err.no Git - linux-2.6/blob - 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         set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 102
 103         if (ipoib_pkey_dev_delay_open(dev))
 104                 return 0;
 105
 106         if (ipoib_ib_dev_open(dev))
 107                 return -EINVAL;
 108
 109         if (ipoib_ib_dev_up(dev)) {
 110                 ipoib_ib_dev_stop(dev, 1);
 111                 return -EINVAL;
 112         }
 113
 114         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 115                 struct ipoib_dev_priv *cpriv;
 116
 117                 /* Bring up any child interfaces too */
 118                 mutex_lock(&priv->vlan_mutex);
 119                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 120                         int flags;
 121
 122                         flags = cpriv->dev->flags;
 123                         if (flags & IFF_UP)
 124                                 continue;
 125
 126                         dev_change_flags(cpriv->dev, flags | IFF_UP);
 127                 }
 128                 mutex_unlock(&priv->vlan_mutex);
 129         }
 130
 131         netif_start_queue(dev);
 132
 133         return 0;
 134 }
 135
 136 static int ipoib_stop(struct net_device *dev)
 137 {
 138         struct ipoib_dev_priv *priv = netdev_priv(dev);
 139
 140         ipoib_dbg(priv, "stopping interface\n");
 141
 142         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 143
 144         netif_stop_queue(dev);
 145
 146         clear_bit(IPOIB_FLAG_NETIF_STOPPED, &priv->flags);
 147
 148         /*
 149          * Now flush workqueue to make sure a scheduled task doesn't
 150          * bring our internal state back up.
 151          */
 152         flush_workqueue(ipoib_workqueue);
 153
 154         ipoib_ib_dev_down(dev, 1);
 155         ipoib_ib_dev_stop(dev, 1);
 156
 157         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 158                 struct ipoib_dev_priv *cpriv;
 159
 160                 /* Bring down any child interfaces too */
 161                 mutex_lock(&priv->vlan_mutex);
 162                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 163                         int flags;
 164
 165                         flags = cpriv->dev->flags;
 166                         if (!(flags & IFF_UP))
 167                                 continue;
 168
 169                         dev_change_flags(cpriv->dev, flags & ~IFF_UP);
 170                 }
 171                 mutex_unlock(&priv->vlan_mutex);
 172         }
 173
 174         return 0;
 175 }
 176
 177 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
 178 {
 179         struct ipoib_dev_priv *priv = netdev_priv(dev);
 180
 181         /* dev->mtu > 2K ==> connected mode */
 182         if (ipoib_cm_admin_enabled(dev) && new_mtu <= IPOIB_CM_MTU) {
 183                 if (new_mtu > priv->mcast_mtu)
 184                         ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
 185                                    priv->mcast_mtu);
 186                 dev->mtu = new_mtu;
 187                 return 0;
 188         }
 189
 190         if (new_mtu > IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN) {
 191                 return -EINVAL;
 192         }
 193
 194         priv->admin_mtu = new_mtu;
 195
 196         dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
 197
 198         return 0;
 199 }
 200
 201 static struct ipoib_path *__path_find(struct net_device *dev, void *gid)
 202 {
 203         struct ipoib_dev_priv *priv = netdev_priv(dev);
 204         struct rb_node *n = priv->path_tree.rb_node;
 205         struct ipoib_path *path;
 206         int ret;
 207
 208         while (n) {
 209                 path = rb_entry(n, struct ipoib_path, rb_node);
 210
 211                 ret = memcmp(gid, path->pathrec.dgid.raw,
 212                              sizeof (union ib_gid));
 213
 214                 if (ret < 0)
 215                         n = n->rb_left;
 216                 else if (ret > 0)
 217                         n = n->rb_right;
 218                 else
 219                         return path;
 220         }
 221
 222         return NULL;
 223 }
 224
 225 static int __path_add(struct net_device *dev, struct ipoib_path *path)
 226 {
 227         struct ipoib_dev_priv *priv = netdev_priv(dev);
 228         struct rb_node **n = &priv->path_tree.rb_node;
 229         struct rb_node *pn = NULL;
 230         struct ipoib_path *tpath;
 231         int ret;
 232
 233         while (*n) {
 234                 pn = *n;
 235                 tpath = rb_entry(pn, struct ipoib_path, rb_node);
 236
 237                 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
 238                              sizeof (union ib_gid));
 239                 if (ret < 0)
 240                         n = &pn->rb_left;
 241                 else if (ret > 0)
 242                         n = &pn->rb_right;
 243                 else
 244                         return -EEXIST;
 245         }
 246
 247         rb_link_node(&path->rb_node, pn, n);
 248         rb_insert_color(&path->rb_node, &priv->path_tree);
 249
 250         list_add_tail(&path->list, &priv->path_list);
 251
 252         return 0;
 253 }
 254
 255 static void path_free(struct net_device *dev, struct ipoib_path *path)
 256 {
 257         struct ipoib_dev_priv *priv = netdev_priv(dev);
 258         struct ipoib_neigh *neigh, *tn;
 259         struct sk_buff *skb;
 260         unsigned long flags;
 261
 262         while ((skb = __skb_dequeue(&path->queue)))
 263                 dev_kfree_skb_irq(skb);
 264
 265         spin_lock_irqsave(&priv->lock, flags);
 266
 267         list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
 268                 /*
 269                  * It's safe to call ipoib_put_ah() inside priv->lock
 270                  * here, because we know that path->ah will always
 271                  * hold one more reference, so ipoib_put_ah() will
 272                  * never do more than decrement the ref count.
