdev_priv(dev);
 492
 493         ipoib_dbg(priv, "Start path record lookup for " IPOIB_GID_FMT "\n",
 494                   IPOIB_GID_ARG(path->pathrec.dgid));
 495
 496         init_completion(&path->done);
 497
 498         path->query_id =
 499                 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
 500                                    &path->pathrec,
 501                                    IB_SA_PATH_REC_DGID          |
 502                                    IB_SA_PATH_REC_SGID          |
 503                                    IB_SA_PATH_REC_NUMB_PATH     |
 504                                    IB_SA_PATH_REC_TRAFFIC_CLASS |
 505                                    IB_SA_PATH_REC_PKEY,
 506                                    1000, GFP_ATOMIC,
 507                                    path_rec_completion,
 508                                    path, &path->query);
 509         if (path->query_id < 0) {
 510                 ipoib_warn(priv, "ib_sa_path_rec_get failed\n");
 511                 path->query = NULL;
 512                 return path->query_id;
 513         }
 514
 515         return 0;
 516 }
 517
 518 static void neigh_add_path(struct sk_buff *skb, struct net_device *dev)
 519 {
 520         struct ipoib_dev_priv *priv = netdev_priv(dev);
 521         struct ipoib_path *path;
 522         struct ipoib_neigh *neigh;
 523
 524         neigh = ipoib_neigh_alloc(skb->dst->neighbour, skb->dev);
 525         if (!neigh) {
 526                 ++dev->stats.tx_dropped;
 527                 dev_kfree_skb_any(skb);
 528                 return;
 529         }
 530
 531         /*
 532          * We can only be called from ipoib_start_xmit, so we're
 533          * inside tx_lock -- no need to save/restore flags.
 534          */
 535         spin_lock(&priv->lock);
 536
 537         path = __path_find(dev, skb->dst->neighbour->ha + 4);
 538         if (!path) {
 539                 path = path_rec_create(dev, skb->dst->neighbour->ha + 4);
 540                 if (!path)
 541                         goto err_path;
 542
 543                 __path_add(dev, path);
 544         }
 545
 546         list_add_tail(&neigh->list, &path->neigh_list);
 547
 548         if (path->ah) {
 549                 kref_get(&path->ah->ref);
 550                 neigh->ah = path->ah;
 551                 memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
 552                        sizeof(union ib_gid));
 553
 554                 if (ipoib_cm_enabled(dev, neigh->neighbour)) {
 555                         if (!ipoib_cm_get(neigh))
 556                                 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
 557                         if (!ipoib_cm_get(neigh)) {
 558                                 list_del(&neigh->list);
 559                                 if (neigh->ah)
 560                                         ipoib_put_ah(neigh->ah);
 561                                 ipoib_neigh_free(dev, neigh);
 562                                 goto err_drop;
 563                         }
 564                         if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE)
 565                                 __skb_queue_tail(&neigh->queue, skb);
 566                         else {
 567                                 ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
 568                                            skb_queue_len(&neigh->queue));
 569                                 goto err_drop;
 570                         }
 571                 } else
 572                         ipoib_send(dev, skb, path->ah, IPOIB_QPN(skb->dst->neighbour->ha));
 573         } else {
 574                 neigh->ah  = NULL;
 575
 576                 if (!path->query && path_rec_start(dev, path))
 577                         goto err_list;
 578
 579                 __skb_queue_tail(&neigh->queue, skb);
 580         }
 581
 582         spin_unlock(&priv->lock);
 583         return;
 584
 585 err_list:
 586         list_del(&neigh->list);
 587
 588 err_path:
 589         ipoib_neigh_free(dev, neigh);
 590 err_drop:
 591         ++dev->stats.tx_dropped;
 592         dev_kfree_skb_any(skb);
 593
 594         spin_unlock(&priv->lock);
 595 }
 596
 597 static void ipoib_path_lookup(struct sk_buff *skb, struct net_device *dev)
 598 {
 599         struct ipoib_dev_priv *priv = netdev_priv(skb->dev);
 600
 601         /* Look up path record for unicasts */
 602         if (skb->dst->neighbour->ha[4] != 0xff) {
 603                 neigh_add_path(skb, dev);
 604                 return;
 605         }
 606
 607         /* Add in the P_Key for multicasts */
 608         skb->dst->neighbour->ha[8] = (priv->pkey >> 8) & 0xff;
 609         skb->dst->neighbour->ha[9] = priv->pkey & 0xff;
 610         ipoib_mcast_send(dev, skb->dst->neighbour->ha + 4, skb);
 611 }
 612
 613 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
 614                              struct ipoib_pseudoheader *phdr)
 615 {
 616         struct ipoib_dev_priv *priv = netdev_priv(dev);
 617         struct ipoib_path *path;
 618
 619         /*
 620          * We can only be called from ipoib_start_xmit, so we're
 621          * inside tx_lock -- no need to save/restore flags.
