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 #include <net/dst.h>
  51
  52 MODULE_AUTHOR("Roland Dreier");
  53 MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
  54 MODULE_LICENSE("Dual BSD/GPL");
  55
  56 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
  57 int ipoib_debug_level;
  58
  59 module_param_named(debug_level, ipoib_debug_level, int, 0644);
  60 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
  61 #endif
  62
  63 struct ipoib_path_iter {
  64         struct net_device *dev;
  65         struct ipoib_path  path;
  66 };
  67
  68 static const u8 ipv4_bcast_addr[] = {
  69         0x00, 0xff, 0xff, 0xff,
  70         0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
  71         0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff
  72 };
  73
  74 struct workqueue_struct *ipoib_workqueue;
  75
  76 static void ipoib_add_one(struct ib_device *device);
  77 static void ipoib_remove_one(struct ib_device *device);
  78
  79 static struct ib_client ipoib_client = {
  80         .name   = "ipoib",
  81         .add    = ipoib_add_one,
  82         .remove = ipoib_remove_one
  83 };
  84
  85 int ipoib_open(struct net_device *dev)
  86 {
  87         struct ipoib_dev_priv *priv = netdev_priv(dev);
  88
  89         ipoib_dbg(priv, "bringing up interface\n");
  90
  91         set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
  92
  93         if (ipoib_pkey_dev_delay_open(dev))
  94                 return 0;
  95
  96         if (ipoib_ib_dev_open(dev))
  97                 return -EINVAL;
  98
  99         if (ipoib_ib_dev_up(dev)) {
 100                 ipoib_ib_dev_stop(dev);
 101                 return -EINVAL;
 102         }
 103
 104         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 105                 struct ipoib_dev_priv *cpriv;
 106
 107                 /* Bring up any child interfaces too */
 108                 mutex_lock(&priv->vlan_mutex);
 109                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 110                         int flags;
 111
 112                         flags = cpriv->dev->flags;
 113                         if (flags & IFF_UP)
 114                                 continue;
 115
 116                         dev_change_flags(cpriv->dev, flags | IFF_UP);
 117                 }
 118                 mutex_unlock(&priv->vlan_mutex);
 119         }
 120
 121         netif_start_queue(dev);
 122
 123         return 0;
 124 }
 125
 126 static int ipoib_stop(struct net_device *dev)
 127 {
 128         struct ipoib_dev_priv *priv = netdev_priv(dev);
 129
 130         ipoib_dbg(priv, "stopping interface\n");
 131
 132         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 133
 134         netif_stop_queue(dev);
 135
 136         ipoib_ib_dev_down(dev);
 137         ipoib_ib_dev_stop(dev);
 138
 139         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 140                 struct ipoib_dev_priv *cpriv;
 141
 142                 /* Bring down any child interfaces too */
 143                 mutex_lock(&priv->vlan_mutex);
 144                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 145                         int flags;
 146
 147                         flags = cpriv->dev->flags;
 148                         if (!(flags & IFF_UP))
 149                                 continue;
 150
 151                         dev_change_flags(cpriv->dev, flags & ~IFF_UP);
 152                 }
 153                 mutex_unlock(&priv->vlan_mutex);
 154         }
 155
 156         return 0;
 157 }
 158
 159 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
 160 {
 161         struct ipoib_dev_priv *priv = netdev_priv(dev);
 162
 163         if (new_mtu > IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN)
 164                 return -EINVAL;
 165
 166         priv->admin_mtu = new_mtu;
 167
 168         dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
 169
 170         return 0;
 171 }
 172
 173 static struct ipoib_path *__path_find(struct net_device *dev,
 174                                       union ib_gid *gid)
 175 {
 176         struct ipoib_dev_priv *priv = netdev_priv(dev);
 177         struct rb_node *n = priv->path_tree.rb_node;
 178         struct ipoib_path *path;
 179         int ret;
 180
 181         while (n) {
 182                 path = rb_entry(n, struct ipoib_path, rb_node);
 183
 184                 ret = memcmp(gid->raw, path->pathrec.dgid.raw,
 185                              sizeof (union ib_gid));
 186
 187                 if (ret < 0)
 188                         n = n->rb_left;
 189                 else if (ret > 0)
 190                         n = n->rb_right;
 191                 else
 192                         return path;
 193         }
 194
 195         return NULL;
 196 }
 197
 198 static int __path_add(struct net_device *dev, struct ipoib_path *path)
 199 {
 200         struct ipoib_dev_priv *priv = netdev_priv(dev);
 201         struct rb_node **n = &priv->path_tree.rb_node;
 202         struct rb_node *pn = NULL;
 203         struct ipoib_path *tpath;
 204         int ret;
 205
 206         while (*n) {
 207                 pn = *n;
 208                 tpath = rb_entry(pn, struct ipoib_path, rb_node);
 209
 210                 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
 211                              sizeof (union ib_gid));
 212                 if (ret < 0)
 213                         n = &pn->rb_left;
 214                 else if (ret > 0)
 215                         n = &pn->rb_right;
 216                 else
 217                         return -EEXIST;
 218         }
 219
 220         rb_link_node(&path->rb_node, pn, n);
 221         rb_insert_color(&path->rb_node, &priv->path_tree);
 222
 223         list_add_tail(&path->list, &priv->path_list);
 224
 225         return 0;
 226 }
 227
 228 static void path_free(struct net_device *dev, struct ipoib_path *path)
 229 {
 230         struct ipoib_dev_priv *priv = netdev_priv(dev);
 231         struct ipoib_neigh *neigh, *tn;
 232         struct sk_buff *skb;
 233         unsigned long flags;
 234
 235         while ((skb = __skb_dequeue(&path->queue)))
 236                 dev_kfree_skb_irq(skb);
 237
 238         spin_lock_irqsave(&priv->lock, flags);
 239
 240         list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
 241                 /*
 242                  * It's safe to call ipoib_put_ah() inside priv->lock
 243                  * here, because we know that path->ah will always
 244                  * hold one more reference, so ipoib_put_ah() will
 245                  * never do more than decrement the ref count.
