SAFE public projects git trees. - safe/jmp/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         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         clear_bit(IPOIB_FLAG_NETIF_STOPPED, &priv->flags);
 152
 153         /*
 154          * Now flush workqueue to make sure a scheduled task doesn't
 155          * bring our internal state back up.
 156          */
 157         flush_workqueue(ipoib_workqueue);
 158
 159         ipoib_ib_dev_down(dev, 1);
 160         ipoib_ib_dev_stop(dev, 1);
 161
 162         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 163                 struct ipoib_dev_priv *cpriv;
 164
 165                 /* Bring down any child interfaces too */
 166                 mutex_lock(&priv->vlan_mutex);
 167                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 168                         int flags;
 169
 170                         flags = cpriv->dev->flags;
 171                         if (!(flags & IFF_UP))
 172                                 continue;
 173
 174                         dev_change_flags(cpriv->dev, flags & ~IFF_UP);
 175                 }
 176                 mutex_unlock(&priv->vlan_mutex);
 177         }
 178
 179         return 0;
 180 }
 181
 182 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
 183 {
 184         struct ipoib_dev_priv *priv = netdev_priv(dev);
 185
 186         /* dev->mtu > 2K ==> connected mode */
 187         if (ipoib_cm_admin_enabled(dev) && new_mtu <= IPOIB_CM_MTU) {
 188                 if (new_mtu > priv->mcast_mtu)
 189                         ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
 190                                    priv->mcast_mtu);
 191                 dev->mtu = new_mtu;
 192                 return 0;
 193         }
 194
 195         if (new_mtu > IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN) {
 196                 return -EINVAL;
 197         }
 198
 199         priv->admin_mtu = new_mtu;
 200
 201         dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
 202
 203         return 0;
 204 }
 205
 206 static struct ipoib_path *__path_find(struct net_device *dev, void *gid)
 207 {
 208         struct ipoib_dev_priv *priv = netdev_priv(dev);
 209         struct rb_node *n = priv->path_tree.rb_node;
 210         struct ipoib_path *path;
 211         int ret;
 212
 213         while (n) {
 214                 path = rb_entry(n, struct ipoib_path, rb_node);
 215
 216                 ret = memcmp(gid, path->pathrec.dgid.raw,
 217                              sizeof (union ib_gid));
 218
 219                 if (ret < 0)
 220                         n = n->rb_left;
 221                 else if (ret > 0)
 222                         n = n->rb_right;
 223                 else
 224                         return path;
 225         }
 226
 227         return NULL;
 228 }
 229
 230 static int __path_add(struct net_device *dev, struct ipoib_path *path)
 231 {
 232         struct ipoib_dev_priv *priv = netdev_priv(dev);
 233         struct rb_node **n = &priv->path_tree.rb_node;
 234         struct rb_node *pn = NULL;
 235         struct ipoib_path *tpath;
 236         int ret;
 237
 238         while (*n) {
 239                 pn = *n;
 240                 tpath = rb_entry(pn, struct ipoib_path, rb_node);
 241
 242                 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
 243                              sizeof (union ib_gid));
 244                 if (ret < 0)
 245                         n = &pn->rb_left;
 246                 else if (ret > 0)
 247                         n = &pn->rb_right;
 248                 else
 249                         return -EEXIST;
 250         }
 251
 252         rb_link_node(&path->rb_node, pn, n);
 253         rb_insert_color(&path->rb_node, &priv->path_tree);
 254
 255         list_add_tail(&path->list, &priv->path_list);
 256
 257         return 0;
 258 }
 259
 260 static void path_free(struct net_device *dev, struct ipoib_path *path)
 261 {
 262         struct ipoib_dev_priv *priv = netdev_priv(dev);
 263         struct ipoib_neigh *neigh, *tn;
 264         struct sk_buff *skb;
 265         unsigned long flags;
 266
 267         while ((skb = __skb_dequeue(&path->queue)))
 268                 dev_kfree_skb_irq(skb);
 269
 270         spin_lock_irqsave(&priv->lock, flags);
 271
 272         list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
 273                 /*
 274                  * It's safe to call ipoib_put_ah() inside priv->lock
 275                  * here, because we know that path->ah will always
 276                  * hold one more reference, so ipoib_put_ah() will
 277                  * never do more than decrement the ref count.
