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 #include <linux/kernel.h>
  45
  46 #include <linux/if_arp.h>       /* For ARPHRD_xxx */
  47
  48 #include <linux/ip.h>
  49 #include <linux/in.h>
  50
  51 #include <net/dst.h>
  52
  53 MODULE_AUTHOR("Roland Dreier");
  54 MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
  55 MODULE_LICENSE("Dual BSD/GPL");
  56
  57 int ipoib_sendq_size __read_mostly = IPOIB_TX_RING_SIZE;
  58 int ipoib_recvq_size __read_mostly = IPOIB_RX_RING_SIZE;
  59
  60 module_param_named(send_queue_size, ipoib_sendq_size, int, 0444);
  61 MODULE_PARM_DESC(send_queue_size, "Number of descriptors in send queue");
  62 module_param_named(recv_queue_size, ipoib_recvq_size, int, 0444);
  63 MODULE_PARM_DESC(recv_queue_size, "Number of descriptors in receive queue");
  64
  65 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
  66 int ipoib_debug_level;
  67
  68 module_param_named(debug_level, ipoib_debug_level, int, 0644);
  69 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
  70 #endif
  71
  72 struct ipoib_path_iter {
  73         struct net_device *dev;
  74         struct ipoib_path  path;
  75 };
  76
  77 static const u8 ipv4_bcast_addr[] = {
  78         0x00, 0xff, 0xff, 0xff,
  79         0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
  80         0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff
  81 };
  82
  83 struct workqueue_struct *ipoib_workqueue;
  84
  85 static void ipoib_add_one(struct ib_device *device);
  86 static void ipoib_remove_one(struct ib_device *device);
  87
  88 static struct ib_client ipoib_client = {
  89         .name   = "ipoib",
  90         .add    = ipoib_add_one,
  91         .remove = ipoib_remove_one
  92 };
  93
  94 int ipoib_open(struct net_device *dev)
  95 {
  96         struct ipoib_dev_priv *priv = netdev_priv(dev);
  97
  98         ipoib_dbg(priv, "bringing up interface\n");
  99
 100         set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 101
 102         if (ipoib_pkey_dev_delay_open(dev))
 103                 return 0;
 104
 105         if (ipoib_ib_dev_open(dev))
 106                 return -EINVAL;
 107
 108         if (ipoib_ib_dev_up(dev)) {
 109                 ipoib_ib_dev_stop(dev);
 110                 return -EINVAL;
 111         }
 112
 113         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 114                 struct ipoib_dev_priv *cpriv;
 115
 116                 /* Bring up any child interfaces too */
 117                 mutex_lock(&priv->vlan_mutex);
 118                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 119                         int flags;
 120
 121                         flags = cpriv->dev->flags;
 122                         if (flags & IFF_UP)
 123                                 continue;
 124
 125                         dev_change_flags(cpriv->dev, flags | IFF_UP);
 126                 }
 127                 mutex_unlock(&priv->vlan_mutex);
 128         }
 129
 130         netif_start_queue(dev);
 131
 132         return 0;
 133 }
 134
 135 static int ipoib_stop(struct net_device *dev)
 136 {
 137         struct ipoib_dev_priv *priv = netdev_priv(dev);
 138
 139         ipoib_dbg(priv, "stopping interface\n");
 140
 141         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 142
 143         netif_stop_queue(dev);
 144
 145         /*
 146          * Now flush workqueue to make sure a scheduled task doesn't
 147          * bring our internal state back up.
 148          */
 149         flush_workqueue(ipoib_workqueue);
 150
 151         ipoib_ib_dev_down(dev, 1);
 152         ipoib_ib_dev_stop(dev);
 153
 154         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 155                 struct ipoib_dev_priv *cpriv;
 156
 157                 /* Bring down any child interfaces too */
 158                 mutex_lock(&priv->vlan_mutex);
 159                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 160                         int flags;
 161
 162                         flags = cpriv->dev->flags;
 163                         if (!(flags & IFF_UP))
 164                                 continue;
 165
 166                         dev_change_flags(cpriv->dev, flags & ~IFF_UP);
 167                 }
 168                 mutex_unlock(&priv->vlan_mutex);
 169         }
 170
 171         return 0;
 172 }
 173
 174 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
 175 {
 176         struct ipoib_dev_priv *priv = netdev_priv(dev);
 177
 178         if (new_mtu > IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN)
 179                 return -EINVAL;
 180
 181         priv->admin_mtu = new_mtu;
 182
 183         dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
 184
 185         return 0;
 186 }
 187
 188 static struct ipoib_path *__path_find(struct net_device *dev,
 189                                       union ib_gid *gid)
 190 {
 191         struct ipoib_dev_priv *priv = netdev_priv(dev);
 192         struct rb_node *n = priv->path_tree.rb_node;
 193         struct ipoib_path *path;
 194         int ret;
 195
 196         while (n) {
 197                 path = rb_entry(n, struct ipoib_path, rb_node);
 198
 199                 ret = memcmp(gid->raw, path->pathrec.dgid.raw,
