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 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
  64 int ipoib_debug_level;
  65
  66 module_param_named(debug_level, ipoib_debug_level, int, 0644);
  67 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
  68 #endif
  69
  70 struct ipoib_path_iter {
  71         struct net_device *dev;
  72         struct ipoib_path  path;
  73 };
  74
  75 static const u8 ipv4_bcast_addr[] = {
  76         0x00, 0xff, 0xff, 0xff,
  77         0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
  78         0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff
  79 };
  80
  81 struct workqueue_struct *ipoib_workqueue;
  82
  83 struct ib_sa_client ipoib_sa_client;
  84
  85 static void ipoib_add_one(struct ib_device *device);
  86 static void ipoib_remove_one(struct ib_device *device);
  87
  88 static struct ib_client ipoib_client = {
  89         .name   = "ipoib",
  90         .add    = ipoib_add_one,
  91         .remove = ipoib_remove_one
  92 };
  93
  94 int ipoib_open(struct net_device *dev)
  95 {
  96         struct ipoib_dev_priv *priv = netdev_priv(dev);
  97
  98         ipoib_dbg(priv, "bringing up interface\n");
  99
 100         set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 101
 102         if (ipoib_pkey_dev_delay_open(dev))
 103                 return 0;
 104
 105         if (ipoib_ib_dev_open(dev))
 106                 goto err_disable;
 107
 108         if (ipoib_ib_dev_up(dev))
 109                 goto err_stop;
 110
 111         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 112                 struct ipoib_dev_priv *cpriv;
 113
 114                 /* Bring up any child interfaces too */
 115                 mutex_lock(&priv->vlan_mutex);
 116                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 117                         int flags;
 118
 119                         flags = cpriv->dev->flags;
 120                         if (flags & IFF_UP)
 121                                 continue;
 122
 123                         dev_change_flags(cpriv->dev, flags | IFF_UP);
 124                 }
 125                 mutex_unlock(&priv->vlan_mutex);
 126         }
 127
 128         netif_start_queue(dev);
 129
 130         return 0;
 131
 132 err_stop:
 133         ipoib_ib_dev_stop(dev, 1);
 134
 135 err_disable:
 136         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 137
 138         return -EINVAL;
 139 }
 140
 141 static int ipoib_stop(struct net_device *dev)
 142 {
 143         struct ipoib_dev_priv *priv = netdev_priv(dev);
 144
 145         ipoib_dbg(priv, "stopping interface\n");
 146
 147         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 148
 149         netif_stop_queue(dev);
 150
 151         ipoib_ib_dev_down(dev, 1);
 152         ipoib_ib_dev_stop(dev, 0);
 153
 154         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 155                 struct ipoib_dev_priv *cpriv;
 156
 157                 /* Bring down any child interfaces too */
 158                 mutex_lock(&priv->vlan_mutex);
 159                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 160                         int flags;
 161
 162                         flags = cpriv->dev->flags;
 163                         if (!(flags & IFF_UP))
 164                                 continue;
 165
 166                         dev_change_flags(cpriv->dev, flags & ~IFF_UP);
 167                 }
 168                 mutex_unlock(&priv->vlan_mutex);
 169         }
 170
 171         return 0;
 172 }
 173
 174 static u32 ipoib_fix_features(struct net_device *dev, u32 features)
 175 {
 176         struct ipoib_dev_priv *priv = netdev_priv(dev);
 177
 178         if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
 179                 features &= ~(NETIF_F_SG | NETIF_F_IP_CSUM | NETIF_F_TSO);
 180
 181         return features;
 182 }
 183
 184 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
 185 {
 186         struct ipoib_dev_priv *priv = netdev_priv(dev);
 187
 188         /* dev->mtu > 2K ==> connected mode */
 189         if (ipoib_cm_admin_enabled(dev)) {
 190                 if (new_mtu > ipoib_cm_max_mtu(dev))
 191                         return -EINVAL;
 192
 193                 if (new_mtu > priv->mcast_mtu)
 194                         ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
 195                                    priv->mcast_mtu);
 196
 197                 dev->mtu = new_mtu;
 198                 return 0;
 199         }
 200
 201         if (new_mtu > IPOIB_UD_MTU(priv->max_ib_mtu))
 202                 return -EINVAL;
 203
 204         priv->admin_mtu = new_mtu;
 205
 206         dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
 207
 208         return 0;
 209 }
 210
 211 struct ipoib_path *__path_find(struct net_device *dev, void *gid)
 212 {
 213         struct ipoib_dev_priv *priv = netdev_priv(dev);
 214         struct rb_node *n = priv->path_tree.rb_node;
 215         struct ipoib_path *path;
 216         int ret;
 217
 218         while (n) {
 219                 path = rb_entry(n, struct ipoib_path, rb_node);
 220
 221                 ret = memcmp(gid, path->pathrec.dgid.raw,
 222                              sizeof (union ib_gid));
 223
 224                 if (ret < 0)
 225                         n = n->rb_left;
 226                 else if (ret > 0)
 227                         n = n->rb_right;
 228                 else
 229                         return path;
 230         }
 231
 232         return NULL;
 233 }
 234
 235 static int __path_add(struct net_device *dev, struct ipoib_path *path)
 236 {
 237         struct ipoib_dev_priv *priv = netdev_priv(dev);
 238         struct rb_node **n = &priv->path_tree.rb_node;
 239         struct rb_node *pn = NULL;
 240         struct ipoib_path *tpath;
 241         int ret;
 242
 243         while (*n) {
 244                 pn = *n;
 245                 tpath = rb_entry(pn, struct ipoib_path, rb_node);
 246
 247                 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
 248                              sizeof (union ib_gid));
 249                 if (ret < 0)
 250                         n = &pn->rb_left;
 251                 else if (ret > 0)
 252                         n = &pn->rb_right;
 253                 else
 254                         return -EEXIST;
 255         }
 256
 257         rb_link_node(&path->rb_node, pn, n);
 258         rb_insert_color(&path->rb_node, &priv->path_tree);
 259
 260         list_add_tail(&path->list, &priv->path_list);
 261
 262         return 0;
 263 }
 264
 265 static void path_free(struct net_device *dev, struct ipoib_path *path)
 266 {
 267         struct ipoib_dev_priv *priv = netdev_priv(dev);
 268         struct ipoib_neigh *neigh, *tn;
 269         struct sk_buff *skb;
 270         unsigned long flags;
 271
 272         while ((skb = __skb_dequeue(&path->queue)))
 273                 dev_kfree_skb_irq(skb);
 274
 275         spin_lock_irqsave(&priv->lock, flags);
 276
 277         list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
 278                 /*
 279                  * It's safe to call ipoib_put_ah() inside priv->lock
 280                  * here, because we know that path->ah will always
 281                  * hold one more reference, so ipoib_put_ah() will
 282                  * never do more than decrement the ref count.