 278                  */
 279                 if (neigh->ah)
 280                         ipoib_put_ah(neigh->ah);
 281
 282                 ipoib_neigh_free(dev, neigh);
 283         }
 284
 285         spin_unlock_irqrestore(&priv->lock, flags);
 286
 287         if (path->ah)
 288                 ipoib_put_ah(path->ah);
 289
 290         kfree(path);
 291 }
 292
 293 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 294
 295 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
 296 {
 297         struct ipoib_path_iter *iter;
 298
 299         iter = kmalloc(sizeof *iter, GFP_KERNEL);
 300         if (!iter)
 301                 return NULL;
 302
 303         iter->dev = dev;
 304         memset(iter->path.pathrec.dgid.raw, 0, 16);
 305
 306         if (ipoib_path_iter_next(iter)) {
 307                 kfree(iter);
 308                 return NULL;
 309         }
 310
 311         return iter;
 312 }
 313
 314 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
 315 {
 316         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
 317         struct rb_node *n;
 318         struct ipoib_path *path;
 319         int ret = 1;
 320
 321         spin_lock_irq(&priv->lock);
 322
 323         n = rb_first(&priv->path_tree);
 324
 325         while (n) {
 326                 path = rb_entry(n, struct ipoib_path, rb_node);
 327
 328                 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
 329                            sizeof (union ib_gid)) < 0) {
 330                         iter->path = *path;
 331                         ret = 0;
 332                         break;
 333                 }
 334
 335                 n = rb_next(n);
 336         }
 337
 338         spin_unlock_irq(&priv->lock);
 339
 340         return ret;
 341 }
 342
 343 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
 344                           struct ipoib_path *path)
 345 {
 346         *path = iter->path;
 347 }
 348
 349 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
 350
 351 void ipoib_flush_paths(struct net_device *dev)
 352 {
 353         struct ipoib_dev_priv *priv = netdev_priv(dev);
 354         struct ipoib_path *path, *tp;
 355         LIST_HEAD(remove_list);
 356
 357         spin_lock_irq(&priv->tx_lock);
 358         spin_lock(&priv->lock);
 359
 360         list_splice(&priv->path_list, &remove_list);
 361         INIT_LIST_HEAD(&priv->path_list);
 362
 363         list_for_each_entry(path, &remove_list, list)
 364                 rb_erase(&path->rb_node, &priv->path_tree);
 365
 366         list_for_each_entry_safe(path, tp, &remove_list, list) {
 367                 if (path->query)
 368                         ib_sa_cancel_query(path->query_id, path->query);
 369                 spin_unlock(&priv->lock);
 370                 spin_unlock_irq(&priv->tx_lock);
 371                 wait_for_completion(&path->done);
 372                 path_free(dev, path);
 373                 spin_lock_irq(&priv->tx_lock);
 374                 spin_lock(&priv->lock);
 375         }
 376         spin_unlock(&priv->lock);
 377         spin_unlock_irq(&priv->tx_lock);
 378 }
 379
 380 static void path_rec_completion(int status,
 381                                 struct ib_sa_path_rec *pathrec,
 382                                 void *path_ptr)
 383 {
 384         struct ipoib_path *path = path_ptr;
 385         struct net_device *dev = path->dev;
 386         struct ipoib_dev_priv *priv = netdev_priv(dev);
 387         struct ipoib_ah *ah = NULL;
 388         struct ipoib_neigh *neigh, *tn;
 389         struct sk_buff_head skqueue;
 390         struct sk_buff *skb;
 391         unsigned long flags;
 392
 393         if (!status)
 394                 ipoib_dbg(priv, "PathRec LID 0x%04x for GID " IPOIB_GID_FMT "\n",
 395                           be16_to_cpu(pathrec->dlid), IPOIB_GID_ARG(pathrec->dgid));
 396         else
 397                 ipoib_dbg(priv, "PathRec status %d for GID " IPOIB_GID_FMT "\n",
 398                           status, IPOIB_GID_ARG(path->pathrec.dgid));
 399
 400         skb_queue_head_init(&skqueue);
 401
 402         if (!status) {
 403                 struct ib_ah_attr av;
 404
 405                 if (!ib_init_ah_from_path(priv->ca, priv->port, pathrec, &av))
 406                         ah = ipoib_create_ah(dev, priv->pd, &av);
 407         }
 408
 409         spin_lock_irqsave(&priv->lock, flags);
 410
 411         path->ah = ah;
 412
 413         if (ah) {
 414                 path->pathrec = *pathrec;
 415
 416                 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
 417                           ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
 418
 419                 while ((skb = __skb_dequeue(&path->queue)))
 420                         __skb_queue_tail(&skqueue, skb);
 421
 422                 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
 423                         kref_get(&path->ah->ref);
 424                         neigh->ah = path->ah;
 425                         memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
 426                                sizeof(union ib_gid));
 427
 428                         if (ipoib_cm_enabled(dev, neigh->neighbour)) {
 429                                 if (!ipoib_cm_get(neigh))
 430                                         ipoib_cm_set(neigh, ipoib_cm_create_tx(dev,
 431                                                                                path,
 432                                                                                neigh));
 433                                 if (!ipoib_cm_get(neigh)) {
 434                                         list_del(&neigh->list);
 435                                         if (neigh->ah)
 436                                                 ipoib_put_ah(neigh->ah);
 437                                         ipoib_neigh_free(dev, neigh);
 438                                         continue;
 439                                 }
 440                         }
 441
 442                         while ((skb = __skb_dequeue(&neigh->queue)))
 443                                 __skb_queue_tail(&skqueue, skb);
 444                 }
 445         }
 446
 447         path->query = NULL;
 448         complete(&path->done);
 449
 450         spin_unlock_irqrestore(&priv->lock, flags);
 451
 452         while ((skb = __skb_dequeue(&skqueue))) {
