drivers/infiniband/ulp/ipoib/ipoib_main.c

   1 /*
   2  * Copyright (c) 2004 Topspin Communications.  All rights reserved.
   3  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
   4  * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
   5  *
   6  * This software is available to you under a choice of one of two
   7  * licenses.  You may choose to be licensed under the terms of the GNU
   8  * General Public License (GPL) Version 2, available from the file
   9  * COPYING in the main directory of this source tree, or the
  10  * OpenIB.org BSD license below:
  11  *
  12  *     Redistribution and use in source and binary forms, with or
  13  *     without modification, are permitted provided that the following
  14  *     conditions are met:
  15  *
  16  *      - Redistributions of source code must retain the above
  17  *        copyright notice, this list of conditions and the following
  18  *        disclaimer.
  19  *
  20  *      - Redistributions in binary form must reproduce the above
  21  *        copyright notice, this list of conditions and the following
  22  *        disclaimer in the documentation and/or other materials
  23  *        provided with the distribution.
  24  *
  25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  32  * SOFTWARE.
  33  */
  34
  35 #include "ipoib.h"
  36
  37 #include <linux/module.h>
  38
  39 #include <linux/init.h>
  40 #include <linux/slab.h>
  41 #include <linux/kernel.h>
  42 #include <linux/vmalloc.h>
  43
  44 #include <linux/if_arp.h>       /* For ARPHRD_xxx */
  45
  46 #include <linux/ip.h>
  47 #include <linux/in.h>
  48
  49 #include <net/dst.h>
  50
  51 MODULE_AUTHOR("Roland Dreier");
  52 MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
  53 MODULE_LICENSE("Dual BSD/GPL");
  54
  55 int ipoib_sendq_size __read_mostly = IPOIB_TX_RING_SIZE;
  56 int ipoib_recvq_size __read_mostly = IPOIB_RX_RING_SIZE;
  57
  58 module_param_named(send_queue_size, ipoib_sendq_size, int, 0444);
  59 MODULE_PARM_DESC(send_queue_size, "Number of descriptors in send queue");
  60 module_param_named(recv_queue_size, ipoib_recvq_size, int, 0444);
  61 MODULE_PARM_DESC(recv_queue_size, "Number of descriptors in receive queue");
  62
  63 static int lro;
  64 module_param(lro, bool, 0444);
  65 MODULE_PARM_DESC(lro,  "Enable LRO (Large Receive Offload)");
  66
  67 static int lro_max_aggr = IPOIB_LRO_MAX_AGGR;
  68 module_param(lro_max_aggr, int, 0644);
  69 MODULE_PARM_DESC(lro_max_aggr, "LRO: Max packets to be aggregated "
  70                 "(default = 64)");
  71
  72 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
  73 int ipoib_debug_level;
  74
  75 module_param_named(debug_level, ipoib_debug_level, int, 0644);
  76 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
  77 #endif
  78
  79 struct ipoib_path_iter {
  80         struct net_device *dev;
  81         struct ipoib_path  path;
  82 };
  83
  84 static const u8 ipv4_bcast_addr[] = {
  85         0x00, 0xff, 0xff, 0xff,
  86         0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
  87         0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff
  88 };
  89
  90 struct workqueue_struct *ipoib_workqueue;
  91
  92 struct ib_sa_client ipoib_sa_client;
  93
  94 static void ipoib_add_one(struct ib_device *device);
  95 static void ipoib_remove_one(struct ib_device *device);
  96
  97 static struct ib_client ipoib_client = {
  98         .name   = "ipoib",
  99         .add    = ipoib_add_one,
 100         .remove = ipoib_remove_one
 101 };
 102
 103 int ipoib_open(struct net_device *dev)
 104 {
 105         struct ipoib_dev_priv *priv = netdev_priv(dev);
 106
 107         ipoib_dbg(priv, "bringing up interface\n");
 108
 109         set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 110
 111         if (ipoib_pkey_dev_delay_open(dev))
 112                 return 0;
 113
 114         if (ipoib_ib_dev_open(dev))
 115                 goto err_disable;
 116
 117         if (ipoib_ib_dev_up(dev))
 118                 goto err_stop;
 119
 120         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 121                 struct ipoib_dev_priv *cpriv;
 122
 123                 /* Bring up any child interfaces too */
 124                 mutex_lock(&priv->vlan_mutex);
 125                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 126                         int flags;
 127
 128                         flags = cpriv->dev->flags;
 129                         if (flags & IFF_UP)
 130                                 continue;
 131
 132                         dev_change_flags(cpriv->dev, flags | IFF_UP);
 133                 }
 134                 mutex_unlock(&priv->vlan_mutex);
 135         }
 136
 137         netif_start_queue(dev);
 138
 139         return 0;
 140
 141 err_stop:
 142         ipoib_ib_dev_stop(dev, 1);
 143
 144 err_disable:
 145         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 146
 147         return -EINVAL;
 148 }
 149
 150 static int ipoib_stop(struct net_device *dev)
 151 {
 152         struct ipoib_dev_priv *priv = netdev_priv(dev);
 153
 154         ipoib_dbg(priv, "stopping interface\n");
 155
 156         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 157
 158         netif_stop_queue(dev);
 159
 160         ipoib_ib_dev_down(dev, 0);
 161         ipoib_ib_dev_stop(dev, 0);
 162
 163         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 164                 struct ipoib_dev_priv *cpriv;
 165
 166                 /* Bring down any child interfaces too */
 167                 mutex_lock(&priv->vlan_mutex);
 168                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 169                         int flags;
 170
 171                         flags = cpriv->dev->flags;
 172                         if (!(flags & IFF_UP))
 173                                 continue;
 174
 175                         dev_change_flags(cpriv->dev, flags & ~IFF_UP);
 176                 }
 177                 mutex_unlock(&priv->vlan_mutex);
 178         }
 179
 180         return 0;
 181 }
 182
 183 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
 184 {
 185         struct ipoib_dev_priv *priv = netdev_priv(dev);
 186
 187         /* dev->mtu > 2K ==> connected mode */
 188         if (ipoib_cm_admin_enabled(dev)) {
 189                 if (new_mtu > ipoib_cm_max_mtu(dev))
 190                         return -EINVAL;
 191
 192                 if (new_mtu > priv->mcast_mtu)
 193                         ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
 194                                    priv->mcast_mtu);
 195
 196                 dev->mtu = new_mtu;
 197                 return 0;
 198         }
 199
 200         if (new_mtu > IPOIB_UD_MTU(priv->max_ib_mtu))
 201                 return -EINVAL;
 202
 203         priv->admin_mtu = new_mtu;
