git.oblomov.eu 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
  35 #include "ipoib.h"
  36
  37 #include <linux/module.h>
  38
  39 #include <linux/init.h>
  40 #include <linux/slab.h>
  41 #include <linux/kernel.h>
  42 #include <linux/vmalloc.h>
  43
  44 #include <linux/if_arp.h>       /* For ARPHRD_xxx */
  45
  46 #include <linux/ip.h>
  47 #include <linux/in.h>
  48
  49 #include <net/dst.h>
  50
  51 MODULE_AUTHOR("Roland Dreier");
  52 MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
  53 MODULE_LICENSE("Dual BSD/GPL");
  54
  55 int ipoib_sendq_size __read_mostly = IPOIB_TX_RING_SIZE;
  56 int ipoib_recvq_size __read_mostly = IPOIB_RX_RING_SIZE;
  57
  58 module_param_named(send_queue_size, ipoib_sendq_size, int, 0444);
  59 MODULE_PARM_DESC(send_queue_size, "Number of descriptors in send queue");
  60 module_param_named(recv_queue_size, ipoib_recvq_size, int, 0444);
  61 MODULE_PARM_DESC(recv_queue_size, "Number of descriptors in receive queue");
  62
  63 static int lro;
  64 module_param(lro, bool, 0444);
  65 MODULE_PARM_DESC(lro,  "Enable LRO (Large Receive Offload)");
  66
  67 static int lro_max_aggr = IPOIB_LRO_MAX_AGGR;
  68 module_param(lro_max_aggr, int, 0644);
  69 MODULE_PARM_DESC(lro_max_aggr, "LRO: Max packets to be aggregated "
  70                 "(default = 64)");
  71
  72 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
  73 int ipoib_debug_level;
  74
  75 module_param_named(debug_level, ipoib_debug_level, int, 0644);
  76 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
  77 #endif
  78
  79 struct ipoib_path_iter {
  80         struct net_device *dev;
  81         struct ipoib_path  path;
  82 };
  83
  84 static const u8 ipv4_bcast_addr[] = {
  85         0x00, 0xff, 0xff, 0xff,
  86         0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
  87         0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff
  88 };
  89
  90 struct workqueue_struct *ipoib_workqueue;
  91
  92 struct ib_sa_client ipoib_sa_client;
  93
  94 static void ipoib_add_one(struct ib_device *device);
  95 static void ipoib_remove_one(struct ib_device *device);
  96
  97 static struct ib_client ipoib_client = {
  98         .name   = "ipoib",
  99         .add    = ipoib_add_one,
 100         .remove = ipoib_remove_one
 101 };
 102
 103 int ipoib_open(struct net_device *dev)
 104 {
 105         struct ipoib_dev_priv *priv = netdev_priv(dev);
 106
 107         ipoib_dbg(priv, "bringing up interface\n");
 108
 109         if (!test_and_set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 110                 napi_enable(&priv->napi);
 111
 112         if (ipoib_pkey_dev_delay_open(dev))
 113                 return 0;
 114
 115         if (ipoib_ib_dev_open(dev))
 116                 goto err_disable;
 117
 118         if (ipoib_ib_dev_up(dev))
 119                 goto err_stop;
 120
 121         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 122                 struct ipoib_dev_priv *cpriv;
 123
 124                 /* Bring up any child interfaces too */
 125                 mutex_lock(&priv->vlan_mutex);
 126                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 127                         int flags;
 128
 129                         flags = cpriv->dev->flags;
 130                         if (flags & IFF_UP)
 131                                 continue;
 132
 133                         dev_change_flags(cpriv->dev, flags | IFF_UP);
 134                 }
 135                 mutex_unlock(&priv->vlan_mutex);
 136         }
 137
 138         netif_start_queue(dev);
 139
 140         return 0;
 141
 142 err_stop:
 143         ipoib_ib_dev_stop(dev, 1);
 144
 145 err_disable:
 146         napi_disable(&priv->napi);
 147         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 148
 149         return -EINVAL;
 150 }
 151
 152 static int ipoib_stop(struct net_device *dev)
 153 {
 154         struct ipoib_dev_priv *priv = netdev_priv(dev);
 155
 156         ipoib_dbg(priv, "stopping interface\n");
 157
 158         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 159         napi_disable(&priv->napi);
 160
 161         netif_stop_queue(dev);
 162
 163         ipoib_ib_dev_down(dev, 0);
 164         ipoib_ib_dev_stop(dev, 0);
 165
 166         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 167                 struct ipoib_dev_priv *cpriv;
 168
 169                 /* Bring down any child interfaces too */
 170                 mutex_lock(&priv->vlan_mutex);
 171                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 172                         int flags;
 173
 174                         flags = cpriv->dev->flags;
 175                         if (!(flags & IFF_UP))
 176                                 continue;
 177
 178                         dev_change_flags(cpriv->dev, flags & ~IFF_UP);
 179                 }
 180                 mutex_unlock(&priv->vlan_mutex);
 181         }
 182
 183         return 0;
 184 }
 185
 186 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
 187 {
 188         struct ipoib_dev_priv *priv = netdev_priv(dev);
 189
 190         /* dev->mtu > 2K ==> connected mode */
 191         if (ipoib_cm_admin_enabled(dev)) {
 192                 if (new_mtu > ipoib_cm_max_mtu(dev))
 193                         return -EINVAL;
 194
 195                 if (new_mtu > priv->mcast_mtu)
 196                         ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
 197                                    priv->mcast_mtu);
 198
 199                 dev->mtu = new_mtu;
 200                 return 0;
 201         }
 202
 203         if (new_mtu > IPOIB_UD_MTU(priv->max_ib_mtu))
 204                 return -EINVAL;
 205
 206         priv->admin_mtu = new_mtu;
 207
 208         dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
 209
 210         return 0;
 211 }
 212
 213 static struct ipoib_path *__path_find(struct net_device *dev, void *gid)
 214 {
 215         struct ipoib_dev_priv *priv = netdev_priv(dev);
 216         struct rb_node *n = priv->path_tree.rb_node;
 217         struct ipoib_path *path;
 218         int ret;
 219
 220         while (n) {
 221                 path = rb_entry(n, struct ipoib_path, rb_node);
 222
 223                 ret = memcmp(gid, path->pathrec.dgid.raw,
 224                              sizeof (union ib_gid));
 225
 226                 if (ret < 0)
 227                         n = n->rb_left;
 228                 else if (ret > 0)
 229                         n = n->rb_right;
 230                 else
 231                         return path;
 232         }
 233
 234         return NULL;
 235 }
 236
 237 static int __path_add(struct net_device *dev, struct ipoib_path *path)
 238 {
 239         struct ipoib_dev_priv *priv = netdev_priv(dev);
 240         struct rb_node **n = &priv->path_tree.rb_node;
 241         struct rb_node *pn = NULL;
 242         struct ipoib_path *tpath;
 243         int ret;
 244
 245         while (*n) {
 246                 pn = *n;
 247                 tpath = rb_entry(pn, struct ipoib_path, rb_node);
 248
 249                 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
 250                              sizeof (union ib_gid));
 251                 if (ret < 0)
 252                         n = &pn->rb_left;
