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         /*
 137          * Now flush workqueue to make sure a scheduled task doesn't
 138          * bring our internal state back up.
 139          */
 140         flush_workqueue(ipoib_workqueue);
 141
 142         ipoib_ib_dev_down(dev, 1);
 143         ipoib_ib_dev_stop(dev);
 144
 145         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 146                 struct ipoib_dev_priv *cpriv;
 147
 148                 /* Bring down any child interfaces too */
 149                 mutex_lock(&priv->vlan_mutex);
 150                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 151                         int flags;
 152
 153                         flags = cpriv->dev->flags;
 154                         if (!(flags & IFF_UP))
 155                                 continue;
 156
 157                         dev_change_flags(cpriv->dev, flags & ~IFF_UP);
 158                 }
 159                 mutex_unlock(&priv->vlan_mutex);
 160         }
 161
 162         return 0;
 163 }
 164
 165 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
 166 {
 167         struct ipoib_dev_priv *priv = netdev_priv(dev);
 168
 169         if (new_mtu > IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN)
 170                 return -EINVAL;
 171
 172         priv->admin_mtu = new_mtu;
 173
 174         dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
 175
 176         return 0;
 177 }
 178
 179 static struct ipoib_path *__path_find(struct net_device *dev,
 180                                       union ib_gid *gid)
 181 {
 182         struct ipoib_dev_priv *priv = netdev_priv(dev);
 183         struct rb_node *n = priv->path_tree.rb_node;
 184         struct ipoib_path *path;
 185         int ret;
 186
 187         while (n) {
 188                 path = rb_entry(n, struct ipoib_path, rb_node);
 189
 190                 ret = memcmp(gid->raw, path->pathrec.dgid.raw,
 191                              sizeof (union ib_gid));
 192
 193                 if (ret < 0)
 194                         n = n->rb_left;
 195                 else if (ret > 0)
 196                         n = n->rb_right;
 197                 else
 198                         return path;
 199         }
 200
 201         return NULL;
 202 }
 203
 204 static int __path_add(struct net_device *dev, struct ipoib_path *path)
 205 {
 206         struct ipoib_dev_priv *priv = netdev_priv(dev);
 207         struct rb_node **n = &priv->path_tree.rb_node;
 208         struct rb_node *pn = NULL;
 209         struct ipoib_path *tpath;
 210         int ret;
 211
 212         while (*n) {
 213                 pn = *n;
 214                 tpath = rb_entry(pn, struct ipoib_path, rb_node);
 215
 216                 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
 217                              sizeof (union ib_gid));
 218                 if (ret < 0)
 219                         n = &pn->rb_left;
 220                 else if (ret > 0)
 221                         n = &pn->rb_right;
 222                 else
 223                         return -EEXIST;
 224         }
 225
 226         rb_link_node(&path->rb_node, pn, n);
 227         rb_insert_color(&path->rb_node, &priv->path_tree);
 228
 229         list_add_tail(&path->list, &priv->path_list);
 230
 231         return 0;
 232 }
 233
 234 static void path_free(struct net_device *dev, struct ipoib_path *path)
 235 {
 236         struct ipoib_dev_priv *priv = netdev_priv(dev);
 237         struct ipoib_neigh *neigh, *tn;
 238         struct sk_buff *skb;
 239         unsigned long flags;
 240
 241         while ((skb = __skb_dequeue(&path->queue)))
 242                 dev_kfree_skb_irq(skb);
 243
 244         spin_lock_irqsave(&priv->lock, flags);
 245
 246         list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
 247                 /*
 248                  * It's safe to call ipoib_put_ah() inside priv->lock
 249                  * here, because we know that path->ah will always
 250                  * hold one more reference, so ipoib_put_ah() will
 251                  * never do more than decrement the ref count.