 246                  */
 247                 if (neigh->ah)
 248                         ipoib_put_ah(neigh->ah);
 249                 *to_ipoib_neigh(neigh->neighbour) = NULL;
 250                 neigh->neighbour->ops->destructor = NULL;
 251                 kfree(neigh);
 252         }
 253
 254         spin_unlock_irqrestore(&priv->lock, flags);
 255
 256         if (path->ah)
 257                 ipoib_put_ah(path->ah);
 258
 259         kfree(path);
 260 }
 261
 262 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 263
 264 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
 265 {
 266         struct ipoib_path_iter *iter;
 267
 268         iter = kmalloc(sizeof *iter, GFP_KERNEL);
 269         if (!iter)
 270                 return NULL;
 271
 272         iter->dev = dev;
 273         memset(iter->path.pathrec.dgid.raw, 0, 16);
 274
 275         if (ipoib_path_iter_next(iter)) {
 276                 kfree(iter);
 277                 return NULL;
 278         }
 279
 280         return iter;
 281 }
 282
 283 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
 284 {
 285         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
 286         struct rb_node *n;
 287         struct ipoib_path *path;
 288         int ret = 1;
 289
 290         spin_lock_irq(&priv->lock);
 291
 292         n = rb_first(&priv->path_tree);
 293
 294         while (n) {
 295                 path = rb_entry(n, struct ipoib_path, rb_node);
 296
 297                 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
 298                            sizeof (union ib_gid)) < 0) {
 299                         iter->path = *path;
 300                         ret = 0;
 301                         break;
 302                 }
 303
 304                 n = rb_next(n);
 305         }
 306
 307         spin_unlock_irq(&priv->lock);
 308
 309         return ret;
 310 }
 311
 312 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
 313                           struct ipoib_path *path)
 314 {
 315         *path = iter->path;
 316 }
 317
 318 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
 319
 320 void ipoib_flush_paths(struct net_device *dev)
 321 {
 322         struct ipoib_dev_priv *priv = netdev_priv(dev);
 323         struct ipoib_path *path, *tp;
 324         LIST_HEAD(remove_list);
 325         unsigned long flags;
 326
 327         spin_lock_irqsave(&priv->lock, flags);
 328
 329         list_splice(&priv->path_list, &remove_list);
 330         INIT_LIST_HEAD(&priv->path_list);
 331
 332         list_for_each_entry(path, &remove_list, list)
 333                 rb_erase(&path->rb_node, &priv->path_tree);
 334
 335         spin_unlock_irqrestore(&priv->lock, flags);
 336
 337         list_for_each_entry_safe(path, tp, &remove_list, list) {
 338                 if (path->query)
 339                         ib_sa_cancel_query(path->query_id, path->query);
 340                 wait_for_completion(&path->done);
 341                 path_free(dev, path);
 342         }
 343 }
 344
 345 static void path_rec_completion(int status,
 346                                 struct ib_sa_path_rec *pathrec,
 347                                 void *path_ptr)
 348 {
 349         struct ipoib_path *path = path_ptr;
 350         struct net_device *dev = path->dev;
 351         struct ipoib_dev_priv *priv = netdev_priv(dev);
 352         struct ipoib_ah *ah = NULL;
 353         struct ipoib_neigh *neigh;
 354         struct sk_buff_head skqueue;
 355         struct sk_buff *skb;
 356         unsigned long flags;
 357
 358         if (pathrec)
 359                 ipoib_dbg(priv, "PathRec LID 0x%04x for GID " IPOIB_GID_FMT "\n",
 360                           be16_to_cpu(pathrec->dlid), IPOIB_GID_ARG(pathrec->dgid));
 361         else
 362                 ipoib_dbg(priv, "PathRec status %d for GID " IPOIB_GID_FMT "\n",
 363                           status, IPOIB_GID_ARG(path->pathrec.dgid));
 364
 365         skb_queue_head_init(&skqueue);
 366
 367         if (!status) {
 368                 struct ib_ah_attr av = {
 369                         .dlid          = be16_to_cpu(pathrec->dlid),
 370                         .sl            = pathrec->sl,
 371                         .port_num      = priv->port
 372                 };
 373                 int path_rate = ib_sa_rate_enum_to_int(pathrec->rate);
 374
 375                 if (path_rate > 0 && priv->local_rate > path_rate)
