git.oblomov.eu Git - linux-2.6/blob - drivers/infiniband/ulp/ipoib/ipoib_main.c

   1 /*
   2  * Copyright (c) 2004 Topspin Communications.  All rights reserved.
   3  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
   4  * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
   5  *
   6  * This software is available to you under a choice of one of two
   7  * licenses.  You may choose to be licensed under the terms of the GNU
   8  * General Public License (GPL) Version 2, available from the file
   9  * COPYING in the main directory of this source tree, or the
  10  * OpenIB.org BSD license below:
  11  *
  12  *     Redistribution and use in source and binary forms, with or
  13  *     without modification, are permitted provided that the following
  14  *     conditions are met:
  15  *
  16  *      - Redistributions of source code must retain the above
  17  *        copyright notice, this list of conditions and the following
  18  *        disclaimer.
  19  *
  20  *      - Redistributions in binary form must reproduce the above
  21  *        copyright notice, this list of conditions and the following
  22  *        disclaimer in the documentation and/or other materials
  23  *        provided with the distribution.
  24  *
  25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  32  * SOFTWARE.
  33  *
  34  * $Id: ipoib_main.c 1377 2004-12-23 19:57:12Z roland $
  35  */
  36
  37 #include "ipoib.h"
  38
  39 #include <linux/module.h>
  40
  41 #include <linux/init.h>
  42 #include <linux/slab.h>
  43 #include <linux/vmalloc.h>
  44
  45 #include <linux/if_arp.h>       /* For ARPHRD_xxx */
  46
  47 #include <linux/ip.h>
  48 #include <linux/in.h>
  49
  50 MODULE_AUTHOR("Roland Dreier");
  51 MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
  52 MODULE_LICENSE("Dual BSD/GPL");
  53
  54 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
  55 int ipoib_debug_level;
  56
  57 module_param_named(debug_level, ipoib_debug_level, int, 0644);
  58 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
  59 #endif
  60
  61 struct ipoib_path_iter {
  62         struct net_device *dev;
  63         struct ipoib_path  path;
  64 };
  65
  66 static const u8 ipv4_bcast_addr[] = {
  67         0x00, 0xff, 0xff, 0xff,
  68         0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
  69         0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff
  70 };
  71
  72 struct workqueue_struct *ipoib_workqueue;
  73
  74 static void ipoib_add_one(struct ib_device *device);
  75 static void ipoib_remove_one(struct ib_device *device);
  76
  77 static struct ib_client ipoib_client = {
  78         .name   = "ipoib",
  79         .add    = ipoib_add_one,
  80         .remove = ipoib_remove_one
  81 };
  82
  83 int ipoib_open(struct net_device *dev)
  84 {
  85         struct ipoib_dev_priv *priv = netdev_priv(dev);
  86
  87         ipoib_dbg(priv, "bringing up interface\n");
  88
  89         set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
  90
  91         if (ipoib_pkey_dev_delay_open(dev))
  92                 return 0;
  93
  94         if (ipoib_ib_dev_open(dev))
  95                 return -EINVAL;
  96
  97         if (ipoib_ib_dev_up(dev)) {
  98                 ipoib_ib_dev_stop(dev);
  99                 return -EINVAL;
 100         }
 101
 102         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 103                 struct ipoib_dev_priv *cpriv;
 104
 105                 /* Bring up any child interfaces too */
 106                 down(&priv->vlan_mutex);
 107                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 108                         int flags;
 109
 110                         flags = cpriv->dev->flags;
 111                         if (flags & IFF_UP)
 112                                 continue;
 113
 114                         dev_change_flags(cpriv->dev, flags | IFF_UP);
 115                 }
 116                 up(&priv->vlan_mutex);
 117         }
 118
 119         netif_start_queue(dev);
 120
 121         return 0;
 122 }
 123
 124 static int ipoib_stop(struct net_device *dev)
 125 {
 126         struct ipoib_dev_priv *priv = netdev_priv(dev);
 127
 128         ipoib_dbg(priv, "stopping interface\n");
 129
 130         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 131
 132         netif_stop_queue(dev);
 133
 134         ipoib_ib_dev_down(dev);
 135         ipoib_ib_dev_stop(dev);
 136
 137         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 138                 struct ipoib_dev_priv *cpriv;
 139
 140                 /* Bring down any child interfaces too */
 141                 down(&priv->vlan_mutex);
 142                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 143                         int flags;
 144
 145                         flags = cpriv->dev->flags;
 146                         if (!(flags & IFF_UP))
 147                                 continue;
 148
 149                         dev_change_flags(cpriv->dev, flags & ~IFF_UP);
 150                 }
 151                 up(&priv->vlan_mutex);
 152         }
 153
 154         return 0;
 155 }
 156
 157 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
 158 {
 159         struct ipoib_dev_priv *priv = netdev_priv(dev);
 160
 161         if (new_mtu > IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN)
 162                 return -EINVAL;
 163
 164         priv->admin_mtu = new_mtu;
 165
 166         dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
 167
 168         return 0;
 169 }
 170
 171 static struct ipoib_path *__path_find(struct net_device *dev,
 172                                       union ib_gid *gid)
 173 {