 253                 else if (ret > 0)
 254                         n = &pn->rb_right;
 255                 else
 256                         return -EEXIST;
 257         }
 258
 259         rb_link_node(&path->rb_node, pn, n);
 260         rb_insert_color(&path->rb_node, &priv->path_tree);
 261
 262         list_add_tail(&path->list, &priv->path_list);
 263
 264         return 0;
 265 }
 266
 267 static void path_free(struct net_device *dev, struct ipoib_path *path)
 268 {
 269         struct ipoib_dev_priv *priv = netdev_priv(dev);
 270         struct ipoib_neigh *neigh, *tn;
 271         struct sk_buff *skb;
 272         unsigned long flags;
 273
 274         while ((skb = __skb_dequeue(&path->queue)))
 275                 dev_kfree_skb_irq(skb);
 276
 277         spin_lock_irqsave(&priv->lock, flags);
 278
 279         list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
 280                 /*
 281                  * It's safe to call ipoib_put_ah() inside priv->lock
 282                  * here, because we know that path->ah will always
 283                  * hold one more reference, so ipoib_put_ah() will
 284                  * never do more than decrement the ref count.
 285                  */
 286                 if (neigh->ah)
 287                         ipoib_put_ah(neigh->ah);
 288
 289                 ipoib_neigh_free(dev, neigh);
 290         }
 291
 292         spin_unlock_irqrestore(&priv->lock, flags);
 293
 294         if (path->ah)
 295                 ipoib_put_ah(path->ah);
 296
 297         kfree(path);
 298 }
 299
 300 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 301
 302 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
 303 {
 304         struct ipoib_path_iter *iter;
 305
 306         iter = kmalloc(sizeof *iter, GFP_KERNEL);
 307         if (!iter)
 308                 return NULL;
 309
 310         iter->dev = dev;
 311         memset(iter->path.pathrec.dgid.raw, 0, 16);
 312
 313         if (ipoib_path_iter_next(iter)) {
 314                 kfree(iter);
 315                 return NULL;
 316         }
 317
 318         return iter;
 319 }
 320
 321 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
 322 {
 323         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
 324         struct rb_node *n;
 325         struct ipoib_path *path;
 326         int ret = 1;
 327
 328         spin_lock_irq(&priv->lock);
 329
 330         n = rb_first(&priv->path_tree);
 331
 332         while (n) {
 333                 path = rb_entry(n, struct ipoib_path, rb_node);
 334
 335                 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
 336                            sizeof (union ib_gid)) < 0) {
 337                         iter->path = *path;
 338                         ret = 0;
 339                         break;
 340                 }
 341
 342                 n = rb_next(n);
 343         }
 344
 345         spin_unlock_irq(&priv->lock);
 346
 347         return ret;
 348 }
 349
 350 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
 351                           struct ipoib_path *path)
 352 {
 353         *path = iter->path;
 354 }
 355
 356 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
 357
 358 void ipoib_mark_paths_invalid(struct net_device *dev)
 359 {
 360         struct ipoib_dev_priv *priv = netdev_priv(dev);
 361         struct ipoib_path *path, *tp;
 362
 363         spin_lock_irq(&priv->lock);
 364
 365         list_for_each_entry_safe(path, tp, &priv->path_list, list) {
 366                 ipoib_dbg(priv, "mark path LID 0x%04x GID %pI6 invalid\n",
 367                         be16_to_cpu(path->pathrec.dlid),
 368                         path->pathrec.dgid.raw);
 369                 path->valid =  0;
 370         }
 371
 372         spin_unlock_irq(&priv->lock);
 373 }
 374
 375 void ipoib_flush_paths(struct net_device *dev)
 376 {
 377         struct ipoib_dev_priv *priv = netdev_priv(dev);
 378         struct ipoib_path *path, *tp;
 379         LIST_HEAD(remove_list);
 380         unsigned long flags;
 381
 382         netif_tx_lock_bh(dev);
 383         spin_lock_irqsave(&priv->lock, flags);
 384
 385         list_splice_init(&priv->path_list, &remove_list);
 386
 387         list_for_each_entry(path, &remove_list, list)
 388                 rb_erase(&path->rb_node, &priv->path_tree);
 389
 390         list_for_each_entry_safe(path, tp, &remove_list, list) {
 391                 if (path->query)
 392                         ib_sa_cancel_query(path->query_id, path->query);
 393                 spin_unlock_irqrestore(&priv->lock, flags);
 394                 netif_tx_unlock_bh(dev);
 395                 wait_for_completion(&path->done);
 396                 path_free(dev, path);
 397                 netif_tx_lock_bh(dev);
 398                 spin_lock_irqsave(&priv->lock, flags);
 399         }
 400
 401         spin_unlock_irqrestore(&priv->lock, flags);
 402         netif_tx_unlock_bh(dev);
 403 }
 404
 405 static void path_rec_completion(int status,
 406                                 struct ib_sa_path_rec *pathrec,
 407                                 void *path_ptr)
 408 {
 409         struct ipoib_path *path = path_ptr;
 410         struct net_device *dev = path->dev;
 411         struct ipoib_dev_priv *priv = netdev_priv(dev);
 412         struct ipoib_ah *ah = NULL;
 413         struct ipoib_ah *old_ah = NULL;
 414         struct ipoib_neigh *neigh, *tn;
 415         struct sk_buff_head skqueue;
 416         struct sk_buff *skb;
 417         unsigned long flags;
 418
 419         if (!status)
 420                 ipoib_dbg(priv, "PathRec LID 0x%04x for GID %pI6\n",
 421                           be16_to_cpu(pathrec->dlid), pathrec->dgid.raw);
 422         else
 423                 ipoib_dbg(priv, "PathRec status %d for GID %pI6\n",
 424                           status, path->pathrec.dgid.raw);
 425
 426         skb_queue_head_init(&skqueue);
 427
 428         if (!status) {
 429                 struct ib_ah_attr av;
 430
 431                 if (!ib_init_ah_from_path(priv->ca, priv->port, pathrec, &av))
 432                         ah = ipoib_create_ah(dev, priv->pd, &av);
 433         }
 434
 435         spin_lock_irqsave(&priv->lock, flags);
 436
 437         if (ah) {
 438                 path->pathrec = *pathrec;
 439
 440                 old_ah   = path->ah;
 441                 path->ah = ah;
 442
 443                 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
 444                           ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
 445
 446                 while ((skb = __skb_dequeue(&path->queue)))
 447                         __skb_queue_tail(&skqueue, skb);
 448
 449                 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
 450                         if (neigh->ah) {
 451                                 WARN_ON(neigh->ah != old_ah);
 452                                 /*
 453                                  * Dropping the ah reference inside
 454                                  * priv->lock is safe here, because we
 455                                  * will hold one more reference from
 456                                  * the original value of path->ah (ie
 457                                  * old_ah).