 252                  */
 253                 if (neigh->ah)
 254                         ipoib_put_ah(neigh->ah);
 255                 *to_ipoib_neigh(neigh->neighbour) = NULL;
 256                 kfree(neigh);
 257         }
 258
 259         spin_unlock_irqrestore(&priv->lock, flags);
 260
 261         if (path->ah)
 262                 ipoib_put_ah(path->ah);
 263
 264         kfree(path);
 265 }
 266
 267 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 268
 269 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
 270 {
 271         struct ipoib_path_iter *iter;
 272
 273         iter = kmalloc(sizeof *iter, GFP_KERNEL);
 274         if (!iter)
 275                 return NULL;
 276
 277         iter->dev = dev;
 278         memset(iter->path.pathrec.dgid.raw, 0, 16);
 279
 280         if (ipoib_path_iter_next(iter)) {
 281                 kfree(iter);
 282                 return NULL;
 283         }
 284
 285         return iter;
 286 }
 287
 288 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
 289 {
 290         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
 291         struct rb_node *n;
 292         struct ipoib_path *path;
 293         int ret = 1;
 294
 295         spin_lock_irq(&priv->lock);
 296
 297         n = rb_first(&priv->path_tree);
 298
 299         while (n) {
 300                 path = rb_entry(n, struct ipoib_path, rb_node);
 301
 302                 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
 303                            sizeof (union ib_gid)) < 0) {
 304                         iter->path = *path;
 305                         ret = 0;
 306                         break;
 307                 }
 308
 309                 n = rb_next(n);
 310         }
 311
 312         spin_unlock_irq(&priv->lock);
 313
 314         return ret;
 315 }
 316
 317 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
 318                           struct ipoib_path *path)
 319 {
 320         *path = iter->path;
 321 }
 322
 323 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
 324
 325 void ipoib_flush_paths(struct net_device *dev)
 326 {
 327         struct ipoib_dev_priv *priv = netdev_priv(dev);
 328         struct ipoib_path *path, *tp;
 329         LIST_HEAD(remove_list);
 330         unsigned long flags;
 331
 332         spin_lock_irqsave(&priv->lock, flags);
 333
 334         list_splice(&priv->path_list, &remove_list);
 335         INIT_LIST_HEAD(&priv->path_list);
 336
 337         list_for_each_entry(path, &remove_list, list)
 338                 rb_erase(&path->rb_node, &priv->path_tree);
 339
 340         spin_unlock_irqrestore(&priv->lock, flags);
 341
 342         list_for_each_entry_safe(path, tp, &remove_list, list) {
 343                 if (path->query)
 344                         ib_sa_cancel_query(path->query_id, path->query);
 345                 wait_for_completion(&path->done);
 346                 path_free(dev, path);
 347         }
 348 }
 349
 350 static void path_rec_completion(int status,
 351                                 struct ib_sa_path_rec *pathrec,
 352                                 void *path_ptr)
 353 {
 354         struct ipoib_path *path = path_ptr;
 355         struct net_device *dev = path->dev;
 356         struct ipoib_dev_priv *priv = netdev_priv(dev);
 357         struct ipoib_ah *ah = NULL;
 358         struct ipoib_neigh *neigh;
 359         struct sk_buff_head skqueue;
 360         struct sk_buff *skb;
 361         unsigned long flags;
 362
 363         if (pathrec)
 364                 ipoib_dbg(priv, "PathRec LID 0x%04x for GID " IPOIB_GID_FMT "\n",
 365                           be16_to_cpu(pathrec->dlid), IPOIB_GID_ARG(pathrec->dgid));
 366         else
 367                 ipoib_dbg(priv, "PathRec status %d for GID " IPOIB_GID_FMT "\n",
 368                           status, IPOIB_GID_ARG(path->pathrec.dgid));
 369
 370         skb_queue_head_init(&skqueue);
 371
 372         if (!status) {