 200                              sizeof (union ib_gid));
 201
 202                 if (ret < 0)
 203                         n = n->rb_left;
 204                 else if (ret > 0)
 205                         n = n->rb_right;
 206                 else
 207                         return path;
 208         }
 209
 210         return NULL;
 211 }
 212
 213 static int __path_add(struct net_device *dev, struct ipoib_path *path)
 214 {
 215         struct ipoib_dev_priv *priv = netdev_priv(dev);
 216         struct rb_node **n = &priv->path_tree.rb_node;
 217         struct rb_node *pn = NULL;
 218         struct ipoib_path *tpath;
 219         int ret;
 220
 221         while (*n) {
 222                 pn = *n;
 223                 tpath = rb_entry(pn, struct ipoib_path, rb_node);
 224
 225                 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
 226                              sizeof (union ib_gid));
 227                 if (ret < 0)
 228                         n = &pn->rb_left;
 229                 else if (ret > 0)
 230                         n = &pn->rb_right;
 231                 else
 232                         return -EEXIST;
 233         }
 234
 235         rb_link_node(&path->rb_node, pn, n);
 236         rb_insert_color(&path->rb_node, &priv->path_tree);
 237
 238         list_add_tail(&path->list, &priv->path_list);
 239
 240         return 0;
 241 }
 242
 243 static void path_free(struct net_device *dev, struct ipoib_path *path)
 244 {
 245         struct ipoib_dev_priv *priv = netdev_priv(dev);
 246         struct ipoib_neigh *neigh, *tn;
 247         struct sk_buff *skb;
 248         unsigned long flags;
 249
 250         while ((skb = __skb_dequeue(&path->queue)))
 251                 dev_kfree_skb_irq(skb);
 252
 253         spin_lock_irqsave(&priv->lock, flags);
 254
 255         list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
 256                 /*
 257                  * It's safe to call ipoib_put_ah() inside priv->lock
 258                  * here, because we know that path->ah will always
 259                  * hold one more reference, so ipoib_put_ah() will
 260                  * never do more than decrement the ref count.
 261                  */
 262                 if (neigh->ah)
 263                         ipoib_put_ah(neigh->ah);
 264
 265                 ipoib_neigh_free(neigh);
 266         }
 267
 268         spin_unlock_irqrestore(&priv->lock, flags);
 269
 270         if (path->ah)
 271                 ipoib_put_ah(path->ah);
 272
 273         kfree(path);
 274 }
 275
 276 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 277
 278 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
 279 {
 280         struct ipoib_path_iter *iter;
 281
 282         iter = kmalloc(sizeof *iter, GFP_KERNEL);
 283         if (!iter)
 284                 return NULL;
 285
 286         iter->dev = dev;
 287         memset(iter->path.pathrec.dgid.raw, 0, 16);
 288
 289         if (ipoib_path_iter_next(iter)) {
 290                 kfree(iter);
 291                 return NULL;
 292         }
 293
 294         return iter;
 295 }
 296
 297 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
 298 {
 299         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
 300         struct rb_node *n;
 301         struct ipoib_path *path;
 302         int ret = 1;
 303
 304         spin_lock_irq(&priv->lock);
 305
 306         n = rb_first(&priv->path_tree);
 307
 308         while (n) {
 309                 path = rb_entry(n, struct ipoib_path, rb_node);
 310
 311                 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
 312                            sizeof (union ib_gid)) < 0) {
 313                         iter->path = *path;
 314                         ret = 0;
 315                         break;
 316                 }
 317
 318                 n = rb_next(n);
 319         }
 320
 321         spin_unlock_irq(&priv->lock);
 322
 323         return ret;
 324 }
 325
 326 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
 327                           struct ipoib_path *path)
 328 {
 329         *path = iter->path;
 330 }
 331
 332 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
 333
 334 void ipoib_flush_paths(struct net_device *dev)
 335 {
 336         struct ipoib_dev_priv *priv = netdev_priv(dev);
 337         struct ipoib_path *path, *tp;
 338         LIST_HEAD(remove_list);
 339
 340         spin_lock_irq(&priv->lock);
 341
 342         list_splice(&priv->path_list, &remove_list);
 343         INIT_LIST_HEAD(&priv->path_list);
 344
 345         list_for_each_entry(path, &remove_list, list)
 346                 rb_erase(&path->rb_node, &priv->path_tree);
 347
 348         list_for_each_entry_safe(path, tp, &remove_list, list) {
 349                 if (path->query)
 350                         ib_sa_cancel_query(path->query_id, path->query);
 351                 spin_unlock_irq(&priv->lock);
 352                 wait_for_completion(&path->done);
 353                 path_free(dev, path);
 354                 spin_lock_irq(&priv->lock);
 355         }
 356         spin_unlock_irq(&priv->lock);
 357 }
 358
 359 static void path_rec_completion(int status,
 360                                 struct ib_sa_path_rec *pathrec,