 283                  */
 284                 if (neigh->ah)
 285                         ipoib_put_ah(neigh->ah);
 286
 287                 ipoib_neigh_free(dev, neigh);
 288         }
 289
 290         spin_unlock_irqrestore(&priv->lock, flags);
 291
 292         if (path->ah)
 293                 ipoib_put_ah(path->ah);
 294
 295         kfree(path);
 296 }
 297
 298 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 299
 300 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
 301 {
 302         struct ipoib_path_iter *iter;
 303
 304         iter = kmalloc(sizeof *iter, GFP_KERNEL);
 305         if (!iter)
 306                 return NULL;
 307
 308         iter->dev = dev;
 309         memset(iter->path.pathrec.dgid.raw, 0, 16);
 310
 311         if (ipoib_path_iter_next(iter)) {
 312                 kfree(iter);
 313                 return NULL;
 314         }
 315
 316         return iter;
 317 }
 318
 319 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
 320 {
 321         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
 322         struct rb_node *n;
 323         struct ipoib_path *path;
 324         int ret = 1;
 325
 326         spin_lock_irq(&priv->lock);
 327
 328         n = rb_first(&priv->path_tree);
 329
 330         while (n) {
 331                 path = rb_entry(n, struct ipoib_path, rb_node);
 332
 333                 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
 334                            sizeof (union ib_gid)) < 0) {
 335                         iter->path = *path;
 336                         ret = 0;
 337                         break;
 338                 }
 339
 340                 n = rb_next(n);
 341         }
 342
 343         spin_unlock_irq(&priv->lock);
 344
 345         return ret;
 346 }
 347
 348 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
 349                           struct ipoib_path *path)
 350 {
 351         *path = iter->path;
 352 }
 353
 354 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
 355
 356 void ipoib_mark_paths_invalid(struct net_device *dev)
 357 {
 358         struct ipoib_dev_priv *priv = netdev_priv(dev);
 359         struct ipoib_path *path, *tp;
 360
 361         spin_lock_irq(&priv->lock);
 362
 363         list_for_each_entry_safe(path, tp, &priv->path_list, list) {
 364                 ipoib_dbg(priv, "mark path LID 0x%04x GID %pI6 invalid\n",
 365                         be16_to_cpu(path->pathrec.dlid),
 366                         path->pathrec.dgid.raw);
 367                 path->valid =  0;
 368         }
 369
 370         spin_unlock_irq(&priv->lock);
 371 }
 372
 373 void ipoib_flush_paths(struct net_device *dev)
 374 {
 375         struct ipoib_dev_priv *priv = netdev_priv(dev);
 376         struct ipoib_path *path, *tp;
 377         LIST_HEAD(remove_list);
 378         unsigned long flags;
 379
 380         netif_tx_lock_bh(dev);
 381         spin_lock_irqsave(&priv->lock, flags);
 382
 383         list_splice_init(&priv->path_list, &remove_list);
 384
 385         list_for_each_entry(path, &remove_list, list)
 386                 rb_erase(&path->rb_node, &priv->path_tree);
 387
 388         list_for_each_entry_safe(path, tp, &remove_list, list) {
 389                 if (path->query)
 390                         ib_sa_cancel_query(path->query_id, path->query);
 391                 spin_unlock_irqrestore(&priv->lock, flags);
 392                 netif_tx_unlock_bh(dev);
 393                 wait_for_completion(&path->done);
 394                 path_free(dev, path);
 395                 netif_tx_lock_bh(dev);
 396                 spin_lock_irqsave(&priv->lock, flags);
 397         }
 398
 399         spin_unlock_irqrestore(&priv->lock, flags);
 400         netif_tx_unlock_bh(dev);
 401 }
 402
 403 static void path_rec_completion(int status,
 404                                 struct ib_sa_path_rec *pathrec,
 405                                 void *path_ptr)
 406 {
 407         struct ipoib_path *path = path_ptr;
 408         struct net_device *dev = path->dev;
 409         struct ipoib_dev_priv *priv = netdev_priv(dev);
 410         struct ipoib_ah *ah = NULL;
 411         struct ipoib_ah *old_ah = NULL;
 412         struct ipoib_neigh *neigh, *tn;
 413         struct sk_buff_head skqueue;
 414         struct sk_buff *skb;
 415         unsigned long flags;
 416
 417         if (!status)
 418                 ipoib_dbg(priv, "PathRec LID 0x%04x for GID %pI6\n",
 419                           be16_to_cpu(pathrec->dlid), pathrec->dgid.raw);
 420         else
 421                 ipoib_dbg(priv, "PathRec status %d for GID %pI6\n",
 422                           status, path->pathrec.dgid.raw);
 423
 424         skb_queue_head_init(&skqueue);
 425
 426         if (!status) {
 427                 struct ib_ah_attr av;
 428
 429                 if (!ib_init_ah_from_path(priv->ca, priv->port, pathrec, &av))
 430                         ah = ipoib_create_ah(dev, priv->pd, &av);
 431         }
 432
 433         spin_lock_irqsave(&priv->lock, flags);
 434
 435         if (!IS_ERR_OR_NULL(ah)) {
 436                 path->pathrec = *pathrec;
 437
 438                 old_ah   = path->ah;
 439                 path->ah = ah;
 440
 441                 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
 442                           ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
 443
 444                 while ((skb = __skb_dequeue(&path->queue)))
 445                         __skb_queue_tail(&skqueue, skb);
 446
 447                 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
 448                         if (neigh->ah) {
 449                                 WARN_ON(neigh->ah != old_ah);
 450                                 /*
 451                                  * Dropping the ah reference inside
 452                                  * priv->lock is safe here, because we
 453                                  * will hold one more reference from
 454                                  * the original value of path->ah (ie
 455                                  * old_ah).
