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, 0);
 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 static 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 (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 static void neigh_add_path(struct sk_buff *skb, struct net_device *dev)
 559 {
 560         struct ipoib_dev_priv *priv = netdev_priv(dev);
 561         struct ipoib_path *path;
 562         struct ipoib_neigh *neigh;
 563         unsigned long flags;
 564
 565         neigh = ipoib_neigh_alloc(skb_dst(skb)->neighbour, skb->dev);
 566         if (!neigh) {
 567                 ++dev->stats.tx_dropped;
 568                 dev_kfree_skb_any(skb);
 569                 return;
 570         }
 571
 572         spin_lock_irqsave(&priv->lock, flags);
 573
 574         path = __path_find(dev, skb_dst(skb)->neighbour->ha + 4);
 575         if (!path) {
 576                 path = path_rec_create(dev, skb_dst(skb)->neighbour->ha + 4);
 577                 if (!path)
 578                         goto err_path;
 579
 580                 __path_add(dev, path);
 581         }
 582
 583         list_add_tail(&neigh->list, &path->neigh_list);
 584
 585         if (path->ah) {
 586                 kref_get(&path->ah->ref);
 587                 neigh->ah = path->ah;
 588                 memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
 589                        sizeof(union ib_gid));
 590
 591                 if (ipoib_cm_enabled(dev, neigh->neighbour)) {
 592                         if (!ipoib_cm_get(neigh))
 593                                 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
 594                         if (!ipoib_cm_get(neigh)) {
 595                                 list_del(&neigh->list);
 596                                 if (neigh->ah)
 597                                         ipoib_put_ah(neigh->ah);
 598                                 ipoib_neigh_free(dev, neigh);
 599                                 goto err_drop;
 600                         }
 601                         if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE)
 602                                 __skb_queue_tail(&neigh->queue, skb);
 603                         else {
 604                                 ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
 605                                            skb_queue_len(&neigh->queue));
 606                                 goto err_drop;
 607                         }
 608                 } else {
 609                         spin_unlock_irqrestore(&priv->lock, flags);
 610                         ipoib_send(dev, skb, path->ah, IPOIB_QPN(skb_dst(skb)->neighbour->ha));
 611                         return;
 612                 }
 613         } else {
 614                 neigh->ah  = NULL;
 615
 616                 if (!path->query && path_rec_start(dev, path))
 617                         goto err_list;
 618
 619                 __skb_queue_tail(&neigh->queue, skb);
 620         }
 621
 622         spin_unlock_irqrestore(&priv->lock, flags);
 623         return;
 624
 625 err_list:
 626         list_del(&neigh->list);
 627
 628 err_path:
 629         ipoib_neigh_free(dev, neigh);
 630 err_drop:
 631         ++dev->stats.tx_dropped;
 632         dev_kfree_skb_any(skb);
 633
 634         spin_unlock_irqrestore(&priv->lock, flags);
 635 }
 636
 637 static void ipoib_path_lookup(struct sk_buff *skb, struct net_device *dev)
 638 {
 639         struct ipoib_dev_priv *priv = netdev_priv(skb->dev);
 640
 641         /* Look up path record for unicasts */
 642         if (skb_dst(skb)->neighbour->ha[4] != 0xff) {
 643                 neigh_add_path(skb, dev);
 644                 return;
 645         }
 646
 647         /* Add in the P_Key for multicasts */
 648         skb_dst(skb)->neighbour->ha[8] = (priv->pkey >> 8) & 0xff;