 361                                 void *path_ptr)
 362 {
 363         struct ipoib_path *path = path_ptr;
 364         struct net_device *dev = path->dev;
 365         struct ipoib_dev_priv *priv = netdev_priv(dev);
 366         struct ipoib_ah *ah = NULL;
 367         struct ipoib_neigh *neigh;
 368         struct sk_buff_head skqueue;
 369         struct sk_buff *skb;
 370         unsigned long flags;
 371
 372         if (pathrec)
 373                 ipoib_dbg(priv, "PathRec LID 0x%04x for GID " IPOIB_GID_FMT "\n",
 374                           be16_to_cpu(pathrec->dlid), IPOIB_GID_ARG(pathrec->dgid));
 375         else
 376                 ipoib_dbg(priv, "PathRec status %d for GID " IPOIB_GID_FMT "\n",
 377                           status, IPOIB_GID_ARG(path->pathrec.dgid));
 378
 379         skb_queue_head_init(&skqueue);
 380
 381         if (!status) {
 382                 struct ib_ah_attr av = {
 383                         .dlid          = be16_to_cpu(pathrec->dlid),
 384                         .sl            = pathrec->sl,
 385                         .port_num      = priv->port,
 386                         .static_rate   = pathrec->rate
 387                 };
 388
 389                 ah = ipoib_create_ah(dev, priv->pd, &av);
 390         }
 391
 392         spin_lock_irqsave(&priv->lock, flags);
 393
 394         path->ah = ah;
 395
 396         if (ah) {
 397                 path->pathrec = *pathrec;
 398
 399                 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
 400                           ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
 401
 402                 while ((skb = __skb_dequeue(&path->queue)))
 403                         __skb_queue_tail(&skqueue, skb);
 404
 405                 list_for_each_entry(neigh, &path->neigh_list, list) {
 406                         kref_get(&path->ah->ref);
 407                         neigh->ah = path->ah;
 408
 409                         while ((skb = __skb_dequeue(&neigh->queue)))
 410                                 __skb_queue_tail(&skqueue, skb);
 411                 }
 412         }
 413
 414         path->query = NULL;
 415         complete(&path->done);
 416
 417         spin_unlock_irqrestore(&priv->lock, flags);
 418
 419         while ((skb = __skb_dequeue(&skqueue))) {
 420                 skb->dev = dev;
 421                 if (dev_queue_xmit(skb))
 422                         ipoib_warn(priv, "dev_queue_xmit failed "
 423                                    "to requeue packet\n");
 424         }
 425 }
 426
 427 static struct ipoib_path *path_rec_create(struct net_device *dev,
 428                                           union ib_gid *gid)
 429 {
 430         struct ipoib_dev_priv *priv = netdev_priv(dev);
 431         struct ipoib_path *path;
 432
 433         path = kzalloc(sizeof *path, GFP_ATOMIC);
 434         if (!path)
 435                 return NULL;
 436
 437         path->dev = dev;
 438
 439         skb_queue_head_init(&path->queue);
 440
 441         INIT_LIST_HEAD(&path->neigh_list);
 442
 443         memcpy(path->pathrec.dgid.raw, gid->raw, sizeof (union ib_gid));
 444         path->pathrec.sgid      = priv->local_gid;
 445         path->pathrec.pkey      = cpu_to_be16(priv->pkey);
 446         path->pathrec.numb_path = 1;
 447
 448         return path;
 449 }
 450
 451 static int path_rec_start(struct net_device *dev,
 452                           struct ipoib_path *path)
 453 {
 454         struct ipoib_dev_priv *priv = netdev_priv(dev);
 455
 456         ipoib_dbg(priv, "Start path record lookup for " IPOIB_GID_FMT "\n",
 457                   IPOIB_GID_ARG(path->pathrec.dgid));
 458
 459         init_completion(&path->done);
 460
 461         path->query_id =
 462                 ib_sa_path_rec_get(priv->ca, priv->port,
 463                                    &path->pathrec,
 464                                    IB_SA_PATH_REC_DGID          |
 465                                    IB_SA_PATH_REC_SGID          |
 466                                    IB_SA_PATH_REC_NUMB_PATH     |
 467                                    IB_SA_PATH_REC_PKEY,
 468                                    1000, GFP_ATOMIC,
 469                                    path_rec_completion,
 470                                    path, &path->query);
 471         if (path->query_id < 0) {
 472                 ipoib_warn(priv, "ib_sa_path_rec_get failed\n");
 473                 path->query = NULL;
 474                 return path->query_id;
 475         }
 476
 477         return 0;
 478 }
 479
 480 static void neigh_add_path(struct sk_buff *skb, struct net_device *dev)
 481 {
 482         struct ipoib_dev_priv *priv = netdev_priv(dev);
 483         struct ipoib_path *path;
 484         struct ipoib_neigh *neigh;
 485
 486         neigh = ipoib_neigh_alloc(skb->dst->neighbour);
 487         if (!neigh) {
 488                 ++priv->stats.tx_dropped;
 489                 dev_kfree_skb_any(skb);
 490                 return;
 491         }
 492
 493         skb_queue_head_init(&neigh->queue);
 494
 495         /*
 496          * We can only be called from ipoib_start_xmit, so we're
 497          * inside tx_lock -- no need to save/restore flags.