 174         struct ipoib_dev_priv *priv = netdev_priv(dev);
 175         struct rb_node *n = priv->path_tree.rb_node;
 176         struct ipoib_path *path;
 177         int ret;
 178
 179         while (n) {
 180                 path = rb_entry(n, struct ipoib_path, rb_node);
 181
 182                 ret = memcmp(gid->raw, path->pathrec.dgid.raw,
 183                              sizeof (union ib_gid));
 184
 185                 if (ret < 0)
 186                         n = n->rb_left;
 187                 else if (ret > 0)
 188                         n = n->rb_right;
 189                 else
 190                         return path;
 191         }
 192
 193         return NULL;
 194 }
 195
 196 static int __path_add(struct net_device *dev, struct ipoib_path *path)
 197 {
 198         struct ipoib_dev_priv *priv = netdev_priv(dev);
 199         struct rb_node **n = &priv->path_tree.rb_node;
 200         struct rb_node *pn = NULL;
 201         struct ipoib_path *tpath;
 202         int ret;
 203
 204         while (*n) {
 205                 pn = *n;
 206                 tpath = rb_entry(pn, struct ipoib_path, rb_node);
 207
 208                 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
 209                              sizeof (union ib_gid));
 210                 if (ret < 0)
 211                         n = &pn->rb_left;
 212                 else if (ret > 0)
 213                         n = &pn->rb_right;
 214                 else
 215                         return -EEXIST;
 216         }
 217
 218         rb_link_node(&path->rb_node, pn, n);
 219         rb_insert_color(&path->rb_node, &priv->path_tree);
 220
 221         list_add_tail(&path->list, &priv->path_list);
 222
 223         return 0;
 224 }
 225
 226 static void path_free(struct net_device *dev, struct ipoib_path *path)
 227 {
 228         struct ipoib_dev_priv *priv = netdev_priv(dev);
 229         struct ipoib_neigh *neigh, *tn;
 230         struct sk_buff *skb;
 231         unsigned long flags;
 232
 233         while ((skb = __skb_dequeue(&path->queue)))
 234                 dev_kfree_skb_irq(skb);
 235
 236         spin_lock_irqsave(&priv->lock, flags);
 237
 238         list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
 239                 /*
 240                  * It's safe to call ipoib_put_ah() inside priv->lock
 241                  * here, because we know that path->ah will always
 242                  * hold one more reference, so ipoib_put_ah() will
 243                  * never do more than decrement the ref count.
 244                  */
 245                 if (neigh->ah)
 246                         ipoib_put_ah(neigh->ah);
 247                 *to_ipoib_neigh(neigh->neighbour) = NULL;
 248                 neigh->neighbour->ops->destructor = NULL;
 249                 kfree(neigh);
 250         }
 251
 252         spin_unlock_irqrestore(&priv->lock, flags);
 253
 254         if (path->ah)
 255                 ipoib_put_ah(path->ah);
 256
 257         kfree(path);
 258 }
 259
 260 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 261
 262 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
 263 {
 264         struct ipoib_path_iter *iter;
 265
 266         iter = kmalloc(sizeof *iter, GFP_KERNEL);
 267         if (!iter)
 268                 return NULL;
 269
 270         iter->dev = dev;
 271         memset(iter->path.pathrec.dgid.raw, 0, 16);
 272
 273         if (ipoib_path_iter_next(iter)) {
 274                 kfree(iter);
 275                 return NULL;
 276         }
 277
 278         return iter;
 279 }
 280
 281 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
 282 {
 283         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
 284         struct rb_node *n;
 285         struct ipoib_path *path;
 286         int ret = 1;
 287
 288         spin_lock_irq(&priv->lock);
 289
 290         n = rb_first(&priv->path_tree);
 291
 292         while (n) {
 293                 path = rb_entry(n, struct ipoib_path, rb_node);
 294
 295                 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
 296                            sizeof (union ib_gid)) < 0) {
 297                         iter->path = *path;
 298                         ret = 0;
 299                         break;
 300                 }
 301
 302                 n = rb_next(n);
 303         }
 304
 305         spin_unlock_irq(&priv->lock);
 306
 307         return ret;
 308 }
 309
 310 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
 311                           struct ipoib_path *path)
 312 {
 313         *path = iter->path;
 314 }
 315
 316 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
 317
 318 void ipoib_flush_paths(struct net_device *dev)
 319 {
 320         struct ipoib_dev_priv *priv = netdev_priv(dev);
 321         struct ipoib_path *path, *tp;
 322         LIST_HEAD(remove_list);
 323         unsigned long flags;
 324
 325         spin_lock_irqsave(&priv->lock, flags);
 326
 327         list_splice(&priv->path_list, &remove_list);
 328         INIT_LIST_HEAD(&priv->path_list);
 329
 330         list_for_each_entry(path, &remove_list, list)
 331                 rb_erase(&path->rb_node, &priv->path_tree);
 332
 333         spin_unlock_irqrestore(&priv->lock, flags);
 334
 335         list_for_each_entry_safe(path, tp, &remove_list, list) {
 336                 if (path->query)
 337                         ib_sa_cancel_query(path->query_id, path->query);
 338                 wait_for_completion(&path->done);
 339                 path_free(dev, path);
 340         }
 341 }
 342
 343 static void path_rec_completion(int status,
 344                                 struct ib_sa_path_rec *pathrec,
 345                                 void *path_ptr)