 649         skb_dst(skb)->neighbour->ha[9] = priv->pkey & 0xff;
 650         ipoib_mcast_send(dev, skb_dst(skb)->neighbour->ha + 4, skb);
 651 }
 652
 653 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
 654                              struct ipoib_pseudoheader *phdr)
 655 {
 656         struct ipoib_dev_priv *priv = netdev_priv(dev);
 657         struct ipoib_path *path;
 658         unsigned long flags;
 659
 660         spin_lock_irqsave(&priv->lock, flags);
 661
 662         path = __path_find(dev, phdr->hwaddr + 4);
 663         if (!path || !path->valid) {
 664                 int new_path = 0;
 665
 666                 if (!path) {
 667                         path = path_rec_create(dev, phdr->hwaddr + 4);
 668                         new_path = 1;
 669                 }
 670                 if (path) {
 671                         /* put pseudoheader back on for next time */
 672                         skb_push(skb, sizeof *phdr);
 673                         __skb_queue_tail(&path->queue, skb);
 674
 675                         if (!path->query && path_rec_start(dev, path)) {
 676                                 spin_unlock_irqrestore(&priv->lock, flags);
 677                                 if (new_path)
 678                                         path_free(dev, path);
 679                                 return;
 680                         } else
 681                                 __path_add(dev, path);
 682                 } else {
 683                         ++dev->stats.tx_dropped;
 684                         dev_kfree_skb_any(skb);
 685                 }
 686
 687                 spin_unlock_irqrestore(&priv->lock, flags);
 688                 return;
 689         }
 690
 691         if (path->ah) {
 692                 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
 693                           be16_to_cpu(path->pathrec.dlid));
 694
 695                 spin_unlock_irqrestore(&priv->lock, flags);
 696                 ipoib_send(dev, skb, path->ah, IPOIB_QPN(phdr->hwaddr));
 697                 return;
 698         } else if ((path->query || !path_rec_start(dev, path)) &&
 699                    skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 700                 /* put pseudoheader back on for next time */
 701                 skb_push(skb, sizeof *phdr);
 702                 __skb_queue_tail(&path->queue, skb);
 703         } else {
 704                 ++dev->stats.tx_dropped;
 705                 dev_kfree_skb_any(skb);
 706         }
 707
 708         spin_unlock_irqrestore(&priv->lock, flags);
 709 }
 710
 711 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
 712 {
 713         struct ipoib_dev_priv *priv = netdev_priv(dev);
 714         struct ipoib_neigh *neigh;
 715         unsigned long flags;
 716
 717         if (likely(skb_dst(skb) && skb_dst(skb)->neighbour)) {
 718                 if (unlikely(!*to_ipoib_neigh(skb_dst(skb)->neighbour))) {
 719                         ipoib_path_lookup(skb, dev);
 720                         return NETDEV_TX_OK;
 721                 }
 722
 723                 neigh = *to_ipoib_neigh(skb_dst(skb)->neighbour);
 724
 725                 if (unlikely((memcmp(&neigh->dgid.raw,
 726                                      skb_dst(skb)->neighbour->ha + 4,
 727                                      sizeof(union ib_gid))) ||
 728                              (neigh->dev != dev))) {
 729                         spin_lock_irqsave(&priv->lock, flags);
 730                         /*
 731                          * It's safe to call ipoib_put_ah() inside
 732                          * priv->lock here, because we know that
 733                          * path->ah will always hold one more reference,
 734                          * so ipoib_put_ah() will never do more than
 735                          * decrement the ref count.