 453                 skb->dev = dev;
 454                 if (dev_queue_xmit(skb))
 455                         ipoib_warn(priv, "dev_queue_xmit failed "
 456                                    "to requeue packet\n");
 457         }
 458 }
 459
 460 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
 461 {
 462         struct ipoib_dev_priv *priv = netdev_priv(dev);
 463         struct ipoib_path *path;
 464
 465         path = kzalloc(sizeof *path, GFP_ATOMIC);
 466         if (!path)
 467                 return NULL;
 468
 469         path->dev = dev;
 470
 471         skb_queue_head_init(&path->queue);
 472
 473         INIT_LIST_HEAD(&path->neigh_list);
 474
 475         memcpy(path->pathrec.dgid.raw, gid, sizeof (union ib_gid));
 476         path->pathrec.sgid          = priv->local_gid;
 477         path->pathrec.pkey          = cpu_to_be16(priv->pkey);
 478         path->pathrec.numb_path     = 1;
 479         path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
 480
 481         return path;
 482 }
 483
 484 static int path_rec_start(struct net_device *dev,
 485                           struct ipoib_path *path)
 486 {
 487         struct ipoib_dev_priv *priv = netdev_priv(dev);
 488
 489         ipoib_dbg(priv, "Start path record lookup for " IPOIB_GID_FMT "\n",
 490                   IPOIB_GID_ARG(path->pathrec.dgid));
 491
 492         init_completion(&path->done);
 493
 494         path->query_id =
 495                 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
 496                                    &path->pathrec,
 497                                    IB_SA_PATH_REC_DGID          |
 498                                    IB_SA_PATH_REC_SGID          |
 499                                    IB_SA_PATH_REC_NUMB_PATH     |
 500                                    IB_SA_PATH_REC_TRAFFIC_CLASS |
 501                                    IB_SA_PATH_REC_PKEY,
 502                                    1000, GFP_ATOMIC,
 503                                    path_rec_completion,
 504                                    path, &path->query);
 505         if (path->query_id < 0) {
 506                 ipoib_warn(priv, "ib_sa_path_rec_get failed\n");
 507                 path->query = NULL;
 508                 return path->query_id;
 509         }
 510
 511         return 0;
 512 }
 513
 514 static void neigh_add_path(struct sk_buff *skb, struct net_device *dev)
 515 {
 516         struct ipoib_dev_priv *priv = netdev_priv(dev);
 517         struct ipoib_path *path;
 518         struct ipoib_neigh *neigh;
 519
 520         neigh = ipoib_neigh_alloc(skb->dst->neighbour, skb->dev);
 521         if (!neigh) {
 522                 ++dev->stats.tx_dropped;
 523                 dev_kfree_skb_any(skb);
 524                 return;
 525         }
 526
 527         /*
 528          * We can only be called from ipoib_start_xmit, so we're
 529          * inside tx_lock -- no need to save/restore flags.
 530          */
 531         spin_lock(&priv->lock);
 532
 533         path = __path_find(dev, skb->dst->neighbour->ha + 4);
 534         if (!path) {
 535                 path = path_rec_create(dev, skb->dst->neighbour->ha + 4);
 536                 if (!path)
 537                         goto err_path;
 538
 539                 __path_add(dev, path);
 540         }
 541
 542         list_add_tail(&neigh->list, &path->neigh_list);
 543
 544         if (path->ah) {
 545                 kref_get(&path->ah->ref);
 546                 neigh->ah = path->ah;
 547                 memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
 548                        sizeof(union ib_gid));
 549
 550                 if (ipoib_cm_enabled(dev, neigh->neighbour)) {
 551                         if (!ipoib_cm_get(neigh))
 552                                 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
 553                         if (!ipoib_cm_get(neigh)) {
 554                                 list_del(&neigh->list);
 555                                 if (neigh->ah)
 556                                         ipoib_put_ah(neigh->ah);
 557                                 ipoib_neigh_free(dev, neigh);
 558                                 goto err_drop;
 559                         }
 560                         if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE)
 561                                 __skb_queue_tail(&neigh->queue, skb);
 562                         else {
 563                                 ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
 564                                            skb_queue_len(&neigh->queue));
 565                                 goto err_drop;
 566                         }
 567                 } else
 568                         ipoib_send(dev, skb, path->ah, IPOIB_QPN(skb->dst->neighbour->ha));
 569         } else {
 570                 neigh->ah  = NULL;
 571
 572                 if (!path->query && path_rec_start(dev, path))
 573                         goto err_list;
 574
 575                 __skb_queue_tail(&neigh->queue, skb);
 576         }
 577
 578         spin_unlock(&priv->lock);
 579         return;
 580
 581 err_list:
 582         list_del(&neigh->list);
 583
 584 err_path:
 585         ipoib_neigh_free(dev, neigh);
 586 err_drop:
 587         ++dev->stats.tx_dropped;
 588         dev_kfree_skb_any(skb);
 589
 590         spin_unlock(&priv->lock);
 591 }
 592
 593 static void ipoib_path_lookup(struct sk_buff *skb, struct net_device *dev)
 594 {
 595         struct ipoib_dev_priv *priv = netdev_priv(skb->dev);
 596
 597         /* Look up path record for unicasts */
 598         if (skb->dst->neighbour->ha[4] != 0xff) {
 599                 neigh_add_path(skb, dev);
 600                 return;
 601         }
 602
 603         /* Add in the P_Key for multicasts */
 604         skb->dst->neighbour->ha[8] = (priv->pkey >> 8) & 0xff;
 605         skb->dst->neighbour->ha[9] = priv->pkey & 0xff;
 606         ipoib_mcast_send(dev, skb->dst->neighbour->ha + 4, skb);
 607 }
 608
 609 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
 610                              struct ipoib_pseudoheader *phdr)
 611 {
 612         struct ipoib_dev_priv *priv = netdev_priv(dev);
 613         struct ipoib_path *path;
 614
 615         /*
 616          * We can only be called from ipoib_start_xmit, so we're
 617          * inside tx_lock -- no need to save/restore flags.