 458                                  */
 459                                 ipoib_put_ah(neigh->ah);
 460                         }
 461                         kref_get(&path->ah->ref);
 462                         neigh->ah = path->ah;
 463                         memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
 464                                sizeof(union ib_gid));
 465
 466                         if (ipoib_cm_enabled(dev, neigh->neighbour)) {
 467                                 if (!ipoib_cm_get(neigh))
 468                                         ipoib_cm_set(neigh, ipoib_cm_create_tx(dev,
 469                                                                                path,
 470                                                                                neigh));
 471                                 if (!ipoib_cm_get(neigh)) {
 472                                         list_del(&neigh->list);
 473                                         if (neigh->ah)
 474                                                 ipoib_put_ah(neigh->ah);
 475                                         ipoib_neigh_free(dev, neigh);
 476                                         continue;
 477                                 }
 478                         }
 479
 480                         while ((skb = __skb_dequeue(&neigh->queue)))
 481                                 __skb_queue_tail(&skqueue, skb);
 482                 }
 483                 path->valid = 1;
 484         }
 485
 486         path->query = NULL;
 487         complete(&path->done);
 488
 489         spin_unlock_irqrestore(&priv->lock, flags);
 490
 491         if (old_ah)
 492                 ipoib_put_ah(old_ah);
 493
 494         while ((skb = __skb_dequeue(&skqueue))) {
 495                 skb->dev = dev;
 496                 if (dev_queue_xmit(skb))
 497                         ipoib_warn(priv, "dev_queue_xmit failed "
 498                                    "to requeue packet\n");
 499         }
 500 }
 501
 502 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
 503 {
 504         struct ipoib_dev_priv *priv = netdev_priv(dev);
 505         struct ipoib_path *path;
 506
 507         if (!priv->broadcast)
 508                 return NULL;
 509
 510         path = kzalloc(sizeof *path, GFP_ATOMIC);
 511         if (!path)
 512                 return NULL;
 513
 514         path->dev = dev;
 515
 516         skb_queue_head_init(&path->queue);
 517
 518         INIT_LIST_HEAD(&path->neigh_list);
 519
 520         memcpy(path->pathrec.dgid.raw, gid, sizeof (union ib_gid));
 521         path->pathrec.sgid          = priv->local_gid;
 522         path->pathrec.pkey          = cpu_to_be16(priv->pkey);
 523         path->pathrec.numb_path     = 1;
 524         path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
 525
 526         return path;
 527 }
 528
 529 static int path_rec_start(struct net_device *dev,
 530                           struct ipoib_path *path)
 531 {
 532         struct ipoib_dev_priv *priv = netdev_priv(dev);
 533
 534         ipoib_dbg(priv, "Start path record lookup for %pI6\n",
 535                   path->pathrec.dgid.raw);
 536
 537         init_completion(&path->done);
 538
 539         path->query_id =
 540                 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
 541                                    &path->pathrec,
 542                                    IB_SA_PATH_REC_DGID          |
 543                                    IB_SA_PATH_REC_SGID          |
 544                                    IB_SA_PATH_REC_NUMB_PATH     |
 545                                    IB_SA_PATH_REC_TRAFFIC_CLASS |
 546                                    IB_SA_PATH_REC_PKEY,
 547                                    1000, GFP_ATOMIC,
 548                                    path_rec_completion,
 549                                    path, &path->query);
 550         if (path->query_id < 0) {
 551                 ipoib_warn(priv, "ib_sa_path_rec_get failed: %d\n", path->query_id);
 552                 path->query = NULL;
 553                 complete(&path->done);
 554                 return path->query_id;
 555         }
 556
 557         return 0;
 558 }
 559
 560 static void neigh_add_path(struct sk_buff *skb, struct net_device *dev)
 561 {
 562         struct ipoib_dev_priv *priv = netdev_priv(dev);
 563         struct ipoib_path *path;
 564         struct ipoib_neigh *neigh;
 565         unsigned long flags;
 566
 567         neigh = ipoib_neigh_alloc(skb->dst->neighbour, skb->dev);
 568         if (!neigh) {
 569                 ++dev->stats.tx_dropped;
 570                 dev_kfree_skb_any(skb);
 571                 return;
 572         }
 573
 574         spin_lock_irqsave(&priv->lock, flags);
 575
 576         path = __path_find(dev, skb->dst->neighbour->ha + 4);
 577         if (!path) {
 578                 path = path_rec_create(dev, skb->dst->neighbour->ha + 4);
 579                 if (!path)
 580                         goto err_path;
 581
 582                 __path_add(dev, path);
 583         }
 584
 585         list_add_tail(&neigh->list, &path->neigh_list);
 586
 587         if (path->ah) {
 588                 kref_get(&path->ah->ref);
 589                 neigh->ah = path->ah;
 590                 memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
 591                        sizeof(union ib_gid));
 592
 593                 if (ipoib_cm_enabled(dev, neigh->neighbour)) {
 594                         if (!ipoib_cm_get(neigh))
 595                                 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