 622          */
 623         spin_lock(&priv->lock);
 624
 625         path = __path_find(dev, phdr->hwaddr + 4);
 626         if (!path) {
 627                 path = path_rec_create(dev, phdr->hwaddr + 4);
 628                 if (path) {
 629                         /* put pseudoheader back on for next time */
 630                         skb_push(skb, sizeof *phdr);
 631                         __skb_queue_tail(&path->queue, skb);
 632
 633                         if (path_rec_start(dev, path)) {
 634                                 spin_unlock(&priv->lock);
 635                                 path_free(dev, path);
 636                                 return;
 637                         } else
 638                                 __path_add(dev, path);
 639                 } else {
 640                         ++dev->stats.tx_dropped;
 641                         dev_kfree_skb_any(skb);
 642                 }
 643
 644                 spin_unlock(&priv->lock);
 645                 return;
 646         }
 647
 648         if (path->ah) {
 649                 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
 650                           be16_to_cpu(path->pathrec.dlid));
 651
 652                 ipoib_send(dev, skb, path->ah, IPOIB_QPN(phdr->hwaddr));
 653         } else if ((path->query || !path_rec_start(dev, path)) &&
 654                    skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 655                 /* put pseudoheader back on for next time */
 656                 skb_push(skb, sizeof *phdr);
 657                 __skb_queue_tail(&path->queue, skb);
 658         } else {
 659                 ++dev->stats.tx_dropped;
 660                 dev_kfree_skb_any(skb);
 661         }
 662
 663         spin_unlock(&priv->lock);
 664 }
 665
 666 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
 667 {
 668         struct ipoib_dev_priv *priv = netdev_priv(dev);
 669         struct ipoib_neigh *neigh;
 670         unsigned long flags;
 671
 672         if (unlikely(!spin_trylock_irqsave(&priv->tx_lock, flags)))
 673                 return NETDEV_TX_LOCKED;
 674
 675         if (likely(skb->dst && skb->dst->neighbour)) {
 676                 if (unlikely(!*to_ipoib_neigh(skb->dst->neighbour))) {
 677                         ipoib_path_lookup(skb, dev);
 678                         goto out;
 679                 }
 680
 681                 neigh = *to_ipoib_neigh(skb->dst->neighbour);
 682
 683                 if (ipoib_cm_get(neigh)) {
 684                         if (ipoib_cm_up(neigh)) {
 685                                 ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
 686                                 goto out;
 687                         }
 688                 } else if (neigh->ah) {
 689                         if (unlikely((memcmp(&neigh->dgid.raw,
 690                                             skb->dst->neighbour->ha + 4,
 691                                             sizeof(union ib_gid))) ||
 692                                          (neigh->dev != dev))) {
 693                                 spin_lock(&priv->lock);
 694                                 /*
 695                                  * It's safe to call ipoib_put_ah() inside
 696                                  * priv->lock here, because we know that
 697                                  * path->ah will always hold one more reference,
 698                                  * so ipoib_put_ah() will never do more than
 699                                  * decrement the ref count.