 204
 205         dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
 206
 207         return 0;
 208 }
 209
 210 static struct ipoib_path *__path_find(struct net_device *dev, void *gid)
 211 {
 212         struct ipoib_dev_priv *priv = netdev_priv(dev);
 213         struct rb_node *n = priv->path_tree.rb_node;
 214         struct ipoib_path *path;
 215         int ret;
 216
 217         while (n) {
 218                 path = rb_entry(n, struct ipoib_path, rb_node);
 219
 220                 ret = memcmp(gid, path->pathrec.dgid.raw,
 221                              sizeof (union ib_gid));
 222
 223                 if (ret < 0)
 224                         n = n->rb_left;
 225                 else if (ret > 0)
 226                         n = n->rb_right;
 227                 else
 228                         return path;
 229         }
 230
 231         return NULL;
 232 }
 233
 234 static int __path_add(struct net_device *dev, struct ipoib_path *path)
 235 {
 236         struct ipoib_dev_priv *priv = netdev_priv(dev);
 237         struct rb_node **n = &priv->path_tree.rb_node;
 238         struct rb_node *pn = NULL;
 239         struct ipoib_path *tpath;
 240         int ret;
 241
 242         while (*n) {
 243                 pn = *n;
 244                 tpath = rb_entry(pn, struct ipoib_path, rb_node);
 245
 246                 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
 247                              sizeof (union ib_gid));
 248                 if (ret < 0)
 249                         n = &pn->rb_left;
 250                 else if (ret > 0)
 251                         n = &pn->rb_right;
 252                 else
 253                         return -EEXIST;
 254         }
 255
 256         rb_link_node(&path->rb_node, pn, n);
 257         rb_insert_color(&path->rb_node, &priv->path_tree);
 258
 259         list_add_tail(&path->list, &priv->path_list);
 260
 261         return 0;
 262 }
 263
 264 static void path_free(struct net_device *dev, struct ipoib_path *path)
 265 {
 266         struct ipoib_dev_priv *priv = netdev_priv(dev);
 267         struct ipoib_neigh *neigh, *tn;
 268         struct sk_buff *skb;
 269         unsigned long flags;
 270
 271         while ((skb = __skb_dequeue(&path->queue)))
 272                 dev_kfree_skb_irq(skb);
 273
 274         spin_lock_irqsave(&priv->lock, flags);
 275
 276         list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
 277                 /*
 278                  * It's safe to call ipoib_put_ah() inside priv->lock
 279                  * here, because we know that path->ah will always
 280                  * hold one more reference, so ipoib_put_ah() will
 281                  * never do more than decrement the ref count.
 282                  */
 283                 if (neigh->ah)
 284                         ipoib_put_ah(neigh->ah);
 285
 286                 ipoib_neigh_free(dev, neigh);
 287         }
 288
 289         spin_unlock_irqrestore(&priv->lock, flags);
 290
 291         if (path->ah)
 292                 ipoib_put_ah(path->ah);
 293
 294         kfree(path);
 295 }
 296
 297 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 298
 299 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
 300 {
 301         struct ipoib_path_iter *iter;
 302
 303         iter = kmalloc(sizeof *iter, GFP_KERNEL);
 304         if (!iter)
 305                 return NULL;
 306
 307         iter->dev = dev;
 308         memset(iter->path.pathrec.dgid.raw, 0, 16);
 309
 310         if (ipoib_path_iter_next(iter)) {
 311                 kfree(iter);
 312                 return NULL;
 313         }
 314
 315         return iter;
 316 }
 317
 318 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
 319 {
 320         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
 321         struct rb_node *n;
 322         struct ipoib_path *path;
 323         int ret = 1;
 324
 325         spin_lock_irq(&priv->lock);
 326
 327         n = rb_first(&priv->path_tree);
 328
 329         while (n) {
 330                 path = rb_entry(n, struct ipoib_path, rb_node);
 331
 332                 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
 333                            sizeof (union ib_gid)) < 0) {
 334                         iter->path = *path;
 335                         ret = 0;
 336                         break;
 337                 }
 338
 339                 n = rb_next(n);
 340         }
 341
 342         spin_unlock_irq(&priv->lock);
 343
 344         return ret;
 345 }
 346
 347 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
 348                           struct ipoib_path *path)
 349 {
 350         *path = iter->path;
 351 }
 352
 353 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
 354
 355 void ipoib_mark_paths_invalid(struct net_device *dev)
 356 {
 357         struct ipoib_dev_priv *priv = netdev_priv(dev);
 358         struct ipoib_path *path, *tp;
 359
 360         spin_lock_irq(&priv->lock);
 361
 362         list_for_each_entry_safe(path, tp, &priv->path_list, list) {
 363                 ipoib_dbg(priv, "mark path LID 0x%04x GID %pI6 invalid\n",
 364                         be16_to_cpu(path->pathrec.dlid),
 365                         path->pathrec.dgid.raw);
 366                 path->valid =  0;
 367         }
 368
 369         spin_unlock_irq(&priv->lock);
 370 }
 371
 372 void ipoib_flush_paths(struct net_device *dev)
 373 {
 374         struct ipoib_dev_priv *priv = netdev_priv(dev);
 375         struct ipoib_path *path, *tp;
 376         LIST_HEAD(remove_list);
 377         unsigned long flags;
 378
 379         netif_tx_lock_bh(dev);
 380         spin_lock_irqsave(&priv->lock, flags);
 381
 382         list_splice_init(&priv->path_list, &remove_list);
 383
 384         list_for_each_entry(path, &remove_list, list)
 385                 rb_erase(&path->rb_node, &priv->path_tree);
 386
 387         list_for_each_entry_safe(path, tp, &remove_list, list) {
 388                 if (path->query)
 389                         ib_sa_cancel_query(path->query_id, path->query);
 390                 spin_unlock_irqrestore(&priv->lock, flags);
 391                 netif_tx_unlock_bh(dev);
 392                 wait_for_completion(&path->done);
 393                 path_free(dev, path);
 394                 netif_tx_lock_bh(dev);
 395                 spin_lock_irqsave(&priv->lock, flags);
 396         }
 397
 398         spin_unlock_irqrestore(&priv->lock, flags);
 399         netif_tx_unlock_bh(dev);
 400 }
 401
 402 static void path_rec_completion(int status,
 403                                 struct ib_sa_path_rec *pathrec,
 404                                 void *path_ptr)
 405 {
 406         struct ipoib_path *path = path_ptr;
 407         struct net_device *dev = path->dev;
 408         struct ipoib_dev_priv *priv = netdev_priv(dev);