 373                 struct ib_ah_attr av = {
 374                         .dlid          = be16_to_cpu(pathrec->dlid),
 375                         .sl            = pathrec->sl,
 376                         .port_num      = priv->port
 377                 };
 378                 int path_rate = ib_sa_rate_enum_to_int(pathrec->rate);
 379
 380                 if (path_rate > 0 && priv->local_rate > path_rate)
 381                         av.static_rate = (priv->local_rate - 1) / path_rate;
 382
 383                 ipoib_dbg(priv, "static_rate %d for local port %dX, path %dX\n",
 384                           av.static_rate, priv->local_rate,
 385                           ib_sa_rate_enum_to_int(pathrec->rate));
 386
 387                 ah = ipoib_create_ah(dev, priv->pd, &av);
 388         }
 389
 390         spin_lock_irqsave(&priv->lock, flags);
 391
 392         path->ah = ah;
 393
 394         if (ah) {
 395                 path->pathrec = *pathrec;
 396
 397                 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
 398                           ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
 399
 400                 while ((skb = __skb_dequeue(&path->queue)))
 401                         __skb_queue_tail(&skqueue, skb);
 402
 403                 list_for_each_entry(neigh, &path->neigh_list, list) {
 404                         kref_get(&path->ah->ref);
 405                         neigh->ah = path->ah;
 406
 407                         while ((skb = __skb_dequeue(&neigh->queue)))
 408                                 __skb_queue_tail(&skqueue, skb);
 409                 }
 410         }
 411
 412         path->query = NULL;
 413         complete(&path->done);
 414
 415         spin_unlock_irqrestore(&priv->lock, flags);
 416
 417         while ((skb = __skb_dequeue(&skqueue))) {
 418                 skb->dev = dev;
 419                 if (dev_queue_xmit(skb))
 420                         ipoib_warn(priv, "dev_queue_xmit failed "
 421                                    "to requeue packet\n");
 422         }
 423 }
 424
 425 static struct ipoib_path *path_rec_create(struct net_device *dev,
 426                                           union ib_gid *gid)
 427 {
 428         struct ipoib_dev_priv *priv = netdev_priv(dev);
 429         struct ipoib_path *path;
 430
 431         path = kzalloc(sizeof *path, GFP_ATOMIC);
 432         if (!path)
 433                 return NULL;
 434
 435         path->dev = dev;
 436
 437         skb_queue_head_init(&path->queue);
 438
 439         INIT_LIST_HEAD(&path->neigh_list);
 440
 441         memcpy(path->pathrec.dgid.raw, gid->raw, sizeof (union ib_gid));
 442         path->pathrec.sgid      = priv->local_gid;
 443         path->pathrec.pkey      = cpu_to_be16(priv->pkey);
 444         path->pathrec.numb_path = 1;
 445
 446         return path;
 447 }
 448
 449 static int path_rec_start(struct net_device *dev,
 450                           struct ipoib_path *path)
 451 {
 452         struct ipoib_dev_priv *priv = netdev_priv(dev);
 453
 454         ipoib_dbg(priv, "Start path record lookup for " IPOIB_GID_FMT "\n",
 455                   IPOIB_GID_ARG(path->pathrec.dgid));
 456
 457         init_completion(&path->done);
 458
 459         path->query_id =
 460                 ib_sa_path_rec_get(priv->ca, priv->port,
 461                                    &path->pathrec,
 462                                    IB_SA_PATH_REC_DGID          |
 463                                    IB_SA_PATH_REC_SGID          |
 464                                    IB_SA_PATH_REC_NUMB_PATH     |
 465                                    IB_SA_PATH_REC_PKEY,
 466                                    1000, GFP_ATOMIC,
 467                                    path_rec_completion,
 468                                    path, &path->query);
 469         if (path->query_id < 0) {
 470                 ipoib_warn(priv, "ib_sa_path_rec_get failed\n");
 471                 path->query = NULL;