 736                          */
 737                         if (neigh->ah)
 738                                 ipoib_put_ah(neigh->ah);
 739                         list_del(&neigh->list);
 740                         ipoib_neigh_free(dev, neigh);
 741                         spin_unlock_irqrestore(&priv->lock, flags);
 742                         ipoib_path_lookup(skb, dev);
 743                         return NETDEV_TX_OK;
 744                 }
 745
 746                 if (ipoib_cm_get(neigh)) {
 747                         if (ipoib_cm_up(neigh)) {
 748                                 ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
 749                                 return NETDEV_TX_OK;
 750                         }
 751                 } else if (neigh->ah) {
 752                         ipoib_send(dev, skb, neigh->ah, IPOIB_QPN(skb_dst(skb)->neighbour->ha));
 753                         return NETDEV_TX_OK;
 754                 }
 755
 756                 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 757                         spin_lock_irqsave(&priv->lock, flags);
 758                         __skb_queue_tail(&neigh->queue, skb);
 759                         spin_unlock_irqrestore(&priv->lock, flags);
 760                 } else {
 761                         ++dev->stats.tx_dropped;
 762                         dev_kfree_skb_any(skb);
 763                 }
 764         } else {
 765                 struct ipoib_pseudoheader *phdr =
 766                         (struct ipoib_pseudoheader *) skb->data;
 767                 skb_pull(skb, sizeof *phdr);
 768
 769                 if (phdr->hwaddr[4] == 0xff) {
 770                         /* Add in the P_Key for multicast*/
 771                         phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
 772                         phdr->hwaddr[9] = priv->pkey & 0xff;
 773
 774                         ipoib_mcast_send(dev, phdr->hwaddr + 4, skb);
 775                 } else {
 776                         /* unicast GID -- should be ARP or RARP reply */
 777
 778                         if ((be16_to_cpup((__be16 *) skb->data) != ETH_P_ARP) &&
 779                             (be16_to_cpup((__be16 *) skb->data) != ETH_P_RARP)) {
 780                                 ipoib_warn(priv, "Unicast, no %s: type %04x, QPN %06x %pI6\n",
 781                                            skb_dst(skb) ? "neigh" : "dst",
 782                                            be16_to_cpup((__be16 *) skb->data),
 783                                            IPOIB_QPN(phdr->hwaddr),
 784                                            phdr->hwaddr + 4);
 785                                 dev_kfree_skb_any(skb);
 786                                 ++dev->stats.tx_dropped;
 787                                 return NETDEV_TX_OK;
 788                         }
 789
 790                         unicast_arp_send(skb, dev, phdr);
 791                 }
 792         }
 793
 794         return NETDEV_TX_OK;
 795 }
 796
 797 static void ipoib_timeout(struct net_device *dev)
 798 {
 799         struct ipoib_dev_priv *priv = netdev_priv(dev);
 800
 801         ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
 802                    jiffies_to_msecs(jiffies - dev->trans_start));
 803         ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
 804                    netif_queue_stopped(dev),
 805                    priv->tx_head, priv->tx_tail);
 806         /* XXX reset QP, etc. */
 807 }
 808
 809 static int ipoib_hard_header(struct sk_buff *skb,
 810                              struct net_device *dev,
 811                              unsigned short type,
 812                              const void *daddr, const void *saddr, unsigned len)
 813 {
 814         struct ipoib_header *header;
 815
 816         header = (struct ipoib_header *) skb_push(skb, sizeof *header);
 817
 818         header->proto = htons(type);
 819         header->reserved = 0;
 820
 821         /*
 822          * If we don't have a neighbour structure, stuff the
 823          * destination address onto the front of the skb so we can
 824          * figure out where to send the packet later.