 409         struct ipoib_ah *ah = NULL;
 410         struct ipoib_ah *old_ah = NULL;
 411         struct ipoib_neigh *neigh, *tn;
 412         struct sk_buff_head skqueue;
 413         struct sk_buff *skb;
 414         unsigned long flags;
 415
 416         if (!status)
 417                 ipoib_dbg(priv, "PathRec LID 0x%04x for GID %pI6\n",
 418                           be16_to_cpu(pathrec->dlid), pathrec->dgid.raw);
 419         else
 420                 ipoib_dbg(priv, "PathRec status %d for GID %pI6\n",
 421                           status, path->pathrec.dgid.raw);
 422
 423         skb_queue_head_init(&skqueue);
 424
 425         if (!status) {
 426                 struct ib_ah_attr av;
 427
 428                 if (!ib_init_ah_from_path(priv->ca, priv->port, pathrec, &av))
 429                         ah = ipoib_create_ah(dev, priv->pd, &av);
 430         }
 431
 432         spin_lock_irqsave(&priv->lock, flags);
 433
 434         if (ah) {
 435                 path->pathrec = *pathrec;
 436
 437                 old_ah   = path->ah;
 438                 path->ah = ah;
 439
 440                 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
 441                           ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
 442
 443                 while ((skb = __skb_dequeue(&path->queue)))
 444                         __skb_queue_tail(&skqueue, skb);
 445
 446                 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
 447                         if (neigh->ah) {
 448                                 WARN_ON(neigh->ah != old_ah);
 449                                 /*
 450                                  * Dropping the ah reference inside
 451                                  * priv->lock is safe here, because we
 452                                  * will hold one more reference from
 453                                  * the original value of path->ah (ie
 454                                  * old_ah).
 455                                  */
 456                                 ipoib_put_ah(neigh->ah);
 457                         }
 458                         kref_get(&path->ah->ref);
 459                         neigh->ah = path->ah;
 460                         memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
 461                                sizeof(union ib_gid));
 462
 463                         if (ipoib_cm_enabled(dev, neigh->neighbour)) {
 464                                 if (!ipoib_cm_get(neigh))
 465                                         ipoib_cm_set(neigh, ipoib_cm_create_tx(dev,
 466                                                                                path,
 467                                                                                neigh));
 468                                 if (!ipoib_cm_get(neigh)) {
 469                                         list_del(&neigh->list);
 470                                         if (neigh->ah)
 471                                                 ipoib_put_ah(neigh->ah);
 472                                         ipoib_neigh_free(dev, neigh);
 473                                         continue;
 474                                 }
 475                         }
 476
 477                         while ((skb = __skb_dequeue(&neigh->queue)))
 478                                 __skb_queue_tail(&skqueue, skb);
 479                 }
 480                 path->valid = 1;
 481         }
 482
 483         path->query = NULL;
 484         complete(&path->done);
 485
 486         spin_unlock_irqrestore(&priv->lock, flags);
 487
 488         if (old_ah)
 489                 ipoib_put_ah(old_ah);
 490
 491         while ((skb = __skb_dequeue(&skqueue))) {
 492                 skb->dev = dev;
 493                 if (dev_queue_xmit(skb))
 494                         ipoib_warn(priv, "dev_queue_xmit failed "
 495                                    "to requeue packet\n");
 496         }
 497 }
 498
 499 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
 500 {
 501         struct ipoib_dev_priv *priv = netdev_priv(dev);
 502         struct ipoib_path *path;
 503
 504         if (!priv->broadcast)
 505                 return NULL;
 506
 507         path = kzalloc(sizeof *path, GFP_ATOMIC);
 508         if (!path)
 509                 return NULL;
 510
 511         path->dev = dev;
 512
 513         skb_queue_head_init(&path->queue);
 514
 515         INIT_LIST_HEAD(&path->neigh_list);
 516
 517         memcpy(path->pathrec.dgid.raw, gid, sizeof (union ib_gid));
 518         path->pathrec.sgid          = priv->local_gid;
 519         path->pathrec.pkey          = cpu_to_be16(priv->pkey);
 520         path->pathrec.numb_path     = 1;
 521         path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
 522
 523         return path;
 524 }
 525
 526 static int path_rec_start(struct net_device *dev,
 527                           struct ipoib_path *path)
 528 {
 529         struct ipoib_dev_priv *priv = netdev_priv(dev);
 530
 531         ipoib_dbg(priv, "Start path record lookup for %pI6\n",
 532                   path->pathrec.dgid.raw);
 533
 534         init_completion(&path->done);
 535
 536         path->query_id =
 537                 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
 538                                    &path->pathrec,
 539                                    IB_SA_PATH_REC_DGID          |
 540                                    IB_SA_PATH_REC_SGID          |
 541                                    IB_SA_PATH_REC_NUMB_PATH     |
 542                                    IB_SA_PATH_REC_TRAFFIC_CLASS |
 543                                    IB_SA_PATH_REC_PKEY,
 544                                    1000, GFP_ATOMIC,
 545                                    path_rec_completion,
 546                                    path, &path->query);
 547         if (path->query_id < 0) {
 548                 ipoib_warn(priv, "ib_sa_path_rec_get failed: %d\n", path->query_id);
 549                 path->query = NULL;
 550                 complete(&path->done);
 551                 return path->query_id;
 552         }
 553
 554         return 0;
 555 }
 556
 557 static void neigh_add_path(struct sk_buff *skb, struct net_device *dev)
 558 {
 559         struct ipoib_dev_priv *priv = netdev_priv(dev);
 560         struct ipoib_path *path;
 561         struct ipoib_neigh *neigh;
 562         unsigned long flags;
 563
 564         neigh = ipoib_neigh_alloc(skb_dst(skb)->neighbour, skb->dev);
 565         if (!neigh) {
 566                 ++dev->stats.tx_dropped;
 567                 dev_kfree_skb_any(skb);
 568                 return;
 569         }
 570
 571         spin_lock_irqsave(&priv->lock, flags);
 572
 573         path = __path_find(dev, skb_dst(skb)->neighbour->ha + 4);
 574         if (!path) {