 700                                  */
 701                                 ipoib_put_ah(neigh->ah);
 702                                 list_del(&neigh->list);
 703                                 ipoib_neigh_free(dev, neigh);
 704                                 spin_unlock(&priv->lock);
 705                                 ipoib_path_lookup(skb, dev);
 706                                 goto out;
 707                         }
 708
 709                         ipoib_send(dev, skb, neigh->ah, IPOIB_QPN(skb->dst->neighbour->ha));
 710                         goto out;
 711                 }
 712
 713                 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 714                         spin_lock(&priv->lock);
 715                         __skb_queue_tail(&neigh->queue, skb);
 716                         spin_unlock(&priv->lock);
 717                 } else {
 718                         ++dev->stats.tx_dropped;
 719                         dev_kfree_skb_any(skb);
 720                 }
 721         } else {
 722                 struct ipoib_pseudoheader *phdr =
 723                         (struct ipoib_pseudoheader *) skb->data;
 724                 skb_pull(skb, sizeof *phdr);
 725
 726                 if (phdr->hwaddr[4] == 0xff) {
 727                         /* Add in the P_Key for multicast*/
 728                         phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
 729                         phdr->hwaddr[9] = priv->pkey & 0xff;
 730
 731                         ipoib_mcast_send(dev, phdr->hwaddr + 4, skb);
 732                 } else {
 733                         /* unicast GID -- should be ARP or RARP reply */
 734
 735                         if ((be16_to_cpup((__be16 *) skb->data) != ETH_P_ARP) &&
 736                             (be16_to_cpup((__be16 *) skb->data) != ETH_P_RARP)) {
 737                                 ipoib_warn(priv, "Unicast, no %s: type %04x, QPN %06x "
 738                                            IPOIB_GID_FMT "\n",
 739                                            skb->dst ? "neigh" : "dst",
 740                                            be16_to_cpup((__be16 *) skb->data),
 741                                            IPOIB_QPN(phdr->hwaddr),
 742                                            IPOIB_GID_RAW_ARG(phdr->hwaddr + 4));
 743                                 dev_kfree_skb_any(skb);
 744                                 ++dev->stats.tx_dropped;
 745                                 goto out;
 746                         }
 747
 748                         unicast_arp_send(skb, dev, phdr);
 749                 }
 750         }
 751
 752 out:
 753         spin_unlock_irqrestore(&priv->tx_lock, flags);
 754
 755         return NETDEV_TX_OK;
 756 }
 757
 758 static void ipoib_timeout(struct net_device *dev)
 759 {
 760         struct ipoib_dev_priv *priv = netdev_priv(dev);
 761
 762         ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
 763                    jiffies_to_msecs(jiffies - dev->trans_start));
 764         ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
 765                    netif_queue_stopped(dev),
 766                    priv->tx_head, priv->tx_tail);
 767         /* XXX reset QP, etc. */
 768 }
 769
 770 static int ipoib_hard_header(struct sk_buff *skb,
 771                              struct net_device *dev,
 772                              unsigned short type,
 773                              const void *daddr, const void *saddr, unsigned len)
 774 {
 775         struct ipoib_header *header;
 776
 777         header = (struct ipoib_header *) skb_push(skb, sizeof *header);
 778
 779         header->proto = htons(type);
 780         header->reserved = 0;
 781
 782         /*
 783          * If we don't have a neighbour structure, stuff the
 784          * destination address onto the front of the skb so we can
 785          * figure out where to send the packet later.