 456                                  */
 457                                 ipoib_put_ah(neigh->ah);
 458                         }
 459                         kref_get(&path->ah->ref);
 460                         neigh->ah = path->ah;
 461                         memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
 462                                sizeof(union ib_gid));
 463
 464                         if (ipoib_cm_enabled(dev, neigh->neighbour)) {
 465                                 if (!ipoib_cm_get(neigh))
 466                                         ipoib_cm_set(neigh, ipoib_cm_create_tx(dev,
 467                                                                                path,
 468                                                                                neigh));
 469                                 if (!ipoib_cm_get(neigh)) {
 470                                         list_del(&neigh->list);
 471                                         if (neigh->ah)
 472                                                 ipoib_put_ah(neigh->ah);
 473                                         ipoib_neigh_free(dev, neigh);
 474                                         continue;
 475                                 }
 476                         }
 477
 478                         while ((skb = __skb_dequeue(&neigh->queue)))
 479                                 __skb_queue_tail(&skqueue, skb);
 480                 }
 481                 path->valid = 1;
 482         }
 483
 484         path->query = NULL;
 485         complete(&path->done);
 486
 487         spin_unlock_irqrestore(&priv->lock, flags);
 488
 489         if (old_ah)
 490                 ipoib_put_ah(old_ah);
 491
 492         while ((skb = __skb_dequeue(&skqueue))) {
 493                 skb->dev = dev;
 494                 if (dev_queue_xmit(skb))
 495                         ipoib_warn(priv, "dev_queue_xmit failed "
 496                                    "to requeue packet\n");
 497         }
 498 }
 499
 500 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
 501 {
 502         struct ipoib_dev_priv *priv = netdev_priv(dev);
 503         struct ipoib_path *path;
 504
 505         if (!priv->broadcast)
 506                 return NULL;
 507
 508         path = kzalloc(sizeof *path, GFP_ATOMIC);
 509         if (!path)
 510                 return NULL;
 511
 512         path->dev = dev;
 513
 514         skb_queue_head_init(&path->queue);
 515
 516         INIT_LIST_HEAD(&path->neigh_list);
 517
 518         memcpy(path->pathrec.dgid.raw, gid, sizeof (union ib_gid));
 519         path->pathrec.sgid          = priv->local_gid;
 520         path->pathrec.pkey          = cpu_to_be16(priv->pkey);
 521         path->pathrec.numb_path     = 1;
 522         path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
 523
 524         return path;
 525 }
 526
 527 static int path_rec_start(struct net_device *dev,
 528                           struct ipoib_path *path)
 529 {
 530         struct ipoib_dev_priv *priv = netdev_priv(dev);
 531
 532         ipoib_dbg(priv, "Start path record lookup for %pI6\n",
 533                   path->pathrec.dgid.raw);
 534
 535         init_completion(&path->done);
 536
 537         path->query_id =
 538                 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
 539                                    &path->pathrec,
 540                                    IB_SA_PATH_REC_DGID          |
 541                                    IB_SA_PATH_REC_SGID          |
 542                                    IB_SA_PATH_REC_NUMB_PATH     |
 543                                    IB_SA_PATH_REC_TRAFFIC_CLASS |
 544                                    IB_SA_PATH_REC_PKEY,
 545                                    1000, GFP_ATOMIC,
 546                                    path_rec_completion,
 547                                    path, &path->query);
 548         if (path->query_id < 0) {
 549                 ipoib_warn(priv, "ib_sa_path_rec_get failed: %d\n", path->query_id);
 550                 path->query = NULL;
 551                 complete(&path->done);
 552                 return path->query_id;
 553         }
 554
 555         return 0;
 556 }
 557
 558 /* called with rcu_read_lock */
 559 static void neigh_add_path(struct sk_buff *skb, struct net_device *dev)
 560 {
 561         struct ipoib_dev_priv *priv = netdev_priv(dev);
 562         struct ipoib_path *path;
 563         struct ipoib_neigh *neigh;
 564         struct neighbour *n;
 565         unsigned long flags;
 566
 567         n = dst_get_neighbour(skb_dst(skb));
 568         neigh = ipoib_neigh_alloc(n, skb->dev);
 569         if (!neigh) {
 570                 ++dev->stats.tx_dropped;
 571                 dev_kfree_skb_any(skb);
 572                 return;
 573         }
 574
 575         spin_lock_irqsave(&priv->lock, flags);
 576
 577         path = __path_find(dev, n->ha + 4);
 578         if (!path) {
 579                 path = path_rec_create(dev, n->ha + 4);
 580                 if (!path)
 581                         goto err_path;
 582
 583                 __path_add(dev, path);
 584         }
 585
 586         list_add_tail(&neigh->list, &path->neigh_list);
 587
 588         if (path->ah) {
 589                 kref_get(&path->ah->ref);
 590                 neigh->ah = path->ah;
 591                 memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
 592                        sizeof(union ib_gid));
 593
 594                 if (ipoib_cm_enabled(dev, neigh->neighbour)) {
 595                         if (!ipoib_cm_get(neigh))
 596                                 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
 597                         if (!ipoib_cm_get(neigh)) {
 598                                 list_del(&neigh->list);