 376                         av.static_rate = (priv->local_rate - 1) / path_rate;
 377
 378                 ipoib_dbg(priv, "static_rate %d for local port %dX, path %dX\n",
 379                           av.static_rate, priv->local_rate,
 380                           ib_sa_rate_enum_to_int(pathrec->rate));
 381
 382                 ah = ipoib_create_ah(dev, priv->pd, &av);
 383         }
 384
 385         spin_lock_irqsave(&priv->lock, flags);
 386
 387         path->ah = ah;
 388
 389         if (ah) {
 390                 path->pathrec = *pathrec;
 391
 392                 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
 393                           ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
 394
 395                 while ((skb = __skb_dequeue(&path->queue)))
 396                         __skb_queue_tail(&skqueue, skb);
 397
 398                 list_for_each_entry(neigh, &path->neigh_list, list) {
 399                         kref_get(&path->ah->ref);
 400                         neigh->ah = path->ah;
 401
 402                         while ((skb = __skb_dequeue(&neigh->queue)))
 403                                 __skb_queue_tail(&skqueue, skb);
 404                 }
 405         }
 406
 407         path->query = NULL;
 408         complete(&path->done);
 409
 410         spin_unlock_irqrestore(&priv->lock, flags);
 411
 412         while ((skb = __skb_dequeue(&skqueue))) {
 413                 skb->dev = dev;
 414                 if (dev_queue_xmit(skb))
 415                         ipoib_warn(priv, "dev_queue_xmit failed "
 416                                    "to requeue packet\n");
 417         }
 418 }
 419
 420 static struct ipoib_path *path_rec_create(struct net_device *dev,
 421                                           union ib_gid *gid)
 422 {
 423         struct ipoib_dev_priv *priv = netdev_priv(dev);
 424         struct ipoib_path *path;
 425
 426         path = kzalloc(sizeof *path, GFP_ATOMIC);
 427         if (!path)
 428                 return NULL;
 429
 430         path->dev = dev;
 431
 432         skb_queue_head_init(&path->queue);
 433
 434         INIT_LIST_HEAD(&path->neigh_list);
 435
 436         memcpy(path->pathrec.dgid.raw, gid->raw, sizeof (union ib_gid));
 437         path->pathrec.sgid      = priv->local_gid;
 438         path->pathrec.pkey      = cpu_to_be16(priv->pkey);
 439         path->pathrec.numb_path = 1;
 440
 441         return path;
 442 }
 443
 444 static int path_rec_start(struct net_device *dev,
 445                           struct ipoib_path *path)
 446 {
 447         struct ipoib_dev_priv *priv = netdev_priv(dev);
 448
 449         ipoib_dbg(priv, "Start path record lookup for " IPOIB_GID_FMT "\n",
 450                   IPOIB_GID_ARG(path->pathrec.dgid));
 451
 452         init_completion(&path->done);
 453
 454         path->query_id =
 455                 ib_sa_path_rec_get(priv->ca, priv->port,
 456                                    &path->pathrec,
 457                                    IB_SA_PATH_REC_DGID          |
 458                                    IB_SA_PATH_REC_SGID          |
 459                                    IB_SA_PATH_REC_NUMB_PATH     |
 460                                    IB_SA_PATH_REC_PKEY,
 461                                    1000, GFP_ATOMIC,
 462                                    path_rec_completion,
 463                                    path, &path->query);
 464         if (path->query_id < 0) {
 465                 ipoib_warn(priv, "ib_sa_path_rec_get failed\n");
 466                 path->query = NULL;
 467                 return path->query_id;
 468         }
 469
 470         return 0;
 471 }
 472
 473 static void neigh_add_path(struct sk_buff *skb, struct net_device *dev)
 474 {
 475         struct ipoib_dev_priv *priv = netdev_priv(dev);
 476         struct ipoib_path *path;
 477         struct ipoib_neigh *neigh;
 478
 479         neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
 480         if (!neigh) {
 481                 ++priv->stats.tx_dropped;
 482                 dev_kfree_skb_any(skb);
 483                 return;
 484         }
 485
 486         skb_queue_head_init(&neigh->queue);
 487         neigh->neighbour = skb->dst->neighbour;
 488         *to_ipoib_neigh(skb->dst->neighbour) = neigh;
 489
 490         /*
 491          * We can only be called from ipoib_start_xmit, so we're
 492          * inside tx_lock -- no need to save/restore flags.