 825          */
 826         if ((!skb_dst(skb) || !skb_dst(skb)->neighbour) && daddr) {
 827                 struct ipoib_pseudoheader *phdr =
 828                         (struct ipoib_pseudoheader *) skb_push(skb, sizeof *phdr);
 829                 memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
 830         }
 831
 832         return 0;
 833 }
 834
 835 static void ipoib_set_mcast_list(struct net_device *dev)
 836 {
 837         struct ipoib_dev_priv *priv = netdev_priv(dev);
 838
 839         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
 840                 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
 841                 return;
 842         }
 843
 844         queue_work(ipoib_workqueue, &priv->restart_task);
 845 }
 846
 847 static void ipoib_neigh_cleanup(struct neighbour *n)
 848 {
 849         struct ipoib_neigh *neigh;
 850         struct ipoib_dev_priv *priv = netdev_priv(n->dev);
 851         unsigned long flags;
 852         struct ipoib_ah *ah = NULL;
 853
 854         neigh = *to_ipoib_neigh(n);
 855         if (neigh)
 856                 priv = netdev_priv(neigh->dev);
 857         else
 858                 return;
 859         ipoib_dbg(priv,
 860                   "neigh_cleanup for %06x %pI6\n",
 861                   IPOIB_QPN(n->ha),
 862                   n->ha + 4);
 863
 864         spin_lock_irqsave(&priv->lock, flags);
 865
 866         if (neigh->ah)
 867                 ah = neigh->ah;
 868         list_del(&neigh->list);
 869         ipoib_neigh_free(n->dev, neigh);
 870
 871         spin_unlock_irqrestore(&priv->lock, flags);
 872
 873         if (ah)
 874                 ipoib_put_ah(ah);
 875 }
 876
 877 struct ipoib_neigh *ipoib_neigh_alloc(struct neighbour *neighbour,
 878                                       struct net_device *dev)
 879 {
 880         struct ipoib_neigh *neigh;
 881
 882         neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
 883         if (!neigh)
 884                 return NULL;
 885
 886         neigh->neighbour = neighbour;
 887         neigh->dev = dev;
 888         memset(&neigh->dgid.raw, 0, sizeof (union ib_gid));
 889         *to_ipoib_neigh(neighbour) = neigh;
 890         skb_queue_head_init(&neigh->queue);
 891         ipoib_cm_set(neigh, NULL);
 892
 893         return neigh;
 894 }
 895
 896 void ipoib_neigh_free(struct net_device *dev, struct ipoib_neigh *neigh)
 897 {
 898         struct sk_buff *skb;
 899         *to_ipoib_neigh(neigh->neighbour) = NULL;
 900         while ((skb = __skb_dequeue(&neigh->queue))) {
 901                 ++dev->stats.tx_dropped;
 902                 dev_kfree_skb_any(skb);
 903         }
 904         if (ipoib_cm_get(neigh))
 905                 ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
 906         kfree(neigh);
 907 }
 908
 909 static int ipoib_neigh_setup_dev(struct net_device *dev, struct neigh_parms *parms)
 910 {
 911         parms->neigh_cleanup = ipoib_neigh_cleanup;
 912
 913         return 0;
 914 }
 915
 916 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
 917 {
 918         struct ipoib_dev_priv *priv = netdev_priv(dev);
 919
 920         /* Allocate RX/TX "rings" to hold queued skbs */
 921         priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
 922                                 GFP_KERNEL);
 923         if (!priv->rx_ring) {
 924                 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
 925                        ca->name, ipoib_recvq_size);
 926                 goto out;
 927         }
 928
 929         priv->tx_ring = vzalloc(ipoib_sendq_size * sizeof *priv->tx_ring);
 930         if (!priv->tx_ring) {
 931                 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
 932                        ca->name, ipoib_sendq_size);
 933                 goto out_rx_ring_cleanup;
 934         }
 935
 936         /* priv->tx_head, tx_tail & tx_outstanding are already 0 */
 937
 938         if (ipoib_ib_dev_init(dev, ca, port))
 939                 goto out_tx_ring_cleanup;
 940
 941         return 0;
 942
 943 out_tx_ring_cleanup:
 944         vfree(priv->tx_ring);
 945
 946 out_rx_ring_cleanup:
 947         kfree(priv->rx_ring);
 948
 949 out:
 950         return -ENOMEM;
 951 }
 952
 953 void ipoib_dev_cleanup(struct net_device *dev)
 954 {
 955         struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
 956
 957         ipoib_delete_debug_files(dev);
 958
 959         /* Delete any child interfaces first */
 960         list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
 961                 unregister_netdev(cpriv->dev);
 962                 ipoib_dev_cleanup(cpriv->dev);
 963                 free_netdev(cpriv->dev);
 964         }
 965
 966         ipoib_ib_dev_cleanup(dev);
 967
 968         kfree(priv->rx_ring);
 969         vfree(priv->tx_ring);
 970
 971         priv->rx_ring = NULL;
 972         priv->tx_ring = NULL;
 973 }
 974
 975 static const struct header_ops ipoib_header_ops = {
 976         .create = ipoib_hard_header,
 977 };
 978
 979 static const struct net_device_ops ipoib_netdev_ops = {