 599                                 if (neigh->ah)
 600                                         ipoib_put_ah(neigh->ah);
 601                                 ipoib_neigh_free(dev, neigh);
 602                                 goto err_drop;
 603                         }
 604                         if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE)
 605                                 __skb_queue_tail(&neigh->queue, skb);
 606                         else {
 607                                 ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
 608                                            skb_queue_len(&neigh->queue));
 609                                 goto err_drop;
 610                         }
 611                 } else {
 612                         spin_unlock_irqrestore(&priv->lock, flags);
 613                         ipoib_send(dev, skb, path->ah, IPOIB_QPN(n->ha));
 614                         return;
 615                 }
 616         } else {
 617                 neigh->ah  = NULL;
 618
 619                 if (!path->query && path_rec_start(dev, path))
 620                         goto err_list;
 621
 622                 __skb_queue_tail(&neigh->queue, skb);
 623         }
 624
 625         spin_unlock_irqrestore(&priv->lock, flags);
 626         return;
 627
 628 err_list:
 629         list_del(&neigh->list);
 630
 631 err_path:
 632         ipoib_neigh_free(dev, neigh);
 633 err_drop:
 634         ++dev->stats.tx_dropped;
 635         dev_kfree_skb_any(skb);
 636
 637         spin_unlock_irqrestore(&priv->lock, flags);
 638 }
 639
 640 /* called with rcu_read_lock */
 641 static void ipoib_path_lookup(struct sk_buff *skb, struct net_device *dev)
 642 {
 643         struct ipoib_dev_priv *priv = netdev_priv(skb->dev);
 644         struct dst_entry *dst = skb_dst(skb);
 645         struct neighbour *n;
 646
 647         /* Look up path record for unicasts */
 648         n = dst_get_neighbour(dst);
 649         if (n->ha[4] != 0xff) {
 650                 neigh_add_path(skb, dev);
 651                 return;
 652         }
 653
 654         /* Add in the P_Key for multicasts */
 655         n->ha[8] = (priv->pkey >> 8) & 0xff;
 656         n->ha[9] = priv->pkey & 0xff;
 657         ipoib_mcast_send(dev, n->ha + 4, skb);
 658 }
 659
 660 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
 661                              struct ipoib_cb *cb)
 662 {
 663         struct ipoib_dev_priv *priv = netdev_priv(dev);
 664         struct ipoib_path *path;
 665         unsigned long flags;
 666
 667         spin_lock_irqsave(&priv->lock, flags);
 668
 669         path = __path_find(dev, cb->hwaddr + 4);
 670         if (!path || !path->valid) {
 671                 int new_path = 0;
 672
 673                 if (!path) {
 674                         path = path_rec_create(dev, cb->hwaddr + 4);
 675                         new_path = 1;
 676                 }
 677                 if (path) {
 678                         __skb_queue_tail(&path->queue, skb);
 679
 680                         if (!path->query && path_rec_start(dev, path)) {
 681                                 spin_unlock_irqrestore(&priv->lock, flags);
 682                                 if (new_path)
 683                                         path_free(dev, path);
 684                                 return;
 685                         } else
 686                                 __path_add(dev, path);
 687                 } else {
 688                         ++dev->stats.tx_dropped;
 689                         dev_kfree_skb_any(skb);
 690                 }
 691
 692                 spin_unlock_irqrestore(&priv->lock, flags);
 693                 return;
 694         }
 695
 696         if (path->ah) {
 697                 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
 698                           be16_to_cpu(path->pathrec.dlid));
 699
 700                 spin_unlock_irqrestore(&priv->lock, flags);
 701                 ipoib_send(dev, skb, path->ah, IPOIB_QPN(cb->hwaddr));
 702                 return;
 703         } else if ((path->query || !path_rec_start(dev, path)) &&
 704                    skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 705                 __skb_queue_tail(&path->queue, skb);
 706         } else {
 707                 ++dev->stats.tx_dropped;
 708                 dev_kfree_skb_any(skb);
 709         }
 710
 711         spin_unlock_irqrestore(&priv->lock, flags);
 712 }
 713
 714 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
 715 {
 716         struct ipoib_dev_priv *priv = netdev_priv(dev);
 717         struct ipoib_neigh *neigh;
 718         struct neighbour *n = NULL;
 719         unsigned long flags;
 720
 721         rcu_read_lock();
 722         if (likely(skb_dst(skb)))
 723                 n = dst_get_neighbour(skb_dst(skb));
 724
 725         if (likely(n)) {
 726                 if (unlikely(!*to_ipoib_neigh(n))) {
 727                         ipoib_path_lookup(skb, dev);
 728                         goto unlock;
 729                 }
 730
 731                 neigh = *to_ipoib_neigh(n);
 732
 733                 if (unlikely((memcmp(&neigh->dgid.raw,
 734                                      n->ha + 4,
 735                                      sizeof(union ib_gid))) ||
 736                              (neigh->dev != dev))) {
 737                         spin_lock_irqsave(&priv->lock, flags);
 738                         /*
 739                          * It's safe to call ipoib_put_ah() inside
 740                          * priv->lock here, because we know that
 741                          * path->ah will always hold one more reference,
 742                          * so ipoib_put_ah() will never do more than
 743                          * decrement the ref count.