 472                 return path->query_id;
 473         }
 474
 475         return 0;
 476 }
 477
 478 static void neigh_add_path(struct sk_buff *skb, struct net_device *dev)
 479 {
 480         struct ipoib_dev_priv *priv = netdev_priv(dev);
 481         struct ipoib_path *path;
 482         struct ipoib_neigh *neigh;
 483
 484         neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
 485         if (!neigh) {
 486                 ++priv->stats.tx_dropped;
 487                 dev_kfree_skb_any(skb);
 488                 return;
 489         }
 490
 491         skb_queue_head_init(&neigh->queue);
 492         neigh->neighbour = skb->dst->neighbour;
 493         *to_ipoib_neigh(skb->dst->neighbour) = neigh;
 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;
 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;
 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         kfree(neigh);
 534
 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) {
 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         queue_work(ipoib_workqueue, &priv->restart_task);
 740 }
 741
 742 static void ipoib_neigh_destructor(struct neighbour *n)
 743 {
 744         struct ipoib_neigh *neigh;
 745         struct ipoib_dev_priv *priv = netdev_priv(n->dev);
 746         unsigned long flags;
 747         struct ipoib_ah *ah = NULL;
 748
 749         ipoib_dbg(priv,
 750                   "neigh_destructor for %06x " IPOIB_GID_FMT "\n",
 751                   be32_to_cpup((__be32 *) n->ha),
 752                   IPOIB_GID_ARG(*((union ib_gid *) (n->ha + 4))));
 753
 754         spin_lock_irqsave(&priv->lock, flags);
 755
 756         neigh = *to_ipoib_neigh(n);
 757         if (neigh) {
 758                 if (neigh->ah)
 759                         ah = neigh->ah;
 760                 list_del(&neigh->list);
 761                 *to_ipoib_neigh(n) = NULL;
 762                 kfree(neigh);
 763         }
 764
 765         spin_unlock_irqrestore(&priv->lock, flags);
 766
 767         if (ah)
 768                 ipoib_put_ah(ah);
 769 }
 770
 771 static int ipoib_neigh_setup_dev(struct net_device *dev, struct neigh_parms *parms)
 772 {
 773         parms->neigh_destructor = ipoib_neigh_destructor;
 774
 775         return 0;
 776 }
 777
 778 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
 779 {
 780         struct ipoib_dev_priv *priv = netdev_priv(dev);
 781
 782         /* Allocate RX/TX "rings" to hold queued skbs */
 783
 784         priv->rx_ring = kzalloc(IPOIB_RX_RING_SIZE * sizeof (struct ipoib_rx_buf),
 785                                 GFP_KERNEL);
 786         if (!priv->rx_ring) {
 787                 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
 788                        ca->name, IPOIB_RX_RING_SIZE);
 789                 goto out;
 790         }
 791
 792         priv->tx_ring = kzalloc(IPOIB_TX_RING_SIZE * sizeof (struct ipoib_tx_buf),
 793                                 GFP_KERNEL);
 794         if (!priv->tx_ring) {
 795                 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
 796                        ca->name, IPOIB_TX_RING_SIZE);
 797                 goto out_rx_ring_cleanup;
 798         }
 799
 800         /* priv->tx_head & tx_tail are already 0 */
 801
 802         if (ipoib_ib_dev_init(dev, ca, port))
 803                 goto out_tx_ring_cleanup;
 804
 805         return 0;
 806
 807 out_tx_ring_cleanup:
 808         kfree(priv->tx_ring);
 809
 810 out_rx_ring_cleanup:
 811         kfree(priv->rx_ring);
 812
 813 out:
 814         return -ENOMEM;
 815 }
 816
 817 void ipoib_dev_cleanup(struct net_device *dev)
 818 {
 819         struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
 820
 821         ipoib_delete_debug_files(dev);
 822
 823         /* Delete any child interfaces first */
 824         list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
 825                 unregister_netdev(cpriv->dev);
 826                 ipoib_dev_cleanup(cpriv->dev);