 618          */
 619         spin_lock(&priv->lock);
 620
 621         path = __path_find(dev, phdr->hwaddr + 4);
 622         if (!path) {
 623                 path = path_rec_create(dev, phdr->hwaddr + 4);
 624                 if (path) {
 625                         /* put pseudoheader back on for next time */
 626                         skb_push(skb, sizeof *phdr);
 627                         __skb_queue_tail(&path->queue, skb);
 628
 629                         if (path_rec_start(dev, path)) {
 630                                 spin_unlock(&priv->lock);
 631                                 path_free(dev, path);
 632                                 return;
 633                         } else
 634                                 __path_add(dev, path);
 635                 } else {
 636                         ++dev->stats.tx_dropped;
 637                         dev_kfree_skb_any(skb);
 638                 }
 639
 640                 spin_unlock(&priv->lock);
 641                 return;
 642         }
 643
 644         if (path->ah) {
 645                 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
 646                           be16_to_cpu(path->pathrec.dlid));
 647
 648                 ipoib_send(dev, skb, path->ah, IPOIB_QPN(phdr->hwaddr));
 649         } else if ((path->query || !path_rec_start(dev, path)) &&
 650                    skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 651                 /* put pseudoheader back on for next time */
 652                 skb_push(skb, sizeof *phdr);
 653                 __skb_queue_tail(&path->queue, skb);
 654         } else {
 655                 ++dev->stats.tx_dropped;
 656                 dev_kfree_skb_any(skb);
 657         }
 658
 659         spin_unlock(&priv->lock);
 660 }
 661
 662 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
 663 {
 664         struct ipoib_dev_priv *priv = netdev_priv(dev);
 665         struct ipoib_neigh *neigh;
 666         unsigned long flags;
 667
 668         if (unlikely(!spin_trylock_irqsave(&priv->tx_lock, flags)))
 669                 return NETDEV_TX_LOCKED;
 670
 671         /*
 672          * Check if our queue is stopped.  Since we have the LLTX bit
 673          * set, we can't rely on netif_stop_queue() preventing our
 674          * xmit function from being called with a full queue.
 675          */
 676         if (unlikely(netif_queue_stopped(dev))) {
 677                 spin_unlock_irqrestore(&priv->tx_lock, flags);
 678                 return NETDEV_TX_BUSY;
 679         }
 680
 681         if (likely(skb->dst && skb->dst->neighbour)) {
 682                 if (unlikely(!*to_ipoib_neigh(skb->dst->neighbour))) {
 683                         ipoib_path_lookup(skb, dev);
 684                         goto out;
 685                 }
 686
 687                 neigh = *to_ipoib_neigh(skb->dst->neighbour);
 688
 689                 if (ipoib_cm_get(neigh)) {
 690                         if (ipoib_cm_up(neigh)) {
 691                                 ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
 692                                 goto out;
 693                         }
 694                 } else if (neigh->ah) {
 695                         if (unlikely(memcmp(&neigh->dgid.raw,
 696                                             skb->dst->neighbour->ha + 4,
 697                                             sizeof(union ib_gid)))) {
 698                                 spin_lock(&priv->lock);
 699                                 /*
 700                                  * It's safe to call ipoib_put_ah() inside
 701                                  * priv->lock here, because we know that
 702                                  * path->ah will always hold one more reference,
 703                                  * so ipoib_put_ah() will never do more than
 704                                  * decrement the ref count.