 273                  */
 274                 if (neigh->ah)
 275                         ipoib_put_ah(neigh->ah);
 276
 277                 ipoib_neigh_free(dev, neigh);
 278         }
 279
 280         spin_unlock_irqrestore(&priv->lock, flags);
 281
 282         if (path->ah)
 283                 ipoib_put_ah(path->ah);
 284
 285         kfree(path);
 286 }
 287
 288 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 289
 290 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
 291 {
 292         struct ipoib_path_iter *iter;
 293
 294         iter = kmalloc(sizeof *iter, GFP_KERNEL);
 295         if (!iter)
 296                 return NULL;
 297
 298         iter->dev = dev;
 299         memset(iter->path.pathrec.dgid.raw, 0, 16);
 300
 301         if (ipoib_path_iter_next(iter)) {
 302                 kfree(iter);
 303                 return NULL;
 304         }
 305
 306         return iter;
 307 }
 308
 309 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
 310 {
 311         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
 312         struct rb_node *n;
 313         struct ipoib_path *path;
 314         int ret = 1;
 315
 316         spin_lock_irq(&priv->lock);
 317
 318         n = rb_first(&priv->path_tree);
 319
 320         while (n) {
 321                 path = rb_entry(n, struct ipoib_path, rb_node);
 322
 323                 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
 324                            sizeof (union ib_gid)) < 0) {
 325                         iter->path = *path;
 326                         ret = 0;
 327                         break;
 328                 }
 329
 330                 n = rb_next(n);
 331         }
 332
 333         spin_unlock_irq(&priv->lock);
 334
 335         return ret;
 336 }
 337
 338 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
 339                           struct ipoib_path *path)
 340 {
 341         *path = iter->path;
 342 }
 343
 344 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
 345
 346 void ipoib_flush_paths(struct net_device *dev)
 347 {
 348         struct ipoib_dev_priv *priv = netdev_priv(dev);
 349         struct ipoib_path *path, *tp;
 350         LIST_HEAD(remove_list);
 351
 352         spin_lock_irq(&priv->tx_lock);
 353         spin_lock(&priv->lock);
 354
 355         list_splice(&priv->path_list, &remove_list);
 356         INIT_LIST_HEAD(&priv->path_list);
 357
 358         list_for_each_entry(path, &remove_list, list)
 359                 rb_erase(&path->rb_node, &priv->path_tree);
 360
 361         list_for_each_entry_safe(path, tp, &remove_list, list) {
 362                 if (path->query)
 363                         ib_sa_cancel_query(path->query_id, path->query);
 364                 spin_unlock(&priv->lock);
 365                 spin_unlock_irq(&priv->tx_lock);
 366                 wait_for_completion(&path->done);
 367                 path_free(dev, path);
 368                 spin_lock_irq(&priv->tx_lock);
 369                 spin_lock(&priv->lock);
 370         }
 371         spin_unlock(&priv->lock);
 372         spin_unlock_irq(&priv->tx_lock);
 373 }
 374
 375 static void path_rec_completion(int status,
 376                                 struct ib_sa_path_rec *pathrec,
 377                                 void *path_ptr)
 378 {
 379         struct ipoib_path *path = path_ptr;
 380         struct net_device *dev = path->dev;
 381         struct ipoib_dev_priv *priv = netdev_priv(dev);
 382         struct ipoib_ah *ah = NULL;
 383         struct ipoib_neigh *neigh, *tn;
 384         struct sk_buff_head skqueue;
 385         struct sk_buff *skb;
 386         unsigned long flags;
 387
 388         if (!status)
 389                 ipoib_dbg(priv, "PathRec LID 0x%04x for GID " IPOIB_GID_FMT "\n",
 390                           be16_to_cpu(pathrec->dlid), IPOIB_GID_ARG(pathrec->dgid));
 391         else
 392                 ipoib_dbg(priv, "PathRec status %d for GID " IPOIB_GID_FMT "\n",
 393                           status, IPOIB_GID_ARG(path->pathrec.dgid));
 394
 395         skb_queue_head_init(&skqueue);
 396
 397         if (!status) {
 398                 struct ib_ah_attr av;
 399
 400                 if (!ib_init_ah_from_path(priv->ca, priv->port, pathrec, &av))
 401                         ah = ipoib_create_ah(dev, priv->pd, &av);
 402         }
 403
 404         spin_lock_irqsave(&priv->lock, flags);
 405
 406         path->ah = ah;
 407
 408         if (ah) {
 409                 path->pathrec = *pathrec;
 410
 411                 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
 412                           ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