 575                 path = path_rec_create(dev, skb_dst(skb)->neighbour->ha + 4);
 576                 if (!path)
 577                         goto err_path;
 578
 579                 __path_add(dev, path);
 580         }
 581
 582         list_add_tail(&neigh->list, &path->neigh_list);
 583
 584         if (path->ah) {
 585                 kref_get(&path->ah->ref);
 586                 neigh->ah = path->ah;
 587                 memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
 588                        sizeof(union ib_gid));
 589
 590                 if (ipoib_cm_enabled(dev, neigh->neighbour)) {
 591                         if (!ipoib_cm_get(neigh))
 592                                 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
 593                         if (!ipoib_cm_get(neigh)) {
 594                                 list_del(&neigh->list);
 595                                 if (neigh->ah)
 596                                         ipoib_put_ah(neigh->ah);
 597                                 ipoib_neigh_free(dev, neigh);
 598                                 goto err_drop;
 599                         }
 600                         if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE)
 601                                 __skb_queue_tail(&neigh->queue, skb);
 602                         else {
 603                                 ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
 604                                            skb_queue_len(&neigh->queue));
 605                                 goto err_drop;
 606                         }
 607                 } else {
 608                         spin_unlock_irqrestore(&priv->lock, flags);
 609                         ipoib_send(dev, skb, path->ah, IPOIB_QPN(skb_dst(skb)->neighbour->ha));
 610                         return;
 611                 }
 612         } else {
 613                 neigh->ah  = NULL;
 614
 615                 if (!path->query && path_rec_start(dev, path))
 616                         goto err_list;
 617
 618                 __skb_queue_tail(&neigh->queue, skb);
 619         }
 620
 621         spin_unlock_irqrestore(&priv->lock, flags);
 622         return;
 623
 624 err_list:
 625         list_del(&neigh->list);
 626
 627 err_path:
 628         ipoib_neigh_free(dev, neigh);
 629 err_drop:
 630         ++dev->stats.tx_dropped;
 631         dev_kfree_skb_any(skb);
 632
 633         spin_unlock_irqrestore(&priv->lock, flags);
 634 }
 635
 636 static void ipoib_path_lookup(struct sk_buff *skb, struct net_device *dev)
 637 {
 638         struct ipoib_dev_priv *priv = netdev_priv(skb->dev);
 639
 640         /* Look up path record for unicasts */
 641         if (skb_dst(skb)->neighbour->ha[4] != 0xff) {
 642                 neigh_add_path(skb, dev);
 643                 return;
 644         }
 645
 646         /* Add in the P_Key for multicasts */
 647         skb_dst(skb)->neighbour->ha[8] = (priv->pkey >> 8) & 0xff;
 648         skb_dst(skb)->neighbour->ha[9] = priv->pkey & 0xff;
 649         ipoib_mcast_send(dev, skb_dst(skb)->neighbour->ha + 4, skb);
 650 }
 651
 652 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
 653                              struct ipoib_pseudoheader *phdr)
 654 {
 655         struct ipoib_dev_priv *priv = netdev_priv(dev);
 656         struct ipoib_path *path;
 657         unsigned long flags;
 658
 659         spin_lock_irqsave(&priv->lock, flags);
 660
 661         path = __path_find(dev, phdr->hwaddr + 4);
 662         if (!path || !path->valid) {
 663                 int new_path = 0;
 664
 665                 if (!path) {
 666                         path = path_rec_create(dev, phdr->hwaddr + 4);
 667                         new_path = 1;
 668                 }
 669                 if (path) {
 670                         /* put pseudoheader back on for next time */
 671                         skb_push(skb, sizeof *phdr);
 672                         __skb_queue_tail(&path->queue, skb);
 673
 674                         if (!path->query && path_rec_start(dev, path)) {
 675                                 spin_unlock_irqrestore(&priv->lock, flags);
 676                                 if (new_path)
 677                                         path_free(dev, path);
 678                                 return;
 679                         } else
 680                                 __path_add(dev, path);
 681                 } else {
 682                         ++dev->stats.tx_dropped;
 683                         dev_kfree_skb_any(skb);
 684                 }
 685
 686                 spin_unlock_irqrestore(&priv->lock, flags);
 687                 return;
 688         }
 689
 690         if (path->ah) {
 691                 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
 692                           be16_to_cpu(path->pathrec.dlid));
 693
 694                 spin_unlock_irqrestore(&priv->lock, flags);
 695                 ipoib_send(dev, skb, path->ah, IPOIB_QPN(phdr->hwaddr));
 696                 return;
 697         } else if ((path->query || !path_rec_start(dev, path)) &&
 698                    skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 699                 /* put pseudoheader back on for next time */
 700                 skb_push(skb, sizeof *phdr);
 701                 __skb_queue_tail(&path->queue, skb);
 702         } else {
 703                 ++dev->stats.tx_dropped;
 704                 dev_kfree_skb_any(skb);
 705         }
 706
 707         spin_unlock_irqrestore(&priv->lock, flags);
 708 }
 709
 710 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
 711 {
 712         struct ipoib_dev_priv *priv = netdev_priv(dev);
 713         struct ipoib_neigh *neigh;
 714         unsigned long flags;
 715
 716         if (likely(skb_dst(skb) && skb_dst(skb)->neighbour)) {
 717                 if (unlikely(!*to_ipoib_neigh(skb_dst(skb)->neighbour))) {
 718                         ipoib_path_lookup(skb, dev);
 719                         return NETDEV_TX_OK;
 720                 }
 721
 722                 neigh = *to_ipoib_neigh(skb_dst(skb)->neighbour);
 723
 724                 if (unlikely((memcmp(&neigh->dgid.raw,
 725                                      skb_dst(skb)->neighbour->ha + 4,
 726                                      sizeof(union ib_gid))) ||
 727                              (neigh->dev != dev))) {
 728                         spin_lock_irqsave(&priv->lock, flags);
 729                         /*
 730                          * It's safe to call ipoib_put_ah() inside
 731                          * priv->lock here, because we know that
 732                          * path->ah will always hold one more reference,
 733                          * so ipoib_put_ah() will never do more than
 734                          * decrement the ref count.