 493          */
 494         spin_lock(&priv->lock);
 495
 496         path = __path_find(dev, (union ib_gid *) (skb->dst->neighbour->ha + 4));
 497         if (!path) {
 498                 path = path_rec_create(dev,
 499                                        (union ib_gid *) (skb->dst->neighbour->ha + 4));
 500                 if (!path)
 501                         goto err;
 502
 503                 __path_add(dev, path);
 504         }
 505
 506         list_add_tail(&neigh->list, &path->neigh_list);
 507
 508         if (path->ah) {
 509                 kref_get(&path->ah->ref);
 510                 neigh->ah = path->ah;
 511
 512                 ipoib_send(dev, skb, path->ah,
 513                            be32_to_cpup((__be32 *) skb->dst->neighbour->ha));
 514         } else {
 515                 neigh->ah  = NULL;
 516                 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 517                         __skb_queue_tail(&neigh->queue, skb);
 518                 } else {
 519                         ++priv->stats.tx_dropped;
 520                         dev_kfree_skb_any(skb);
 521                 }
 522
 523                 if (!path->query && path_rec_start(dev, path))
 524                         goto err;
 525         }
 526
 527         spin_unlock(&priv->lock);
 528         return;
 529
 530 err:
 531         *to_ipoib_neigh(skb->dst->neighbour) = NULL;
 532         list_del(&neigh->list);
 533         neigh->neighbour->ops->destructor = NULL;
 534         kfree(neigh);
 535
 536         ++priv->stats.tx_dropped;
 537         dev_kfree_skb_any(skb);
 538
 539         spin_unlock(&priv->lock);
 540 }
 541
 542 static void ipoib_path_lookup(struct sk_buff *skb, struct net_device *dev)
 543 {
 544         struct ipoib_dev_priv *priv = netdev_priv(skb->dev);
 545
 546         /* Look up path record for unicasts */
 547         if (skb->dst->neighbour->ha[4] != 0xff) {
 548                 neigh_add_path(skb, dev);
 549                 return;
 550         }
 551
 552         /* Add in the P_Key for multicasts */
 553         skb->dst->neighbour->ha[8] = (priv->pkey >> 8) & 0xff;
 554         skb->dst->neighbour->ha[9] = priv->pkey & 0xff;
 555         ipoib_mcast_send(dev, (union ib_gid *) (skb->dst->neighbour->ha + 4), skb);
 556 }
 557
 558 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
 559                              struct ipoib_pseudoheader *phdr)
 560 {
 561         struct ipoib_dev_priv *priv = netdev_priv(dev);
 562         struct ipoib_path *path;
 563
 564         /*
 565          * We can only be called from ipoib_start_xmit, so we're
 566          * inside tx_lock -- no need to save/restore flags.
 567          */
 568         spin_lock(&priv->lock);
 569
 570         path = __path_find(dev, (union ib_gid *) (phdr->hwaddr + 4));
 571         if (!path) {
 572                 path = path_rec_create(dev,
 573                                        (union ib_gid *) (phdr->hwaddr + 4));
 574                 if (path) {
 575                         /* put pseudoheader back on for next time */
 576                         skb_push(skb, sizeof *phdr);
 577                         __skb_queue_tail(&path->queue, skb);
 578
 579                         if (path_rec_start(dev, path)) {
 580                                 spin_unlock(&priv->lock);
 581                                 path_free(dev, path);
 582                                 return;
 583                         } else
 584                                 __path_add(dev, path);
 585                 } else {
 586                         ++priv->stats.tx_dropped;
 587                         dev_kfree_skb_any(skb);
 588                 }
 589
 590                 spin_unlock(&priv->lock);
 591                 return;
 592         }
 593
 594         if (path->ah) {
 595                 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
 596                           be16_to_cpu(path->pathrec.dlid));
 597
 598                 ipoib_send(dev, skb, path->ah,
 599                            be32_to_cpup((__be32 *) phdr->hwaddr));
 600         } else if ((path->query || !path_rec_start(dev, path)) &&
 601                    skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 602                 /* put pseudoheader back on for next time */
 603                 skb_push(skb, sizeof *phdr);
 604                 __skb_queue_tail(&path->queue, skb);
 605         } else {
 606                 ++priv->stats.tx_dropped;
 607                 dev_kfree_skb_any(skb);
 608         }
 609
 610         spin_unlock(&priv->lock);
 611 }
 612
 613 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
 614 {
 615         struct ipoib_dev_priv *priv = netdev_priv(dev);
 616         struct ipoib_neigh *neigh;
 617         unsigned long flags;
 618
 619         if (!spin_trylock_irqsave(&priv->tx_lock, flags))
 620                 return NETDEV_TX_LOCKED;
 621
 622         /*
 623          * Check if our queue is stopped.  Since we have the LLTX bit
 624          * set, we can't rely on netif_stop_queue() preventing our
 625          * xmit function from being called with a full queue.