 744                          */
 745                         if (neigh->ah)
 746                                 ipoib_put_ah(neigh->ah);
 747                         list_del(&neigh->list);
 748                         ipoib_neigh_free(dev, neigh);
 749                         spin_unlock_irqrestore(&priv->lock, flags);
 750                         ipoib_path_lookup(skb, dev);
 751                         goto unlock;
 752                 }
 753
 754                 if (ipoib_cm_get(neigh)) {
 755                         if (ipoib_cm_up(neigh)) {
 756                                 ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
 757                                 goto unlock;
 758                         }
 759                 } else if (neigh->ah) {
 760                         ipoib_send(dev, skb, neigh->ah, IPOIB_QPN(n->ha));
 761                         goto unlock;
 762                 }
 763
 764                 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 765                         spin_lock_irqsave(&priv->lock, flags);
 766                         __skb_queue_tail(&neigh->queue, skb);
 767                         spin_unlock_irqrestore(&priv->lock, flags);
 768                 } else {
 769                         ++dev->stats.tx_dropped;
 770                         dev_kfree_skb_any(skb);
 771                 }
 772         } else {
 773                 struct ipoib_cb *cb = (struct ipoib_cb *) skb->cb;
 774
 775                 if (cb->hwaddr[4] == 0xff) {
 776                         /* Add in the P_Key for multicast*/
 777                         cb->hwaddr[8] = (priv->pkey >> 8) & 0xff;
 778                         cb->hwaddr[9] = priv->pkey & 0xff;
 779
 780                         ipoib_mcast_send(dev, cb->hwaddr + 4, skb);
 781                 } else {
 782                         /* unicast GID -- should be ARP or RARP reply */
 783
 784                         if ((be16_to_cpup((__be16 *) skb->data) != ETH_P_ARP) &&
 785                             (be16_to_cpup((__be16 *) skb->data) != ETH_P_RARP)) {
 786                                 ipoib_warn(priv, "Unicast, no %s: type %04x, QPN %06x %pI6\n",
 787                                            skb_dst(skb) ? "neigh" : "dst",
 788                                            be16_to_cpup((__be16 *) skb->data),
 789                                            IPOIB_QPN(cb->hwaddr),
 790                                            cb->hwaddr + 4);
 791                                 dev_kfree_skb_any(skb);
 792                                 ++dev->stats.tx_dropped;
 793                                 goto unlock;
 794                         }
 795
 796                         unicast_arp_send(skb, dev, cb);
 797                 }
 798         }
 799 unlock:
 800         rcu_read_unlock();
 801         return NETDEV_TX_OK;
 802 }
 803
 804 static void ipoib_timeout(struct net_device *dev)
 805 {
 806         struct ipoib_dev_priv *priv = netdev_priv(dev);
 807
 808         ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
 809                    jiffies_to_msecs(jiffies - dev->trans_start));
 810         ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
 811                    netif_queue_stopped(dev),
 812                    priv->tx_head, priv->tx_tail);
 813         /* XXX reset QP, etc. */
 814 }
 815
 816 static int ipoib_hard_header(struct sk_buff *skb,
 817                              struct net_device *dev,
 818                              unsigned short type,
 819                              const void *daddr, const void *saddr, unsigned len)
 820 {
 821         struct ipoib_header *header;
 822
 823         header = (struct ipoib_header *) skb_push(skb, sizeof *header);
 824
 825         header->proto = htons(type);
 826         header->reserved = 0;
 827
 828         /*
 829          * If we don't have a dst_entry structure, stuff the
 830          * destination address into skb->cb so we can figure out where
 831          * to send the packet later.
 832          */
 833         if (!skb_dst(skb)) {
 834                 struct ipoib_cb *cb = (struct ipoib_cb *) skb->cb;
 835                 memcpy(cb->hwaddr, daddr, INFINIBAND_ALEN);
 836         }
 837
 838         return 0;
 839 }
 840
 841 static void ipoib_set_mcast_list(struct net_device *dev)
 842 {
 843         struct ipoib_dev_priv *priv = netdev_priv(dev);
 844
 845         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
 846                 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
 847                 return;
 848         }
 849
 850         queue_work(ipoib_workqueue, &priv->restart_task);
 851 }
 852
 853 static void ipoib_neigh_cleanup(struct neighbour *n)
 854 {
 855         struct ipoib_neigh *neigh;
 856         struct ipoib_dev_priv *priv = netdev_priv(n->dev);
 857         unsigned long flags;
 858         struct ipoib_ah *ah = NULL;
 859
 860         neigh = *to_ipoib_neigh(n);
 861         if (neigh)
 862                 priv = netdev_priv(neigh->dev);
 863         else
 864                 return;
 865         ipoib_dbg(priv,
 866                   "neigh_cleanup for %06x %pI6\n",
 867                   IPOIB_QPN(n->ha),
 868                   n->ha + 4);
 869
 870         spin_lock_irqsave(&priv->lock, flags);
 871
 872         if (neigh->ah)
 873                 ah = neigh->ah;
 874         list_del(&neigh->list);
 875         ipoib_neigh_free(n->dev, neigh);
 876
 877         spin_unlock_irqrestore(&priv->lock, flags);
 878
 879         if (ah)
 880                 ipoib_put_ah(ah);
 881 }
 882
 883 struct ipoib_neigh *ipoib_neigh_alloc(struct neighbour *neighbour,
 884                                       struct net_device *dev)
 885 {
 886         struct ipoib_neigh *neigh;
 887
 888         neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
 889         if (!neigh)
 890                 return NULL;
 891
 892         neigh->neighbour = neighbour;
 893         neigh->dev = dev;
 894         memset(&neigh->dgid.raw, 0, sizeof (union ib_gid));
 895         *to_ipoib_neigh(neighbour) = neigh;
 896         skb_queue_head_init(&neigh->queue);
 897         ipoib_cm_set(neigh, NULL);
 898
 899         return neigh;
 900 }
 901
 902 void ipoib_neigh_free(struct net_device *dev, struct ipoib_neigh *neigh)
 903 {
 904         struct sk_buff *skb;
 905         *to_ipoib_neigh(neigh->neighbour) = NULL;
 906         while ((skb = __skb_dequeue(&neigh->queue))) {
 907                 ++dev->stats.tx_dropped;
 908                 dev_kfree_skb_any(skb);
 909         }
 910         if (ipoib_cm_get(neigh))
 911                 ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
 912         kfree(neigh);
 913 }
 914
 915 static int ipoib_neigh_setup_dev(struct net_device *dev, struct neigh_parms *parms)
 916 {
 917         parms->neigh_cleanup = ipoib_neigh_cleanup;
 918
 919         return 0;
 920 }
 921