 413
 414                 while ((skb = __skb_dequeue(&path->queue)))
 415                         __skb_queue_tail(&skqueue, skb);
 416
 417                 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
 418                         kref_get(&path->ah->ref);
 419                         neigh->ah = path->ah;
 420                         memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
 421                                sizeof(union ib_gid));
 422
 423                         if (ipoib_cm_enabled(dev, neigh->neighbour)) {
 424                                 if (!ipoib_cm_get(neigh))
 425                                         ipoib_cm_set(neigh, ipoib_cm_create_tx(dev,
 426                                                                                path,
 427                                                                                neigh));
 428                                 if (!ipoib_cm_get(neigh)) {
 429                                         list_del(&neigh->list);
 430                                         if (neigh->ah)
 431                                                 ipoib_put_ah(neigh->ah);
 432                                         ipoib_neigh_free(dev, neigh);
 433                                         continue;
 434                                 }
 435                         }
 436
 437                         while ((skb = __skb_dequeue(&neigh->queue)))
 438                                 __skb_queue_tail(&skqueue, skb);
 439                 }
 440         }
 441
 442         path->query = NULL;
 443         complete(&path->done);
 444
 445         spin_unlock_irqrestore(&priv->lock, flags);
 446
 447         while ((skb = __skb_dequeue(&skqueue))) {
 448                 skb->dev = dev;
 449                 if (dev_queue_xmit(skb))
 450                         ipoib_warn(priv, "dev_queue_xmit failed "
 451                                    "to requeue packet\n");
 452         }
 453 }
 454
 455 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
 456 {
 457         struct ipoib_dev_priv *priv = netdev_priv(dev);
 458         struct ipoib_path *path;
 459
 460         path = kzalloc(sizeof *path, GFP_ATOMIC);
 461         if (!path)
 462                 return NULL;
 463
 464         path->dev = dev;
 465
 466         skb_queue_head_init(&path->queue);
 467
 468         INIT_LIST_HEAD(&path->neigh_list);
 469
 470         memcpy(path->pathrec.dgid.raw, gid, sizeof (union ib_gid));
 471         path->pathrec.sgid          = priv->local_gid;
 472         path->pathrec.pkey          = cpu_to_be16(priv->pkey);
 473         path->pathrec.numb_path     = 1;
 474         path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
 475
 476         return path;
 477 }
 478
 479 static int path_rec_start(struct net_device *dev,
 480                           struct ipoib_path *path)
 481 {
 482         struct ipoib_dev_priv *priv = netdev_priv(dev);
 483
 484         ipoib_dbg(priv, "Start path record lookup for " IPOIB_GID_FMT "\n",
 485                   IPOIB_GID_ARG(path->pathrec.dgid));
 486
 487         init_completion(&path->done);
 488
 489         path->query_id =
 490                 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
 491                                    &path->pathrec,
 492                                    IB_SA_PATH_REC_DGID          |
 493                                    IB_SA_PATH_REC_SGID          |
 494                                    IB_SA_PATH_REC_NUMB_PATH     |
 495                                    IB_SA_PATH_REC_TRAFFIC_CLASS |
 496                                    IB_SA_PATH_REC_PKEY,
 497                                    1000, GFP_ATOMIC,
 498                                    path_rec_completion,
 499                                    path, &path->query);
 500         if (path->query_id < 0) {
 501                 ipoib_warn(priv, "ib_sa_path_rec_get failed\n");
 502                 path->query = NULL;
 503                 return path->query_id;
 504         }
 505
 506         return 0;
 507 }
 508
 509 static void neigh_add_path(struct sk_buff *skb, struct net_device *dev)
 510 {
 511         struct ipoib_dev_priv *priv = netdev_priv(dev);
 512         struct ipoib_path *path;
 513         struct ipoib_neigh *neigh;
 514
 515         neigh = ipoib_neigh_alloc(skb->dst->neighbour);
 516         if (!neigh) {
 517                 ++priv->stats.tx_dropped;
 518                 dev_kfree_skb_any(skb);
 519                 return;
 520         }
 521
 522         /*
 523          * We can only be called from ipoib_start_xmit, so we're
 524          * inside tx_lock -- no need to save/restore flags.