 626          */
 627         if (unlikely(netif_queue_stopped(dev))) {
 628                 spin_unlock_irqrestore(&priv->tx_lock, flags);
 629                 return NETDEV_TX_BUSY;
 630         }
 631
 632         if (skb->dst && skb->dst->neighbour) {
 633                 if (unlikely(!*to_ipoib_neigh(skb->dst->neighbour))) {
 634                         ipoib_path_lookup(skb, dev);
 635                         goto out;
 636                 }
 637
 638                 neigh = *to_ipoib_neigh(skb->dst->neighbour);
 639
 640                 if (likely(neigh->ah)) {
 641                         ipoib_send(dev, skb, neigh->ah,
 642                                    be32_to_cpup((__be32 *) skb->dst->neighbour->ha));
 643                         goto out;
 644                 }
 645
 646                 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 647                         spin_lock(&priv->lock);
 648                         __skb_queue_tail(&neigh->queue, skb);
 649                         spin_unlock(&priv->lock);
 650                 } else {
 651                         ++priv->stats.tx_dropped;
 652                         dev_kfree_skb_any(skb);
 653                 }
 654         } else {
 655                 struct ipoib_pseudoheader *phdr =
 656                         (struct ipoib_pseudoheader *) skb->data;
 657                 skb_pull(skb, sizeof *phdr);
 658
 659                 if (phdr->hwaddr[4] == 0xff) {
 660                         /* Add in the P_Key for multicast*/
 661                         phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
 662                         phdr->hwaddr[9] = priv->pkey & 0xff;
 663
 664                         ipoib_mcast_send(dev, (union ib_gid *) (phdr->hwaddr + 4), skb);
 665                 } else {
 666                         /* unicast GID -- should be ARP or RARP reply */
 667
 668                         if ((be16_to_cpup((__be16 *) skb->data) != ETH_P_ARP) &&
 669                             (be16_to_cpup((__be16 *) skb->data) != ETH_P_RARP)) {
 670                                 ipoib_warn(priv, "Unicast, no %s: type %04x, QPN %06x "
 671                                            IPOIB_GID_FMT "\n",
 672                                            skb->dst ? "neigh" : "dst",
 673                                            be16_to_cpup((__be16 *) skb->data),
 674                                            be32_to_cpup((__be32 *) phdr->hwaddr),
 675                                            IPOIB_GID_ARG(*(union ib_gid *) (phdr->hwaddr + 4)));
 676                                 dev_kfree_skb_any(skb);
 677                                 ++priv->stats.tx_dropped;
 678                                 goto out;
 679                         }
 680
 681                         unicast_arp_send(skb, dev, phdr);
 682                 }
 683         }
 684
 685 out:
 686         spin_unlock_irqrestore(&priv->tx_lock, flags);
 687
 688         return NETDEV_TX_OK;
 689 }
 690
 691 static struct net_device_stats *ipoib_get_stats(struct net_device *dev)
 692 {
 693         struct ipoib_dev_priv *priv = netdev_priv(dev);
 694
 695         return &priv->stats;
 696 }
 697
 698 static void ipoib_timeout(struct net_device *dev)
 699 {
 700         struct ipoib_dev_priv *priv = netdev_priv(dev);
 701
 702         ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
 703                    jiffies_to_msecs(jiffies - dev->trans_start));
 704         ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
 705                    netif_queue_stopped(dev),
 706                    priv->tx_head, priv->tx_tail);
 707         /* XXX reset QP, etc. */
 708 }
 709
 710 static int ipoib_hard_header(struct sk_buff *skb,
 711                              struct net_device *dev,
 712                              unsigned short type,
 713                              void *daddr, void *saddr, unsigned len)
 714 {
 715         struct ipoib_header *header;
 716
 717         header = (struct ipoib_header *) skb_push(skb, sizeof *header);
 718
 719         header->proto = htons(type);
 720         header->reserved = 0;
 721
 722         /*
 723          * If we don't have a neighbour structure, stuff the
 724          * destination address onto the front of the skb so we can
 725          * figure out where to send the packet later.