 705                                  */
 706                                 ipoib_put_ah(neigh->ah);
 707                                 list_del(&neigh->list);
 708                                 ipoib_neigh_free(dev, neigh);
 709                                 spin_unlock(&priv->lock);
 710                                 ipoib_path_lookup(skb, dev);
 711                                 goto out;
 712                         }
 713
 714                         ipoib_send(dev, skb, neigh->ah, IPOIB_QPN(skb->dst->neighbour->ha));
 715                         goto out;
 716                 }
 717
 718                 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 719                         spin_lock(&priv->lock);
 720                         __skb_queue_tail(&neigh->queue, skb);
 721                         spin_unlock(&priv->lock);
 722                 } else {
 723                         ++dev->stats.tx_dropped;
 724                         dev_kfree_skb_any(skb);
 725                 }
 726         } else {
 727                 struct ipoib_pseudoheader *phdr =
 728                         (struct ipoib_pseudoheader *) skb->data;
 729                 skb_pull(skb, sizeof *phdr);
 730
 731                 if (phdr->hwaddr[4] == 0xff) {
 732                         /* Add in the P_Key for multicast*/
 733                         phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
 734                         phdr->hwaddr[9] = priv->pkey & 0xff;
 735
 736                         ipoib_mcast_send(dev, phdr->hwaddr + 4, skb);
 737                 } else {
 738                         /* unicast GID -- should be ARP or RARP reply */
 739
 740                         if ((be16_to_cpup((__be16 *) skb->data) != ETH_P_ARP) &&
 741                             (be16_to_cpup((__be16 *) skb->data) != ETH_P_RARP)) {
 742                                 ipoib_warn(priv, "Unicast, no %s: type %04x, QPN %06x "
 743                                            IPOIB_GID_FMT "\n",
 744                                            skb->dst ? "neigh" : "dst",
 745                                            be16_to_cpup((__be16 *) skb->data),
 746                                            IPOIB_QPN(phdr->hwaddr),
 747                                            IPOIB_GID_RAW_ARG(phdr->hwaddr + 4));
 748                                 dev_kfree_skb_any(skb);
 749                                 ++dev->stats.tx_dropped;
 750                                 goto out;
 751                         }
 752
 753                         unicast_arp_send(skb, dev, phdr);
 754                 }
 755         }
 756
 757 out:
 758         spin_unlock_irqrestore(&priv->tx_lock, flags);
 759
 760         return NETDEV_TX_OK;
 761 }
 762
 763 static void ipoib_timeout(struct net_device *dev)
 764 {
 765         struct ipoib_dev_priv *priv = netdev_priv(dev);
 766
 767         ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
 768                    jiffies_to_msecs(jiffies - dev->trans_start));
 769         ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
 770                    netif_queue_stopped(dev),
 771                    priv->tx_head, priv->tx_tail);
 772         /* XXX reset QP, etc. */
 773 }
 774
 775 static int ipoib_hard_header(struct sk_buff *skb,
 776                              struct net_device *dev,
 777                              unsigned short type,
 778                              const void *daddr, const void *saddr, unsigned len)
 779 {
 780         struct ipoib_header *header;
 781
 782         header = (struct ipoib_header *) skb_push(skb, sizeof *header);
 783
 784         header->proto = htons(type);
 785         header->reserved = 0;
 786
 787         /*
 788          * If we don't have a neighbour structure, stuff the
 789          * destination address onto the front of the skb so we can
 790          * figure out where to send the packet later.
 791          */
 792         if ((!skb->dst || !skb->dst->neighbour) && daddr) {
 793                 struct ipoib_pseudoheader *phdr =
 794                         (struct ipoib_pseudoheader *) skb_push(skb, sizeof *phdr);
 795                 memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
 796         }
 797
 798         return 0;
 799 }
 800
 801 static void ipoib_set_mcast_list(struct net_device *dev)
 802 {
 803         struct ipoib_dev_priv *priv = netdev_priv(dev);
 804
 805         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
 806                 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
 807                 return;
 808         }
 809
 810         queue_work(ipoib_workqueue, &priv->restart_task);
 811 }
 812
 813 static void ipoib_neigh_cleanup(struct neighbour *n)
 814 {
 815         struct ipoib_neigh *neigh;
 816         struct ipoib_dev_priv *priv = netdev_priv(n->dev);
 817         unsigned long flags;
 818         struct ipoib_ah *ah = NULL;
 819
 820         neigh = *to_ipoib_neigh(n);
 821         if (neigh) {