 980         .ndo_open                = ipoib_open,
 981         .ndo_stop                = ipoib_stop,
 982         .ndo_change_mtu          = ipoib_change_mtu,
 983         .ndo_fix_features        = ipoib_fix_features,
 984         .ndo_start_xmit          = ipoib_start_xmit,
 985         .ndo_tx_timeout          = ipoib_timeout,
 986         .ndo_set_multicast_list  = ipoib_set_mcast_list,
 987         .ndo_neigh_setup         = ipoib_neigh_setup_dev,
 988 };
 989
 990 static void ipoib_setup(struct net_device *dev)
 991 {
 992         struct ipoib_dev_priv *priv = netdev_priv(dev);
 993
 994         dev->netdev_ops          = &ipoib_netdev_ops;
 995         dev->header_ops          = &ipoib_header_ops;
 996
 997         ipoib_set_ethtool_ops(dev);
 998
 999         netif_napi_add(dev, &priv->napi, ipoib_poll, 100);
1000
1001         dev->watchdog_timeo      = HZ;
1002
1003         dev->flags              |= IFF_BROADCAST | IFF_MULTICAST;
1004
1005         /*
1006          * We add in INFINIBAND_ALEN to allow for the destination
1007          * address "pseudoheader" for skbs without neighbour struct.
1008          */
1009         dev->hard_header_len     = IPOIB_ENCAP_LEN + INFINIBAND_ALEN;
1010         dev->addr_len            = INFINIBAND_ALEN;
1011         dev->type                = ARPHRD_INFINIBAND;
1012         dev->tx_queue_len        = ipoib_sendq_size * 2;
1013         dev->features            = (NETIF_F_VLAN_CHALLENGED     |
1014                                     NETIF_F_HIGHDMA);
1015         dev->priv_flags         &= ~IFF_XMIT_DST_RELEASE;
1016
1017         memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
1018
1019         netif_carrier_off(dev);
1020
1021         priv->dev = dev;
1022
1023         spin_lock_init(&priv->lock);
1024
1025         mutex_init(&priv->vlan_mutex);
1026
1027         INIT_LIST_HEAD(&priv->path_list);
1028         INIT_LIST_HEAD(&priv->child_intfs);
1029         INIT_LIST_HEAD(&priv->dead_ahs);
1030         INIT_LIST_HEAD(&priv->multicast_list);
1031
1032         INIT_DELAYED_WORK(&priv->pkey_poll_task, ipoib_pkey_poll);
1033         INIT_DELAYED_WORK(&priv->mcast_task,   ipoib_mcast_join_task);
1034         INIT_WORK(&priv->carrier_on_task, ipoib_mcast_carrier_on_task);
1035         INIT_WORK(&priv->flush_light,   ipoib_ib_dev_flush_light);
1036         INIT_WORK(&priv->flush_normal,   ipoib_ib_dev_flush_normal);
1037         INIT_WORK(&priv->flush_heavy,   ipoib_ib_dev_flush_heavy);
1038         INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
1039         INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
1040 }
1041
1042 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
1043 {
1044         struct net_device *dev;
1045
1046         dev = alloc_netdev((int) sizeof (struct ipoib_dev_priv), name,
1047                            ipoib_setup);
1048         if (!dev)
1049                 return NULL;
1050
1051         return netdev_priv(dev);
1052 }
1053
1054 static ssize_t show_pkey(struct device *dev,
1055                          struct device_attribute *attr, char *buf)
1056 {
1057         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1058
1059         return sprintf(buf, "0x%04x\n", priv->pkey);
1060 }
1061 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1062
1063 static ssize_t show_umcast(struct device *dev,
1064                            struct device_attribute *attr, char *buf)
1065 {
1066         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1067
1068         return sprintf(buf, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
1069 }
1070
1071 static ssize_t set_umcast(struct device *dev,
1072                           struct device_attribute *attr,
1073                           const char *buf, size_t count)
1074 {
1075         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1076         unsigned long umcast_val = simple_strtoul(buf, NULL, 0);
1077
1078         if (umcast_val > 0) {
1079                 set_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1080                 ipoib_warn(priv, "ignoring multicast groups joined directly "
1081                                 "by userspace\n");
1082         } else
1083                 clear_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1084
1085         return count;
1086 }
1087 static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast);
1088
1089 int ipoib_add_umcast_attr(struct net_device *dev)
1090 {
1091         return device_create_file(&dev->dev, &dev_attr_umcast);
1092 }
1093
1094 static ssize_t create_child(struct device *dev,
1095                             struct device_attribute *attr,
1096                             const char *buf, size_t count)
1097 {
1098         int pkey;
1099         int ret;
1100
1101         if (sscanf(buf, "%i", &pkey) != 1)
1102                 return -EINVAL;
1103
1104         if (pkey < 0 || pkey > 0xffff)
1105                 return -EINVAL;
1106
1107         /*
1108          * Set the full membership bit, so that we join the right
1109          * broadcast group, etc.
