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

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