 726          */
 727         if (!skb->dst || !skb->dst->neighbour) {
 728                 struct ipoib_pseudoheader *phdr =
 729                         (struct ipoib_pseudoheader *) skb_push(skb, sizeof *phdr);
 730                 memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
 731         }
 732
 733         return 0;
 734 }
 735
 736 static void ipoib_set_mcast_list(struct net_device *dev)
 737 {
 738         struct ipoib_dev_priv *priv = netdev_priv(dev);
 739
 740         queue_work(ipoib_workqueue, &priv->restart_task);
 741 }
 742
 743 static void ipoib_neigh_destructor(struct neighbour *n)
 744 {
 745         struct ipoib_neigh *neigh;
 746         struct ipoib_dev_priv *priv = netdev_priv(n->dev);
 747         unsigned long flags;
 748         struct ipoib_ah *ah = NULL;
 749
 750         ipoib_dbg(priv,
 751                   "neigh_destructor for %06x " IPOIB_GID_FMT "\n",
 752                   be32_to_cpup((__be32 *) n->ha),
 753                   IPOIB_GID_ARG(*((union ib_gid *) (n->ha + 4))));
 754
 755         spin_lock_irqsave(&priv->lock, flags);
 756
 757         neigh = *to_ipoib_neigh(n);
 758         if (neigh) {
 759                 if (neigh->ah)
 760                         ah = neigh->ah;
 761                 list_del(&neigh->list);
 762                 *to_ipoib_neigh(n) = NULL;
 763                 kfree(neigh);
 764         }
 765
 766         spin_unlock_irqrestore(&priv->lock, flags);
 767
 768         if (ah)
 769                 ipoib_put_ah(ah);
 770 }
 771
 772 static int ipoib_neigh_setup(struct neighbour *neigh)
 773 {
 774         /*
 775          * Is this kosher?  I can't find anybody in the kernel that
 776          * sets neigh->destructor, so we should be able to set it here
 777          * without trouble.
 778          */
 779         neigh->ops->destructor = ipoib_neigh_destructor;
 780
 781         return 0;
 782 }
 783
 784 static int ipoib_neigh_setup_dev(struct net_device *dev, struct neigh_parms *parms)
 785 {
 786         parms->neigh_setup = ipoib_neigh_setup;
 787
 788         return 0;
 789 }
 790
 791 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
 792 {
 793         struct ipoib_dev_priv *priv = netdev_priv(dev);
 794
 795         /* Allocate RX/TX "rings" to hold queued skbs */
 796
 797         priv->rx_ring = kzalloc(IPOIB_RX_RING_SIZE * sizeof (struct ipoib_rx_buf),
 798                                 GFP_KERNEL);
 799         if (!priv->rx_ring) {
 800                 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
 801                        ca->name, IPOIB_RX_RING_SIZE);
 802                 goto out;
 803         }
 804
 805         priv->tx_ring = kzalloc(IPOIB_TX_RING_SIZE * sizeof (struct ipoib_tx_buf),
 806                                 GFP_KERNEL);
 807         if (!priv->tx_ring) {
 808                 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
 809                        ca->name, IPOIB_TX_RING_SIZE);
 810                 goto out_rx_ring_cleanup;
 811         }
 812
 813         /* priv->tx_head & tx_tail are already 0 */
 814
 815         if (ipoib_ib_dev_init(dev, ca, port))
 816                 goto out_tx_ring_cleanup;
 817
 818         return 0;
 819
 820 out_tx_ring_cleanup:
 821         kfree(priv->tx_ring);
 822
 823 out_rx_ring_cleanup:
 824         kfree(priv->rx_ring);
 825
 826 out:
 827         return -ENOMEM;
 828 }
 829
 830 void ipoib_dev_cleanup(struct net_device *dev)
 831 {
 832         struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
 833
 834         ipoib_delete_debug_files(dev);
 835
 836         /* Delete any child interfaces first */
 837         list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
 838                 unregister_netdev(cpriv->dev);
 839                 ipoib_dev_cleanup(cpriv->dev);
 840                 free_netdev(cpriv->dev);
 841         }
 842
 843         ipoib_ib_dev_cleanup(dev);
 844
 845         kfree(priv->rx_ring);
 846         kfree(priv->tx_ring);
 847
 848         priv->rx_ring = NULL;
 849         priv->tx_ring = NULL;
 850 }
 851
 852 static void ipoib_setup(struct net_device *dev)
 853 {
 854         struct ipoib_dev_priv *priv = netdev_priv(dev);
 855
 856         dev->open                = ipoib_open;
 857         dev->stop                = ipoib_stop;
 858         dev->change_mtu          = ipoib_change_mtu;
 859         dev->hard_start_xmit     = ipoib_start_xmit;
 860         dev->get_stats           = ipoib_get_stats;
 861         dev->tx_timeout          = ipoib_timeout;
 862         dev->hard_header         = ipoib_hard_header;
 863         dev->set_multicast_list  = ipoib_set_mcast_list;
 864         dev->neigh_setup         = ipoib_neigh_setup_dev;
 865
 866         dev->watchdog_timeo      = HZ;
 867
 868         dev->flags              |= IFF_BROADCAST | IFF_MULTICAST;
 869
 870         /*
 871          * We add in INFINIBAND_ALEN to allow for the destination
 872          * address "pseudoheader" for skbs without neighbour struct.