1110          */
1111         pkey |= 0x8000;
1112
1113         ret = ipoib_vlan_add(to_net_dev(dev), pkey);
1114
1115         return ret ? ret : count;
1116 }
1117 static DEVICE_ATTR(create_child, S_IWUSR, NULL, create_child);
1118
1119 static ssize_t delete_child(struct device *dev,
1120                             struct device_attribute *attr,
1121                             const char *buf, size_t count)
1122 {
1123         int pkey;
1124         int ret;
1125
1126         if (sscanf(buf, "%i", &pkey) != 1)
1127                 return -EINVAL;
1128
1129         if (pkey < 0 || pkey > 0xffff)
1130                 return -EINVAL;
1131
1132         ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
1133
1134         return ret ? ret : count;
1135
1136 }
1137 static DEVICE_ATTR(delete_child, S_IWUSR, NULL, delete_child);
1138
1139 int ipoib_add_pkey_attr(struct net_device *dev)
1140 {
1141         return device_create_file(&dev->dev, &dev_attr_pkey);
1142 }
1143
1144 int ipoib_set_dev_features(struct ipoib_dev_priv *priv, struct ib_device *hca)
1145 {
1146         struct ib_device_attr *device_attr;
1147         int result = -ENOMEM;
1148
1149         device_attr = kmalloc(sizeof *device_attr, GFP_KERNEL);
1150         if (!device_attr) {
1151                 printk(KERN_WARNING "%s: allocation of %zu bytes failed\n",
1152                        hca->name, sizeof *device_attr);
1153                 return result;
1154         }
1155
1156         result = ib_query_device(hca, device_attr);
1157         if (result) {
1158                 printk(KERN_WARNING "%s: ib_query_device failed (ret = %d)\n",
1159                        hca->name, result);
1160                 kfree(device_attr);
1161                 return result;
1162         }
1163         priv->hca_caps = device_attr->device_cap_flags;
1164
1165         kfree(device_attr);
1166
1167         if (priv->hca_caps & IB_DEVICE_UD_IP_CSUM) {
1168                 priv->dev->hw_features = NETIF_F_SG |
1169                         NETIF_F_IP_CSUM | NETIF_F_RXCSUM;
1170
1171                 if (priv->hca_caps & IB_DEVICE_UD_TSO)
1172                         priv->dev->hw_features |= NETIF_F_TSO;
1173
1174                 priv->dev->features |= priv->dev->hw_features;
1175         }
1176
1177         return 0;
1178 }
1179
1180 static struct net_device *ipoib_add_port(const char *format,
1181                                          struct ib_device *hca, u8 port)
1182 {
1183         struct ipoib_dev_priv *priv;
1184         struct ib_port_attr attr;
1185         int result = -ENOMEM;
1186
1187         priv = ipoib_intf_alloc(format);
1188         if (!priv)
1189                 goto alloc_mem_failed;
1190
1191         SET_NETDEV_DEV(priv->dev, hca->dma_device);
1192         priv->dev->dev_id = port - 1;
1193
1194         if (!ib_query_port(hca, port, &attr))
1195                 priv->max_ib_mtu = ib_mtu_enum_to_int(attr.max_mtu);
1196         else {
1197                 printk(KERN_WARNING "%s: ib_query_port %d failed\n",
1198                        hca->name, port);
1199                 goto device_init_failed;
1200         }
1201
1202         /* MTU will be reset when mcast join happens */
1203         priv->dev->mtu  = IPOIB_UD_MTU(priv->max_ib_mtu);
1204         priv->mcast_mtu  = priv->admin_mtu = priv->dev->mtu;
1205
1206         result = ib_query_pkey(hca, port, 0, &priv->pkey);
1207         if (result) {
1208                 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
1209                        hca->name, port, result);
1210                 goto device_init_failed;
1211         }
1212
1213         if (ipoib_set_dev_features(priv, hca))
