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) && 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                 *to_ipoib_neigh(n) = NULL;
 767                 kfree(neigh);
 768         }
 769
 770         spin_unlock_irqrestore(&priv->lock, flags);
 771
 772         if (ah)
 773                 ipoib_put_ah(ah);
 774 }
 775
 776 static int ipoib_neigh_setup_dev(struct net_device *dev, struct neigh_parms *parms)
 777 {
 778         parms->neigh_destructor = ipoib_neigh_destructor;
 779
 780         return 0;
 781 }
 782
 783 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
 784 {
 785         struct ipoib_dev_priv *priv = netdev_priv(dev);
 786
 787         /* Allocate RX/TX "rings" to hold queued skbs */
 788
 789         priv->rx_ring = kzalloc(IPOIB_RX_RING_SIZE * sizeof (struct ipoib_rx_buf),
 790                                 GFP_KERNEL);
 791         if (!priv->rx_ring) {
 792                 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
 793                        ca->name, IPOIB_RX_RING_SIZE);
 794                 goto out;
 795         }
 796
 797         priv->tx_ring = kzalloc(IPOIB_TX_RING_SIZE * sizeof (struct ipoib_tx_buf),
 798                                 GFP_KERNEL);
 799         if (!priv->tx_ring) {
 800                 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
 801                        ca->name, IPOIB_TX_RING_SIZE);
 802                 goto out_rx_ring_cleanup;
 803         }
 804
 805         /* priv->tx_head & tx_tail are already 0 */
 806
 807         if (ipoib_ib_dev_init(dev, ca, port))
 808                 goto out_tx_ring_cleanup;
 809
 810         return 0;
 811
 812 out_tx_ring_cleanup:
 813         kfree(priv->tx_ring);
 814
 815 out_rx_ring_cleanup:
 816         kfree(priv->rx_ring);
 817
 818 out:
 819         return -ENOMEM;
 820 }
 821
 822 void ipoib_dev_cleanup(struct net_device *dev)
 823 {
 824         struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
 825
 826         ipoib_delete_debug_files(dev);
 827
 828         /* Delete any child interfaces first */
 829         list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
 830                 unregister_netdev(cpriv->dev);
 831                 ipoib_dev_cleanup(cpriv->dev);
 832                 free_netdev(cpriv->dev);
 833         }
 834
 835         ipoib_ib_dev_cleanup(dev);
 836
 837         kfree(priv->rx_ring);
 838         kfree(priv->tx_ring);
 839
 840         priv->rx_ring = NULL;
 841         priv->tx_ring = NULL;
 842 }
 843
 844 static void ipoib_setup(struct net_device *dev)
 845 {
 846         struct ipoib_dev_priv *priv = netdev_priv(dev);
 847
 848         dev->open                = ipoib_open;
 849         dev->stop                = ipoib_stop;
 850         dev->change_mtu          = ipoib_change_mtu;
 851         dev->hard_start_xmit     = ipoib_start_xmit;
 852         dev->get_stats           = ipoib_get_stats;
 853         dev->tx_timeout          = ipoib_timeout;
 854         dev->hard_header         = ipoib_hard_header;
 855         dev->set_multicast_list  = ipoib_set_mcast_list;
 856         dev->neigh_setup         = ipoib_neigh_setup_dev;
 857
 858         dev->watchdog_timeo      = HZ;
 859
 860         dev->flags              |= IFF_BROADCAST | IFF_MULTICAST;
 861
 862         /*
 863          * We add in INFINIBAND_ALEN to allow for the destination
 864          * address "pseudoheader" for skbs without neighbour struct.