 498          */
 499         spin_lock(&priv->lock);
 500
 501         path = __path_find(dev, (union ib_gid *) (skb->dst->neighbour->ha + 4));
 502         if (!path) {
 503                 path = path_rec_create(dev,
 504                                        (union ib_gid *) (skb->dst->neighbour->ha + 4));
 505                 if (!path)
 506                         goto err_path;
 507
 508                 __path_add(dev, path);
 509         }
 510
 511         list_add_tail(&neigh->list, &path->neigh_list);
 512
 513         if (path->ah) {
 514                 kref_get(&path->ah->ref);
 515                 neigh->ah = path->ah;
 516
 517                 ipoib_send(dev, skb, path->ah,
 518                            be32_to_cpup((__be32 *) skb->dst->neighbour->ha));
 519         } else {
 520                 neigh->ah  = NULL;
 521                 __skb_queue_tail(&neigh->queue, skb);
 522
 523                 if (!path->query && path_rec_start(dev, path))
 524                         goto err_list;
 525         }
 526
 527         spin_unlock(&priv->lock);
 528         return;
 529
 530 err_list:
 531         list_del(&neigh->list);
 532
 533 err_path:
 534         ipoib_neigh_free(neigh);
 535         ++priv->stats.tx_dropped;
 536         dev_kfree_skb_any(skb);
 537
 538         spin_unlock(&priv->lock);
 539 }
 540
 541 static void ipoib_path_lookup(struct sk_buff *skb, struct net_device *dev)
 542 {
 543         struct ipoib_dev_priv *priv = netdev_priv(skb->dev);
 544
 545         /* Look up path record for unicasts */
 546         if (skb->dst->neighbour->ha[4] != 0xff) {
 547                 neigh_add_path(skb, dev);
 548                 return;
 549         }
 550
 551         /* Add in the P_Key for multicasts */
 552         skb->dst->neighbour->ha[8] = (priv->pkey >> 8) & 0xff;
 553         skb->dst->neighbour->ha[9] = priv->pkey & 0xff;
 554         ipoib_mcast_send(dev, (union ib_gid *) (skb->dst->neighbour->ha + 4), skb);
 555 }
 556
 557 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
 558                              struct ipoib_pseudoheader *phdr)
 559 {
 560         struct ipoib_dev_priv *priv = netdev_priv(dev);
 561         struct ipoib_path *path;
 562
 563         /*
 564          * We can only be called from ipoib_start_xmit, so we're
 565          * inside tx_lock -- no need to save/restore flags.
 566          */
 567         spin_lock(&priv->lock);
 568
 569         path = __path_find(dev, (union ib_gid *) (phdr->hwaddr + 4));
 570         if (!path) {
 571                 path = path_rec_create(dev,
 572                                        (union ib_gid *) (phdr->hwaddr + 4));
 573                 if (path) {
 574                         /* put pseudoheader back on for next time */
 575                         skb_push(skb, sizeof *phdr);
 576                         __skb_queue_tail(&path->queue, skb);
 577
 578                         if (path_rec_start(dev, path)) {
 579                                 spin_unlock(&priv->lock);
 580                                 path_free(dev, path);
 581                                 return;
 582                         } else
 583                                 __path_add(dev, path);
 584                 } else {
 585                         ++priv->stats.tx_dropped;
 586                         dev_kfree_skb_any(skb);
 587                 }
 588
 589                 spin_unlock(&priv->lock);
 590                 return;
 591         }
 592
 593         if (path->ah) {
 594                 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
 595                           be16_to_cpu(path->pathrec.dlid));
 596
 597                 ipoib_send(dev, skb, path->ah,
 598                            be32_to_cpup((__be32 *) phdr->hwaddr));
 599         } else if ((path->query || !path_rec_start(dev, path)) &&
 600                    skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 601                 /* put pseudoheader back on for next time */
 602                 skb_push(skb, sizeof *phdr);
 603                 __skb_queue_tail(&path->queue, skb);
 604         } else {
 605                 ++priv->stats.tx_dropped;
 606                 dev_kfree_skb_any(skb);
 607         }
 608
 609         spin_unlock(&priv->lock);
 610 }
 611
 612 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
 613 {
 614         struct ipoib_dev_priv *priv = netdev_priv(dev);
 615         struct ipoib_neigh *neigh;
 616         unsigned long flags;
 617
 618         if (!spin_trylock_irqsave(&priv->tx_lock, flags))
 619                 return NETDEV_TX_LOCKED;
 620
 621         /*
 622          * Check if our queue is stopped.  Since we have the LLTX bit
 623          * set, we can't rely on netif_stop_queue() preventing our
 624          * xmit function from being called with a full queue.