1214                 goto device_init_failed;
1215
1216         /*
1217          * Set the full membership bit, so that we join the right
1218          * broadcast group, etc.
1219          */
1220         priv->pkey |= 0x8000;
1221
1222         priv->dev->broadcast[8] = priv->pkey >> 8;
1223         priv->dev->broadcast[9] = priv->pkey & 0xff;
1224
1225         result = ib_query_gid(hca, port, 0, &priv->local_gid);
1226         if (result) {
1227                 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1228                        hca->name, port, result);
1229                 goto device_init_failed;
1230         } else
1231                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1232
1233         result = ipoib_dev_init(priv->dev, hca, port);
1234         if (result < 0) {
1235                 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1236                        hca->name, port, result);
1237                 goto device_init_failed;
1238         }
1239
1240         INIT_IB_EVENT_HANDLER(&priv->event_handler,
1241                               priv->ca, ipoib_event);
1242         result = ib_register_event_handler(&priv->event_handler);
1243         if (result < 0) {
1244                 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1245                        "port %d (ret = %d)\n",
1246                        hca->name, port, result);
1247                 goto event_failed;
1248         }
1249
1250         result = register_netdev(priv->dev);
1251         if (result) {
1252                 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1253                        hca->name, port, result);
1254                 goto register_failed;
1255         }
1256
1257         ipoib_create_debug_files(priv->dev);
1258
1259         if (ipoib_cm_add_mode_attr(priv->dev))
1260                 goto sysfs_failed;
1261         if (ipoib_add_pkey_attr(priv->dev))
1262                 goto sysfs_failed;
1263         if (ipoib_add_umcast_attr(priv->dev))
1264                 goto sysfs_failed;
1265         if (device_create_file(&priv->dev->dev, &dev_attr_create_child))
1266                 goto sysfs_failed;
1267         if (device_create_file(&priv->dev->dev, &dev_attr_delete_child))
1268                 goto sysfs_failed;
1269
1270         return priv->dev;
1271
1272 sysfs_failed:
1273         ipoib_delete_debug_files(priv->dev);
1274         unregister_netdev(priv->dev);
1275
1276 register_failed:
1277         ib_unregister_event_handler(&priv->event_handler);
1278         flush_workqueue(ipoib_workqueue);
1279
1280 event_failed:
1281         ipoib_dev_cleanup(priv->dev);
1282
1283 device_init_failed:
1284         free_netdev(priv->dev);
1285
1286 alloc_mem_failed:
1287         return ERR_PTR(result);
1288 }
1289
1290 static void ipoib_add_one(struct ib_device *device)
1291 {
1292         struct list_head *dev_list;
1293         struct net_device *dev;
1294         struct ipoib_dev_priv *priv;
1295         int s, e, p;
1296
1297         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1298                 return;
1299
1300         dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1301         if (!dev_list)
1302                 return;
1303
1304         INIT_LIST_HEAD(dev_list);
1305
1306         if (device->node_type == RDMA_NODE_IB_SWITCH) {
1307                 s = 0;
1308                 e = 0;
1309         } else {
1310                 s = 1;
1311                 e = device->phys_port_cnt;
1312         }
1313
1314         for (p = s; p <= e; ++p) {
1315                 if (rdma_port_get_link_layer(device, p) != IB_LINK_LAYER_INFINIBAND)