 865          */
 866         dev->hard_header_len     = IPOIB_ENCAP_LEN + INFINIBAND_ALEN;
 867         dev->addr_len            = INFINIBAND_ALEN;
 868         dev->type                = ARPHRD_INFINIBAND;
 869         dev->tx_queue_len        = IPOIB_TX_RING_SIZE * 2;
 870         dev->features            = NETIF_F_VLAN_CHALLENGED | NETIF_F_LLTX;
 871
 872         /* MTU will be reset when mcast join happens */
 873         dev->mtu                 = IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN;
 874         priv->mcast_mtu          = priv->admin_mtu = dev->mtu;
 875
 876         memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
 877
 878         netif_carrier_off(dev);
 879
 880         SET_MODULE_OWNER(dev);
 881
 882         priv->dev = dev;
 883
 884         spin_lock_init(&priv->lock);
 885         spin_lock_init(&priv->tx_lock);
 886
 887         mutex_init(&priv->mcast_mutex);
 888         mutex_init(&priv->vlan_mutex);
 889
 890         INIT_LIST_HEAD(&priv->path_list);
 891         INIT_LIST_HEAD(&priv->child_intfs);
 892         INIT_LIST_HEAD(&priv->dead_ahs);
 893         INIT_LIST_HEAD(&priv->multicast_list);
 894
 895         INIT_WORK(&priv->pkey_task,    ipoib_pkey_poll,          priv->dev);
 896         INIT_WORK(&priv->mcast_task,   ipoib_mcast_join_task,    priv->dev);
 897         INIT_WORK(&priv->flush_task,   ipoib_ib_dev_flush,       priv->dev);
 898         INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task, priv->dev);
 899         INIT_WORK(&priv->ah_reap_task, ipoib_reap_ah,            priv->dev);
 900 }
 901
 902 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
 903 {
 904         struct net_device *dev;
 905
 906         dev = alloc_netdev((int) sizeof (struct ipoib_dev_priv), name,
 907                            ipoib_setup);
 908         if (!dev)
 909                 return NULL;
 910
 911         return netdev_priv(dev);
 912 }
 913
 914 static ssize_t show_pkey(struct class_device *cdev, char *buf)
 915 {
 916         struct ipoib_dev_priv *priv =
 917                 netdev_priv(container_of(cdev, struct net_device, class_dev));
 918
 919         return sprintf(buf, "0x%04x\n", priv->pkey);
 920 }
 921 static CLASS_DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
 922
 923 static ssize_t create_child(struct class_device *cdev,
 924                             const char *buf, size_t count)
 925 {
 926         int pkey;
 927         int ret;
 928
 929         if (sscanf(buf, "%i", &pkey) != 1)
 930                 return -EINVAL;
 931
 932         if (pkey < 0 || pkey > 0xffff)
 933                 return -EINVAL;
 934
 935         /*
 936          * Set the full membership bit, so that we join the right
 937          * broadcast group, etc.
 938          */
 939         pkey |= 0x8000;
 940
 941         ret = ipoib_vlan_add(container_of(cdev, struct net_device, class_dev),
 942                              pkey);
 943
 944         return ret ? ret : count;
 945 }
 946 static CLASS_DEVICE_ATTR(create_child, S_IWUGO, NULL, create_child);
 947
 948 static ssize_t delete_child(struct class_device *cdev,
 949                             const char *buf, size_t count)
 950 {
 951         int pkey;
 952         int ret;
 953
 954         if (sscanf(buf, "%i", &pkey) != 1)
 955                 return -EINVAL;
 956
 957         if (pkey < 0 || pkey > 0xffff)
 958                 return -EINVAL;
 959
 960         ret = ipoib_vlan_delete(container_of(cdev, struct net_device, class_dev),
 961                                 pkey);
 962
 963         return ret ? ret : count;
 964
 965 }
 966 static CLASS_DEVICE_ATTR(delete_child, S_IWUGO, NULL, delete_child);
 967
 968 int ipoib_add_pkey_attr(struct net_device *dev)
 969 {
 970         return class_device_create_file(&dev->class_dev,
 971                                         &class_device_attr_pkey);
 972 }
 973
 974 static struct net_device *ipoib_add_port(const char *format,
 975                                          struct ib_device *hca, u8 port)
 976 {
 977         struct ipoib_dev_priv *priv;
 978         int result = -ENOMEM;
 979
 980         priv = ipoib_intf_alloc(format);
 981         if (!priv)
 982                 goto alloc_mem_failed;
 983
 984         SET_NETDEV_DEV(priv->dev, hca->dma_device);
 985
 986         result = ib_query_pkey(hca, port, 0, &priv->pkey);
 987         if (result) {
 988                 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
 989                        hca->name, port, result);
 990                 goto alloc_mem_failed;
 991         }
 992
 993         /*
 994          * Set the full membership bit, so that we join the right
 995          * broadcast group, etc.