 625          */
 626         if (unlikely(netif_queue_stopped(dev))) {
 627                 spin_unlock_irqrestore(&priv->tx_lock, flags);
 628                 return NETDEV_TX_BUSY;
 629         }
 630
 631         if (skb->dst && skb->dst->neighbour) {
 632                 if (unlikely(!*to_ipoib_neigh(skb->dst->neighbour))) {
 633                         ipoib_path_lookup(skb, dev);
 634                         goto out;
 635                 }
 636
 637                 neigh = *to_ipoib_neigh(skb->dst->neighbour);
 638
 639                 if (likely(neigh->ah)) {
 640                         ipoib_send(dev, skb, neigh->ah,
 641                                    be32_to_cpup((__be32 *) skb->dst->neighbour->ha));
 642                         goto out;
 643                 }
 644
 645                 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 646                         spin_lock(&priv->lock);
 647                         __skb_queue_tail(&neigh->queue, skb);
 648                         spin_unlock(&priv->lock);
 649                 } else {
 650                         ++priv->stats.tx_dropped;
 651                         dev_kfree_skb_any(skb);
 652                 }
 653         } else {
 654                 struct ipoib_pseudoheader *phdr =
 655                         (struct ipoib_pseudoheader *) skb->data;
 656                 skb_pull(skb, sizeof *phdr);
 657
 658                 if (phdr->hwaddr[4] == 0xff) {
 659                         /* Add in the P_Key for multicast*/
 660                         phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
 661                         phdr->hwaddr[9] = priv->pkey & 0xff;
 662
 663                         ipoib_mcast_send(dev, (union ib_gid *) (phdr->hwaddr + 4), skb);
 664                 } else {
 665                         /* unicast GID -- should be ARP or RARP reply */
 666
 667                         if ((be16_to_cpup((__be16 *) skb->data) != ETH_P_ARP) &&
 668                             (be16_to_cpup((__be16 *) skb->data) != ETH_P_RARP)) {
 669                                 ipoib_warn(priv, "Unicast, no %s: type %04x, QPN %06x "
 670                                            IPOIB_GID_FMT "\n",
 671                                            skb->dst ? "neigh" : "dst",
 672                                            be16_to_cpup((__be16 *) skb->data),
 673                                            be32_to_cpup((__be32 *) phdr->hwaddr),
 674                                            IPOIB_GID_ARG(*(union ib_gid *) (phdr->hwaddr + 4)));
 675                                 dev_kfree_skb_any(skb);
 676                                 ++priv->stats.tx_dropped;
 677                                 goto out;
 678                         }
 679
 680                         unicast_arp_send(skb, dev, phdr);
 681                 }
 682         }
 683
 684 out:
 685         spin_unlock_irqrestore(&priv->tx_lock, flags);
 686
 687         return NETDEV_TX_OK;
 688 }
 689
 690 static struct net_device_stats *ipoib_get_stats(struct net_device *dev)
 691 {
 692         struct ipoib_dev_priv *priv = netdev_priv(dev);
 693
 694         return &priv->stats;
 695 }
 696
 697 static void ipoib_timeout(struct net_device *dev)
 698 {
 699         struct ipoib_dev_priv *priv = netdev_priv(dev);
 700
 701         ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
 702                    jiffies_to_msecs(jiffies - dev->trans_start));
 703         ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
 704                    netif_queue_stopped(dev),
 705                    priv->tx_head, priv->tx_tail);
 706         /* XXX reset QP, etc. */
 707 }
 708
 709 static int ipoib_hard_header(struct sk_buff *skb,
 710                              struct net_device *dev,
 711                              unsigned short type,
 712                              void *daddr, void *saddr, unsigned len)
 713 {
 714         struct ipoib_header *header;
 715
 716         header = (struct ipoib_header *) skb_push(skb, sizeof *header);
 717
 718         header->proto = htons(type);
 719         header->reserved = 0;
 720
 721         /*
 722          * If we don't have a neighbour structure, stuff the
 723          * destination address onto the front of the skb so we can
 724          * figure out where to send the packet later.