 735                          */
 736                         if (neigh->ah)
 737                                 ipoib_put_ah(neigh->ah);
 738                         list_del(&neigh->list);
 739                         ipoib_neigh_free(dev, neigh);
 740                         spin_unlock_irqrestore(&priv->lock, flags);
 741                         ipoib_path_lookup(skb, dev);
 742                         return NETDEV_TX_OK;
 743                 }
 744
 745                 if (ipoib_cm_get(neigh)) {
 746                         if (ipoib_cm_up(neigh)) {
 747                                 ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
 748                                 return NETDEV_TX_OK;
 749                         }
 750                 } else if (neigh->ah) {
 751                         ipoib_send(dev, skb, neigh->ah, IPOIB_QPN(skb_dst(skb)->neighbour->ha));
 752                         return NETDEV_TX_OK;
 753                 }
 754
 755                 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 756                         spin_lock_irqsave(&priv->lock, flags);
 757                         __skb_queue_tail(&neigh->queue, skb);
 758                         spin_unlock_irqrestore(&priv->lock, flags);
 759                 } else {
 760                         ++dev->stats.tx_dropped;
 761                         dev_kfree_skb_any(skb);
 762                 }
 763         } else {
 764                 struct ipoib_pseudoheader *phdr =
 765                         (struct ipoib_pseudoheader *) skb->data;
 766                 skb_pull(skb, sizeof *phdr);
 767
 768                 if (phdr->hwaddr[4] == 0xff) {
 769                         /* Add in the P_Key for multicast*/
 770                         phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
 771                         phdr->hwaddr[9] = priv->pkey & 0xff;
 772
 773                         ipoib_mcast_send(dev, phdr->hwaddr + 4, skb);
 774                 } else {
 775                         /* unicast GID -- should be ARP or RARP reply */
 776
 777                         if ((be16_to_cpup((__be16 *) skb->data) != ETH_P_ARP) &&
 778                             (be16_to_cpup((__be16 *) skb->data) != ETH_P_RARP)) {
 779                                 ipoib_warn(priv, "Unicast, no %s: type %04x, QPN %06x %pI6\n",
 780                                            skb_dst(skb) ? "neigh" : "dst",
 781                                            be16_to_cpup((__be16 *) skb->data),
 782                                            IPOIB_QPN(phdr->hwaddr),
 783                                            phdr->hwaddr + 4);
 784                                 dev_kfree_skb_any(skb);
 785                                 ++dev->stats.tx_dropped;
 786                                 return NETDEV_TX_OK;
 787                         }
 788
 789                         unicast_arp_send(skb, dev, phdr);
 790                 }
 791         }
 792
 793         return NETDEV_TX_OK;
 794 }
 795
 796 static void ipoib_timeout(struct net_device *dev)
 797 {
 798         struct ipoib_dev_priv *priv = netdev_priv(dev);
 799
 800         ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
 801                    jiffies_to_msecs(jiffies - dev->trans_start));
 802         ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
 803                    netif_queue_stopped(dev),
 804                    priv->tx_head, priv->tx_tail);
 805         /* XXX reset QP, etc. */
 806 }
 807
 808 static int ipoib_hard_header(struct sk_buff *skb,
 809                              struct net_device *dev,
 810                              unsigned short type,
 811                              const void *daddr, const void *saddr, unsigned len)
 812 {
 813         struct ipoib_header *header;
 814
 815         header = (struct ipoib_header *) skb_push(skb, sizeof *header);
 816
 817         header->proto = htons(type);
 818         header->reserved = 0;
 819
 820         /*
 821          * If we don't have a neighbour structure, stuff the
 822          * destination address onto the front of the skb so we can
 823          * figure out where to send the packet later.
 824          */
 825         if ((!skb_dst(skb) || !skb_dst(skb)->neighbour) && daddr) {
 826                 struct ipoib_pseudoheader *phdr =
 827                         (struct ipoib_pseudoheader *) skb_push(skb, sizeof *phdr);
 828                 memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
 829         }
 830
 831         return 0;
 832 }
 833
 834 static void ipoib_set_mcast_list(struct net_device *dev)
 835 {
 836         struct ipoib_dev_priv *priv = netdev_priv(dev);
 837
 838         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
 839                 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
 840                 return;
 841         }
 842
 843         queue_work(ipoib_workqueue, &priv->restart_task);
 844 }
 845
 846 static void ipoib_neigh_cleanup(struct neighbour *n)
 847 {
 848         struct ipoib_neigh *neigh;
 849         struct ipoib_dev_priv *priv = netdev_priv(n->dev);
 850         unsigned long flags;
 851         struct ipoib_ah *ah = NULL;
 852
 853         neigh = *to_ipoib_neigh(n);
 854         if (neigh)
 855                 priv = netdev_priv(neigh->dev);
 856         else
 857                 return;
 858         ipoib_dbg(priv,
 859                   "neigh_cleanup for %06x %pI6\n",
 860                   IPOIB_QPN(n->ha),
 861                   n->ha + 4);
 862
 863         spin_lock_irqsave(&priv->lock, flags);
 864
 865         if (neigh->ah)
 866                 ah = neigh->ah;
 867         list_del(&neigh->list);
 868         ipoib_neigh_free(n->dev, neigh);
 869
 870         spin_unlock_irqrestore(&priv->lock, flags);
 871
 872         if (ah)
 873                 ipoib_put_ah(ah);
 874 }
 875
 876 struct ipoib_neigh *ipoib_neigh_alloc(struct neighbour *neighbour,
 877                                       struct net_device *dev)
 878 {
 879         struct ipoib_neigh *neigh;
 880
 881         neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
 882         if (!neigh)
 883                 return NULL;
 884
 885         neigh->neighbour = neighbour;
 886         neigh->dev = dev;
 887         memset(&neigh->dgid.raw, 0, sizeof (union ib_gid));
 888         *to_ipoib_neigh(neighbour) = neigh;
 889         skb_queue_head_init(&neigh->queue);
 890         ipoib_cm_set(neigh, NULL);
 891
 892         return neigh;
 893 }
 894
 895 void ipoib_neigh_free(struct net_device *dev, struct ipoib_neigh *neigh)
 896 {
 897         struct sk_buff *skb;
 898         *to_ipoib_neigh(neigh->neighbour) = NULL;
 899         while ((skb = __skb_dequeue(&neigh->queue))) {
 900                 ++dev->stats.tx_dropped;
 901                 dev_kfree_skb_any(skb);
 902         }
 903         if (ipoib_cm_get(neigh))
 904                 ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
 905         kfree(neigh);
 906 }
 907
 908 static int ipoib_neigh_setup_dev(struct net_device *dev, struct neigh_parms *parms)
 909 {
 910         parms->neigh_cleanup = ipoib_neigh_cleanup;
 911
 912         return 0;
 913 }
 914
 915 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
 916 {
 917         struct ipoib_dev_priv *priv = netdev_priv(dev);
 918
 919         /* Allocate RX/TX "rings" to hold queued skbs */
 920         priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
 921                                 GFP_KERNEL);
 922         if (!priv->rx_ring) {
 923                 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
 924                        ca->name, ipoib_recvq_size);
 925                 goto out;
 926         }
 927
 928         priv->tx_ring = vmalloc(ipoib_sendq_size * sizeof *priv->tx_ring);
 929         if (!priv->tx_ring) {
 930                 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
 931                        ca->name, ipoib_sendq_size);
 932                 goto out_rx_ring_cleanup;
 933         }
 934         memset(priv->tx_ring, 0, ipoib_sendq_size * sizeof *priv->tx_ring);
 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 int get_skb_hdr(struct sk_buff *skb, void **iphdr,
 980                        void **tcph, u64 *hdr_flags, void *priv)
 981 {
 982         unsigned int ip_len;
 983         struct iphdr *iph;
 984
 985         if (unlikely(skb->protocol != htons(ETH_P_IP)))
 986                 return -1;
 987
 988         /*
 989          * In the future we may add an else clause that verifies the
 990          * checksum and allows devices which do not calculate checksum
 991          * to use LRO.