 346 {
 347         struct ipoib_path *path = path_ptr;
 348         struct net_device *dev = path->dev;
 349         struct ipoib_dev_priv *priv = netdev_priv(dev);
 350         struct ipoib_ah *ah = NULL;
 351         struct ipoib_neigh *neigh;
 352         struct sk_buff_head skqueue;
 353         struct sk_buff *skb;
 354         unsigned long flags;
 355
 356         if (pathrec)
 357                 ipoib_dbg(priv, "PathRec LID 0x%04x for GID " IPOIB_GID_FMT "\n",
 358                           be16_to_cpu(pathrec->dlid), IPOIB_GID_ARG(pathrec->dgid));
 359         else
 360                 ipoib_dbg(priv, "PathRec status %d for GID " IPOIB_GID_FMT "\n",
 361                           status, IPOIB_GID_ARG(path->pathrec.dgid));
 362
 363         skb_queue_head_init(&skqueue);
 364
 365         if (!status) {
 366                 struct ib_ah_attr av = {
 367                         .dlid          = be16_to_cpu(pathrec->dlid),
 368                         .sl            = pathrec->sl,
 369                         .port_num      = priv->port
 370                 };
 371                 int path_rate = ib_sa_rate_enum_to_int(pathrec->rate);
 372
 373                 if (path_rate > 0 && priv->local_rate > path_rate)
 374                         av.static_rate = (priv->local_rate - 1) / path_rate;
 375
 376                 ipoib_dbg(priv, "static_rate %d for local port %dX, path %dX\n",
 377                           av.static_rate, priv->local_rate,
 378                           ib_sa_rate_enum_to_int(pathrec->rate));
 379
 380                 ah = ipoib_create_ah(dev, priv->pd, &av);
 381         }
 382
 383         spin_lock_irqsave(&priv->lock, flags);
 384
 385         path->ah = ah;
 386
 387         if (ah) {
 388                 path->pathrec = *pathrec;
 389
 390                 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
 391                           ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
 392
 393                 while ((skb = __skb_dequeue(&path->queue)))
 394                         __skb_queue_tail(&skqueue, skb);
 395
 396                 list_for_each_entry(neigh, &path->neigh_list, list) {
 397                         kref_get(&path->ah->ref);
 398                         neigh->ah = path->ah;
 399
 400                         while ((skb = __skb_dequeue(&neigh->queue)))
 401                                 __skb_queue_tail(&skqueue, skb);
 402                 }
 403         }
 404
 405         path->query = NULL;
 406         complete(&path->done);
 407
 408         spin_unlock_irqrestore(&priv->lock, flags);
 409
 410         while ((skb = __skb_dequeue(&skqueue))) {
 411                 skb->dev = dev;
 412                 if (dev_queue_xmit(skb))
 413                         ipoib_warn(priv, "dev_queue_xmit failed "
 414                                    "to requeue packet\n");
 415         }
 416 }
 417
 418 static struct ipoib_path *path_rec_create(struct net_device *dev,
 419                                           union ib_gid *gid)
 420 {
 421         struct ipoib_dev_priv *priv = netdev_priv(dev);
 422         struct ipoib_path *path;
 423
 424         path = kzalloc(sizeof *path, GFP_ATOMIC);
 425         if (!path)
 426                 return NULL;
 427
 428         path->dev = dev;
 429
 430         skb_queue_head_init(&path->queue);
 431
 432         INIT_LIST_HEAD(&path->neigh_list);
 433
 434         memcpy(path->pathrec.dgid.raw, gid->raw, sizeof (union ib_gid));
 435         path->pathrec.sgid      = priv->local_gid;
 436         path->pathrec.pkey      = cpu_to_be16(priv->pkey);
 437         path->pathrec.numb_path = 1;
 438
 439         return path;
 440 }
 441
 442 static int path_rec_start(struct net_device *dev,
 443                           struct ipoib_path *path)
 444 {
 445         struct ipoib_dev_priv *priv = netdev_priv(dev);
 446
 447         ipoib_dbg(priv, "Start path record lookup for " IPOIB_GID_FMT "\n",
 448                   IPOIB_GID_ARG(path->pathrec.dgid));
 449
 450         init_completion(&path->done);
 451
 452         path->query_id =
 453                 ib_sa_path_rec_get(priv->ca, priv->port,
 454                                    &path->pathrec,
 455                                    IB_SA_PATH_REC_DGID          |
 456                                    IB_SA_PATH_REC_SGID          |
 457                                    IB_SA_PATH_REC_NUMB_PATH     |
 458                                    IB_SA_PATH_REC_PKEY,
 459                                    1000, GFP_ATOMIC,
 460                                    path_rec_completion,
 461                                    path, &path->query);
 462         if (path->query_id < 0) {
 463                 ipoib_warn(priv, "ib_sa_path_rec_get failed\n");
 464                 path->query = NULL;
 465                 return path->query_id;
 466         }
 467
 468         return 0;
 469 }
 470
 471 static void neigh_add_path(struct sk_buff *skb, struct net_device *dev)
 472 {
 473         struct ipoib_dev_priv *priv = netdev_priv(dev);
 474         struct ipoib_path *path;
 475         struct ipoib_neigh *neigh;
 476
 477         neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
 478         if (!neigh) {
 479                 ++priv->stats.tx_dropped;
 480                 dev_kfree_skb_any(skb);
 481                 return;
 482         }
 483
 484         skb_queue_head_init(&neigh->queue);
 485         neigh->neighbour = skb->dst->neighbour;
 486         *to_ipoib_neigh(skb->dst->neighbour) = neigh;
 487
 488         /*
 489          * We can only be called from ipoib_start_xmit, so we're
 490          * inside tx_lock -- no need to save/restore flags.