 725          */
 726         if ((!skb->dst || !skb->dst->neighbour) && daddr) {
 727                 struct ipoib_pseudoheader *phdr =
 728                         (struct ipoib_pseudoheader *) skb_push(skb, sizeof *phdr);
 729                 memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
 730         }
 731
 732         return 0;
 733 }
 734
 735 static void ipoib_set_mcast_list(struct net_device *dev)
 736 {
 737         struct ipoib_dev_priv *priv = netdev_priv(dev);
 738
 739         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
 740                 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
 741                 return;
 742         }
 743
 744         queue_work(ipoib_workqueue, &priv->restart_task);
 745 }
 746
 747 static void ipoib_neigh_destructor(struct neighbour *n)
 748 {
 749         struct ipoib_neigh *neigh;
 750         struct ipoib_dev_priv *priv = netdev_priv(n->dev);
 751         unsigned long flags;
 752         struct ipoib_ah *ah = NULL;
 753
 754         ipoib_dbg(priv,
 755                   "neigh_destructor for %06x " IPOIB_GID_FMT "\n",
 756                   be32_to_cpup((__be32 *) n->ha),
 757                   IPOIB_GID_ARG(*((union ib_gid *) (n->ha + 4))));
 758
 759         spin_lock_irqsave(&priv->lock, flags);
 760
 761         neigh = *to_ipoib_neigh(n);
 762         if (neigh) {
 763                 if (neigh->ah)
 764                         ah = neigh->ah;
 765                 list_del(&neigh->list);
 766                 ipoib_neigh_free(neigh);
 767         }
 768
 769         spin_unlock_irqrestore(&priv->lock, flags);
 770
 771         if (ah)
 772                 ipoib_put_ah(ah);
 773 }
 774
 775 struct ipoib_neigh *ipoib_neigh_alloc(struct neighbour *neighbour)
 776 {
 777         struct ipoib_neigh *neigh;
 778
 779         neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
 780         if (!neigh)
 781                 return NULL;
 782
 783         neigh->neighbour = neighbour;
 784         *to_ipoib_neigh(neighbour) = neigh;
 785
 786         return neigh;
 787 }
 788
 789 void ipoib_neigh_free(struct ipoib_neigh *neigh)
 790 {
 791         *to_ipoib_neigh(neigh->neighbour) = NULL;
 792         kfree(neigh);
 793 }
 794
 795 static int ipoib_neigh_setup_dev(struct net_device *dev, struct neigh_parms *parms)
 796 {
 797         parms->neigh_destructor = ipoib_neigh_destructor;
 798
 799         return 0;
 800 }
 801
 802 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
 803 {
 804         struct ipoib_dev_priv *priv = netdev_priv(dev);
 805
 806         /* Allocate RX/TX "rings" to hold queued skbs */
 807         priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
 808                                 GFP_KERNEL);
 809         if (!priv->rx_ring) {
 810                 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
 811                        ca->name, ipoib_recvq_size);
 812                 goto out;
 813         }
 814
 815         priv->tx_ring = kzalloc(ipoib_sendq_size * sizeof *priv->tx_ring,
 816                                 GFP_KERNEL);
 817         if (!priv->tx_ring) {
 818                 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
 819                        ca->name, ipoib_sendq_size);
 820                 goto out_rx_ring_cleanup;
 821         }
 822
 823         /* priv->tx_head & tx_tail are already 0 */
 824
 825         if (ipoib_ib_dev_init(dev, ca, port))
 826                 goto out_tx_ring_cleanup;
 827
 828         return 0;
 829
 830 out_tx_ring_cleanup:
 831         kfree(priv->tx_ring);
 832
 833 out_rx_ring_cleanup:
 834         kfree(priv->rx_ring);
 835
 836 out:
 837         return -ENOMEM;
 838 }
 839
 840 void ipoib_dev_cleanup(struct net_device *dev)
 841 {
 842         struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
 843
 844         ipoib_delete_debug_files(dev);
 845
 846         /* Delete any child interfaces first */
 847         list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
 848                 unregister_netdev(cpriv->dev);
 849                 ipoib_dev_cleanup(cpriv->dev);
 850                 free_netdev(cpriv->dev);
 851         }
 852
 853         ipoib_ib_dev_cleanup(dev);
 854
 855         kfree(priv->rx_ring);
 856         kfree(priv->tx_ring);
 857
 858         priv->rx_ring = NULL;
 859         priv->tx_ring = NULL;
 860 }
 861
 862 static void ipoib_setup(struct net_device *dev)
 863 {
 864         struct ipoib_dev_priv *priv = netdev_priv(dev);
 865
 866         dev->open                = ipoib_open;
 867         dev->stop                = ipoib_stop;
 868         dev->change_mtu          = ipoib_change_mtu;
 869         dev->hard_start_xmit     = ipoib_start_xmit;
 870         dev->get_stats           = ipoib_get_stats;
 871         dev->tx_timeout          = ipoib_timeout;
 872         dev->hard_header         = ipoib_hard_header;
 873         dev->set_multicast_list  = ipoib_set_mcast_list;
 874         dev->neigh_setup         = ipoib_neigh_setup_dev;
 875
 876         dev->watchdog_timeo      = HZ;
 877
 878         dev->flags              |= IFF_BROADCAST | IFF_MULTICAST;
 879
 880         /*
 881          * We add in INFINIBAND_ALEN to allow for the destination
 882          * address "pseudoheader" for skbs without neighbour struct.