 786          */
 787         if ((!skb->dst || !skb->dst->neighbour) && daddr) {
 788                 struct ipoib_pseudoheader *phdr =
 789                         (struct ipoib_pseudoheader *) skb_push(skb, sizeof *phdr);
 790                 memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
 791         }
 792
 793         return 0;
 794 }
 795
 796 static void ipoib_set_mcast_list(struct net_device *dev)
 797 {
 798         struct ipoib_dev_priv *priv = netdev_priv(dev);
 799
 800         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
 801                 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
 802                 return;
 803         }
 804
 805         queue_work(ipoib_workqueue, &priv->restart_task);
 806 }
 807
 808 static void ipoib_neigh_cleanup(struct neighbour *n)
 809 {
 810         struct ipoib_neigh *neigh;
 811         struct ipoib_dev_priv *priv = netdev_priv(n->dev);
 812         unsigned long flags;
 813         struct ipoib_ah *ah = NULL;
 814
 815         neigh = *to_ipoib_neigh(n);
 816         if (neigh) {
 817                 priv = netdev_priv(neigh->dev);
 818                 ipoib_dbg(priv, "neigh_destructor for bonding device: %s\n",
 819                           n->dev->name);
 820         } else
 821                 return;
 822         ipoib_dbg(priv,
 823                   "neigh_cleanup for %06x " IPOIB_GID_FMT "\n",
 824                   IPOIB_QPN(n->ha),
 825                   IPOIB_GID_RAW_ARG(n->ha + 4));
 826
 827         spin_lock_irqsave(&priv->lock, flags);
 828
 829         if (neigh->ah)
 830                 ah = neigh->ah;
 831         list_del(&neigh->list);
 832         ipoib_neigh_free(n->dev, neigh);
 833
 834         spin_unlock_irqrestore(&priv->lock, flags);
 835
 836         if (ah)
 837                 ipoib_put_ah(ah);
 838 }
 839
 840 struct ipoib_neigh *ipoib_neigh_alloc(struct neighbour *neighbour,
 841                                       struct net_device *dev)
 842 {
 843         struct ipoib_neigh *neigh;
 844
 845         neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
 846         if (!neigh)
 847                 return NULL;
 848
 849         neigh->neighbour = neighbour;
 850         neigh->dev = dev;
 851         *to_ipoib_neigh(neighbour) = neigh;
 852         skb_queue_head_init(&neigh->queue);
 853         ipoib_cm_set(neigh, NULL);
 854
 855         return neigh;
 856 }
 857
 858 void ipoib_neigh_free(struct net_device *dev, struct ipoib_neigh *neigh)
 859 {
 860         struct sk_buff *skb;
 861         *to_ipoib_neigh(neigh->neighbour) = NULL;
 862         while ((skb = __skb_dequeue(&neigh->queue))) {
 863                 ++dev->stats.tx_dropped;
 864                 dev_kfree_skb_any(skb);
 865         }
 866         if (ipoib_cm_get(neigh))
 867                 ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
 868         kfree(neigh);
 869 }
 870
 871 static int ipoib_neigh_setup_dev(struct net_device *dev, struct neigh_parms *parms)
 872 {
 873         parms->neigh_cleanup = ipoib_neigh_cleanup;
 874
 875         return 0;
 876 }
 877
 878 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
 879 {
 880         struct ipoib_dev_priv *priv = netdev_priv(dev);
 881
 882         /* Allocate RX/TX "rings" to hold queued skbs */
 883         priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
 884                                 GFP_KERNEL);
 885         if (!priv->rx_ring) {
 886                 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
 887                        ca->name, ipoib_recvq_size);
 888                 goto out;
 889         }
 890
 891         priv->tx_ring = kzalloc(ipoib_sendq_size * sizeof *priv->tx_ring,
 892                                 GFP_KERNEL);
 893         if (!priv->tx_ring) {
 894                 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
 895                        ca->name, ipoib_sendq_size);
 896                 goto out_rx_ring_cleanup;
 897         }
 898
 899         /* priv->tx_head, tx_tail & tx_outstanding are already 0 */
 900
 901         if (ipoib_ib_dev_init(dev, ca, port))
 902                 goto out_tx_ring_cleanup;
 903
 904         return 0;
 905
 906 out_tx_ring_cleanup:
 907         kfree(priv->tx_ring);
 908
 909 out_rx_ring_cleanup:
 910         kfree(priv->rx_ring);
 911
 912 out:
 913         return -ENOMEM;
 914 }
 915
 916 void ipoib_dev_cleanup(struct net_device *dev)
 917 {
 918         struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
 919
 920         ipoib_delete_debug_files(dev);
 921
 922         /* Delete any child interfaces first */
 923         list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
 924                 unregister_netdev(cpriv->dev);
 925                 ipoib_dev_cleanup(cpriv->dev);
 926                 free_netdev(cpriv->dev);
 927         }
 928
 929         ipoib_ib_dev_cleanup(dev);
 930
 931         kfree(priv->rx_ring);
 932         kfree(priv->tx_ring);
 933
 934         priv->rx_ring = NULL;
 935         priv->tx_ring = NULL;
 936 }
 937
 938 static const struct header_ops ipoib_header_ops = {
 939         .create = ipoib_hard_header,
 940 };
 941
 942 static void ipoib_setup(struct net_device *dev)
 943 {
 944         struct ipoib_dev_priv *priv = netdev_priv(dev);
 945
 946         dev->open                = ipoib_open;
 947         dev->stop                = ipoib_stop;
 948         dev->change_mtu          = ipoib_change_mtu;
 949         dev->hard_start_xmit     = ipoib_start_xmit;
 950         dev->tx_timeout          = ipoib_timeout;
 951         dev->header_ops          = &ipoib_header_ops;
 952         dev->set_multicast_list  = ipoib_set_mcast_list;
 953         dev->neigh_setup         = ipoib_neigh_setup_dev;
 954
 955         netif_napi_add(dev, &priv->napi, ipoib_poll, 100);
 956
 957         dev->watchdog_timeo      = HZ;
 958
 959         dev->flags              |= IFF_BROADCAST | IFF_MULTICAST;
 960
 961         /*
 962          * We add in INFINIBAND_ALEN to allow for the destination
 963          * address "pseudoheader" for skbs without neighbour struct.