 525          */
 526         spin_lock(&priv->lock);
 527
 528         path = __path_find(dev, skb->dst->neighbour->ha + 4);
 529         if (!path) {
 530                 path = path_rec_create(dev, skb->dst->neighbour->ha + 4);
 531                 if (!path)
 532                         goto err_path;
 533
 534                 __path_add(dev, path);
 535         }
 536
 537         list_add_tail(&neigh->list, &path->neigh_list);
 538
 539         if (path->ah) {
 540                 kref_get(&path->ah->ref);
 541                 neigh->ah = path->ah;
 542                 memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
 543                        sizeof(union ib_gid));
 544
 545                 if (ipoib_cm_enabled(dev, neigh->neighbour)) {
 546                         if (!ipoib_cm_get(neigh))
 547                                 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
 548                         if (!ipoib_cm_get(neigh)) {
 549                                 list_del(&neigh->list);
 550                                 if (neigh->ah)
 551                                         ipoib_put_ah(neigh->ah);
 552                                 ipoib_neigh_free(dev, neigh);
 553                                 goto err_drop;
 554                         }
 555                         if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE)
 556                                 __skb_queue_tail(&neigh->queue, skb);
 557                         else {
 558                                 ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
 559                                            skb_queue_len(&neigh->queue));
 560                                 goto err_drop;
 561                         }
 562                 } else
 563                         ipoib_send(dev, skb, path->ah, IPOIB_QPN(skb->dst->neighbour->ha));
 564         } else {
 565                 neigh->ah  = NULL;
 566
 567                 if (!path->query && path_rec_start(dev, path))
 568                         goto err_list;
 569
 570                 __skb_queue_tail(&neigh->queue, skb);
 571         }
 572
 573         spin_unlock(&priv->lock);
 574         return;
 575
 576 err_list:
 577         list_del(&neigh->list);
 578
 579 err_path:
 580         ipoib_neigh_free(dev, neigh);
 581 err_drop:
 582         ++priv->stats.tx_dropped;
 583         dev_kfree_skb_any(skb);
 584
 585         spin_unlock(&priv->lock);
 586 }
 587
 588 static void ipoib_path_lookup(struct sk_buff *skb, struct net_device *dev)
 589 {
 590         struct ipoib_dev_priv *priv = netdev_priv(skb->dev);
 591
 592         /* Look up path record for unicasts */
 593         if (skb->dst->neighbour->ha[4] != 0xff) {
 594                 neigh_add_path(skb, dev);
 595                 return;
 596         }
 597
 598         /* Add in the P_Key for multicasts */
 599         skb->dst->neighbour->ha[8] = (priv->pkey >> 8) & 0xff;
 600         skb->dst->neighbour->ha[9] = priv->pkey & 0xff;
 601         ipoib_mcast_send(dev, skb->dst->neighbour->ha + 4, skb);
 602 }
 603
 604 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
 605                              struct ipoib_pseudoheader *phdr)
 606 {
 607         struct ipoib_dev_priv *priv = netdev_priv(dev);
 608         struct ipoib_path *path;
 609
 610         /*
 611          * We can only be called from ipoib_start_xmit, so we're
 612          * inside tx_lock -- no need to save/restore flags.
 613          */
 614         spin_lock(&priv->lock);
 615
 616         path = __path_find(dev, phdr->hwaddr + 4);
 617         if (!path) {
 618                 path = path_rec_create(dev, phdr->hwaddr + 4);
 619                 if (path) {
 620                         /* put pseudoheader back on for next time */
 621                         skb_push(skb, sizeof *phdr);
 622                         __skb_queue_tail(&path->queue, skb);
 623
 624                         if (path_rec_start(dev, path)) {
 625                                 spin_unlock(&priv->lock);
 626                                 path_free(dev, path);
 627                                 return;
 628                         } else
 629                                 __path_add(dev, path);
 630                 } else {
 631                         ++priv->stats.tx_dropped;
 632                         dev_kfree_skb_any(skb);
 633                 }
 634
 635                 spin_unlock(&priv->lock);
 636                 return;
 637         }
 638
 639         if (path->ah) {
 640                 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
 641                           be16_to_cpu(path->pathrec.dlid));
 642
 643                 ipoib_send(dev, skb, path->ah, IPOIB_QPN(phdr->hwaddr));
 644         } else if ((path->query || !path_rec_start(dev, path)) &&
 645                    skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 646                 /* put pseudoheader back on for next time */
 647                 skb_push(skb, sizeof *phdr);
 648                 __skb_queue_tail(&path->queue, skb);
 649         } else {
 650                 ++priv->stats.tx_dropped;
 651                 dev_kfree_skb_any(skb);
 652         }
 653
 654         spin_unlock(&priv->lock);
 655 }
 656
 657 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
 658 {
 659         struct ipoib_dev_priv *priv = netdev_priv(dev);
 660         struct ipoib_neigh *neigh;
 661         unsigned long flags;
 662
 663         if (unlikely(!spin_trylock_irqsave(&priv->tx_lock, flags)))
 664                 return NETDEV_TX_LOCKED;
 665
 666         /*
 667          * Check if our queue is stopped.  Since we have the LLTX bit
 668          * set, we can't rely on netif_stop_queue() preventing our
 669          * xmit function from being called with a full queue.