 922 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
 923 {
 924         struct ipoib_dev_priv *priv = netdev_priv(dev);
 925
 926         /* Allocate RX/TX "rings" to hold queued skbs */
 927         priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
 928                                 GFP_KERNEL);
 929         if (!priv->rx_ring) {
 930                 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
 931                        ca->name, ipoib_recvq_size);
 932                 goto out;
 933         }
 934
 935         priv->tx_ring = vzalloc(ipoib_sendq_size * sizeof *priv->tx_ring);
 936         if (!priv->tx_ring) {
 937                 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
 938                        ca->name, ipoib_sendq_size);
 939                 goto out_rx_ring_cleanup;
 940         }
 941
 942         /* priv->tx_head, tx_tail & tx_outstanding are already 0 */
 943
 944         if (ipoib_ib_dev_init(dev, ca, port))
 945                 goto out_tx_ring_cleanup;
 946
 947         return 0;
 948
 949 out_tx_ring_cleanup:
 950         vfree(priv->tx_ring);
 951
 952 out_rx_ring_cleanup:
 953         kfree(priv->rx_ring);
 954
 955 out:
 956         return -ENOMEM;
 957 }
 958
 959 void ipoib_dev_cleanup(struct net_device *dev)
 960 {
 961         struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
 962
 963         ipoib_delete_debug_files(dev);
 964
 965         /* Delete any child interfaces first */
 966         list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
 967                 unregister_netdev(cpriv->dev);
 968                 ipoib_dev_cleanup(cpriv->dev);
 969                 free_netdev(cpriv->dev);
 970         }
 971
 972         ipoib_ib_dev_cleanup(dev);
 973
 974         kfree(priv->rx_ring);
 975         vfree(priv->tx_ring);
 976
 977         priv->rx_ring = NULL;
 978         priv->tx_ring = NULL;
 979 }
 980
 981 static const struct header_ops ipoib_header_ops = {
 982         .create = ipoib_hard_header,
 983 };
 984
 985 static const struct net_device_ops ipoib_netdev_ops = {
 986         .ndo_open                = ipoib_open,
 987         .ndo_stop                = ipoib_stop,
 988         .ndo_change_mtu          = ipoib_change_mtu,
 989         .ndo_fix_features        = ipoib_fix_features,
 990         .ndo_start_xmit          = ipoib_start_xmit,
 991         .ndo_tx_timeout          = ipoib_timeout,
 992         .ndo_set_rx_mode         = ipoib_set_mcast_list,
 993         .ndo_neigh_setup         = ipoib_neigh_setup_dev,
 994 };
 995
 996 static void ipoib_setup(struct net_device *dev)
 997 {
 998         struct ipoib_dev_priv *priv = netdev_priv(dev);
 999
1000         dev->netdev_ops          = &ipoib_netdev_ops;
1001         dev->header_ops          = &ipoib_header_ops;
1002
1003         ipoib_set_ethtool_ops(dev);
1004
1005         netif_napi_add(dev, &priv->napi, ipoib_poll, 100);
1006
1007         dev->watchdog_timeo      = HZ;
1008
1009         dev->flags              |= IFF_BROADCAST | IFF_MULTICAST;
1010
1011         dev->hard_header_len     = IPOIB_ENCAP_LEN;
1012         dev->addr_len            = INFINIBAND_ALEN;
1013         dev->type                = ARPHRD_INFINIBAND;
1014         dev->tx_queue_len        = ipoib_sendq_size * 2;
1015         dev->features            = (NETIF_F_VLAN_CHALLENGED     |
1016                                     NETIF_F_HIGHDMA);
1017         dev->priv_flags         &= ~IFF_XMIT_DST_RELEASE;
1018
1019         memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
1020
1021         netif_carrier_off(dev);
1022
1023         priv->dev = dev;
1024
1025         spin_lock_init(&priv->lock);
1026
1027         mutex_init(&priv->vlan_mutex);
1028
1029         INIT_LIST_HEAD(&priv->path_list);
1030         INIT_LIST_HEAD(&priv->child_intfs);
1031         INIT_LIST_HEAD(&priv->dead_ahs);
1032         INIT_LIST_HEAD(&priv->multicast_list);
1033
1034         INIT_DELAYED_WORK(&priv->pkey_poll_task, ipoib_pkey_poll);
1035         INIT_DELAYED_WORK(&priv->mcast_task,   ipoib_mcast_join_task);
1036         INIT_WORK(&priv->carrier_on_task, ipoib_mcast_carrier_on_task);
1037         INIT_WORK(&priv->flush_light,   ipoib_ib_dev_flush_light);
1038         INIT_WORK(&priv->flush_normal,   ipoib_ib_dev_flush_normal);
1039         INIT_WORK(&priv->flush_heavy,   ipoib_ib_dev_flush_heavy);
1040         INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
1041         INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
1042 }
1043
1044 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
1045 {
1046         struct net_device *dev;
1047
1048         dev = alloc_netdev((int) sizeof (struct ipoib_dev_priv), name,
1049                            ipoib_setup);
1050         if (!dev)
1051                 return NULL;
1052
1053         return netdev_priv(dev);
1054 }
1055
1056 static ssize_t show_pkey(struct device *dev,
1057                          struct device_attribute *attr, char *buf)
1058 {
1059         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1060
1061         return sprintf(buf, "0x%04x\n", priv->pkey);
1062 }
1063 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1064
1065 static ssize_t show_umcast(struct device *dev,
1066                            struct device_attribute *attr, char *buf)
1067 {
1068         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1069
1070         return sprintf(buf, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
1071 }
1072
1073 static ssize_t set_umcast(struct device *dev,
1074                           struct device_attribute *attr,
1075                           const char *buf, size_t count)
1076 {
1077         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1078         unsigned long umcast_val = simple_strtoul(buf, NULL, 0);
1079
1080         if (umcast_val > 0) {
1081                 set_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1082                 ipoib_warn(priv, "ignoring multicast groups joined directly "
1083                                 "by userspace\n");
1084         } else
1085                 clear_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1086
1087         return count;
1088 }
1089 static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast);
1090
1091 int ipoib_add_umcast_attr(struct net_device *dev)
1092 {
1093         return device_create_file(&dev->dev, &dev_attr_umcast);
1094 }
1095
1096 static ssize_t create_child(struct device *dev,
1097                             struct device_attribute *attr,
1098                             const char *buf, size_t count)
1099 {
1100         int pkey;
1101         int ret;
1102
1103         if (sscanf(buf, "%i", &pkey) != 1)
1104                 return -EINVAL;
1105
1106         if (pkey < 0 || pkey > 0xffff)
1107                 return -EINVAL;
1108
1109         /*
1110          * Set the full membership bit, so that we join the right
1111          * broadcast group, etc.