 992          */
 993         if (unlikely(skb->ip_summed != CHECKSUM_UNNECESSARY))
 994                 return -1;
 995
 996         /* Check for non-TCP packet */
 997         skb_reset_network_header(skb);
 998         iph = ip_hdr(skb);
 999         if (iph->protocol != IPPROTO_TCP)
1000                 return -1;
1001
1002         ip_len = ip_hdrlen(skb);
1003         skb_set_transport_header(skb, ip_len);
1004         *tcph = tcp_hdr(skb);
1005
1006         /* check if IP header and TCP header are complete */
1007         if (ntohs(iph->tot_len) < ip_len + tcp_hdrlen(skb))
1008                 return -1;
1009
1010         *hdr_flags = LRO_IPV4 | LRO_TCP;
1011         *iphdr = iph;
1012
1013         return 0;
1014 }
1015
1016 static void ipoib_lro_setup(struct ipoib_dev_priv *priv)
1017 {
1018         priv->lro.lro_mgr.max_aggr       = lro_max_aggr;
1019         priv->lro.lro_mgr.max_desc       = IPOIB_MAX_LRO_DESCRIPTORS;
1020         priv->lro.lro_mgr.lro_arr        = priv->lro.lro_desc;
1021         priv->lro.lro_mgr.get_skb_header = get_skb_hdr;
1022         priv->lro.lro_mgr.features       = LRO_F_NAPI;
1023         priv->lro.lro_mgr.dev            = priv->dev;
1024         priv->lro.lro_mgr.ip_summed_aggr = CHECKSUM_UNNECESSARY;
1025 }
1026
1027 static const struct net_device_ops ipoib_netdev_ops = {
1028         .ndo_open                = ipoib_open,
1029         .ndo_stop                = ipoib_stop,
1030         .ndo_change_mtu          = ipoib_change_mtu,
1031         .ndo_start_xmit          = ipoib_start_xmit,
1032         .ndo_tx_timeout          = ipoib_timeout,
1033         .ndo_set_multicast_list  = ipoib_set_mcast_list,
1034         .ndo_neigh_setup         = ipoib_neigh_setup_dev,
1035 };
1036
1037 static void ipoib_setup(struct net_device *dev)
1038 {
1039         struct ipoib_dev_priv *priv = netdev_priv(dev);
1040
1041         dev->netdev_ops          = &ipoib_netdev_ops;
1042         dev->header_ops          = &ipoib_header_ops;
1043
1044         ipoib_set_ethtool_ops(dev);
1045
1046         netif_napi_add(dev, &priv->napi, ipoib_poll, 100);
1047
1048         dev->watchdog_timeo      = HZ;
1049
1050         dev->flags              |= IFF_BROADCAST | IFF_MULTICAST;
1051
1052         /*
1053          * We add in INFINIBAND_ALEN to allow for the destination
1054          * address "pseudoheader" for skbs without neighbour struct.