 670          */
 671         if (unlikely(netif_queue_stopped(dev))) {
 672                 spin_unlock_irqrestore(&priv->tx_lock, flags);
 673                 return NETDEV_TX_BUSY;
 674         }
 675
 676         if (likely(skb->dst && skb->dst->neighbour)) {
 677                 if (unlikely(!*to_ipoib_neigh(skb->dst->neighbour))) {
 678                         ipoib_path_lookup(skb, dev);
 679                         goto out;
 680                 }
 681
 682                 neigh = *to_ipoib_neigh(skb->dst->neighbour);
 683
 684                 if (ipoib_cm_get(neigh)) {
 685                         if (ipoib_cm_up(neigh)) {
 686                                 ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
 687                                 goto out;
 688                         }
 689                 } else if (neigh->ah) {
 690                         if (unlikely(memcmp(&neigh->dgid.raw,
 691                                             skb->dst->neighbour->ha + 4,
 692                                             sizeof(union ib_gid)))) {
 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                         ++priv->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                                 ++priv->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 struct net_device_stats *ipoib_get_stats(struct net_device *dev)
 759 {
 760         struct ipoib_dev_priv *priv = netdev_priv(dev);
 761
 762         return &priv->stats;
 763 }
 764
 765 static void ipoib_timeout(struct net_device *dev)
 766 {
 767         struct ipoib_dev_priv *priv = netdev_priv(dev);
 768
 769         ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
 770                    jiffies_to_msecs(jiffies - dev->trans_start));
 771         ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
 772                    netif_queue_stopped(dev),
 773                    priv->tx_head, priv->tx_tail);
 774         /* XXX reset QP, etc. */
 775 }
 776
 777 static int ipoib_hard_header(struct sk_buff *skb,
 778                              struct net_device *dev,
 779                              unsigned short type,
 780                              void *daddr, void *saddr, unsigned len)
 781 {
 782         struct ipoib_header *header;
 783
 784         header = (struct ipoib_header *) skb_push(skb, sizeof *header);
 785
 786         header->proto = htons(type);
 787         header->reserved = 0;
 788
 789         /*
 790          * If we don't have a neighbour structure, stuff the
 791          * destination address onto the front of the skb so we can
 792          * figure out where to send the packet later.
 793          */
 794         if ((!skb->dst || !skb->dst->neighbour) && daddr) {
 795                 struct ipoib_pseudoheader *phdr =
 796                         (struct ipoib_pseudoheader *) skb_push(skb, sizeof *phdr);
 797                 memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
 798         }
 799
 800         return 0;
 801 }
 802
 803 static void ipoib_set_mcast_list(struct net_device *dev)
 804 {
 805         struct ipoib_dev_priv *priv = netdev_priv(dev);
 806
 807         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
 808                 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
 809                 return;
 810         }
 811
 812         queue_work(ipoib_workqueue, &priv->restart_task);
 813 }
 814
 815 static void ipoib_neigh_cleanup(struct neighbour *n)
 816 {
 817         struct ipoib_neigh *neigh;
 818         struct ipoib_dev_priv *priv = netdev_priv(n->dev);
 819         unsigned long flags;
 820         struct ipoib_ah *ah = NULL;
 821
 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         neigh = *to_ipoib_neigh(n);
 830         if (neigh) {
 831                 if (neigh->ah)
 832                         ah = neigh->ah;
 833                 list_del(&neigh->list);
 834                 ipoib_neigh_free(n->dev, neigh);
 835         }
 836
 837         spin_unlock_irqrestore(&priv->lock, flags);
 838
 839         if (ah)
 840                 ipoib_put_ah(ah);
 841 }
 842
 843 struct ipoib_neigh *ipoib_neigh_alloc(struct neighbour *neighbour)
 844 {
 845         struct ipoib_neigh *neigh;
 846
 847         neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
 848         if (!neigh)
 849                 return NULL;
 850
 851         neigh->neighbour = neighbour;
 852         *to_ipoib_neigh(neighbour) = neigh;
 853         skb_queue_head_init(&neigh->queue);
 854         ipoib_cm_set(neigh, NULL);
 855
 856         return neigh;
 857 }
 858
 859 void ipoib_neigh_free(struct net_device *dev, struct ipoib_neigh *neigh)
 860 {
 861         struct ipoib_dev_priv *priv = netdev_priv(dev);