 491          */
 492         spin_lock(&priv->lock);
 493
 494         path = __path_find(dev, (union ib_gid *) (skb->dst->neighbour->ha + 4));
 495         if (!path) {
 496                 path = path_rec_create(dev,
 497                                        (union ib_gid *) (skb->dst->neighbour->ha + 4));
 498                 if (!path)
 499                         goto err;
 500
 501                 __path_add(dev, path);
 502         }
 503
 504         list_add_tail(&neigh->list, &path->neigh_list);
 505
 506         if (path->pathrec.dlid) {
 507                 kref_get(&path->ah->ref);
 508                 neigh->ah = path->ah;
 509
 510                 ipoib_send(dev, skb, path->ah,
 511                            be32_to_cpup((__be32 *) skb->dst->neighbour->ha));
 512         } else {
 513                 neigh->ah  = NULL;
 514                 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 515                         __skb_queue_tail(&neigh->queue, skb);
 516                 } else {
 517                         ++priv->stats.tx_dropped;
 518                         dev_kfree_skb_any(skb);
 519                 }
 520
 521                 if (!path->query && path_rec_start(dev, path))
 522                         goto err;
 523         }
 524
 525         spin_unlock(&priv->lock);
 526         return;
 527
 528 err:
 529         *to_ipoib_neigh(skb->dst->neighbour) = NULL;
 530         list_del(&neigh->list);
 531         neigh->neighbour->ops->destructor = NULL;
 532         kfree(neigh);
 533
 534         ++priv->stats.tx_dropped;
 535         dev_kfree_skb_any(skb);
 536
 537         spin_unlock(&priv->lock);
 538 }
 539
 540 static void ipoib_path_lookup(struct sk_buff *skb, struct net_device *dev)
 541 {
 542         struct ipoib_dev_priv *priv = netdev_priv(skb->dev);
 543
 544         /* Look up path record for unicasts */
 545         if (skb->dst->neighbour->ha[4] != 0xff) {
 546                 neigh_add_path(skb, dev);
 547                 return;
 548         }
 549
 550         /* Add in the P_Key for multicasts */
 551         skb->dst->neighbour->ha[8] = (priv->pkey >> 8) & 0xff;
 552         skb->dst->neighbour->ha[9] = priv->pkey & 0xff;
 553         ipoib_mcast_send(dev, (union ib_gid *) (skb->dst->neighbour->ha + 4), skb);
 554 }
 555
 556 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
 557                              struct ipoib_pseudoheader *phdr)
 558 {
 559         struct ipoib_dev_priv *priv = netdev_priv(dev);
 560         struct ipoib_path *path;
 561
 562         /*
 563          * We can only be called from ipoib_start_xmit, so we're
 564          * inside tx_lock -- no need to save/restore flags.
 565          */
 566         spin_lock(&priv->lock);
 567
 568         path = __path_find(dev, (union ib_gid *) (phdr->hwaddr + 4));
 569         if (!path) {
 570                 path = path_rec_create(dev,
 571                                        (union ib_gid *) (phdr->hwaddr + 4));
 572                 if (path) {
 573                         /* put pseudoheader back on for next time */
 574                         skb_push(skb, sizeof *phdr);
 575                         __skb_queue_tail(&path->queue, skb);
 576
 577                         if (path_rec_start(dev, path)) {
 578                                 spin_unlock(&priv->lock);
 579                                 path_free(dev, path);
 580                                 return;
 581                         } else
 582                                 __path_add(dev, path);
 583                 } else {
 584                         ++priv->stats.tx_dropped;
 585                         dev_kfree_skb_any(skb);
 586                 }
 587
 588                 spin_unlock(&priv->lock);
 589                 return;
 590         }
 591
 592         if (path->pathrec.dlid) {
 593                 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
 594                           be16_to_cpu(path->pathrec.dlid));
 595
 596                 ipoib_send(dev, skb, path->ah,
 597                            be32_to_cpup((__be32 *) phdr->hwaddr));
 598         } else if ((path->query || !path_rec_start(dev, path)) &&
 599                    skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 600                 /* put pseudoheader back on for next time */
 601                 skb_push(skb, sizeof *phdr);
 602                 __skb_queue_tail(&path->queue, skb);
 603         } else {
 604                 ++priv->stats.tx_dropped;
 605                 dev_kfree_skb_any(skb);
 606         }
 607
 608         spin_unlock(&priv->lock);
 609 }
 610
 611 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
 612 {
 613         struct ipoib_dev_priv *priv = netdev_priv(dev);
 614         struct ipoib_neigh *neigh;
 615         unsigned long flags;
 616
 617         if (!spin_trylock_irqsave(&priv->tx_lock, flags))
 618                 return NETDEV_TX_LOCKED;
 619
 620         /*
 621          * Check if our queue is stopped.  Since we have the LLTX bit
 622          * set, we can't rely on netif_stop_queue() preventing our
 623          * xmit function from being called with a full queue.