 883          */
 884         dev->hard_header_len     = IPOIB_ENCAP_LEN + INFINIBAND_ALEN;
 885         dev->addr_len            = INFINIBAND_ALEN;
 886         dev->type                = ARPHRD_INFINIBAND;
 887         dev->tx_queue_len        = ipoib_sendq_size * 2;
 888         dev->features            = NETIF_F_VLAN_CHALLENGED | NETIF_F_LLTX;
 889
 890         /* MTU will be reset when mcast join happens */
 891         dev->mtu                 = IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN;
 892         priv->mcast_mtu          = priv->admin_mtu = dev->mtu;
 893
 894         memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
 895
 896         netif_carrier_off(dev);
 897
 898         SET_MODULE_OWNER(dev);
 899
 900         priv->dev = dev;
 901
 902         spin_lock_init(&priv->lock);
 903         spin_lock_init(&priv->tx_lock);
 904
 905         mutex_init(&priv->mcast_mutex);
 906         mutex_init(&priv->vlan_mutex);
 907
 908         INIT_LIST_HEAD(&priv->path_list);
 909         INIT_LIST_HEAD(&priv->child_intfs);
 910         INIT_LIST_HEAD(&priv->dead_ahs);
 911         INIT_LIST_HEAD(&priv->multicast_list);
 912
 913         INIT_WORK(&priv->pkey_task,    ipoib_pkey_poll,          priv->dev);
 914         INIT_WORK(&priv->mcast_task,   ipoib_mcast_join_task,    priv->dev);
 915         INIT_WORK(&priv->flush_task,   ipoib_ib_dev_flush,       priv->dev);
 916         INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task, priv->dev);
 917         INIT_WORK(&priv->ah_reap_task, ipoib_reap_ah,            priv->dev);
 918 }
 919
 920 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
 921 {
 922         struct net_device *dev;
 923
 924         dev = alloc_netdev((int) sizeof (struct ipoib_dev_priv), name,
 925                            ipoib_setup);
 926         if (!dev)
 927                 return NULL;
 928
 929         return netdev_priv(dev);
 930 }
 931
 932 static ssize_t show_pkey(struct class_device *cdev, char *buf)
 933 {
 934         struct ipoib_dev_priv *priv =
 935                 netdev_priv(container_of(cdev, struct net_device, class_dev));
 936
 937         return sprintf(buf, "0x%04x\n", priv->pkey);
 938 }
 939 static CLASS_DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
 940
 941 static ssize_t create_child(struct class_device *cdev,
 942                             const char *buf, size_t count)
 943 {
 944         int pkey;
 945         int ret;
 946
 947         if (sscanf(buf, "%i", &pkey) != 1)
 948                 return -EINVAL;
 949
 950         if (pkey < 0 || pkey > 0xffff)
 951                 return -EINVAL;
 952
 953         /*
 954          * Set the full membership bit, so that we join the right
 955          * broadcast group, etc.
 956          */
 957         pkey |= 0x8000;
 958
 959         ret = ipoib_vlan_add(container_of(cdev, struct net_device, class_dev),
 960                              pkey);
 961
 962         return ret ? ret : count;
 963 }
 964 static CLASS_DEVICE_ATTR(create_child, S_IWUGO, NULL, create_child);
 965
 966 static ssize_t delete_child(struct class_device *cdev,
 967                             const char *buf, size_t count)
 968 {
 969         int pkey;
 970         int ret;
 971
 972         if (sscanf(buf, "%i", &pkey) != 1)
 973                 return -EINVAL;
 974
 975         if (pkey < 0 || pkey > 0xffff)
 976                 return -EINVAL;
 977
 978         ret = ipoib_vlan_delete(container_of(cdev, struct net_device, class_dev),
 979                                 pkey);
 980
 981         return ret ? ret : count;
 982
 983 }
 984 static CLASS_DEVICE_ATTR(delete_child, S_IWUGO, NULL, delete_child);
 985
 986 int ipoib_add_pkey_attr(struct net_device *dev)
 987 {
 988         return class_device_create_file(&dev->class_dev,
 989                                         &class_device_attr_pkey);
 990 }
 991
 992 static struct net_device *ipoib_add_port(const char *format,
 993                                          struct ib_device *hca, u8 port)
 994 {
 995         struct ipoib_dev_priv *priv;
 996         int result = -ENOMEM;
 997
 998         priv = ipoib_intf_alloc(format);
 999         if (!priv)
1000                 goto alloc_mem_failed;
1001
1002         SET_NETDEV_DEV(priv->dev, hca->dma_device);
1003
1004         result = ib_query_pkey(hca, port, 0, &priv->pkey);
1005         if (result) {
1006                 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
1007                        hca->name, port, result);
1008                 goto alloc_mem_failed;
1009         }
1010
1011         /*
1012          * Set the full membership bit, so that we join the right
1013          * broadcast group, etc.