1112          */
1113         pkey |= 0x8000;
1114
1115         ret = ipoib_vlan_add(to_net_dev(dev), pkey);
1116
1117         return ret ? ret : count;
1118 }
1119 static DEVICE_ATTR(create_child, S_IWUSR, NULL, create_child);
1120
1121 static ssize_t delete_child(struct device *dev,
1122                             struct device_attribute *attr,
1123                             const char *buf, size_t count)
1124 {
1125         int pkey;
1126         int ret;
1127
1128         if (sscanf(buf, "%i", &pkey) != 1)
1129                 return -EINVAL;
1130
1131         if (pkey < 0 || pkey > 0xffff)
1132                 return -EINVAL;
1133
1134         ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
1135
1136         return ret ? ret : count;
1137
1138 }
1139 static DEVICE_ATTR(delete_child, S_IWUSR, NULL, delete_child);
1140
1141 int ipoib_add_pkey_attr(struct net_device *dev)
1142 {
1143         return device_create_file(&dev->dev, &dev_attr_pkey);
1144 }
1145
1146 int ipoib_set_dev_features(struct ipoib_dev_priv *priv, struct ib_device *hca)
1147 {
1148         struct ib_device_attr *device_attr;
1149         int result = -ENOMEM;
1150
1151         device_attr = kmalloc(sizeof *device_attr, GFP_KERNEL);
1152         if (!device_attr) {
1153                 printk(KERN_WARNING "%s: allocation of %zu bytes failed\n",
1154                        hca->name, sizeof *device_attr);
1155                 return result;
1156         }
1157
1158         result = ib_query_device(hca, device_attr);
1159         if (result) {
1160                 printk(KERN_WARNING "%s: ib_query_device failed (ret = %d)\n",
1161                        hca->name, result);
1162                 kfree(device_attr);
1163                 return result;
1164         }
1165         priv->hca_caps = device_attr->device_cap_flags;
1166
1167         kfree(device_attr);
1168
1169         if (priv->hca_caps & IB_DEVICE_UD_IP_CSUM) {
1170                 priv->dev->hw_features = NETIF_F_SG |
1171                         NETIF_F_IP_CSUM | NETIF_F_RXCSUM;
1172
1173                 if (priv->hca_caps & IB_DEVICE_UD_TSO)
1174                         priv->dev->hw_features |= NETIF_F_TSO;
1175
1176                 priv->dev->features |= priv->dev->hw_features;
1177         }
1178
1179         return 0;
1180 }
1181
1182 static struct net_device *ipoib_add_port(const char *format,
1183                                          struct ib_device *hca, u8 port)
1184 {
1185         struct ipoib_dev_priv *priv;
1186         struct ib_port_attr attr;
1187         int result = -ENOMEM;
1188
1189         priv = ipoib_intf_alloc(format);
1190         if (!priv)
1191                 goto alloc_mem_failed;
1192
1193         SET_NETDEV_DEV(priv->dev, hca->dma_device);
1194         priv->dev->dev_id = port - 1;
1195
1196         if (!ib_query_port(hca, port, &attr))
1197                 priv->max_ib_mtu = ib_mtu_enum_to_int(attr.max_mtu);
1198         else {
1199                 printk(KERN_WARNING "%s: ib_query_port %d failed\n",
1200                        hca->name, port);
1201                 goto device_init_failed;
1202         }
1203
1204         /* MTU will be reset when mcast join happens */
1205         priv->dev->mtu  = IPOIB_UD_MTU(priv->max_ib_mtu);
1206         priv->mcast_mtu  = priv->admin_mtu = priv->dev->mtu;
1207
1208         result = ib_query_pkey(hca, port, 0, &priv->pkey);
1209         if (result) {
1210                 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
1211                        hca->name, port, result);
1212                 goto device_init_failed;
1213         }
1214
1215         if (ipoib_set_dev_features(priv, hca))
1216                 goto device_init_failed;
1217
1218         /*
1219          * Set the full membership bit, so that we join the right
1220          * broadcast group, etc.