1055          */
1056         dev->hard_header_len     = IPOIB_ENCAP_LEN + INFINIBAND_ALEN;
1057         dev->addr_len            = INFINIBAND_ALEN;
1058         dev->type                = ARPHRD_INFINIBAND;
1059         dev->tx_queue_len        = ipoib_sendq_size * 2;
1060         dev->features            = (NETIF_F_VLAN_CHALLENGED     |
1061                                     NETIF_F_HIGHDMA);
1062         dev->priv_flags         &= ~IFF_XMIT_DST_RELEASE;
1063
1064         memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
1065
1066         netif_carrier_off(dev);
1067
1068         priv->dev = dev;
1069
1070         ipoib_lro_setup(priv);
1071
1072         spin_lock_init(&priv->lock);
1073
1074         mutex_init(&priv->vlan_mutex);
1075
1076         INIT_LIST_HEAD(&priv->path_list);
1077         INIT_LIST_HEAD(&priv->child_intfs);
1078         INIT_LIST_HEAD(&priv->dead_ahs);
1079         INIT_LIST_HEAD(&priv->multicast_list);
1080
1081         INIT_DELAYED_WORK(&priv->pkey_poll_task, ipoib_pkey_poll);
1082         INIT_DELAYED_WORK(&priv->mcast_task,   ipoib_mcast_join_task);
1083         INIT_WORK(&priv->carrier_on_task, ipoib_mcast_carrier_on_task);
1084         INIT_WORK(&priv->flush_light,   ipoib_ib_dev_flush_light);
1085         INIT_WORK(&priv->flush_normal,   ipoib_ib_dev_flush_normal);
1086         INIT_WORK(&priv->flush_heavy,   ipoib_ib_dev_flush_heavy);
1087         INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
1088         INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
1089 }
1090
1091 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
1092 {
1093         struct net_device *dev;
1094
1095         dev = alloc_netdev((int) sizeof (struct ipoib_dev_priv), name,
1096                            ipoib_setup);
1097         if (!dev)
1098                 return NULL;
1099
1100         return netdev_priv(dev);
1101 }
1102
1103 static ssize_t show_pkey(struct device *dev,
1104                          struct device_attribute *attr, char *buf)
1105 {
1106         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1107
1108         return sprintf(buf, "0x%04x\n", priv->pkey);
1109 }
1110 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1111
1112 static ssize_t show_umcast(struct device *dev,
1113                            struct device_attribute *attr, char *buf)
1114 {
1115         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1116
1117         return sprintf(buf, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
1118 }
1119
1120 static ssize_t set_umcast(struct device *dev,
1121                           struct device_attribute *attr,
1122                           const char *buf, size_t count)
1123 {
1124         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1125         unsigned long umcast_val = simple_strtoul(buf, NULL, 0);
1126
1127         if (umcast_val > 0) {
1128                 set_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1129                 ipoib_warn(priv, "ignoring multicast groups joined directly "
1130                                 "by userspace\n");
1131         } else
1132                 clear_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1133
1134         return count;
1135 }
1136 static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast);
1137
1138 int ipoib_add_umcast_attr(struct net_device *dev)
1139 {
1140         return device_create_file(&dev->dev, &dev_attr_umcast);
1141 }
1142
1143 static ssize_t create_child(struct device *dev,
1144                             struct device_attribute *attr,
1145                             const char *buf, size_t count)
1146 {
1147         int pkey;
1148         int ret;
1149
1150         if (sscanf(buf, "%i", &pkey) != 1)
1151                 return -EINVAL;
1152
1153         if (pkey < 0 || pkey > 0xffff)
1154                 return -EINVAL;
1155
1156         /*
1157          * Set the full membership bit, so that we join the right
1158          * broadcast group, etc.
1159          */
1160         pkey |= 0x8000;
1161
1162         ret = ipoib_vlan_add(to_net_dev(dev), pkey);
1163
1164         return ret ? ret : count;
1165 }
1166 static DEVICE_ATTR(create_child, S_IWUSR, NULL, create_child);
1167
1168 static ssize_t delete_child(struct device *dev,
1169                             struct device_attribute *attr,
1170                             const char *buf, size_t count)
1171 {
1172         int pkey;
1173         int ret;
1174
1175         if (sscanf(buf, "%i", &pkey) != 1)
1176                 return -EINVAL;
1177
1178         if (pkey < 0 || pkey > 0xffff)
1179                 return -EINVAL;
1180
1181         ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
1182
1183         return ret ? ret : count;
1184
1185 }
1186 static DEVICE_ATTR(delete_child, S_IWUSR, NULL, delete_child);
1187
1188 int ipoib_add_pkey_attr(struct net_device *dev)
1189 {
1190         return device_create_file(&dev->dev, &dev_attr_pkey);
1191 }
1192
1193 int ipoib_set_dev_features(struct ipoib_dev_priv *priv, struct ib_device *hca)
1194 {
1195         struct ib_device_attr *device_attr;
1196         int result = -ENOMEM;
1197
1198         device_attr = kmalloc(sizeof *device_attr, GFP_KERNEL);
1199         if (!device_attr) {
1200                 printk(KERN_WARNING "%s: allocation of %zu bytes failed\n",
1201                        hca->name, sizeof *device_attr);
1202                 return result;
1203         }
1204
1205         result = ib_query_device(hca, device_attr);
1206         if (result) {
1207                 printk(KERN_WARNING "%s: ib_query_device failed (ret = %d)\n",
1208                        hca->name, result);
1209                 kfree(device_attr);
1210                 return result;
1211         }
1212         priv->hca_caps = device_attr->device_cap_flags;
1213
1214         kfree(device_attr);
1215
1216         if (priv->hca_caps & IB_DEVICE_UD_IP_CSUM) {
1217                 set_bit(IPOIB_FLAG_CSUM, &priv->flags);
1218                 priv->dev->features |= NETIF_F_SG | NETIF_F_IP_CSUM;
1219         }
1220
1221         if (lro)
1222                 priv->dev->features |= NETIF_F_LRO;
1223
1224         if (priv->dev->features & NETIF_F_SG && priv->hca_caps & IB_DEVICE_UD_TSO)
1225                 priv->dev->features |= NETIF_F_TSO;
1226
1227         return 0;
1228 }
1229
1230
1231 static struct net_device *ipoib_add_port(const char *format,
1232                                          struct ib_device *hca, u8 port)
1233 {
1234         struct ipoib_dev_priv *priv;
1235         struct ib_port_attr attr;
1236         int result = -ENOMEM;
1237
1238         priv = ipoib_intf_alloc(format);
1239         if (!priv)
1240                 goto alloc_mem_failed;
1241
1242         SET_NETDEV_DEV(priv->dev, hca->dma_device);
1243
1244         if (!ib_query_port(hca, port, &attr))
1245                 priv->max_ib_mtu = ib_mtu_enum_to_int(attr.max_mtu);
1246         else {
1247                 printk(KERN_WARNING "%s: ib_query_port %d failed\n",
1248                        hca->name, port);
1249                 goto device_init_failed;
1250         }
1251
1252         /* MTU will be reset when mcast join happens */
1253         priv->dev->mtu  = IPOIB_UD_MTU(priv->max_ib_mtu);
1254         priv->mcast_mtu  = priv->admin_mtu = priv->dev->mtu;
1255
1256         result = ib_query_pkey(hca, port, 0, &priv->pkey);
1257         if (result) {
1258                 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
1259                        hca->name, port, result);
1260                 goto device_init_failed;
1261         }
1262
1263         if (ipoib_set_dev_features(priv, hca))
1264                 goto device_init_failed;
1265
1266         /*
1267          * Set the full membership bit, so that we join the right
1268          * broadcast group, etc.