 624          */
 625         if (unlikely(netif_queue_stopped(dev))) {
 626                 spin_unlock_irqrestore(&priv->tx_lock, flags);
 627                 return NETDEV_TX_BUSY;
 628         }
 629
 630         if (skb->dst && skb->dst->neighbour) {
 631                 if (unlikely(!*to_ipoib_neigh(skb->dst->neighbour))) {
 632                         ipoib_path_lookup(skb, dev);
 633                         goto out;
 634                 }
 635
 636                 neigh = *to_ipoib_neigh(skb->dst->neighbour);
 637
 638                 if (likely(neigh->ah)) {
 639                         ipoib_send(dev, skb, neigh->ah,
 640                                    be32_to_cpup((__be32 *) skb->dst->neighbour->ha));
 641                         goto out;
 642                 }
 643
 644                 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 645                         spin_lock(&priv->lock);
 646                         __skb_queue_tail(&neigh->queue, skb);
 647                         spin_unlock(&priv->lock);
 648                 } else {
 649                         ++priv->stats.tx_dropped;
 650                         dev_kfree_skb_any(skb);
 651                 }
 652         } else {
 653                 struct ipoib_pseudoheader *phdr =
 654                         (struct ipoib_pseudoheader *) skb->data;
 655                 skb_pull(skb, sizeof *phdr);
 656
 657                 if (phdr->hwaddr[4] == 0xff) {
 658                         /* Add in the P_Key for multicast*/
 659                         phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
 660                         phdr->hwaddr[9] = priv->pkey & 0xff;
 661
 662                         ipoib_mcast_send(dev, (union ib_gid *) (phdr->hwaddr + 4), skb);
 663                 } else {
 664                         /* unicast GID -- should be ARP or RARP reply */
 665
 666                         if ((be16_to_cpup((__be16 *) skb->data) != ETH_P_ARP) &&
 667                             (be16_to_cpup((__be16 *) skb->data) != ETH_P_RARP)) {
 668                                 ipoib_warn(priv, "Unicast, no %s: type %04x, QPN %06x "
 669                                            IPOIB_GID_FMT "\n",
 670                                            skb->dst ? "neigh" : "dst",
 671                                            be16_to_cpup((__be16 *) skb->data),
 672                                            be32_to_cpup((__be32 *) phdr->hwaddr),
 673                                            IPOIB_GID_ARG(*(union ib_gid *) (phdr->hwaddr + 4)));
 674                                 dev_kfree_skb_any(skb);
 675                                 ++priv->stats.tx_dropped;
 676                                 goto out;
 677                         }
 678
 679                         unicast_arp_send(skb, dev, phdr);
 680                 }
 681         }
 682
 683 out:
 684         spin_unlock_irqrestore(&priv->tx_lock, flags);
 685
 686         return NETDEV_TX_OK;
 687 }
 688
 689 static struct net_device_stats *ipoib_get_stats(struct net_device *dev)
 690 {
 691         struct ipoib_dev_priv *priv = netdev_priv(dev);
 692
 693         return &priv->stats;
 694 }
 695
 696 static void ipoib_timeout(struct net_device *dev)
 697 {
 698         struct ipoib_dev_priv *priv = netdev_priv(dev);
 699
 700         ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
 701                    jiffies_to_msecs(jiffies - dev->trans_start));
 702         ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
 703                    netif_queue_stopped(dev),
 704                    priv->tx_head, priv->tx_tail);
 705         /* XXX reset QP, etc. */
 706 }
 707
 708 static int ipoib_hard_header(struct sk_buff *skb,
 709                              struct net_device *dev,
 710                              unsigned short type,
 711                              void *daddr, void *saddr, unsigned len)
 712 {
 713         struct ipoib_header *header;
 714
 715         header = (struct ipoib_header *) skb_push(skb, sizeof *header);
 716
 717         header->proto = htons(type);
 718         header->reserved = 0;
 719
 720         /*
 721          * If we don't have a neighbour structure, stuff the
 722          * destination address onto the front of the skb so we can
 723          * figure out where to send the packet later.