 827                 free_netdev(cpriv->dev);
 828         }
 829
 830         ipoib_ib_dev_cleanup(dev);
 831
 832         kfree(priv->rx_ring);
 833         kfree(priv->tx_ring);
 834
 835         priv->rx_ring = NULL;
 836         priv->tx_ring = NULL;
 837 }
 838
 839 static void ipoib_setup(struct net_device *dev)
 840 {
 841         struct ipoib_dev_priv *priv = netdev_priv(dev);
 842
 843         dev->open                = ipoib_open;
 844         dev->stop                = ipoib_stop;
 845         dev->change_mtu          = ipoib_change_mtu;
 846         dev->hard_start_xmit     = ipoib_start_xmit;
 847         dev->get_stats           = ipoib_get_stats;
 848         dev->tx_timeout          = ipoib_timeout;
 849         dev->hard_header         = ipoib_hard_header;
 850         dev->set_multicast_list  = ipoib_set_mcast_list;
 851         dev->neigh_setup         = ipoib_neigh_setup_dev;
 852
 853         dev->watchdog_timeo      = HZ;
 854
 855         dev->flags              |= IFF_BROADCAST | IFF_MULTICAST;
 856
 857         /*
 858          * We add in INFINIBAND_ALEN to allow for the destination
 859          * address "pseudoheader" for skbs without neighbour struct.
 860          */
 861         dev->hard_header_len     = IPOIB_ENCAP_LEN + INFINIBAND_ALEN;
 862         dev->addr_len            = INFINIBAND_ALEN;
 863         dev->type                = ARPHRD_INFINIBAND;
 864         dev->tx_queue_len        = IPOIB_TX_RING_SIZE * 2;
 865         dev->features            = NETIF_F_VLAN_CHALLENGED | NETIF_F_LLTX;
 866
 867         /* MTU will be reset when mcast join happens */
 868         dev->mtu                 = IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN;
 869         priv->mcast_mtu          = priv->admin_mtu = dev->mtu;
 870
 871         memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
 872
 873         netif_carrier_off(dev);
 874
 875         SET_MODULE_OWNER(dev);
 876
 877         priv->dev = dev;
 878
 879         spin_lock_init(&priv->lock);
 880         spin_lock_init(&priv->tx_lock);
 881
 882         mutex_init(&priv->mcast_mutex);
 883         mutex_init(&priv->vlan_mutex);
 884
 885         INIT_LIST_HEAD(&priv->path_list);
 886         INIT_LIST_HEAD(&priv->child_intfs);
 887         INIT_LIST_HEAD(&priv->dead_ahs);
 888         INIT_LIST_HEAD(&priv->multicast_list);
 889
 890         INIT_WORK(&priv->pkey_task,    ipoib_pkey_poll,          priv->dev);
 891         INIT_WORK(&priv->mcast_task,   ipoib_mcast_join_task,    priv->dev);
 892         INIT_WORK(&priv->flush_task,   ipoib_ib_dev_flush,       priv->dev);
 893         INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task, priv->dev);
 894         INIT_WORK(&priv->ah_reap_task, ipoib_reap_ah,            priv->dev);
 895 }
 896
 897 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
 898 {
 899         struct net_device *dev;
 900
 901         dev = alloc_netdev((int) sizeof (struct ipoib_dev_priv), name,
 902                            ipoib_setup);
 903         if (!dev)
 904                 return NULL;
 905
 906         return netdev_priv(dev);
 907 }
 908
 909 static ssize_t show_pkey(struct class_device *cdev, char *buf)
 910 {
 911         struct ipoib_dev_priv *priv =
 912                 netdev_priv(container_of(cdev, struct net_device, class_dev));
 913
 914         return sprintf(buf, "0x%04x\n", priv->pkey);
 915 }
 916 static CLASS_DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
 917
 918 static ssize_t create_child(struct class_device *cdev,
 919                             const char *buf, size_t count)
 920 {
 921         int pkey;
 922         int ret;
 923
 924         if (sscanf(buf, "%i", &pkey) != 1)
 925                 return -EINVAL;
 926
 927         if (pkey < 0 || pkey > 0xffff)
 928                 return -EINVAL;
 929
 930         /*
 931          * Set the full membership bit, so that we join the right
 932          * broadcast group, etc.