1269          */
1270         priv->pkey |= 0x8000;
1271
1272         priv->dev->broadcast[8] = priv->pkey >> 8;
1273         priv->dev->broadcast[9] = priv->pkey & 0xff;
1274
1275         result = ib_query_gid(hca, port, 0, &priv->local_gid);
1276         if (result) {
1277                 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1278                        hca->name, port, result);
1279                 goto device_init_failed;
1280         } else
1281                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1282
1283         result = ipoib_dev_init(priv->dev, hca, port);
1284         if (result < 0) {
1285                 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1286                        hca->name, port, result);
1287                 goto device_init_failed;
1288         }
1289
1290         INIT_IB_EVENT_HANDLER(&priv->event_handler,
1291                               priv->ca, ipoib_event);
1292         result = ib_register_event_handler(&priv->event_handler);
1293         if (result < 0) {
1294                 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1295                        "port %d (ret = %d)\n",
1296                        hca->name, port, result);
1297                 goto event_failed;
1298         }
1299
1300         result = register_netdev(priv->dev);
1301         if (result) {
1302                 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1303                        hca->name, port, result);
1304                 goto register_failed;
1305         }
1306
1307         ipoib_create_debug_files(priv->dev);
1308
1309         if (ipoib_cm_add_mode_attr(priv->dev))
1310                 goto sysfs_failed;
1311         if (ipoib_add_pkey_attr(priv->dev))
1312                 goto sysfs_failed;
1313         if (ipoib_add_umcast_attr(priv->dev))
1314                 goto sysfs_failed;
1315         if (device_create_file(&priv->dev->dev, &dev_attr_create_child))
1316                 goto sysfs_failed;
1317         if (device_create_file(&priv->dev->dev, &dev_attr_delete_child))
1318                 goto sysfs_failed;
1319
1320         return priv->dev;
1321
1322 sysfs_failed:
1323         ipoib_delete_debug_files(priv->dev);
1324         unregister_netdev(priv->dev);
1325
1326 register_failed:
1327         ib_unregister_event_handler(&priv->event_handler);
1328         flush_workqueue(ipoib_workqueue);
1329
1330 event_failed:
1331         ipoib_dev_cleanup(priv->dev);
1332
1333 device_init_failed:
1334         free_netdev(priv->dev);
1335
1336 alloc_mem_failed:
1337         return ERR_PTR(result);
1338 }
1339
1340 static void ipoib_add_one(struct ib_device *device)
1341 {
1342         struct list_head *dev_list;
1343         struct net_device *dev;
1344         struct ipoib_dev_priv *priv;
1345         int s, e, p;
1346
1347         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1348                 return;
1349
1350         dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1351         if (!dev_list)
1352                 return;
1353
1354         INIT_LIST_HEAD(dev_list);
1355
1356         if (device->node_type == RDMA_NODE_IB_SWITCH) {
1357                 s = 0;
1358                 e = 0;
1359         } else {
1360                 s = 1;
1361                 e = device->phys_port_cnt;
1362         }
1363
1364         for (p = s; p <= e; ++p) {
1365                 dev = ipoib_add_port("ib%d", device, p);
1366                 if (!IS_ERR(dev)) {
1367                         priv = netdev_priv(dev);
1368                         list_add_tail(&priv->list, dev_list);
1369                 }
1370         }
1371
1372         ib_set_client_data(device, &ipoib_client, dev_list);
1373 }
1374
1375 static void ipoib_remove_one(struct ib_device *device)
1376 {
1377         struct ipoib_dev_priv *priv, *tmp;
1378         struct list_head *dev_list;
1379
1380         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1381                 return;
1382
1383         dev_list = ib_get_client_data(device, &ipoib_client);
1384
1385         list_for_each_entry_safe(priv, tmp, dev_list, list) {
1386                 ib_unregister_event_handler(&priv->event_handler);
1387
1388                 rtnl_lock();
1389                 dev_change_flags(priv->dev, priv->dev->flags & ~IFF_UP);
1390                 rtnl_unlock();
1391
1392                 flush_workqueue(ipoib_workqueue);
1393
1394                 unregister_netdev(priv->dev);
1395                 ipoib_dev_cleanup(priv->dev);
1396                 free_netdev(priv->dev);
1397         }
1398
1399         kfree(dev_list);
1400 }
1401
1402 static int __init ipoib_init_module(void)
1403 {
1404         int ret;
1405
1406         ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1407         ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1408         ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1409
1410         ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1411         ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1412         ipoib_sendq_size = max(ipoib_sendq_size, max(2 * MAX_SEND_CQE,
1413                                                      IPOIB_MIN_QUEUE_SIZE));
1414 #ifdef CONFIG_INFINIBAND_IPOIB_CM
1415         ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
1416 #endif
1417
1418         /*
1419          * When copying small received packets, we only copy from the
1420          * linear data part of the SKB, so we rely on this condition.
1421          */
1422         BUILD_BUG_ON(IPOIB_CM_COPYBREAK > IPOIB_CM_HEAD_SIZE);
1423
1424         ret = ipoib_register_debugfs();
1425         if (ret)
1426                 return ret;
1427
1428         /*
1429          * We create our own workqueue mainly because we want to be
1430          * able to flush it when devices are being removed.  We can't
1431          * use schedule_work()/flush_scheduled_work() because both
1432          * unregister_netdev() and linkwatch_event take the rtnl lock,
1433          * so flush_scheduled_work() can deadlock during device
1434          * removal.
1435          */
1436         ipoib_workqueue = create_singlethread_workqueue("ipoib");
1437         if (!ipoib_workqueue) {
1438                 ret = -ENOMEM;
1439                 goto err_fs;
1440         }
1441
1442         ib_sa_register_client(&ipoib_sa_client);
1443
1444         ret = ib_register_client(&ipoib_client);
1445         if (ret)
1446                 goto err_sa;
1447
1448         return 0;
1449
1450 err_sa:
1451         ib_sa_unregister_client(&ipoib_sa_client);
1452         destroy_workqueue(ipoib_workqueue);
1453
1454 err_fs:
1455         ipoib_unregister_debugfs();
1456
1457         return ret;
1458 }
1459
1460 static void __exit ipoib_cleanup_module(void)
1461 {
1462         ib_unregister_client(&ipoib_client);
1463         ib_sa_unregister_client(&ipoib_sa_client);
1464         ipoib_unregister_debugfs();
1465         destroy_workqueue(ipoib_workqueue);
1466 }
1467
1468 module_init(ipoib_init_module);
1469 module_exit(ipoib_cleanup_module);