 596                         if (!ipoib_cm_get(neigh)) {
 597                                 list_del(&neigh->list);
 598                                 if (neigh->ah)
 599                                         ipoib_put_ah(neigh->ah);
 600                                 ipoib_neigh_free(dev, neigh);
 601                                 goto err_drop;
 602                         }
 603                         if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE)
 604                                 __skb_queue_tail(&neigh->queue, skb);
 605                         else {
 606                                 ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
 607                                            skb_queue_len(&neigh->queue));
 608                                 goto err_drop;
 609                         }
 610                 } else
 611                         ipoib_send(dev, skb, path->ah, IPOIB_QPN(skb->dst->neighbour->ha));
 612         } else {
 613                 neigh->ah  = NULL;
 614
 615                 if (!path->query && path_rec_start(dev, path))
 616                         goto err_list;
 617
 618                 __skb_queue_tail(&neigh->queue, skb);
 619         }
 620
 621         spin_unlock_irqrestore(&priv->lock, flags);
 622         return;
 623
 624 err_list:
 625         list_del(&neigh->list);
 626
 627 err_path:
 628         ipoib_neigh_free(dev, neigh);
 629 err_drop:
 630         ++dev->stats.tx_dropped;
 631         dev_kfree_skb_any(skb);
 632
 633         spin_unlock_irqrestore(&priv->lock, flags);
 634 }
 635
 636 static void ipoib_path_lookup(struct sk_buff *skb, struct net_device *dev)
 637 {
 638         struct ipoib_dev_priv *priv = netdev_priv(skb->dev);
 639
 640         /* Look up path record for unicasts */
 641         if (skb->dst->neighbour->ha[4] != 0xff) {
 642                 neigh_add_path(skb, dev);
 643                 return;
 644         }
 645
 646         /* Add in the P_Key for multicasts */
 647         skb->dst->neighbour->ha[8] = (priv->pkey >> 8) & 0xff;
 648         skb->dst->neighbour->ha[9] = priv->pkey & 0xff;
 649         ipoib_mcast_send(dev, skb->dst->neighbour->ha + 4, skb);
 650 }
 651
 652 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
 653                              struct ipoib_pseudoheader *phdr)
 654 {
 655         struct ipoib_dev_priv *priv = netdev_priv(dev);
 656         struct ipoib_path *path;
 657         unsigned long flags;
 658
 659         spin_lock_irqsave(&priv->lock, flags);
 660
 661         path = __path_find(dev, phdr->hwaddr + 4);
 662         if (!path || !path->valid) {
 663                 if (!path)
 664                         path = path_rec_create(dev, phdr->hwaddr + 4);
 665                 if (path) {
 666                         /* put pseudoheader back on for next time */
 667                         skb_push(skb, sizeof *phdr);
 668                         __skb_queue_tail(&path->queue, skb);
 669
 670                         if (!path->query && path_rec_start(dev, path)) {
 671                                 spin_unlock_irqrestore(&priv->lock, flags);
 672                                 path_free(dev, path);
 673                                 return;
 674                         } else
 675                                 __path_add(dev, path);
 676                 } else {
 677                         ++dev->stats.tx_dropped;
 678                         dev_kfree_skb_any(skb);
 679                 }
 680
 681                 spin_unlock_irqrestore(&priv->lock, flags);
 682                 return;
 683         }
 684
 685         if (path->ah) {
 686                 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
 687                           be16_to_cpu(path->pathrec.dlid));
 688
 689                 ipoib_send(dev, skb, path->ah, IPOIB_QPN(phdr->hwaddr));
 690         } else if ((path->query || !path_rec_start(dev, path)) &&
 691                    skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 692                 /* put pseudoheader back on for next time */
 693                 skb_push(skb, sizeof *phdr);
 694                 __skb_queue_tail(&path->queue, skb);
 695         } else {
 696                 ++dev->stats.tx_dropped;
 697                 dev_kfree_skb_any(skb);
 698         }
 699
 700         spin_unlock_irqrestore(&priv->lock, flags);
 701 }
 702
 703 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
 704 {
 705         struct ipoib_dev_priv *priv = netdev_priv(dev);
 706         struct ipoib_neigh *neigh;
 707         unsigned long flags;
 708
 709         if (likely(skb->dst && skb->dst->neighbour)) {
 710                 if (unlikely(!*to_ipoib_neigh(skb->dst->neighbour))) {
 711                         ipoib_path_lookup(skb, dev);
 712                         return NETDEV_TX_OK;
 713                 }
 714
 715                 neigh = *to_ipoib_neigh(skb->dst->neighbour);
 716
 717                 if (unlikely((memcmp(&neigh->dgid.raw,
 718                                      skb->dst->neighbour->ha + 4,
 719                                      sizeof(union ib_gid))) ||
 720                              (neigh->dev != dev))) {
 721                         spin_lock_irqsave(&priv->lock, flags);
 722                         /*
 723                          * It's safe to call ipoib_put_ah() inside
 724                          * priv->lock here, because we know that
 725                          * path->ah will always hold one more reference,
 726                          * so ipoib_put_ah() will never do more than
 727                          * decrement the ref count.