 873          */
 874         dev->hard_header_len     = IPOIB_ENCAP_LEN + INFINIBAND_ALEN;
 875         dev->addr_len            = INFINIBAND_ALEN;
 876         dev->type                = ARPHRD_INFINIBAND;
 877         dev->tx_queue_len        = IPOIB_TX_RING_SIZE * 2;
 878         dev->features            = NETIF_F_VLAN_CHALLENGED | NETIF_F_LLTX;
 879
 880         /* MTU will be reset when mcast join happens */
 881         dev->mtu                 = IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN;
 882         priv->mcast_mtu          = priv->admin_mtu = dev->mtu;
 883
 884         memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
 885
 886         netif_carrier_off(dev);
 887
 888         SET_MODULE_OWNER(dev);
 889
 890         priv->dev = dev;
 891
 892         spin_lock_init(&priv->lock);
 893         spin_lock_init(&priv->tx_lock);
 894
 895         mutex_init(&priv->mcast_mutex);
 896         mutex_init(&priv->vlan_mutex);
 897
 898         INIT_LIST_HEAD(&priv->path_list);
 899         INIT_LIST_HEAD(&priv->child_intfs);
 900         INIT_LIST_HEAD(&priv->dead_ahs);
 901         INIT_LIST_HEAD(&priv->multicast_list);
 902
 903         INIT_WORK(&priv->pkey_task,    ipoib_pkey_poll,          priv->dev);
 904         INIT_WORK(&priv->mcast_task,   ipoib_mcast_join_task,    priv->dev);
 905         INIT_WORK(&priv->flush_task,   ipoib_ib_dev_flush,       priv->dev);
 906         INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task, priv->dev);
 907         INIT_WORK(&priv->ah_reap_task, ipoib_reap_ah,            priv->dev);
 908 }
 909
 910 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
 911 {
 912         struct net_device *dev;
 913
 914         dev = alloc_netdev((int) sizeof (struct ipoib_dev_priv), name,
 915                            ipoib_setup);
 916         if (!dev)
 917                 return NULL;
 918
 919         return netdev_priv(dev);
 920 }
 921
 922 static ssize_t show_pkey(struct class_device *cdev, char *buf)
 923 {
 924         struct ipoib_dev_priv *priv =
 925                 netdev_priv(container_of(cdev, struct net_device, class_dev));
 926
 927         return sprintf(buf, "0x%04x\n", priv->pkey);
 928 }
 929 static CLASS_DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
 930
 931 static ssize_t create_child(struct class_device *cdev,
 932                             const char *buf, size_t count)
 933 {
 934         int pkey;
 935         int ret;
 936
 937         if (sscanf(buf, "%i", &pkey) != 1)
 938                 return -EINVAL;
 939
 940         if (pkey < 0 || pkey > 0xffff)
 941                 return -EINVAL;
 942
 943         /*
 944          * Set the full membership bit, so that we join the right
 945          * broadcast group, etc.
 946          */
 947         pkey |= 0x8000;
 948
 949         ret = ipoib_vlan_add(container_of(cdev, struct net_device, class_dev),
 950                              pkey);
 951
 952         return ret ? ret : count;
 953 }
 954 static CLASS_DEVICE_ATTR(create_child, S_IWUGO, NULL, create_child);
 955
 956 static ssize_t delete_child(struct class_device *cdev,
 957                             const char *buf, size_t count)
 958 {
 959         int pkey;
 960         int ret;
 961
 962         if (sscanf(buf, "%i", &pkey) != 1)
 963                 return -EINVAL;
 964
 965         if (pkey < 0 || pkey > 0xffff)
 966                 return -EINVAL;
 967
 968         ret = ipoib_vlan_delete(container_of(cdev, struct net_device, class_dev),
 969                                 pkey);
 970
 971         return ret ? ret : count;
 972
 973 }
 974 static CLASS_DEVICE_ATTR(delete_child, S_IWUGO, NULL, delete_child);
 975
 976 int ipoib_add_pkey_attr(struct net_device *dev)
 977 {
 978         return class_device_create_file(&dev->class_dev,
 979                                         &class_device_attr_pkey);
 980 }
 981
 982 static struct net_device *ipoib_add_port(const char *format,
 983                                          struct ib_device *hca, u8 port)
 984 {
 985         struct ipoib_dev_priv *priv;
 986         int result = -ENOMEM;
 987
 988         priv = ipoib_intf_alloc(format);
 989         if (!priv)
 990                 goto alloc_mem_failed;
 991
 992         SET_NETDEV_DEV(priv->dev, hca->dma_device);
 993
 994         result = ib_query_pkey(hca, port, 0, &priv->pkey);
 995         if (result) {
 996                 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
 997                        hca->name, port, result);
 998                 goto alloc_mem_failed;
 999         }
1000
1001         /*
1002          * Set the full membership bit, so that we join the right
1003          * broadcast group, etc.