1221          */
1222         priv->pkey |= 0x8000;
1223
1224         priv->dev->broadcast[8] = priv->pkey >> 8;
1225         priv->dev->broadcast[9] = priv->pkey & 0xff;
1226
1227         result = ib_query_gid(hca, port, 0, &priv->local_gid);
1228         if (result) {
1229                 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1230                        hca->name, port, result);
1231                 goto device_init_failed;
1232         } else
1233                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1234
1235         result = ipoib_dev_init(priv->dev, hca, port);
1236         if (result < 0) {
1237                 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1238                        hca->name, port, result);
1239                 goto device_init_failed;
1240         }
1241
1242         INIT_IB_EVENT_HANDLER(&priv->event_handler,
1243                               priv->ca, ipoib_event);
1244         result = ib_register_event_handler(&priv->event_handler);
1245         if (result < 0) {
1246                 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1247                        "port %d (ret = %d)\n",
1248                        hca->name, port, result);
1249                 goto event_failed;
1250         }
1251
1252         result = register_netdev(priv->dev);
1253         if (result) {
1254                 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1255                        hca->name, port, result);
1256                 goto register_failed;
1257         }
1258
1259         ipoib_create_debug_files(priv->dev);
1260
1261         if (ipoib_cm_add_mode_attr(priv->dev))
1262                 goto sysfs_failed;
1263         if (ipoib_add_pkey_attr(priv->dev))
1264                 goto sysfs_failed;
1265         if (ipoib_add_umcast_attr(priv->dev))
1266                 goto sysfs_failed;
1267         if (device_create_file(&priv->dev->dev, &dev_attr_create_child))
1268                 goto sysfs_failed;
1269         if (device_create_file(&priv->dev->dev, &dev_attr_delete_child))
1270                 goto sysfs_failed;
1271
1272         return priv->dev;
1273
1274 sysfs_failed:
1275         ipoib_delete_debug_files(priv->dev);
1276         unregister_netdev(priv->dev);
1277
1278 register_failed:
1279         ib_unregister_event_handler(&priv->event_handler);
1280         flush_workqueue(ipoib_workqueue);
1281
1282 event_failed:
1283         ipoib_dev_cleanup(priv->dev);
1284
1285 device_init_failed:
1286         free_netdev(priv->dev);
1287
1288 alloc_mem_failed:
1289         return ERR_PTR(result);
1290 }
1291
1292 static void ipoib_add_one(struct ib_device *device)
1293 {
1294         struct list_head *dev_list;
1295         struct net_device *dev;
1296         struct ipoib_dev_priv *priv;
1297         int s, e, p;
1298
1299         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1300                 return;
1301
1302         dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1303         if (!dev_list)
1304                 return;
1305
1306         INIT_LIST_HEAD(dev_list);
1307
1308         if (device->node_type == RDMA_NODE_IB_SWITCH) {
1309                 s = 0;
1310                 e = 0;
1311         } else {
1312                 s = 1;
1313                 e = device->phys_port_cnt;
1314         }
1315
1316         for (p = s; p <= e; ++p) {
1317                 if (rdma_port_get_link_layer(device, p) != IB_LINK_LAYER_INFINIBAND)
1318                         continue;
1319                 dev = ipoib_add_port("ib%d", device, p);
1320                 if (!IS_ERR(dev)) {
1321                         priv = netdev_priv(dev);
1322                         list_add_tail(&priv->list, dev_list);
1323                 }
1324         }
1325
1326         ib_set_client_data(device, &ipoib_client, dev_list);
1327 }
1328
1329 static void ipoib_remove_one(struct ib_device *device)
1330 {
1331         struct ipoib_dev_priv *priv, *tmp;
1332         struct list_head *dev_list;
1333
1334         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1335                 return;
1336
1337         dev_list = ib_get_client_data(device, &ipoib_client);
1338
1339         list_for_each_entry_safe(priv, tmp, dev_list, list) {
1340                 ib_unregister_event_handler(&priv->event_handler);
1341
1342                 rtnl_lock();
1343                 dev_change_flags(priv->dev, priv->dev->flags & ~IFF_UP);
1344                 rtnl_unlock();
1345
1346                 flush_workqueue(ipoib_workqueue);
1347
1348                 unregister_netdev(priv->dev);
1349                 ipoib_dev_cleanup(priv->dev);
1350                 free_netdev(priv->dev);
1351         }
1352
1353         kfree(dev_list);
1354 }
1355
1356 static int __init ipoib_init_module(void)
1357 {
1358         int ret;
1359
1360         ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1361         ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1362         ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1363
1364         ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1365         ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1366         ipoib_sendq_size = max3(ipoib_sendq_size, 2 * MAX_SEND_CQE, IPOIB_MIN_QUEUE_SIZE);
1367 #ifdef CONFIG_INFINIBAND_IPOIB_CM
1368         ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
1369 #endif
1370
1371         /*
1372          * When copying small received packets, we only copy from the
1373          * linear data part of the SKB, so we rely on this condition.
1374          */
1375         BUILD_BUG_ON(IPOIB_CM_COPYBREAK > IPOIB_CM_HEAD_SIZE);
1376
1377         ret = ipoib_register_debugfs();
1378         if (ret)
1379                 return ret;
1380
1381         /*
1382          * We create our own workqueue mainly because we want to be
1383          * able to flush it when devices are being removed.  We can't
1384          * use schedule_work()/flush_scheduled_work() because both
1385          * unregister_netdev() and linkwatch_event take the rtnl lock,
1386          * so flush_scheduled_work() can deadlock during device
1387          * removal.
1388          */
1389         ipoib_workqueue = create_singlethread_workqueue("ipoib");
1390         if (!ipoib_workqueue) {
1391                 ret = -ENOMEM;
1392                 goto err_fs;
1393         }
1394
1395         ib_sa_register_client(&ipoib_sa_client);
1396
1397         ret = ib_register_client(&ipoib_client);
1398         if (ret)
1399                 goto err_sa;
1400
1401         return 0;
1402
1403 err_sa:
1404         ib_sa_unregister_client(&ipoib_sa_client);
1405         destroy_workqueue(ipoib_workqueue);
1406
1407 err_fs:
1408         ipoib_unregister_debugfs();
1409
1410         return ret;
1411 }
1412
1413 static void __exit ipoib_cleanup_module(void)
1414 {
1415         ib_unregister_client(&ipoib_client);
1416         ib_sa_unregister_client(&ipoib_sa_client);
1417         ipoib_unregister_debugfs();
1418         destroy_workqueue(ipoib_workqueue);
1419 }
1420
1421 module_init(ipoib_init_module);
1422 module_exit(ipoib_cleanup_module);