 724          */
 725         if (!skb->dst || !skb->dst->neighbour) {
 726                 struct ipoib_pseudoheader *phdr =
 727                         (struct ipoib_pseudoheader *) skb_push(skb, sizeof *phdr);
 728                 memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
 729         }
 730
 731         return 0;
 732 }
 733
 734 static void ipoib_set_mcast_list(struct net_device *dev)
 735 {
 736         struct ipoib_dev_priv *priv = netdev_priv(dev);
 737
 738         queue_work(ipoib_workqueue, &priv->restart_task);
 739 }
 740
 741 static void ipoib_neigh_destructor(struct neighbour *n)
 742 {
 743         struct ipoib_neigh *neigh;
 744         struct ipoib_dev_priv *priv = netdev_priv(n->dev);
 745         unsigned long flags;
 746         struct ipoib_ah *ah = NULL;
 747
 748         ipoib_dbg(priv,
 749                   "neigh_destructor for %06x " IPOIB_GID_FMT "\n",
 750                   be32_to_cpup((__be32 *) n->ha),
 751                   IPOIB_GID_ARG(*((union ib_gid *) (n->ha + 4))));
 752
 753         spin_lock_irqsave(&priv->lock, flags);
 754
 755         neigh = *to_ipoib_neigh(n);
 756         if (neigh) {
 757                 if (neigh->ah)
 758                         ah = neigh->ah;
 759                 list_del(&neigh->list);
 760                 *to_ipoib_neigh(n) = NULL;
 761                 kfree(neigh);
 762         }
 763
 764         spin_unlock_irqrestore(&priv->lock, flags);
 765
 766         if (ah)
 767                 ipoib_put_ah(ah);
 768 }
 769
 770 static int ipoib_neigh_setup(struct neighbour *neigh)
 771 {
 772         /*
 773          * Is this kosher?  I can't find anybody in the kernel that
 774          * sets neigh->destructor, so we should be able to set it here
 775          * without trouble.
 776          */
 777         neigh->ops->destructor = ipoib_neigh_destructor;
 778
 779         return 0;
 780 }
 781
 782 static int ipoib_neigh_setup_dev(struct net_device *dev, struct neigh_parms *parms)
 783 {
 784         parms->neigh_setup = ipoib_neigh_setup;
 785
 786         return 0;
 787 }
 788
 789 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
 790 {
 791         struct ipoib_dev_priv *priv = netdev_priv(dev);
 792
 793         /* Allocate RX/TX "rings" to hold queued skbs */
 794
 795         priv->rx_ring = kzalloc(IPOIB_RX_RING_SIZE * sizeof (struct ipoib_rx_buf),
 796                                 GFP_KERNEL);
 797         if (!priv->rx_ring) {
 798                 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
 799                        ca->name, IPOIB_RX_RING_SIZE);
 800                 goto out;
 801         }
 802
 803         priv->tx_ring = kzalloc(IPOIB_TX_RING_SIZE * sizeof (struct ipoib_tx_buf),
 804                                 GFP_KERNEL);
 805         if (!priv->tx_ring) {
 806                 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
 807                        ca->name, IPOIB_TX_RING_SIZE);
 808                 goto out_rx_ring_cleanup;
 809         }
 810
 811         /* priv->tx_head & tx_tail are already 0 */
 812
 813         if (ipoib_ib_dev_init(dev, ca, port))
 814                 goto out_tx_ring_cleanup;
 815
 816         return 0;
 817
 818 out_tx_ring_cleanup:
 819         kfree(priv->tx_ring);
 820
 821 out_rx_ring_cleanup:
 822         kfree(priv->rx_ring);
 823
 824 out:
 825         return -ENOMEM;
 826 }
 827
 828 void ipoib_dev_cleanup(struct net_device *dev)
 829 {
 830         struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
 831
 832         ipoib_delete_debug_files(dev);
 833
 834         /* Delete any child interfaces first */
 835         list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
 836                 unregister_netdev(cpriv->dev);
 837                 ipoib_dev_cleanup(cpriv->dev);
 838                 free_netdev(cpriv->dev);
 839         }
 840
 841         ipoib_ib_dev_cleanup(dev);
 842
 843         kfree(priv->rx_ring);
 844         kfree(priv->tx_ring);
 845
 846         priv->rx_ring = NULL;
 847         priv->tx_ring = NULL;
 848 }
 849
 850 static void ipoib_setup(struct net_device *dev)
 851 {
 852         struct ipoib_dev_priv *priv = netdev_priv(dev);
 853
 854         dev->open                = ipoib_open;
 855         dev->stop                = ipoib_stop;
 856         dev->change_mtu          = ipoib_change_mtu;
 857         dev->hard_start_xmit     = ipoib_start_xmit;
 858         dev->get_stats           = ipoib_get_stats;
 859         dev->tx_timeout          = ipoib_timeout;
 860         dev->hard_header         = ipoib_hard_header;
 861         dev->set_multicast_list  = ipoib_set_mcast_list;
 862         dev->neigh_setup         = ipoib_neigh_setup_dev;
 863
 864         dev->watchdog_timeo      = HZ;
 865
 866         dev->flags              |= IFF_BROADCAST | IFF_MULTICAST;
 867
 868         /*
 869          * We add in INFINIBAND_ALEN to allow for the destination
 870          * address "pseudoheader" for skbs without neighbour struct.