 964          */
 965         dev->hard_header_len     = IPOIB_ENCAP_LEN + INFINIBAND_ALEN;
 966         dev->addr_len            = INFINIBAND_ALEN;
 967         dev->type                = ARPHRD_INFINIBAND;
 968         dev->tx_queue_len        = ipoib_sendq_size * 2;
 969         dev->features            = NETIF_F_VLAN_CHALLENGED | NETIF_F_LLTX;
 970
 971         /* MTU will be reset when mcast join happens */
 972         dev->mtu                 = IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN;
 973         priv->mcast_mtu          = priv->admin_mtu = dev->mtu;
 974
 975         memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
 976
 977         netif_carrier_off(dev);
 978
 979         priv->dev = dev;
 980
 981         spin_lock_init(&priv->lock);
 982         spin_lock_init(&priv->tx_lock);
 983
 984         mutex_init(&priv->mcast_mutex);
 985         mutex_init(&priv->vlan_mutex);
 986
 987         INIT_LIST_HEAD(&priv->path_list);
 988         INIT_LIST_HEAD(&priv->child_intfs);
 989         INIT_LIST_HEAD(&priv->dead_ahs);
 990         INIT_LIST_HEAD(&priv->multicast_list);
 991
 992         INIT_DELAYED_WORK(&priv->pkey_poll_task, ipoib_pkey_poll);
 993         INIT_WORK(&priv->pkey_event_task, ipoib_pkey_event);
 994         INIT_DELAYED_WORK(&priv->mcast_task,   ipoib_mcast_join_task);
 995         INIT_WORK(&priv->flush_task,   ipoib_ib_dev_flush);
 996         INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
 997         INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
 998 }
 999
1000 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
1001 {
1002         struct net_device *dev;
1003
1004         dev = alloc_netdev((int) sizeof (struct ipoib_dev_priv), name,
1005                            ipoib_setup);
1006         if (!dev)
1007                 return NULL;
1008
1009         return netdev_priv(dev);
1010 }
1011
1012 static ssize_t show_pkey(struct device *dev,
1013                          struct device_attribute *attr, char *buf)
1014 {
1015         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1016
1017         return sprintf(buf, "0x%04x\n", priv->pkey);
1018 }
1019 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1020
1021 static ssize_t show_umcast(struct device *dev,
1022                            struct device_attribute *attr, char *buf)
1023 {
1024         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1025
1026         return sprintf(buf, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
1027 }
1028
1029 static ssize_t set_umcast(struct device *dev,
1030                           struct device_attribute *attr,
1031                           const char *buf, size_t count)
1032 {
1033         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1034         unsigned long umcast_val = simple_strtoul(buf, NULL, 0);
1035
1036         if (umcast_val > 0) {
1037                 set_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1038                 ipoib_warn(priv, "ignoring multicast groups joined directly "
1039                                 "by userspace\n");
1040         } else
1041                 clear_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1042
1043         return count;
1044 }
1045 static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast);
1046
1047 int ipoib_add_umcast_attr(struct net_device *dev)
1048 {
1049         return device_create_file(&dev->dev, &dev_attr_umcast);
1050 }
1051
1052 static ssize_t create_child(struct device *dev,
1053                             struct device_attribute *attr,
1054                             const char *buf, size_t count)
1055 {
1056         int pkey;
1057         int ret;
1058
1059         if (sscanf(buf, "%i", &pkey) != 1)
1060                 return -EINVAL;
1061
1062         if (pkey < 0 || pkey > 0xffff)
1063                 return -EINVAL;
1064
1065         /*
1066          * Set the full membership bit, so that we join the right
1067          * broadcast group, etc.