 728                          */
 729                         if (neigh->ah)
 730                                 ipoib_put_ah(neigh->ah);
 731                         list_del(&neigh->list);
 732                         ipoib_neigh_free(dev, neigh);
 733                         spin_unlock_irqrestore(&priv->lock, flags);
 734                         ipoib_path_lookup(skb, dev);
 735                         return NETDEV_TX_OK;
 736                 }
 737
 738                 if (ipoib_cm_get(neigh)) {
 739                         if (ipoib_cm_up(neigh)) {
 740                                 ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
 741                                 return NETDEV_TX_OK;
 742                         }
 743                 } else if (neigh->ah) {
 744                         ipoib_send(dev, skb, neigh->ah, IPOIB_QPN(skb->dst->neighbour->ha));
 745                         return NETDEV_TX_OK;
 746                 }
 747
 748                 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 749                         spin_lock_irqsave(&priv->lock, flags);
 750                         __skb_queue_tail(&neigh->queue, skb);
 751                         spin_unlock_irqrestore(&priv->lock, flags);
 752                 } else {
 753                         ++dev->stats.tx_dropped;
 754                         dev_kfree_skb_any(skb);
 755                 }
 756         } else {
 757                 struct ipoib_pseudoheader *phdr =
 758                         (struct ipoib_pseudoheader *) skb->data;
 759                 skb_pull(skb, sizeof *phdr);
 760
 761                 if (phdr->hwaddr[4] == 0xff) {
 762                         /* Add in the P_Key for multicast*/
 763                         phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
 764                         phdr->hwaddr[9] = priv->pkey & 0xff;
 765
 766                         ipoib_mcast_send(dev, phdr->hwaddr + 4, skb);
 767                 } else {
 768                         /* unicast GID -- should be ARP or RARP reply */
 769
 770                         if ((be16_to_cpup((__be16 *) skb->data) != ETH_P_ARP) &&
 771                             (be16_to_cpup((__be16 *) skb->data) != ETH_P_RARP)) {
 772                                 ipoib_warn(priv, "Unicast, no %s: type %04x, QPN %06x %pI6\n",
 773                                            skb->dst ? "neigh" : "dst",
 774                                            be16_to_cpup((__be16 *) skb->data),
 775                                            IPOIB_QPN(phdr->hwaddr),
 776                                            phdr->hwaddr + 4);
 777                                 dev_kfree_skb_any(skb);
 778                                 ++dev->stats.tx_dropped;
 779                                 return NETDEV_TX_OK;
 780                         }
 781
 782                         unicast_arp_send(skb, dev, phdr);
 783                 }
 784         }
 785
 786         return NETDEV_TX_OK;
 787 }
 788
 789 static void ipoib_timeout(struct net_device *dev)
 790 {
 791         struct ipoib_dev_priv *priv = netdev_priv(dev);
 792
 793         ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
 794                    jiffies_to_msecs(jiffies - dev->trans_start));
 795         ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
 796                    netif_queue_stopped(dev),
 797                    priv->tx_head, priv->tx_tail);
 798         /* XXX reset QP, etc. */
 799 }
 800
 801 static int ipoib_hard_header(struct sk_buff *skb,
 802                              struct net_device *dev,
 803                              unsigned short type,
 804                              const void *daddr, const void *saddr, unsigned len)
 805 {
 806         struct ipoib_header *header;
 807
 808         header = (struct ipoib_header *) skb_push(skb, sizeof *header);
 809
 810         header->proto = htons(type);
 811         header->reserved = 0;
 812
 813         /*
 814          * If we don't have a neighbour structure, stuff the
 815          * destination address onto the front of the skb so we can
 816          * figure out where to send the packet later.
 817          */
 818         if ((!skb->dst || !skb->dst->neighbour) && daddr) {
 819                 struct ipoib_pseudoheader *phdr =
 820                         (struct ipoib_pseudoheader *) skb_push(skb, sizeof *phdr);
 821                 memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
 822         }
 823
 824         return 0;
 825 }
 826
 827 static void ipoib_set_mcast_list(struct net_device *dev)
 828 {
 829         struct ipoib_dev_priv *priv = netdev_priv(dev);
 830
 831         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
 832                 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
 833                 return;
 834         }
 835
 836         queue_work(ipoib_workqueue, &priv->restart_task);
 837 }
 838
 839 static void ipoib_neigh_cleanup(struct neighbour *n)
 840 {
 841         struct ipoib_neigh *neigh;
 842         struct ipoib_dev_priv *priv = netdev_priv(n->dev);
 843         unsigned long flags;
 844         struct ipoib_ah *ah = NULL;
 845
 846         neigh = *to_ipoib_neigh(n);
 847         if (neigh)
 848                 priv = netdev_priv(neigh->dev);
 849         else
 850                 return;
 851         ipoib_dbg(priv,
 852                   "neigh_cleanup for %06x %pI6\n",
 853                   IPOIB_QPN(n->ha),
 854                   n->ha + 4);
 855
 856         spin_lock_irqsave(&priv->lock, flags);
 857
 858         if (neigh->ah)
 859                 ah = neigh->ah;
 860         list_del(&neigh->list);
 861         ipoib_neigh_free(n->dev, neigh);
 862
 863         spin_unlock_irqrestore(&priv->lock, flags);
 864
 865         if (ah)
 866                 ipoib_put_ah(ah);
 867 }
 868
 869 struct ipoib_neigh *ipoib_neigh_alloc(struct neighbour *neighbour,
 870                                       struct net_device *dev)