1004          */
1005         priv->pkey |= 0x8000;
1006
1007         priv->dev->broadcast[8] = priv->pkey >> 8;
1008         priv->dev->broadcast[9] = priv->pkey & 0xff;
1009
1010         result = ib_query_gid(hca, port, 0, &priv->local_gid);
1011         if (result) {
1012                 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1013                        hca->name, port, result);
1014                 goto alloc_mem_failed;
1015         } else
1016                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1017
1018
1019         result = ipoib_dev_init(priv->dev, hca, port);
1020         if (result < 0) {
1021                 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1022                        hca->name, port, result);
1023                 goto device_init_failed;
1024         }
1025
1026         INIT_IB_EVENT_HANDLER(&priv->event_handler,
1027                               priv->ca, ipoib_event);
1028         result = ib_register_event_handler(&priv->event_handler);
1029         if (result < 0) {
1030                 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1031                        "port %d (ret = %d)\n",
1032                        hca->name, port, result);
1033                 goto event_failed;
1034         }
1035
1036         result = register_netdev(priv->dev);
1037         if (result) {
1038                 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1039                        hca->name, port, result);
1040                 goto register_failed;
1041         }
1042
1043         ipoib_create_debug_files(priv->dev);
1044
1045         if (ipoib_add_pkey_attr(priv->dev))
1046                 goto sysfs_failed;
1047         if (class_device_create_file(&priv->dev->class_dev,
1048                                      &class_device_attr_create_child))
1049                 goto sysfs_failed;
1050         if (class_device_create_file(&priv->dev->class_dev,
1051                                      &class_device_attr_delete_child))
1052                 goto sysfs_failed;
1053
1054         return priv->dev;
1055
1056 sysfs_failed:
1057         ipoib_delete_debug_files(priv->dev);
1058         unregister_netdev(priv->dev);
1059
1060 register_failed:
1061         ib_unregister_event_handler(&priv->event_handler);
1062         flush_scheduled_work();
1063
1064 event_failed:
1065         ipoib_dev_cleanup(priv->dev);
1066
1067 device_init_failed:
1068         free_netdev(priv->dev);
1069
1070 alloc_mem_failed:
1071         return ERR_PTR(result);
1072 }
1073
1074 static void ipoib_add_one(struct ib_device *device)
1075 {
1076         struct list_head *dev_list;
1077         struct net_device *dev;
1078         struct ipoib_dev_priv *priv;
1079         int s, e, p;
1080
1081         dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1082         if (!dev_list)
1083                 return;
1084
1085         INIT_LIST_HEAD(dev_list);
1086
1087         if (device->node_type == IB_NODE_SWITCH) {
1088                 s = 0;
1089                 e = 0;
1090         } else {
1091                 s = 1;
1092                 e = device->phys_port_cnt;
1093         }
1094
1095         for (p = s; p <= e; ++p) {
1096                 dev = ipoib_add_port("ib%d", device, p);
1097                 if (!IS_ERR(dev)) {
1098                         priv = netdev_priv(dev);
1099                         list_add_tail(&priv->list, dev_list);
1100                 }
1101         }
1102
1103         ib_set_client_data(device, &ipoib_client, dev_list);
1104 }
1105
1106 static void ipoib_remove_one(struct ib_device *device)
1107 {
1108         struct ipoib_dev_priv *priv, *tmp;
1109         struct list_head *dev_list;
1110
1111         dev_list = ib_get_client_data(device, &ipoib_client);
1112
1113         list_for_each_entry_safe(priv, tmp, dev_list, list) {
1114                 ib_unregister_event_handler(&priv->event_handler);
1115                 flush_scheduled_work();
1116
1117                 unregister_netdev(priv->dev);
1118                 ipoib_dev_cleanup(priv->dev);
1119                 free_netdev(priv->dev);
1120         }
1121
1122         kfree(dev_list);
1123 }
1124
1125 static int __init ipoib_init_module(void)
1126 {
1127         int ret;
1128
1129         ret = ipoib_register_debugfs();
1130         if (ret)
1131                 return ret;
1132
1133         /*
1134          * We create our own workqueue mainly because we want to be
1135          * able to flush it when devices are being removed.  We can't
1136          * use schedule_work()/flush_scheduled_work() because both
1137          * unregister_netdev() and linkwatch_event take the rtnl lock,
1138          * so flush_scheduled_work() can deadlock during device
1139          * removal.
1140          */
1141         ipoib_workqueue = create_singlethread_workqueue("ipoib");
1142         if (!ipoib_workqueue) {
1143                 ret = -ENOMEM;
1144                 goto err_fs;
1145         }
1146
1147         ret = ib_register_client(&ipoib_client);
1148         if (ret)
1149                 goto err_wq;
1150
1151         return 0;
1152
1153 err_wq:
1154         destroy_workqueue(ipoib_workqueue);
1155
1156 err_fs:
1157         ipoib_unregister_debugfs();
1158
1159         return ret;
1160 }
1161
1162 static void __exit ipoib_cleanup_module(void)
1163 {
1164         ib_unregister_client(&ipoib_client);
1165         ipoib_unregister_debugfs();
1166         destroy_workqueue(ipoib_workqueue);
1167 }
1168
1169 module_init(ipoib_init_module);
1170 module_exit(ipoib_cleanup_module);