 933          */
 934         pkey |= 0x8000;
 935
 936         ret = ipoib_vlan_add(container_of(cdev, struct net_device, class_dev),
 937                              pkey);
 938
 939         return ret ? ret : count;
 940 }
 941 static CLASS_DEVICE_ATTR(create_child, S_IWUGO, NULL, create_child);
 942
 943 static ssize_t delete_child(struct class_device *cdev,
 944                             const char *buf, size_t count)
 945 {
 946         int pkey;
 947         int ret;
 948
 949         if (sscanf(buf, "%i", &pkey) != 1)
 950                 return -EINVAL;
 951
 952         if (pkey < 0 || pkey > 0xffff)
 953                 return -EINVAL;
 954
 955         ret = ipoib_vlan_delete(container_of(cdev, struct net_device, class_dev),
 956                                 pkey);
 957
 958         return ret ? ret : count;
 959
 960 }
 961 static CLASS_DEVICE_ATTR(delete_child, S_IWUGO, NULL, delete_child);
 962
 963 int ipoib_add_pkey_attr(struct net_device *dev)
 964 {
 965         return class_device_create_file(&dev->class_dev,
 966                                         &class_device_attr_pkey);
 967 }
 968
 969 static struct net_device *ipoib_add_port(const char *format,
 970                                          struct ib_device *hca, u8 port)
 971 {
 972         struct ipoib_dev_priv *priv;
 973         int result = -ENOMEM;
 974
 975         priv = ipoib_intf_alloc(format);
 976         if (!priv)
 977                 goto alloc_mem_failed;
 978
 979         SET_NETDEV_DEV(priv->dev, hca->dma_device);
 980
 981         result = ib_query_pkey(hca, port, 0, &priv->pkey);
 982         if (result) {
 983                 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
 984                        hca->name, port, result);
 985                 goto alloc_mem_failed;
 986         }
 987
 988         /*
 989          * Set the full membership bit, so that we join the right
 990          * broadcast group, etc.
 991          */
 992         priv->pkey |= 0x8000;
 993
 994         priv->dev->broadcast[8] = priv->pkey >> 8;
 995         priv->dev->broadcast[9] = priv->pkey & 0xff;
 996
 997         result = ib_query_gid(hca, port, 0, &priv->local_gid);
 998         if (result) {
 999                 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1000                        hca->name, port, result);
1001                 goto alloc_mem_failed;
1002         } else
1003                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1004
1005
1006         result = ipoib_dev_init(priv->dev, hca, port);
1007         if (result < 0) {
1008                 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1009                        hca->name, port, result);
1010                 goto device_init_failed;
1011         }
1012
1013         INIT_IB_EVENT_HANDLER(&priv->event_handler,
1014                               priv->ca, ipoib_event);
1015         result = ib_register_event_handler(&priv->event_handler);
1016         if (result < 0) {
1017                 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1018                        "port %d (ret = %d)\n",
1019                        hca->name, port, result);
1020                 goto event_failed;
1021         }
1022
1023         result = register_netdev(priv->dev);
1024         if (result) {
1025                 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1026                        hca->name, port, result);
1027                 goto register_failed;
1028         }
1029
1030         ipoib_create_debug_files(priv->dev);
1031
1032         if (ipoib_add_pkey_attr(priv->dev))
1033                 goto sysfs_failed;
1034         if (class_device_create_file(&priv->dev->class_dev,
1035                                      &class_device_attr_create_child))