 862         struct sk_buff *skb;
 863         *to_ipoib_neigh(neigh->neighbour) = NULL;
 864         while ((skb = __skb_dequeue(&neigh->queue))) {
 865                 ++priv->stats.tx_dropped;
 866                 dev_kfree_skb_any(skb);
 867         }
 868         if (ipoib_cm_get(neigh))
 869                 ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
 870         kfree(neigh);
 871 }
 872
 873 static int ipoib_neigh_setup_dev(struct net_device *dev, struct neigh_parms *parms)
 874 {
 875         parms->neigh_cleanup = ipoib_neigh_cleanup;
 876
 877         return 0;
 878 }
 879
 880 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
 881 {
 882         struct ipoib_dev_priv *priv = netdev_priv(dev);
 883
 884         /* Allocate RX/TX "rings" to hold queued skbs */
 885         priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
 886                                 GFP_KERNEL);
 887         if (!priv->rx_ring) {
 888                 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
 889                        ca->name, ipoib_recvq_size);
 890                 goto out;
 891         }
 892
 893         priv->tx_ring = kzalloc(ipoib_sendq_size * sizeof *priv->tx_ring,
 894                                 GFP_KERNEL);
 895         if (!priv->tx_ring) {
 896                 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
 897                        ca->name, ipoib_sendq_size);
 898                 goto out_rx_ring_cleanup;
 899         }
 900
 901         /* priv->tx_head & tx_tail are already 0 */
 902
 903         if (ipoib_ib_dev_init(dev, ca, port))
 904                 goto out_tx_ring_cleanup;
 905
 906         return 0;
 907
 908 out_tx_ring_cleanup:
 909         kfree(priv->tx_ring);
 910
 911 out_rx_ring_cleanup:
 912         kfree(priv->rx_ring);
 913
 914 out:
 915         return -ENOMEM;
 916 }
 917
 918 void ipoib_dev_cleanup(struct net_device *dev)
 919 {
 920         struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
 921
 922         ipoib_delete_debug_files(dev);
 923
 924         /* Delete any child interfaces first */
 925         list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
 926                 unregister_netdev(cpriv->dev);
 927                 ipoib_dev_cleanup(cpriv->dev);
 928                 free_netdev(cpriv->dev);
 929         }
 930
 931         ipoib_ib_dev_cleanup(dev);
 932
 933         kfree(priv->rx_ring);
 934         kfree(priv->tx_ring);
 935
 936         priv->rx_ring = NULL;
 937         priv->tx_ring = NULL;
 938 }
 939
 940 static void ipoib_setup(struct net_device *dev)
 941 {
 942         struct ipoib_dev_priv *priv = netdev_priv(dev);
 943
 944         dev->open                = ipoib_open;
 945         dev->stop                = ipoib_stop;
 946         dev->change_mtu          = ipoib_change_mtu;
 947         dev->hard_start_xmit     = ipoib_start_xmit;
 948         dev->get_stats           = ipoib_get_stats;
 949         dev->tx_timeout          = ipoib_timeout;
 950         dev->hard_header         = ipoib_hard_header;
 951         dev->set_multicast_list  = ipoib_set_mcast_list;
 952         dev->neigh_setup         = ipoib_neigh_setup_dev;
 953         dev->poll                = ipoib_poll;
 954         dev->weight              = 100;
 955
 956         dev->watchdog_timeo      = HZ;
 957
 958         dev->flags              |= IFF_BROADCAST | IFF_MULTICAST;
 959
 960         /*
 961          * We add in INFINIBAND_ALEN to allow for the destination
 962          * address "pseudoheader" for skbs without neighbour struct.
 963          */
 964         dev->hard_header_len     = IPOIB_ENCAP_LEN + INFINIBAND_ALEN;
 965         dev->addr_len            = INFINIBAND_ALEN;
 966         dev->type                = ARPHRD_INFINIBAND;
 967         dev->tx_queue_len        = ipoib_sendq_size * 2;
 968         dev->features            = NETIF_F_VLAN_CHALLENGED | NETIF_F_LLTX;
 969
 970         /* MTU will be reset when mcast join happens */
 971         dev->mtu                 = IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN;
 972         priv->mcast_mtu          = priv->admin_mtu = dev->mtu;
 973
 974         memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
 975
 976         netif_carrier_off(dev);
 977
 978         SET_MODULE_OWNER(dev);
 979
 980         priv->dev = dev;
 981
 982         spin_lock_init(&priv->lock);
 983         spin_lock_init(&priv->tx_lock);
 984
 985         mutex_init(&priv->mcast_mutex);
 986         mutex_init(&priv->vlan_mutex);
 987
 988         INIT_LIST_HEAD(&priv->path_list);
 989         INIT_LIST_HEAD(&priv->child_intfs);
 990         INIT_LIST_HEAD(&priv->dead_ahs);