1316                         continue;
1317                 dev = ipoib_add_port("ib%d", device, p);
1318                 if (!IS_ERR(dev)) {
1319                         priv = netdev_priv(dev);
1320                         list_add_tail(&priv->list, dev_list);
1321                 }
1322         }
1323
1324         ib_set_client_data(device, &ipoib_client, dev_list);
1325 }
1326
1327 static void ipoib_remove_one(struct ib_device *device)
1328 {
1329         struct ipoib_dev_priv *priv, *tmp;
1330         struct list_head *dev_list;
1331
1332         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1333                 return;
1334
1335         dev_list = ib_get_client_data(device, &ipoib_client);
1336
1337         list_for_each_entry_safe(priv, tmp, dev_list, list) {
1338                 ib_unregister_event_handler(&priv->event_handler);
1339
1340                 rtnl_lock();
1341                 dev_change_flags(priv->dev, priv->dev->flags & ~IFF_UP);
1342                 rtnl_unlock();
1343
1344                 flush_workqueue(ipoib_workqueue);
1345
1346                 unregister_netdev(priv->dev);
1347                 ipoib_dev_cleanup(priv->dev);
1348                 free_netdev(priv->dev);
1349         }
1350
1351         kfree(dev_list);
1352 }
1353
1354 static int __init ipoib_init_module(void)
1355 {
1356         int ret;
1357
1358         ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1359         ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1360         ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1361
1362         ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1363         ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1364         ipoib_sendq_size = max3(ipoib_sendq_size, 2 * MAX_SEND_CQE, IPOIB_MIN_QUEUE_SIZE);
1365 #ifdef CONFIG_INFINIBAND_IPOIB_CM
1366         ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
1367 #endif
1368
1369         /*
1370          * When copying small received packets, we only copy from the
1371          * linear data part of the SKB, so we rely on this condition.
1372          */
1373         BUILD_BUG_ON(IPOIB_CM_COPYBREAK > IPOIB_CM_HEAD_SIZE);
1374
1375         ret = ipoib_register_debugfs();
1376         if (ret)
1377                 return ret;
1378
1379         /*
1380          * We create our own workqueue mainly because we want to be
1381          * able to flush it when devices are being removed.  We can't
1382          * use schedule_work()/flush_scheduled_work() because both
1383          * unregister_netdev() and linkwatch_event take the rtnl lock,
1384          * so flush_scheduled_work() can deadlock during device
1385          * removal.
1386          */
1387         ipoib_workqueue = create_singlethread_workqueue("ipoib");
1388         if (!ipoib_workqueue) {
1389                 ret = -ENOMEM;
1390                 goto err_fs;
1391         }
1392
1393         ib_sa_register_client(&ipoib_sa_client);
1394
1395         ret = ib_register_client(&ipoib_client);
1396         if (ret)
1397                 goto err_sa;
1398
1399         return 0;
1400
1401 err_sa:
1402         ib_sa_unregister_client(&ipoib_sa_client);
1403         destroy_workqueue(ipoib_workqueue);
1404
1405 err_fs:
1406         ipoib_unregister_debugfs();
1407
1408         return ret;
1409 }
1410
1411 static void __exit ipoib_cleanup_module(void)
1412 {
1413         ib_unregister_client(&ipoib_client);
1414         ib_sa_unregister_client(&ipoib_sa_client);
1415         ipoib_unregister_debugfs();
1416         destroy_workqueue(ipoib_workqueue);
1417 }
1418
1419 module_init(ipoib_init_module);
1420 module_exit(ipoib_cleanup_module);