 871 {
 872         struct ipoib_neigh *neigh;
 873
 874         neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
 875         if (!neigh)
 876                 return NULL;
 877
 878         neigh->neighbour = neighbour;
 879         neigh->dev = dev;
 880         *to_ipoib_neigh(neighbour) = neigh;
 881         skb_queue_head_init(&neigh->queue);
 882         ipoib_cm_set(neigh, NULL);
 883
 884         return neigh;
 885 }
 886
 887 void ipoib_neigh_free(struct net_device *dev, struct ipoib_neigh *neigh)
 888 {
 889         struct sk_buff *skb;
 890         *to_ipoib_neigh(neigh->neighbour) = NULL;
 891         while ((skb = __skb_dequeue(&neigh->queue))) {
 892                 ++dev->stats.tx_dropped;
 893                 dev_kfree_skb_any(skb);
 894         }
 895         if (ipoib_cm_get(neigh))
 896                 ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
 897         kfree(neigh);
 898 }
 899
 900 static int ipoib_neigh_setup_dev(struct net_device *dev, struct neigh_parms *parms)
 901 {
 902         parms->neigh_cleanup = ipoib_neigh_cleanup;
 903
 904         return 0;
 905 }
 906
 907 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
 908 {
 909         struct ipoib_dev_priv *priv = netdev_priv(dev);
 910
 911         /* Allocate RX/TX "rings" to hold queued skbs */
 912         priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
 913                                 GFP_KERNEL);
 914         if (!priv->rx_ring) {
 915                 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
 916                        ca->name, ipoib_recvq_size);
 917                 goto out;
 918         }
 919
 920         priv->tx_ring = vmalloc(ipoib_sendq_size * sizeof *priv->tx_ring);
 921         if (!priv->tx_ring) {
 922                 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
 923                        ca->name, ipoib_sendq_size);
 924                 goto out_rx_ring_cleanup;
 925         }
 926         memset(priv->tx_ring, 0, ipoib_sendq_size * sizeof *priv->tx_ring);
 927
 928         /* priv->tx_head, tx_tail & tx_outstanding are already 0 */
 929
 930         if (ipoib_ib_dev_init(dev, ca, port))
 931                 goto out_tx_ring_cleanup;
 932
 933         return 0;
 934
 935 out_tx_ring_cleanup:
 936         vfree(priv->tx_ring);
 937
 938 out_rx_ring_cleanup:
 939         kfree(priv->rx_ring);
 940
 941 out:
 942         return -ENOMEM;
 943 }
 944
 945 void ipoib_dev_cleanup(struct net_device *dev)
 946 {
 947         struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
 948
 949         ipoib_delete_debug_files(dev);
 950
 951         /* Delete any child interfaces first */
 952         list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
 953                 unregister_netdev(cpriv->dev);
 954                 ipoib_dev_cleanup(cpriv->dev);
 955                 free_netdev(cpriv->dev);
 956         }
 957
 958         ipoib_ib_dev_cleanup(dev);
 959
 960         kfree(priv->rx_ring);
 961         vfree(priv->tx_ring);
 962
 963         priv->rx_ring = NULL;
 964         priv->tx_ring = NULL;
 965 }
 966
 967 static const struct header_ops ipoib_header_ops = {
 968         .create = ipoib_hard_header,
 969 };
 970
 971 static int get_skb_hdr(struct sk_buff *skb, void **iphdr,
 972                        void **tcph, u64 *hdr_flags, void *priv)
 973 {
 974         unsigned int ip_len;
 975         struct iphdr *iph;
 976
 977         if (unlikely(skb->protocol != htons(ETH_P_IP)))
 978                 return -1;
 979
 980         /*
 981          * In the future we may add an else clause that verifies the
 982          * checksum and allows devices which do not calculate checksum
 983          * to use LRO.
 984          */
 985         if (unlikely(skb->ip_summed != CHECKSUM_UNNECESSARY))
 986                 return -1;
 987
 988         /* Check for non-TCP packet */
 989         skb_reset_network_header(skb);
 990         iph = ip_hdr(skb);
 991         if (iph->protocol != IPPROTO_TCP)
 992                 return -1;
 993
 994         ip_len = ip_hdrlen(skb);
 995         skb_set_transport_header(skb, ip_len);
 996         *tcph = tcp_hdr(skb);
 997
 998         /* check if IP header and TCP header are complete */
 999         if (ntohs(iph->tot_len) < ip_len + tcp_hdrlen(skb))
1000                 return -1;
1001
1002         *hdr_flags = LRO_IPV4 | LRO_TCP;
1003         *iphdr = iph;
1004
1005         return 0;
1006 }
1007
1008 static void ipoib_lro_setup(struct ipoib_dev_priv *priv)
1009 {
1010         priv->lro.lro_mgr.max_aggr       = lro_max_aggr;
1011         priv->lro.lro_mgr.max_desc       = IPOIB_MAX_LRO_DESCRIPTORS;
1012         priv->lro.lro_mgr.lro_arr        = priv->lro.lro_desc;
1013         priv->lro.lro_mgr.get_skb_header = get_skb_hdr;
1014         priv->lro.lro_mgr.features       = LRO_F_NAPI;
1015         priv->lro.lro_mgr.dev            = priv->dev;
1016         priv->lro.lro_mgr.ip_summed_aggr = CHECKSUM_UNNECESSARY;
1017 }
1018
1019 static void ipoib_setup(struct net_device *dev)
1020 {
1021         struct ipoib_dev_priv *priv = netdev_priv(dev);
1022
1023         dev->open                = ipoib_open;
1024         dev->stop                = ipoib_stop;
1025         dev->change_mtu          = ipoib_change_mtu;
1026         dev->hard_start_xmit     = ipoib_start_xmit;
1027         dev->tx_timeout          = ipoib_timeout;
1028         dev->header_ops          = &ipoib_header_ops;
1029         dev->set_multicast_list  = ipoib_set_mcast_list;
1030         dev->neigh_setup         = ipoib_neigh_setup_dev;
1031
1032         ipoib_set_ethtool_ops(dev);
1033
1034         netif_napi_add(dev, &priv->napi, ipoib_poll, 100);
1035
1036         dev->watchdog_timeo      = HZ;
1037
1038         dev->flags              |= IFF_BROADCAST | IFF_MULTICAST;
1039
1040         /*
1041          * We add in INFINIBAND_ALEN to allow for the destination
1042          * address "pseudoheader" for skbs without neighbour struct.