1068          */
1069         pkey |= 0x8000;
1070
1071         ret = ipoib_vlan_add(to_net_dev(dev), pkey);
1072
1073         return ret ? ret : count;
1074 }
1075 static DEVICE_ATTR(create_child, S_IWUGO, NULL, create_child);
1076
1077 static ssize_t delete_child(struct device *dev,
1078                             struct device_attribute *attr,
1079                             const char *buf, size_t count)
1080 {
1081         int pkey;
1082         int ret;
1083
1084         if (sscanf(buf, "%i", &pkey) != 1)
1085                 return -EINVAL;
1086
1087         if (pkey < 0 || pkey > 0xffff)
1088                 return -EINVAL;
1089
1090         ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
1091
1092         return ret ? ret : count;
1093
1094 }
1095 static DEVICE_ATTR(delete_child, S_IWUGO, NULL, delete_child);
1096
1097 int ipoib_add_pkey_attr(struct net_device *dev)
1098 {
1099         return device_create_file(&dev->dev, &dev_attr_pkey);
1100 }
1101
1102 static struct net_device *ipoib_add_port(const char *format,
1103                                          struct ib_device *hca, u8 port)
1104 {
1105         struct ipoib_dev_priv *priv;
1106         int result = -ENOMEM;
1107
1108         priv = ipoib_intf_alloc(format);
1109         if (!priv)
1110                 goto alloc_mem_failed;
1111
1112         SET_NETDEV_DEV(priv->dev, hca->dma_device);
1113
1114         result = ib_query_pkey(hca, port, 0, &priv->pkey);
1115         if (result) {
1116                 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
1117                        hca->name, port, result);
1118                 goto device_init_failed;
1119         }
1120
1121         /*
1122          * Set the full membership bit, so that we join the right
1123          * broadcast group, etc.
1124          */
1125         priv->pkey |= 0x8000;
1126
1127         priv->dev->broadcast[8] = priv->pkey >> 8;
1128         priv->dev->broadcast[9] = priv->pkey & 0xff;
1129
1130         result = ib_query_gid(hca, port, 0, &priv->local_gid);
1131         if (result) {
1132                 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1133                        hca->name, port, result);
1134                 goto device_init_failed;
1135         } else
1136                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1137
1138
1139         result = ipoib_dev_init(priv->dev, hca, port);
1140         if (result < 0) {
1141                 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1142                        hca->name, port, result);
1143                 goto device_init_failed;
1144         }
1145
1146         INIT_IB_EVENT_HANDLER(&priv->event_handler,
1147                               priv->ca, ipoib_event);
1148         result = ib_register_event_handler(&priv->event_handler);
1149         if (result < 0) {
1150                 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1151                        "port %d (ret = %d)\n",
1152                        hca->name, port, result);
1153                 goto event_failed;
1154         }
1155
1156         result = register_netdev(priv->dev);
1157         if (result) {
1158                 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1159                        hca->name, port, result);
1160                 goto register_failed;
1161         }
1162
1163         ipoib_create_debug_files(priv->dev);
1164
1165         if (ipoib_cm_add_mode_attr(priv->dev))
1166                 goto sysfs_failed;
1167         if (ipoib_add_pkey_attr(priv->dev))
1168                 goto sysfs_failed;
1169         if (ipoib_add_umcast_attr(priv->dev))
1170                 goto sysfs_failed;
1171         if (device_create_file(&priv->dev->dev, &dev_attr_create_child))
1172                 goto sysfs_failed;
1173         if (device_create_file(&priv->dev->dev, &dev_attr_delete_child))
1174                 goto sysfs_failed;
1175
1176         return priv->dev;
1177
1178 sysfs_failed:
1179         ipoib_delete_debug_files(priv->dev);
1180         unregister_netdev(priv->dev);
1181
1182 register_failed:
1183         ib_unregister_event_handler(&priv->event_handler);