1014          */
1015         priv->pkey |= 0x8000;
1016
1017         priv->dev->broadcast[8] = priv->pkey >> 8;
1018         priv->dev->broadcast[9] = priv->pkey & 0xff;
1019
1020         result = ib_query_gid(hca, port, 0, &priv->local_gid);
1021         if (result) {
1022                 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1023                        hca->name, port, result);
1024                 goto alloc_mem_failed;
1025         } else
1026                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1027
1028
1029         result = ipoib_dev_init(priv->dev, hca, port);
1030         if (result < 0) {
1031                 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1032                        hca->name, port, result);
1033                 goto device_init_failed;
1034         }
1035
1036         INIT_IB_EVENT_HANDLER(&priv->event_handler,
1037                               priv->ca, ipoib_event);
1038         result = ib_register_event_handler(&priv->event_handler);
1039         if (result < 0) {
1040                 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1041                        "port %d (ret = %d)\n",
1042                        hca->name, port, result);
1043                 goto event_failed;
1044         }
1045
1046         result = register_netdev(priv->dev);
1047         if (result) {
1048                 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1049                        hca->name, port, result);
1050                 goto register_failed;
1051         }
1052
1053         ipoib_create_debug_files(priv->dev);
1054
1055         if (ipoib_add_pkey_attr(priv->dev))
1056                 goto sysfs_failed;
1057         if (class_device_create_file(&priv->dev->class_dev,
1058                                      &class_device_attr_create_child))
1059                 goto sysfs_failed;
1060         if (class_device_create_file(&priv->dev->class_dev,
1061                                      &class_device_attr_delete_child))
1062                 goto sysfs_failed;
1063
1064         return priv->dev;
1065
1066 sysfs_failed:
1067         ipoib_delete_debug_files(priv->dev);
1068         unregister_netdev(priv->dev);
1069
1070 register_failed:
1071         ib_unregister_event_handler(&priv->event_handler);
1072         flush_scheduled_work();
1073
1074 event_failed:
1075         ipoib_dev_cleanup(priv->dev);
1076
1077 device_init_failed:
1078         free_netdev(priv->dev);
1079
1080 alloc_mem_failed:
1081         return ERR_PTR(result);
1082 }
1083
1084 static void ipoib_add_one(struct ib_device *device)
1085 {
1086         struct list_head *dev_list;
1087         struct net_device *dev;
1088         struct ipoib_dev_priv *priv;
1089         int s, e, p;
1090
1091         dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1092         if (!dev_list)
1093                 return;
1094
1095         INIT_LIST_HEAD(dev_list);
1096
1097         if (device->node_type == IB_NODE_SWITCH) {
1098                 s = 0;
1099                 e = 0;
1100         } else {
1101                 s = 1;
1102                 e = device->phys_port_cnt;
1103         }
1104
1105         for (p = s; p <= e; ++p) {
1106                 dev = ipoib_add_port("ib%d", device, p);
1107                 if (!IS_ERR(dev)) {
1108                         priv = netdev_priv(dev);
1109                         list_add_tail(&priv->list, dev_list);
1110                 }
1111         }
1112
1113         ib_set_client_data(device, &ipoib_client, dev_list);
1114 }
1115
1116 static void ipoib_remove_one(struct ib_device *device)
1117 {
1118         struct ipoib_dev_priv *priv, *tmp;
1119         struct list_head *dev_list;
1120
1121         dev_list = ib_get_client_data(device, &ipoib_client);
1122
1123         list_for_each_entry_safe(priv, tmp, dev_list, list) {
1124                 ib_unregister_event_handler(&priv->event_handler);
1125                 flush_scheduled_work();
1126
1127                 unregister_netdev(priv->dev);
1128                 ipoib_dev_cleanup(priv->dev);
1129                 free_netdev(priv->dev);
1130         }
1131
1132         kfree(dev_list);
1133 }
1134
1135 static int __init ipoib_init_module(void)
1136 {
1137         int ret;
1138
1139         ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1140         ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1141         ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1142
1143         ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1144         ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1145         ipoib_sendq_size = max(ipoib_sendq_size, IPOIB_MIN_QUEUE_SIZE);
1146
1147         ret = ipoib_register_debugfs();
1148         if (ret)
1149                 return ret;
1150
1151         /*
1152          * We create our own workqueue mainly because we want to be
1153          * able to flush it when devices are being removed.  We can't
1154          * use schedule_work()/flush_scheduled_work() because both
1155          * unregister_netdev() and linkwatch_event take the rtnl lock,
1156          * so flush_scheduled_work() can deadlock during device
1157          * removal.
1158          */
1159         ipoib_workqueue = create_singlethread_workqueue("ipoib");
1160         if (!ipoib_workqueue) {
1161                 ret = -ENOMEM;
1162                 goto err_fs;
1163         }
1164
1165         ret = ib_register_client(&ipoib_client);
1166         if (ret)
1167                 goto err_wq;
1168
1169         return 0;
1170
1171 err_wq:
1172         destroy_workqueue(ipoib_workqueue);
1173
1174 err_fs:
1175         ipoib_unregister_debugfs();
1176
1177         return ret;
1178 }
1179
1180 static void __exit ipoib_cleanup_module(void)
1181 {
1182         ib_unregister_client(&ipoib_client);
1183         ipoib_unregister_debugfs();
1184         destroy_workqueue(ipoib_workqueue);
1185 }
1186
1187 module_init(ipoib_init_module);
1188 module_exit(ipoib_cleanup_module);