1043          */
1044         dev->hard_header_len     = IPOIB_ENCAP_LEN + INFINIBAND_ALEN;
1045         dev->addr_len            = INFINIBAND_ALEN;
1046         dev->type                = ARPHRD_INFINIBAND;
1047         dev->tx_queue_len        = ipoib_sendq_size * 2;
1048         dev->features            = (NETIF_F_VLAN_CHALLENGED     |
1049                                     NETIF_F_HIGHDMA);
1050
1051         memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
1052
1053         netif_carrier_off(dev);
1054
1055         priv->dev = dev;
1056
1057         ipoib_lro_setup(priv);
1058
1059         spin_lock_init(&priv->lock);
1060
1061         mutex_init(&priv->vlan_mutex);
1062
1063         INIT_LIST_HEAD(&priv->path_list);
1064         INIT_LIST_HEAD(&priv->child_intfs);
1065         INIT_LIST_HEAD(&priv->dead_ahs);
1066         INIT_LIST_HEAD(&priv->multicast_list);
1067
1068         INIT_DELAYED_WORK(&priv->pkey_poll_task, ipoib_pkey_poll);
1069         INIT_DELAYED_WORK(&priv->mcast_task,   ipoib_mcast_join_task);
1070         INIT_WORK(&priv->carrier_on_task, ipoib_mcast_carrier_on_task);
1071         INIT_WORK(&priv->flush_light,   ipoib_ib_dev_flush_light);
1072         INIT_WORK(&priv->flush_normal,   ipoib_ib_dev_flush_normal);
1073         INIT_WORK(&priv->flush_heavy,   ipoib_ib_dev_flush_heavy);
1074         INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
1075         INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
1076 }
1077
1078 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
1079 {
1080         struct net_device *dev;
1081
1082         dev = alloc_netdev((int) sizeof (struct ipoib_dev_priv), name,
1083                            ipoib_setup);
1084         if (!dev)
1085                 return NULL;
1086
1087         return netdev_priv(dev);
1088 }
1089
1090 static ssize_t show_pkey(struct device *dev,
1091                          struct device_attribute *attr, char *buf)
1092 {
1093         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1094
1095         return sprintf(buf, "0x%04x\n", priv->pkey);
1096 }
1097 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1098
1099 static ssize_t show_umcast(struct device *dev,
1100                            struct device_attribute *attr, char *buf)
1101 {
1102         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1103
1104         return sprintf(buf, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
1105 }
1106
1107 static ssize_t set_umcast(struct device *dev,
1108                           struct device_attribute *attr,
1109                           const char *buf, size_t count)
1110 {
1111         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1112         unsigned long umcast_val = simple_strtoul(buf, NULL, 0);
1113
1114         if (umcast_val > 0) {
1115                 set_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1116                 ipoib_warn(priv, "ignoring multicast groups joined directly "
1117                                 "by userspace\n");
1118         } else
1119                 clear_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1120
1121         return count;
1122 }
1123 static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast);
1124
1125 int ipoib_add_umcast_attr(struct net_device *dev)
1126 {
1127         return device_create_file(&dev->dev, &dev_attr_umcast);
1128 }
1129
1130 static ssize_t create_child(struct device *dev,
1131                             struct device_attribute *attr,
1132                             const char *buf, size_t count)
1133 {
1134         int pkey;
1135         int ret;
1136
1137         if (sscanf(buf, "%i", &pkey) != 1)
1138                 return -EINVAL;
1139
1140         if (pkey < 0 || pkey > 0xffff)
1141                 return -EINVAL;
1142
1143         /*
1144          * Set the full membership bit, so that we join the right
1145          * broadcast group, etc.
1146          */
1147         pkey |= 0x8000;
1148
1149         ret = ipoib_vlan_add(to_net_dev(dev), pkey);
1150
1151         return ret ? ret : count;
1152 }
1153 static DEVICE_ATTR(create_child, S_IWUGO, NULL, create_child);
1154
1155 static ssize_t delete_child(struct device *dev,
1156                             struct device_attribute *attr,
1157                             const char *buf, size_t count)
1158 {
1159         int pkey;
1160         int ret;
1161
1162         if (sscanf(buf, "%i", &pkey) != 1)
1163                 return -EINVAL;
1164
1165         if (pkey < 0 || pkey > 0xffff)
1166                 return -EINVAL;
1167
1168         ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
1169
1170         return ret ? ret : count;
1171
1172 }
1173 static DEVICE_ATTR(delete_child, S_IWUGO, NULL, delete_child);
1174
1175 int ipoib_add_pkey_attr(struct net_device *dev)
1176 {
1177         return device_create_file(&dev->dev, &dev_attr_pkey);
1178 }
1179
1180 int ipoib_set_dev_features(struct ipoib_dev_priv *priv, struct ib_device *hca)
1181 {
1182         struct ib_device_attr *device_attr;
1183         int result = -ENOMEM;
1184
1185         device_attr = kmalloc(sizeof *device_attr, GFP_KERNEL);
1186         if (!device_attr) {
1187                 printk(KERN_WARNING "%s: allocation of %zu bytes failed\n",
1188                        hca->name, sizeof *device_attr);
1189                 return result;
1190         }
1191
1192         result = ib_query_device(hca, device_attr);
1193         if (result) {
1194                 printk(KERN_WARNING "%s: ib_query_device failed (ret = %d)\n",
1195                        hca->name, result);
1196                 kfree(device_attr);
1197                 return result;
1198         }
1199         priv->hca_caps = device_attr->device_cap_flags;
1200
1201         kfree(device_attr);
1202
1203         if (priv->hca_caps & IB_DEVICE_UD_IP_CSUM) {
1204                 set_bit(IPOIB_FLAG_CSUM, &priv->flags);
1205                 priv->dev->features |= NETIF_F_SG | NETIF_F_IP_CSUM;
1206         }
1207
1208         if (lro)
1209                 priv->dev->features |= NETIF_F_LRO;
1210
1211         if (priv->dev->features & NETIF_F_SG && priv->hca_caps & IB_DEVICE_UD_TSO)
1212                 priv->dev->features |= NETIF_F_TSO;
1213
1214         return 0;
1215 }
1216
1217
1218 static struct net_device *ipoib_add_port(const char *format,
1219                                          struct ib_device *hca, u8 port)
1220 {
1221         struct ipoib_dev_priv *priv;
1222         struct ib_port_attr attr;
1223         int result = -ENOMEM;
1224
1225         priv = ipoib_intf_alloc(format);
1226         if (!priv)
1227                 goto alloc_mem_failed;
1228
1229         SET_NETDEV_DEV(priv->dev, hca->dma_device);
1230
1231         if (!ib_query_port(hca, port, &attr))
1232                 priv->max_ib_mtu = ib_mtu_enum_to_int(attr.max_mtu);
1233         else {
1234                 printk(KERN_WARNING "%s: ib_query_port %d failed\n",
1235                        hca->name, port);
1236                 goto device_init_failed;
1237         }
1238
1239         /* MTU will be reset when mcast join happens */
1240         priv->dev->mtu  = IPOIB_UD_MTU(priv->max_ib_mtu);
1241         priv->mcast_mtu  = priv->admin_mtu = priv->dev->mtu;
1242
1243         result = ib_query_pkey(hca, port, 0, &priv->pkey);
1244         if (result) {
1245                 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
1246                        hca->name, port, result);
1247                 goto device_init_failed;
1248         }
1249
1250         if (ipoib_set_dev_features(priv, hca))
1251                 goto device_init_failed;
1252
1253         /*
1254          * Set the full membership bit, so that we join the right
1255          * broadcast group, etc.