 822                 priv = netdev_priv(neigh->dev);
 823                 ipoib_dbg(priv, "neigh_destructor for bonding device: %s\n",
 824                           n->dev->name);
 825         } else
 826                 return;
 827         ipoib_dbg(priv,
 828                   "neigh_cleanup for %06x " IPOIB_GID_FMT "\n",
 829                   IPOIB_QPN(n->ha),
 830                   IPOIB_GID_RAW_ARG(n->ha + 4));
 831
 832         spin_lock_irqsave(&priv->lock, flags);
 833
 834         if (neigh->ah)
 835                 ah = neigh->ah;
 836         list_del(&neigh->list);
 837         ipoib_neigh_free(n->dev, neigh);
 838
 839         spin_unlock_irqrestore(&priv->lock, flags);
 840
 841         if (ah)
 842                 ipoib_put_ah(ah);
 843 }
 844
 845 struct ipoib_neigh *ipoib_neigh_alloc(struct neighbour *neighbour,
 846                                       struct net_device *dev)
 847 {
 848         struct ipoib_neigh *neigh;
 849
 850         neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
 851         if (!neigh)
 852                 return NULL;
 853
 854         neigh->neighbour = neighbour;
 855         neigh->dev = dev;
 856         *to_ipoib_neigh(neighbour) = neigh;
 857         skb_queue_head_init(&neigh->queue);
 858         ipoib_cm_set(neigh, NULL);
 859
 860         return neigh;
 861 }
 862
 863 void ipoib_neigh_free(struct net_device *dev, struct ipoib_neigh *neigh)
 864 {
 865         struct sk_buff *skb;
 866         *to_ipoib_neigh(neigh->neighbour) = NULL;
 867         while ((skb = __skb_dequeue(&neigh->queue))) {
 868                 ++dev->stats.tx_dropped;
 869                 dev_kfree_skb_any(skb);
 870         }
 871         if (ipoib_cm_get(neigh))
 872                 ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
 873         kfree(neigh);
 874 }
 875
 876 static int ipoib_neigh_setup_dev(struct net_device *dev, struct neigh_parms *parms)
 877 {
 878         parms->neigh_cleanup = ipoib_neigh_cleanup;
 879
 880         return 0;
 881 }
 882
 883 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
 884 {
 885         struct ipoib_dev_priv *priv = netdev_priv(dev);
 886
 887         /* Allocate RX/TX "rings" to hold queued skbs */
 888         priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
 889                                 GFP_KERNEL);
 890         if (!priv->rx_ring) {
 891                 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
 892                        ca->name, ipoib_recvq_size);
 893                 goto out;
 894         }
 895
 896         priv->tx_ring = kzalloc(ipoib_sendq_size * sizeof *priv->tx_ring,
 897                                 GFP_KERNEL);
 898         if (!priv->tx_ring) {
 899                 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
 900                        ca->name, ipoib_sendq_size);
 901                 goto out_rx_ring_cleanup;
 902         }
 903
 904         /* priv->tx_head & tx_tail are already 0 */
 905
 906         if (ipoib_ib_dev_init(dev, ca, port))
 907                 goto out_tx_ring_cleanup;
 908
 909         return 0;
 910
 911 out_tx_ring_cleanup:
 912         kfree(priv->tx_ring);
 913
 914 out_rx_ring_cleanup:
 915         kfree(priv->rx_ring);
 916
 917 out:
 918         return -ENOMEM;
 919 }
 920
 921 void ipoib_dev_cleanup(struct net_device *dev)
 922 {
 923         struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
 924
 925         ipoib_delete_debug_files(dev);
 926
 927         /* Delete any child interfaces first */
 928         list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
 929                 unregister_netdev(cpriv->dev);
 930                 ipoib_dev_cleanup(cpriv->dev);
 931                 free_netdev(cpriv->dev);
 932         }
 933
 934         ipoib_ib_dev_cleanup(dev);
 935
 936         kfree(priv->rx_ring);
 937         kfree(priv->tx_ring);
 938
 939         priv->rx_ring = NULL;
 940         priv->tx_ring = NULL;
 941 }
 942
 943 static const struct header_ops ipoib_header_ops = {
 944         .create = ipoib_hard_header,
 945 };
 946
 947 static void ipoib_setup(struct net_device *dev)
 948 {
 949         struct ipoib_dev_priv *priv = netdev_priv(dev);
 950
 951         dev->open                = ipoib_open;
 952         dev->stop                = ipoib_stop;
 953         dev->change_mtu          = ipoib_change_mtu;
 954         dev->hard_start_xmit     = ipoib_start_xmit;
 955         dev->tx_timeout          = ipoib_timeout;
 956         dev->header_ops          = &ipoib_header_ops;
 957         dev->set_multicast_list  = ipoib_set_mcast_list;
 958         dev->neigh_setup         = ipoib_neigh_setup_dev;
 959
 960         netif_napi_add(dev, &priv->napi, ipoib_poll, 100);
 961
 962         dev->watchdog_timeo      = HZ;
 963
 964         dev->flags              |= IFF_BROADCAST | IFF_MULTICAST;
 965
 966         /*
 967          * We add in INFINIBAND_ALEN to allow for the destination
 968          * address "pseudoheader" for skbs without neighbour struct.