 871          */
 872         dev->hard_header_len     = IPOIB_ENCAP_LEN + INFINIBAND_ALEN;
 873         dev->addr_len            = INFINIBAND_ALEN;
 874         dev->type                = ARPHRD_INFINIBAND;
 875         dev->tx_queue_len        = IPOIB_TX_RING_SIZE * 2;
 876         dev->features            = NETIF_F_VLAN_CHALLENGED | NETIF_F_LLTX;
 877
 878         /* MTU will be reset when mcast join happens */
 879         dev->mtu                 = IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN;
 880         priv->mcast_mtu          = priv->admin_mtu = dev->mtu;
 881
 882         memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
 883
 884         netif_carrier_off(dev);
 885
 886         SET_MODULE_OWNER(dev);
 887
 888         priv->dev = dev;
 889
 890         spin_lock_init(&priv->lock);
 891         spin_lock_init(&priv->tx_lock);
 892
 893         init_MUTEX(&priv->mcast_mutex);
 894         init_MUTEX(&priv->vlan_mutex);
 895
 896         INIT_LIST_HEAD(&priv->path_list);
 897         INIT_LIST_HEAD(&priv->child_intfs);
 898         INIT_LIST_HEAD(&priv->dead_ahs);
 899         INIT_LIST_HEAD(&priv->multicast_list);
 900
 901         INIT_WORK(&priv->pkey_task,    ipoib_pkey_poll,          priv->dev);
 902         INIT_WORK(&priv->mcast_task,   ipoib_mcast_join_task,    priv->dev);
 903         INIT_WORK(&priv->flush_task,   ipoib_ib_dev_flush,       priv->dev);
 904         INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task, priv->dev);
 905         INIT_WORK(&priv->ah_reap_task, ipoib_reap_ah,            priv->dev);
 906 }
 907
 908 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
 909 {
 910         struct net_device *dev;
 911
 912         dev = alloc_netdev((int) sizeof (struct ipoib_dev_priv), name,
 913                            ipoib_setup);
 914         if (!dev)
 915                 return NULL;
 916
 917         return netdev_priv(dev);
 918 }
 919
 920 static ssize_t show_pkey(struct class_device *cdev, char *buf)
 921 {
 922         struct ipoib_dev_priv *priv =
 923                 netdev_priv(container_of(cdev, struct net_device, class_dev));
 924
 925         return sprintf(buf, "0x%04x\n", priv->pkey);
 926 }
 927 static CLASS_DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
 928
 929 static ssize_t create_child(struct class_device *cdev,
 930                             const char *buf, size_t count)
 931 {
 932         int pkey;
 933         int ret;
 934
 935         if (sscanf(buf, "%i", &pkey) != 1)
 936                 return -EINVAL;
 937
 938         if (pkey < 0 || pkey > 0xffff)
 939                 return -EINVAL;
 940
 941         /*
 942          * Set the full membership bit, so that we join the right
 943          * broadcast group, etc.
 944          */
 945         pkey |= 0x8000;
 946
 947         ret = ipoib_vlan_add(container_of(cdev, struct net_device, class_dev),
 948                              pkey);
 949
 950         return ret ? ret : count;
 951 }
 952 static CLASS_DEVICE_ATTR(create_child, S_IWUGO, NULL, create_child);
 953
 954 static ssize_t delete_child(struct class_device *cdev,
 955                             const char *buf, size_t count)
 956 {
 957         int pkey;
 958         int ret;
 959
 960         if (sscanf(buf, "%i", &pkey) != 1)
 961                 return -EINVAL;
 962
 963         if (pkey < 0 || pkey > 0xffff)
 964                 return -EINVAL;
 965
 966         ret = ipoib_vlan_delete(container_of(cdev, struct net_device, class_dev),
 967                                 pkey);
 968
 969         return ret ? ret : count;
 970
 971 }
 972 static CLASS_DEVICE_ATTR(delete_child, S_IWUGO, NULL, delete_child);
 973
 974 int ipoib_add_pkey_attr(struct net_device *dev)
 975 {
 976         return class_device_create_file(&dev->class_dev,
 977                                         &class_device_attr_pkey);
 978 }
 979
 980 static struct net_device *ipoib_add_port(const char *format,
 981                                          struct ib_device *hca, u8 port)
 982 {
 983         struct ipoib_dev_priv *priv;
 984         int result = -ENOMEM;
 985
 986         priv = ipoib_intf_alloc(format);
 987         if (!priv)
 988                 goto alloc_mem_failed;
 989
 990         SET_NETDEV_DEV(priv->dev, hca->dma_device);
 991
 992         result = ib_query_pkey(hca, port, 0, &priv->pkey);
 993         if (result) {
 994                 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
 995                        hca->name, port, result);
 996                 goto alloc_mem_failed;
 997         }
 998
 999         /*
1000          * Set the full membership bit, so that we join the right
1001          * broadcast group, etc.