1036                 goto sysfs_failed;
1037         if (class_device_create_file(&priv->dev->class_dev,
1038                                      &class_device_attr_delete_child))
1039                 goto sysfs_failed;
1040
1041         return priv->dev;
1042
1043 sysfs_failed:
1044         ipoib_delete_debug_files(priv->dev);
1045         unregister_netdev(priv->dev);
1046
1047 register_failed:
1048         ib_unregister_event_handler(&priv->event_handler);
1049         flush_scheduled_work();
1050
1051 event_failed:
1052         ipoib_dev_cleanup(priv->dev);
1053
1054 device_init_failed:
1055         free_netdev(priv->dev);
1056
1057 alloc_mem_failed:
1058         return ERR_PTR(result);
1059 }
1060
1061 static void ipoib_add_one(struct ib_device *device)
1062 {
1063         struct list_head *dev_list;
1064         struct net_device *dev;
1065         struct ipoib_dev_priv *priv;
1066         int s, e, p;
1067
1068         dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1069         if (!dev_list)
1070                 return;
1071
1072         INIT_LIST_HEAD(dev_list);
1073
1074         if (device->node_type == IB_NODE_SWITCH) {
1075                 s = 0;
1076                 e = 0;
1077         } else {
1078                 s = 1;
1079                 e = device->phys_port_cnt;
1080         }
1081
1082         for (p = s; p <= e; ++p) {
1083                 dev = ipoib_add_port("ib%d", device, p);
1084                 if (!IS_ERR(dev)) {
1085                         priv = netdev_priv(dev);
1086                         list_add_tail(&priv->list, dev_list);
1087                 }
1088         }
1089
1090         ib_set_client_data(device, &ipoib_client, dev_list);
1091 }
1092
1093 static void ipoib_remove_one(struct ib_device *device)
1094 {
1095         struct ipoib_dev_priv *priv, *tmp;
1096         struct list_head *dev_list;
1097
1098         dev_list = ib_get_client_data(device, &ipoib_client);
1099
1100         list_for_each_entry_safe(priv, tmp, dev_list, list) {
1101                 ib_unregister_event_handler(&priv->event_handler);
1102                 flush_scheduled_work();
1103
1104                 unregister_netdev(priv->dev);
1105                 ipoib_dev_cleanup(priv->dev);
1106                 free_netdev(priv->dev);
1107         }
1108
1109         kfree(dev_list);
1110 }
1111
1112 static int __init ipoib_init_module(void)
1113 {
1114         int ret;
1115
1116         ret = ipoib_register_debugfs();
1117         if (ret)
1118                 return ret;
1119
1120         /*
1121          * We create our own workqueue mainly because we want to be
1122          * able to flush it when devices are being removed.  We can't
1123          * use schedule_work()/flush_scheduled_work() because both
1124          * unregister_netdev() and linkwatch_event take the rtnl lock,
1125          * so flush_scheduled_work() can deadlock during device
1126          * removal.
1127          */
1128         ipoib_workqueue = create_singlethread_workqueue("ipoib");
1129         if (!ipoib_workqueue) {
1130                 ret = -ENOMEM;
1131                 goto err_fs;
1132         }
1133
1134         ret = ib_register_client(&ipoib_client);
1135         if (ret)
1136                 goto err_wq;
1137
1138         return 0;
1139
1140 err_wq:
1141         destroy_workqueue(ipoib_workqueue);
1142
1143 err_fs:
1144         ipoib_unregister_debugfs();
1145
1146         return ret;
1147 }
1148
1149 static void __exit ipoib_cleanup_module(void)
1150 {
1151         ib_unregister_client(&ipoib_client);
1152         ipoib_unregister_debugfs();
1153         destroy_workqueue(ipoib_workqueue);
1154 }
1155
1156 module_init(ipoib_init_module);
1157 module_exit(ipoib_cleanup_module);