 969          */
 970         dev->hard_header_len     = IPOIB_ENCAP_LEN + INFINIBAND_ALEN;
 971         dev->addr_len            = INFINIBAND_ALEN;
 972         dev->type                = ARPHRD_INFINIBAND;
 973         dev->tx_queue_len        = ipoib_sendq_size * 2;
 974         dev->features            = NETIF_F_VLAN_CHALLENGED | NETIF_F_LLTX;
 975
 976         /* MTU will be reset when mcast join happens */
 977         dev->mtu                 = IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN;
 978         priv->mcast_mtu          = priv->admin_mtu = dev->mtu;
 979
 980         memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
 981
 982         netif_carrier_off(dev);
 983
 984         priv->dev = dev;
 985
 986         spin_lock_init(&priv->lock);
 987         spin_lock_init(&priv->tx_lock);
 988
 989         mutex_init(&priv->mcast_mutex);
 990         mutex_init(&priv->vlan_mutex);
 991
 992         INIT_LIST_HEAD(&priv->path_list);
 993         INIT_LIST_HEAD(&priv->child_intfs);
 994         INIT_LIST_HEAD(&priv->dead_ahs);
 995         INIT_LIST_HEAD(&priv->multicast_list);
 996
 997         INIT_DELAYED_WORK(&priv->pkey_poll_task, ipoib_pkey_poll);
 998         INIT_WORK(&priv->pkey_event_task, ipoib_pkey_event);
 999         INIT_DELAYED_WORK(&priv->mcast_task,   ipoib_mcast_join_task);
1000         INIT_WORK(&priv->flush_task,   ipoib_ib_dev_flush);
1001         INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
1002         INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
1003 }
1004
1005 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
1006 {
1007         struct net_device *dev;
1008
1009         dev = alloc_netdev((int) sizeof (struct ipoib_dev_priv), name,
1010                            ipoib_setup);
1011         if (!dev)
1012                 return NULL;
1013
1014         return netdev_priv(dev);
1015 }
1016
1017 static ssize_t show_pkey(struct device *dev,
1018                          struct device_attribute *attr, char *buf)
1019 {
1020         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1021
1022         return sprintf(buf, "0x%04x\n", priv->pkey);
1023 }
1024 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1025
1026 static ssize_t show_umcast(struct device *dev,
1027                            struct device_attribute *attr, char *buf)
1028 {
1029         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1030
1031         return sprintf(buf, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
1032 }
1033
1034 static ssize_t set_umcast(struct device *dev,
1035                           struct device_attribute *attr,
1036                           const char *buf, size_t count)
1037 {
1038         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1039         unsigned long umcast_val = simple_strtoul(buf, NULL, 0);
1040
1041         if (umcast_val > 0) {
1042                 set_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1043                 ipoib_warn(priv, "ignoring multicast groups joined directly "
1044                                 "by userspace\n");
1045         } else
1046                 clear_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1047
1048         return count;
1049 }
1050 static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast);
1051
1052 int ipoib_add_umcast_attr(struct net_device *dev)
1053 {
1054         return device_create_file(&dev->dev, &dev_attr_umcast);
1055 }
1056
1057 static ssize_t create_child(struct device *dev,
1058                             struct device_attribute *attr,
1059                             const char *buf, size_t count)
1060 {
1061         int pkey;
1062         int ret;
1063
1064         if (sscanf(buf, "%i", &pkey) != 1)
1065                 return -EINVAL;
1066
1067         if (pkey < 0 || pkey > 0xffff)
1068                 return -EINVAL;
1069
1070         /*
1071          * Set the full membership bit, so that we join the right
1072          * broadcast group, etc.
1073          */
1074         pkey |= 0x8000;
1075
1076         ret = ipoib_vlan_add(to_net_dev(dev), pkey);
1077
1078         return ret ? ret : count;
1079 }
1080 static DEVICE_ATTR(create_child, S_IWUGO, NULL, create_child);
1081
1082 static ssize_t delete_child(struct device *dev,
1083                             struct device_attribute *attr,
1084                             const char *buf, size_t count)
1085 {
1086         int pkey;
1087         int ret;
1088
1089         if (sscanf(buf, "%i", &pkey) != 1)
1090                 return -EINVAL;
1091
1092         if (pkey < 0 || pkey > 0xffff)
1093                 return -EINVAL;
1094
1095         ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
1096
1097         return ret ? ret : count;
1098
1099 }
1100 static DEVICE_ATTR(delete_child, S_IWUGO, NULL, delete_child);
1101
1102 int ipoib_add_pkey_attr(struct net_device *dev)
1103 {
1104         return device_create_file(&dev->dev, &dev_attr_pkey);
1105 }
1106
1107 static struct net_device *ipoib_add_port(const char *format,
1108                                          struct ib_device *hca, u8 port)
1109 {
1110         struct ipoib_dev_priv *priv;
1111         int result = -ENOMEM;
1112
1113         priv = ipoib_intf_alloc(format);
1114         if (!priv)
1115                 goto alloc_mem_failed;
1116
1117         SET_NETDEV_DEV(priv->dev, hca->dma_device);
1118
1119         result = ib_query_pkey(hca, port, 0, &priv->pkey);
1120         if (result) {
1121                 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
1122                        hca->name, port, result);
1123                 goto device_init_failed;
1124         }
1125
1126         /*
1127          * Set the full membership bit, so that we join the right
1128          * broadcast group, etc.