 996          */
 997         priv->pkey |= 0x8000;
 998
 999         priv->dev->broadcast[8] = priv->pkey >> 8;
1000         priv->dev->broadcast[9] = priv->pkey & 0xff;
1001
1002         result = ib_query_gid(hca, port, 0, &priv->local_gid);
1003         if (result) {
1004                 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1005                        hca->name, port, result);
1006                 goto alloc_mem_failed;
1007         } else
1008                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1009
1010
1011         result = ipoib_dev_init(priv->dev, hca, port);
1012         if (result < 0) {
1013                 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1014                        hca->name, port, result);
1015                 goto device_init_failed;
1016         }
1017
1018         INIT_IB_EVENT_HANDLER(&priv->event_handler,
1019                               priv->ca, ipoib_event);
1020         result = ib_register_event_handler(&priv->event_handler);
1021         if (result < 0) {
1022                 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1023                        "port %d (ret = %d)\n",
1024                        hca->name, port, result);
1025                 goto event_failed;
1026         }
1027
1028         result = register_netdev(priv->dev);
1029         if (result) {
1030                 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1031                        hca->name, port, result);
1032                 goto register_failed;
1033         }
1034
1035         ipoib_create_debug_files(priv->dev);
1036
1037         if (ipoib_add_pkey_attr(priv->dev))
1038                 goto sysfs_failed;
1039         if (class_device_create_file(&priv->dev->class_dev,
1040                                      &class_device_attr_create_child))
1041                 goto sysfs_failed;
1042         if (class_device_create_file(&priv->dev->class_dev,
1043                                      &class_device_attr_delete_child))
1044                 goto sysfs_failed;
1045
1046         return priv->dev;
1047
1048 sysfs_failed:
1049         ipoib_delete_debug_files(priv->dev);
1050         unregister_netdev(priv->dev);
1051
1052 register_failed:
1053         ib_unregister_event_handler(&priv->event_handler);
1054         flush_scheduled_work();
1055
1056 event_failed:
1057         ipoib_dev_cleanup(priv->dev);
1058
1059 device_init_failed:
1060         free_netdev(priv->dev);
1061
1062 alloc_mem_failed:
1063         return ERR_PTR(result);
1064 }
1065
1066 static void ipoib_add_one(struct ib_device *device)
1067 {
1068         struct list_head *dev_list;
1069         struct net_device *dev;
1070         struct ipoib_dev_priv *priv;
1071         int s, e, p;
1072
1073         dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1074         if (!dev_list)
1075                 return;
1076
1077         INIT_LIST_HEAD(dev_list);
1078
1079         if (device->node_type == IB_NODE_SWITCH) {
1080                 s = 0;
1081                 e = 0;
1082         } else {
1083                 s = 1;
1084                 e = device->phys_port_cnt;
1085         }
1086
1087         for (p = s; p <= e; ++p) {
1088                 dev = ipoib_add_port("ib%d", device, p);
1089                 if (!IS_ERR(dev)) {
1090                         priv = netdev_priv(dev);
1091                         list_add_tail(&priv->list, dev_list);
1092                 }
1093         }
1094
1095         ib_set_client_data(device, &ipoib_client, dev_list);
1096 }
1097
1098 static void ipoib_remove_one(struct ib_device *device)
1099 {
1100         struct ipoib_dev_priv *priv, *tmp;
1101         struct list_head *dev_list;
1102
1103         dev_list = ib_get_client_data(device, &ipoib_client);
1104
1105         list_for_each_entry_safe(priv, tmp, dev_list, list) {
1106                 ib_unregister_event_handler(&priv->event_handler);
1107                 flush_scheduled_work();
1108
1109                 unregister_netdev(priv->dev);
1110                 ipoib_dev_cleanup(priv->dev);
1111                 free_netdev(priv->dev);
1112         }
1113
1114         kfree(dev_list);
1115 }
1116
1117 static int __init ipoib_init_module(void)
1118 {
1119         int ret;
1120
1121         ret = ipoib_register_debugfs();
1122         if (ret)
1123                 return ret;
1124
1125         /*
1126          * We create our own workqueue mainly because we want to be
1127          * able to flush it when devices are being removed.  We can't
1128          * use schedule_work()/flush_scheduled_work() because both
1129          * unregister_netdev() and linkwatch_event take the rtnl lock,
1130          * so flush_scheduled_work() can deadlock during device
1131          * removal.
1132          */
1133         ipoib_workqueue = create_singlethread_workqueue("ipoib");
1134         if (!ipoib_workqueue) {
1135                 ret = -ENOMEM;
1136                 goto err_fs;
1137         }
1138
1139         ret = ib_register_client(&ipoib_client);
1140         if (ret)
1141                 goto err_wq;
1142
1143         return 0;
1144
1145 err_wq:
1146         destroy_workqueue(ipoib_workqueue);
1147
1148 err_fs:
1149         ipoib_unregister_debugfs();
1150
1151         return ret;
1152 }
1153
1154 static void __exit ipoib_cleanup_module(void)
1155 {
1156         ib_unregister_client(&ipoib_client);
1157         ipoib_unregister_debugfs();
1158         destroy_workqueue(ipoib_workqueue);
1159 }
1160
1161 module_init(ipoib_init_module);
1162 module_exit(ipoib_cleanup_module);