 991         INIT_LIST_HEAD(&priv->multicast_list);
 992
 993         INIT_DELAYED_WORK(&priv->pkey_poll_task, ipoib_pkey_poll);
 994         INIT_WORK(&priv->pkey_event_task, ipoib_pkey_event);
 995         INIT_DELAYED_WORK(&priv->mcast_task,   ipoib_mcast_join_task);
 996         INIT_WORK(&priv->flush_task,   ipoib_ib_dev_flush);
 997         INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
 998         INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
 999 }
1000
1001 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
1002 {
1003         struct net_device *dev;
1004
1005         dev = alloc_netdev((int) sizeof (struct ipoib_dev_priv), name,
1006                            ipoib_setup);
1007         if (!dev)
1008                 return NULL;
1009
1010         return netdev_priv(dev);
1011 }
1012
1013 static ssize_t show_pkey(struct device *dev,
1014                          struct device_attribute *attr, char *buf)
1015 {
1016         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1017
1018         return sprintf(buf, "0x%04x\n", priv->pkey);
1019 }
1020 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1021
1022 static ssize_t show_umcast(struct device *dev,
1023                            struct device_attribute *attr, char *buf)
1024 {
1025         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1026
1027         return sprintf(buf, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
1028 }
1029
1030 static ssize_t set_umcast(struct device *dev,
1031                           struct device_attribute *attr,
1032                           const char *buf, size_t count)
1033 {
1034         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1035         unsigned long umcast_val = simple_strtoul(buf, NULL, 0);
1036
1037         if (umcast_val > 0) {
1038                 set_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1039                 ipoib_warn(priv, "ignoring multicast groups joined directly "
1040                                 "by userspace\n");
1041         } else
1042                 clear_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1043
1044         return count;
1045 }
1046 static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast);
1047
1048 int ipoib_add_umcast_attr(struct net_device *dev)
1049 {
1050         return device_create_file(&dev->dev, &dev_attr_umcast);
1051 }
1052
1053 static ssize_t create_child(struct device *dev,
1054                             struct device_attribute *attr,
1055                             const char *buf, size_t count)
1056 {
1057         int pkey;
1058         int ret;
1059
1060         if (sscanf(buf, "%i", &pkey) != 1)
1061                 return -EINVAL;
1062
1063         if (pkey < 0 || pkey > 0xffff)
1064                 return -EINVAL;
1065
1066         /*
1067          * Set the full membership bit, so that we join the right
1068          * broadcast group, etc.
1069          */
1070         pkey |= 0x8000;
1071
1072         ret = ipoib_vlan_add(to_net_dev(dev), pkey);
1073
1074         return ret ? ret : count;
1075 }
1076 static DEVICE_ATTR(create_child, S_IWUGO, NULL, create_child);
1077
1078 static ssize_t delete_child(struct device *dev,
1079                             struct device_attribute *attr,
1080                             const char *buf, size_t count)
1081 {
1082         int pkey;
1083         int ret;
1084
1085         if (sscanf(buf, "%i", &pkey) != 1)
1086                 return -EINVAL;
1087
1088         if (pkey < 0 || pkey > 0xffff)
1089                 return -EINVAL;
1090
1091         ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
1092
1093         return ret ? ret : count;
1094
1095 }
1096 static DEVICE_ATTR(delete_child, S_IWUGO, NULL, delete_child);
1097
1098 int ipoib_add_pkey_attr(struct net_device *dev)
1099 {
1100         return device_create_file(&dev->dev, &dev_attr_pkey);
1101 }
1102
1103 static struct net_device *ipoib_add_port(const char *format,
1104                                          struct ib_device *hca, u8 port)
1105 {
1106         struct ipoib_dev_priv *priv;
1107         int result = -ENOMEM;
1108
1109         priv = ipoib_intf_alloc(format);
1110         if (!priv)
1111                 goto alloc_mem_failed;
1112
1113         SET_NETDEV_DEV(priv->dev, hca->dma_device);
1114
1115         result = ib_query_pkey(hca, port, 0, &priv->pkey);
1116         if (result) {
1117                 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
1118                        hca->name, port, result);
1119                 goto device_init_failed;
1120         }
1121
1122         /*
1123          * Set the full membership bit, so that we join the right
1124          * broadcast group, etc.