1256          */
1257         priv->pkey |= 0x8000;
1258
1259         priv->dev->broadcast[8] = priv->pkey >> 8;
1260         priv->dev->broadcast[9] = priv->pkey & 0xff;
1261
1262         result = ib_query_gid(hca, port, 0, &priv->local_gid);
1263         if (result) {
1264                 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1265                        hca->name, port, result);
1266                 goto device_init_failed;
1267         } else
1268                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1269
1270         result = ipoib_dev_init(priv->dev, hca, port);
1271         if (result < 0) {
1272                 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1273                        hca->name, port, result);
1274                 goto device_init_failed;
1275         }
1276
1277         INIT_IB_EVENT_HANDLER(&priv->event_handler,
1278                               priv->ca, ipoib_event);
1279         result = ib_register_event_handler(&priv->event_handler);
1280         if (result < 0) {
1281                 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1282                        "port %d (ret = %d)\n",
1283                        hca->name, port, result);
1284                 goto event_failed;
1285         }
1286
1287         result = register_netdev(priv->dev);
1288         if (result) {
1289                 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1290                        hca->name, port, result);
1291                 goto register_failed;
1292         }
1293
1294         ipoib_create_debug_files(priv->dev);
1295
1296         if (ipoib_cm_add_mode_attr(priv->dev))
1297                 goto sysfs_failed;
1298         if (ipoib_add_pkey_attr(priv->dev))
1299                 goto sysfs_failed;
1300         if (ipoib_add_umcast_attr(priv->dev))
1301                 goto sysfs_failed;
1302         if (device_create_file(&priv->dev->dev, &dev_attr_create_child))
1303                 goto sysfs_failed;
1304         if (device_create_file(&priv->dev->dev, &dev_attr_delete_child))
1305                 goto sysfs_failed;
1306
1307         return priv->dev;
1308
1309 sysfs_failed:
1310         ipoib_delete_debug_files(priv->dev);
1311         unregister_netdev(priv->dev);
1312
1313 register_failed:
1314         ib_unregister_event_handler(&priv->event_handler);
1315         flush_workqueue(ipoib_workqueue);
1316
1317 event_failed:
1318         ipoib_dev_cleanup(priv->dev);
1319
1320 device_init_failed:
1321         free_netdev(priv->dev);
1322
1323 alloc_mem_failed:
1324         return ERR_PTR(result);
1325 }
1326
1327 static void ipoib_add_one(struct ib_device *device)
1328 {
1329         struct list_head *dev_list;
1330         struct net_device *dev;
1331         struct ipoib_dev_priv *priv;
1332         int s, e, p;
1333
1334         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1335                 return;
1336
1337         dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1338         if (!dev_list)
1339                 return;
1340
1341         INIT_LIST_HEAD(dev_list);
1342
1343         if (device->node_type == RDMA_NODE_IB_SWITCH) {
1344                 s = 0;
1345                 e = 0;
1346         } else {
1347                 s = 1;
1348                 e = device->phys_port_cnt;
1349         }
1350
1351         for (p = s; p <= e; ++p) {
1352                 dev = ipoib_add_port("ib%d", device, p);
1353                 if (!IS_ERR(dev)) {
1354                         priv = netdev_priv(dev);
1355                         list_add_tail(&priv->list, dev_list);
1356                 }
1357         }
1358
1359         ib_set_client_data(device, &ipoib_client, dev_list);
1360 }
1361
1362 static void ipoib_remove_one(struct ib_device *device)
1363 {
1364         struct ipoib_dev_priv *priv, *tmp;
1365         struct list_head *dev_list;
1366
1367         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1368                 return;
1369
1370         dev_list = ib_get_client_data(device, &ipoib_client);
1371
1372         list_for_each_entry_safe(priv, tmp, dev_list, list) {
1373                 ib_unregister_event_handler(&priv->event_handler);
1374
1375                 rtnl_lock();
1376                 dev_change_flags(priv->dev, priv->dev->flags & ~IFF_UP);
1377                 rtnl_unlock();
1378
1379                 flush_workqueue(ipoib_workqueue);
1380
1381                 unregister_netdev(priv->dev);
1382                 ipoib_dev_cleanup(priv->dev);
1383                 free_netdev(priv->dev);
1384         }
1385
1386         kfree(dev_list);
1387 }
1388
1389 static int __init ipoib_init_module(void)
1390 {
1391         int ret;
1392
1393         ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1394         ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1395         ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1396
1397         ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1398         ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1399         ipoib_sendq_size = max(ipoib_sendq_size, max(2 * MAX_SEND_CQE,
1400                                                      IPOIB_MIN_QUEUE_SIZE));
1401 #ifdef CONFIG_INFINIBAND_IPOIB_CM
1402         ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
1403 #endif
1404
1405         /*
1406          * When copying small received packets, we only copy from the
1407          * linear data part of the SKB, so we rely on this condition.
1408          */
1409         BUILD_BUG_ON(IPOIB_CM_COPYBREAK > IPOIB_CM_HEAD_SIZE);
1410
1411         ret = ipoib_register_debugfs();
1412         if (ret)
1413                 return ret;
1414
1415         /*
1416          * We create our own workqueue mainly because we want to be
1417          * able to flush it when devices are being removed.  We can't
1418          * use schedule_work()/flush_scheduled_work() because both
1419          * unregister_netdev() and linkwatch_event take the rtnl lock,
1420          * so flush_scheduled_work() can deadlock during device
1421          * removal.
1422          */
1423         ipoib_workqueue = create_singlethread_workqueue("ipoib");
1424         if (!ipoib_workqueue) {
1425                 ret = -ENOMEM;
1426                 goto err_fs;
1427         }
1428
1429         ib_sa_register_client(&ipoib_sa_client);
1430
1431         ret = ib_register_client(&ipoib_client);
1432         if (ret)
1433                 goto err_sa;
1434
1435         return 0;
1436
1437 err_sa:
1438         ib_sa_unregister_client(&ipoib_sa_client);
1439         destroy_workqueue(ipoib_workqueue);
1440
1441 err_fs:
1442         ipoib_unregister_debugfs();
1443
1444         return ret;
1445 }
1446
1447 static void __exit ipoib_cleanup_module(void)
1448 {
1449         ib_unregister_client(&ipoib_client);
1450         ib_sa_unregister_client(&ipoib_sa_client);
1451         ipoib_unregister_debugfs();
1452         destroy_workqueue(ipoib_workqueue);
1453 }
1454
1455 module_init(ipoib_init_module);
1456 module_exit(ipoib_cleanup_module);