1129          */
1130         priv->pkey |= 0x8000;
1131
1132         priv->dev->broadcast[8] = priv->pkey >> 8;
1133         priv->dev->broadcast[9] = priv->pkey & 0xff;
1134
1135         result = ib_query_gid(hca, port, 0, &priv->local_gid);
1136         if (result) {
1137                 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1138                        hca->name, port, result);
1139                 goto device_init_failed;
1140         } else
1141                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1142
1143
1144         result = ipoib_dev_init(priv->dev, hca, port);
1145         if (result < 0) {
1146                 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1147                        hca->name, port, result);
1148                 goto device_init_failed;
1149         }
1150
1151         INIT_IB_EVENT_HANDLER(&priv->event_handler,
1152                               priv->ca, ipoib_event);
1153         result = ib_register_event_handler(&priv->event_handler);
1154         if (result < 0) {
1155                 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1156                        "port %d (ret = %d)\n",
1157                        hca->name, port, result);
1158                 goto event_failed;
1159         }
1160
1161         result = register_netdev(priv->dev);
1162         if (result) {
1163                 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1164                        hca->name, port, result);
1165                 goto register_failed;
1166         }
1167
1168         ipoib_create_debug_files(priv->dev);
1169
1170         if (ipoib_cm_add_mode_attr(priv->dev))
1171                 goto sysfs_failed;
1172         if (ipoib_add_pkey_attr(priv->dev))
1173                 goto sysfs_failed;
1174         if (ipoib_add_umcast_attr(priv->dev))
1175                 goto sysfs_failed;
1176         if (device_create_file(&priv->dev->dev, &dev_attr_create_child))
1177                 goto sysfs_failed;
1178         if (device_create_file(&priv->dev->dev, &dev_attr_delete_child))
1179                 goto sysfs_failed;
1180
1181         return priv->dev;
1182
1183 sysfs_failed:
1184         ipoib_delete_debug_files(priv->dev);
1185         unregister_netdev(priv->dev);
1186
1187 register_failed:
1188         ib_unregister_event_handler(&priv->event_handler);
1189         flush_scheduled_work();
1190
1191 event_failed:
1192         ipoib_dev_cleanup(priv->dev);
1193
1194 device_init_failed:
1195         free_netdev(priv->dev);
1196
1197 alloc_mem_failed:
1198         return ERR_PTR(result);
1199 }
1200
1201 static void ipoib_add_one(struct ib_device *device)
1202 {
1203         struct list_head *dev_list;
1204         struct net_device *dev;
1205         struct ipoib_dev_priv *priv;
1206         int s, e, p;
1207
1208         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1209                 return;
1210
1211         dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1212         if (!dev_list)
1213                 return;
1214
1215         INIT_LIST_HEAD(dev_list);
1216
1217         if (device->node_type == RDMA_NODE_IB_SWITCH) {
1218                 s = 0;
1219                 e = 0;
1220         } else {
1221                 s = 1;
1222                 e = device->phys_port_cnt;
1223         }
1224
1225         for (p = s; p <= e; ++p) {
1226                 dev = ipoib_add_port("ib%d", device, p);
1227                 if (!IS_ERR(dev)) {
1228                         priv = netdev_priv(dev);
1229                         list_add_tail(&priv->list, dev_list);
1230                 }
1231         }
1232
1233         ib_set_client_data(device, &ipoib_client, dev_list);
1234 }
1235
1236 static void ipoib_remove_one(struct ib_device *device)
1237 {
1238         struct ipoib_dev_priv *priv, *tmp;
1239         struct list_head *dev_list;
1240
1241         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1242                 return;
1243
1244         dev_list = ib_get_client_data(device, &ipoib_client);
1245
1246         list_for_each_entry_safe(priv, tmp, dev_list, list) {
1247                 ib_unregister_event_handler(&priv->event_handler);
1248                 flush_scheduled_work();
1249
1250                 unregister_netdev(priv->dev);
1251                 ipoib_dev_cleanup(priv->dev);
1252                 free_netdev(priv->dev);
1253         }
1254
1255         kfree(dev_list);
1256 }
1257
1258 static int __init ipoib_init_module(void)
1259 {
1260         int ret;
1261
1262         ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1263         ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1264         ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1265
1266         ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1267         ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1268         ipoib_sendq_size = max(ipoib_sendq_size, IPOIB_MIN_QUEUE_SIZE);
1269
1270         ret = ipoib_register_debugfs();
1271         if (ret)
1272                 return ret;
1273
1274         /*
1275          * We create our own workqueue mainly because we want to be
1276          * able to flush it when devices are being removed.  We can't
1277          * use schedule_work()/flush_scheduled_work() because both
1278          * unregister_netdev() and linkwatch_event take the rtnl lock,
1279          * so flush_scheduled_work() can deadlock during device
1280          * removal.
1281          */
1282         ipoib_workqueue = create_singlethread_workqueue("ipoib");
1283         if (!ipoib_workqueue) {
1284                 ret = -ENOMEM;
1285                 goto err_fs;
1286         }
1287
1288         ib_sa_register_client(&ipoib_sa_client);
1289
1290         ret = ib_register_client(&ipoib_client);
1291         if (ret)
1292                 goto err_sa;
1293
1294         return 0;
1295
1296 err_sa:
1297         ib_sa_unregister_client(&ipoib_sa_client);
1298         destroy_workqueue(ipoib_workqueue);
1299
1300 err_fs:
1301         ipoib_unregister_debugfs();
1302
1303         return ret;
1304 }
1305
1306 static void __exit ipoib_cleanup_module(void)
1307 {
1308         ib_unregister_client(&ipoib_client);
1309         ib_sa_unregister_client(&ipoib_sa_client);
1310         ipoib_unregister_debugfs();
1311         destroy_workqueue(ipoib_workqueue);
1312 }
1313
1314 module_init(ipoib_init_module);
1315 module_exit(ipoib_cleanup_module);