1125          */
1126         priv->pkey |= 0x8000;
1127
1128         priv->dev->broadcast[8] = priv->pkey >> 8;
1129         priv->dev->broadcast[9] = priv->pkey & 0xff;
1130
1131         result = ib_query_gid(hca, port, 0, &priv->local_gid);
1132         if (result) {
1133                 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1134                        hca->name, port, result);
1135                 goto device_init_failed;
1136         } else
1137                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1138
1139
1140         result = ipoib_dev_init(priv->dev, hca, port);
1141         if (result < 0) {
1142                 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1143                        hca->name, port, result);
1144                 goto device_init_failed;
1145         }
1146
1147         INIT_IB_EVENT_HANDLER(&priv->event_handler,
1148                               priv->ca, ipoib_event);
1149         result = ib_register_event_handler(&priv->event_handler);
1150         if (result < 0) {
1151                 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1152                        "port %d (ret = %d)\n",
1153                        hca->name, port, result);
1154                 goto event_failed;
1155         }
1156
1157         result = register_netdev(priv->dev);
1158         if (result) {
1159                 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1160                        hca->name, port, result);
1161                 goto register_failed;
1162         }
1163
1164         ipoib_create_debug_files(priv->dev);
1165
1166         if (ipoib_cm_add_mode_attr(priv->dev))
1167                 goto sysfs_failed;
1168         if (ipoib_add_pkey_attr(priv->dev))
1169                 goto sysfs_failed;
1170         if (ipoib_add_umcast_attr(priv->dev))
1171                 goto sysfs_failed;
1172         if (device_create_file(&priv->dev->dev, &dev_attr_create_child))
1173                 goto sysfs_failed;
1174         if (device_create_file(&priv->dev->dev, &dev_attr_delete_child))
1175                 goto sysfs_failed;
1176
1177         return priv->dev;
1178
1179 sysfs_failed:
1180         ipoib_delete_debug_files(priv->dev);
1181         unregister_netdev(priv->dev);
1182
1183 register_failed:
1184         ib_unregister_event_handler(&priv->event_handler);
1185         flush_scheduled_work();
1186
1187 event_failed:
1188         ipoib_dev_cleanup(priv->dev);
1189
1190 device_init_failed:
1191         free_netdev(priv->dev);
1192
1193 alloc_mem_failed:
1194         return ERR_PTR(result);
1195 }
1196
1197 static void ipoib_add_one(struct ib_device *device)
1198 {
1199         struct list_head *dev_list;
1200         struct net_device *dev;
1201         struct ipoib_dev_priv *priv;
1202         int s, e, p;
1203
1204         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1205                 return;
1206
1207         dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1208         if (!dev_list)
1209                 return;
1210
1211         INIT_LIST_HEAD(dev_list);
1212
1213         if (device->node_type == RDMA_NODE_IB_SWITCH) {
1214                 s = 0;
1215                 e = 0;
1216         } else {
1217                 s = 1;
1218                 e = device->phys_port_cnt;
1219         }
1220
1221         for (p = s; p <= e; ++p) {
1222                 dev = ipoib_add_port("ib%d", device, p);
1223                 if (!IS_ERR(dev)) {
1224                         priv = netdev_priv(dev);
1225                         list_add_tail(&priv->list, dev_list);
1226                 }
1227         }
1228
1229         ib_set_client_data(device, &ipoib_client, dev_list);
1230 }
1231
1232 static void ipoib_remove_one(struct ib_device *device)
1233 {
1234         struct ipoib_dev_priv *priv, *tmp;
1235         struct list_head *dev_list;
1236
1237         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1238                 return;
1239
1240         dev_list = ib_get_client_data(device, &ipoib_client);
1241
1242         list_for_each_entry_safe(priv, tmp, dev_list, list) {
1243                 ib_unregister_event_handler(&priv->event_handler);
1244                 flush_scheduled_work();
1245
1246                 unregister_netdev(priv->dev);
1247                 ipoib_dev_cleanup(priv->dev);
1248                 free_netdev(priv->dev);
1249         }
1250
1251         kfree(dev_list);
1252 }
1253
1254 static int __init ipoib_init_module(void)
1255 {
1256         int ret;
1257
1258         ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1259         ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1260         ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1261
1262         ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1263         ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1264         ipoib_sendq_size = max(ipoib_sendq_size, IPOIB_MIN_QUEUE_SIZE);
1265
1266         ret = ipoib_register_debugfs();
1267         if (ret)
1268                 return ret;
1269
1270         /*
1271          * We create our own workqueue mainly because we want to be
1272          * able to flush it when devices are being removed.  We can't
1273          * use schedule_work()/flush_scheduled_work() because both
1274          * unregister_netdev() and linkwatch_event take the rtnl lock,
1275          * so flush_scheduled_work() can deadlock during device
1276          * removal.
1277          */
1278         ipoib_workqueue = create_singlethread_workqueue("ipoib");
1279         if (!ipoib_workqueue) {
1280                 ret = -ENOMEM;
1281                 goto err_fs;
1282         }
1283
1284         ib_sa_register_client(&ipoib_sa_client);
1285
1286         ret = ib_register_client(&ipoib_client);
1287         if (ret)
1288                 goto err_sa;
1289
1290         return 0;
1291
1292 err_sa:
1293         ib_sa_unregister_client(&ipoib_sa_client);
1294         destroy_workqueue(ipoib_workqueue);
1295
1296 err_fs:
1297         ipoib_unregister_debugfs();
1298
1299         return ret;
1300 }
1301
1302 static void __exit ipoib_cleanup_module(void)
1303 {
1304         ib_unregister_client(&ipoib_client);
1305         ib_sa_unregister_client(&ipoib_sa_client);
1306         ipoib_unregister_debugfs();
1307         destroy_workqueue(ipoib_workqueue);
1308 }
1309
1310 module_init(ipoib_init_module);
1311 module_exit(ipoib_cleanup_module);