1184         flush_scheduled_work();
1185
1186 event_failed:
1187         ipoib_dev_cleanup(priv->dev);
1188
1189 device_init_failed:
1190         free_netdev(priv->dev);
1191
1192 alloc_mem_failed:
1193         return ERR_PTR(result);
1194 }
1195
1196 static void ipoib_add_one(struct ib_device *device)
1197 {
1198         struct list_head *dev_list;
1199         struct net_device *dev;
1200         struct ipoib_dev_priv *priv;
1201         int s, e, p;
1202
1203         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1204                 return;
1205
1206         dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1207         if (!dev_list)
1208                 return;
1209
1210         INIT_LIST_HEAD(dev_list);
1211
1212         if (device->node_type == RDMA_NODE_IB_SWITCH) {
1213                 s = 0;
1214                 e = 0;
1215         } else {
1216                 s = 1;
1217                 e = device->phys_port_cnt;
1218         }
1219
1220         for (p = s; p <= e; ++p) {
1221                 dev = ipoib_add_port("ib%d", device, p);
1222                 if (!IS_ERR(dev)) {
1223                         priv = netdev_priv(dev);
1224                         list_add_tail(&priv->list, dev_list);
1225                 }
1226         }
1227
1228         ib_set_client_data(device, &ipoib_client, dev_list);
1229 }
1230
1231 static void ipoib_remove_one(struct ib_device *device)
1232 {
1233         struct ipoib_dev_priv *priv, *tmp;
1234         struct list_head *dev_list;
1235
1236         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1237                 return;
1238
1239         dev_list = ib_get_client_data(device, &ipoib_client);
1240
1241         list_for_each_entry_safe(priv, tmp, dev_list, list) {
1242                 ib_unregister_event_handler(&priv->event_handler);
1243                 flush_scheduled_work();
1244
1245                 unregister_netdev(priv->dev);
1246                 ipoib_dev_cleanup(priv->dev);
1247                 free_netdev(priv->dev);
1248         }
1249
1250         kfree(dev_list);
1251 }
1252
1253 static int __init ipoib_init_module(void)
1254 {
1255         int ret;
1256
1257         ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1258         ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1259         ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1260
1261         ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1262         ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1263         ipoib_sendq_size = max(ipoib_sendq_size, IPOIB_MIN_QUEUE_SIZE);
1264 #ifdef CONFIG_INFINIBAND_IPOIB_CM
1265         ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
1266 #endif
1267
1268         ret = ipoib_register_debugfs();
1269         if (ret)
1270                 return ret;
1271
1272         /*
1273          * We create our own workqueue mainly because we want to be
1274          * able to flush it when devices are being removed.  We can't
1275          * use schedule_work()/flush_scheduled_work() because both
1276          * unregister_netdev() and linkwatch_event take the rtnl lock,
1277          * so flush_scheduled_work() can deadlock during device
1278          * removal.
1279          */
1280         ipoib_workqueue = create_singlethread_workqueue("ipoib");
1281         if (!ipoib_workqueue) {
1282                 ret = -ENOMEM;
1283                 goto err_fs;
1284         }
1285
1286         ib_sa_register_client(&ipoib_sa_client);
1287
1288         ret = ib_register_client(&ipoib_client);
1289         if (ret)
1290                 goto err_sa;
1291
1292         return 0;
1293
1294 err_sa:
1295         ib_sa_unregister_client(&ipoib_sa_client);
1296         destroy_workqueue(ipoib_workqueue);
1297
1298 err_fs:
1299         ipoib_unregister_debugfs();
1300
1301         return ret;
1302 }
1303
1304 static void __exit ipoib_cleanup_module(void)
1305 {
1306         ib_unregister_client(&ipoib_client);
1307         ib_sa_unregister_client(&ipoib_sa_client);
1308         ipoib_unregister_debugfs();
1309         destroy_workqueue(ipoib_workqueue);
1310 }
1311
1312 module_init(ipoib_init_module);
1313 module_exit(ipoib_cleanup_module);