1002          */
1003         priv->pkey |= 0x8000;
1004
1005         priv->dev->broadcast[8] = priv->pkey >> 8;
1006         priv->dev->broadcast[9] = priv->pkey & 0xff;
1007
1008         result = ib_query_gid(hca, port, 0, &priv->local_gid);
1009         if (result) {
1010                 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1011                        hca->name, port, result);
1012                 goto alloc_mem_failed;
1013         } else
1014                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1015
1016
1017         result = ipoib_dev_init(priv->dev, hca, port);
1018         if (result < 0) {
1019                 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1020                        hca->name, port, result);
1021                 goto device_init_failed;
1022         }
1023
1024         INIT_IB_EVENT_HANDLER(&priv->event_handler,
1025                               priv->ca, ipoib_event);
1026         result = ib_register_event_handler(&priv->event_handler);
1027         if (result < 0) {
1028                 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1029                        "port %d (ret = %d)\n",
1030                        hca->name, port, result);
1031                 goto event_failed;
1032         }
1033
1034         result = register_netdev(priv->dev);
1035         if (result) {
1036                 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1037                        hca->name, port, result);
1038                 goto register_failed;
1039         }
1040
1041         ipoib_create_debug_files(priv->dev);
1042
1043         if (ipoib_add_pkey_attr(priv->dev))
1044                 goto sysfs_failed;
1045         if (class_device_create_file(&priv->dev->class_dev,
1046                                      &class_device_attr_create_child))
1047                 goto sysfs_failed;
1048         if (class_device_create_file(&priv->dev->class_dev,
1049                                      &class_device_attr_delete_child))
1050                 goto sysfs_failed;
1051
1052         return priv->dev;
1053
1054 sysfs_failed:
1055         ipoib_delete_debug_files(priv->dev);
1056         unregister_netdev(priv->dev);
1057
1058 register_failed:
1059         ib_unregister_event_handler(&priv->event_handler);
1060         flush_scheduled_work();
1061
1062 event_failed:
1063         ipoib_dev_cleanup(priv->dev);
1064
1065 device_init_failed:
1066         free_netdev(priv->dev);
1067
1068 alloc_mem_failed:
1069         return ERR_PTR(result);
1070 }
1071
1072 static void ipoib_add_one(struct ib_device *device)
1073 {
1074         struct list_head *dev_list;
1075         struct net_device *dev;
1076         struct ipoib_dev_priv *priv;
1077         int s, e, p;
1078
1079         dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1080         if (!dev_list)
1081                 return;
1082
1083         INIT_LIST_HEAD(dev_list);
1084
1085         if (device->node_type == IB_NODE_SWITCH) {
1086                 s = 0;
1087                 e = 0;
1088         } else {
1089                 s = 1;
1090                 e = device->phys_port_cnt;
1091         }
1092
1093         for (p = s; p <= e; ++p) {
1094                 dev = ipoib_add_port("ib%d", device, p);
1095                 if (!IS_ERR(dev)) {
1096                         priv = netdev_priv(dev);
1097                         list_add_tail(&priv->list, dev_list);
1098                 }
1099         }
1100
1101         ib_set_client_data(device, &ipoib_client, dev_list);
1102 }
1103
1104 static void ipoib_remove_one(struct ib_device *device)
1105 {
1106         struct ipoib_dev_priv *priv, *tmp;
1107         struct list_head *dev_list;
1108
1109         dev_list = ib_get_client_data(device, &ipoib_client);
1110
1111         list_for_each_entry_safe(priv, tmp, dev_list, list) {
1112                 ib_unregister_event_handler(&priv->event_handler);
1113                 flush_scheduled_work();
1114
1115                 unregister_netdev(priv->dev);
1116                 ipoib_dev_cleanup(priv->dev);
1117                 free_netdev(priv->dev);
1118         }
1119
1120         kfree(dev_list);
1121 }
1122
1123 static int __init ipoib_init_module(void)
1124 {
1125         int ret;
1126
1127         ret = ipoib_register_debugfs();
1128         if (ret)
1129                 return ret;
1130
1131         /*
1132          * We create our own workqueue mainly because we want to be
1133          * able to flush it when devices are being removed.  We can't
1134          * use schedule_work()/flush_scheduled_work() because both
1135          * unregister_netdev() and linkwatch_event take the rtnl lock,
1136          * so flush_scheduled_work() can deadlock during device
1137          * removal.
1138          */
1139         ipoib_workqueue = create_singlethread_workqueue("ipoib");
1140         if (!ipoib_workqueue) {
1141                 ret = -ENOMEM;
1142                 goto err_fs;
1143         }
1144
1145         ret = ib_register_client(&ipoib_client);
1146         if (ret)
1147                 goto err_wq;
1148
1149         return 0;
1150
1151 err_wq:
1152         destroy_workqueue(ipoib_workqueue);
1153
1154 err_fs:
1155         ipoib_unregister_debugfs();
1156
1157         return ret;
1158 }
1159
1160 static void __exit ipoib_cleanup_module(void)
1161 {
1162         ib_unregister_client(&ipoib_client);
1163         ipoib_unregister_debugfs();
1164         destroy_workqueue(ipoib_workqueue);
1165 }
1166
1167 module_init(ipoib_init_module);
1168 module_exit(ipoib_cleanup_module);