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 = kmalloc(sizeof *path, GFP_ATOMIC);
 360         if (!path)
 361                 return NULL;
 362
 363         path->dev          = dev;
 364         path->pathrec.dlid = 0;
 365         path->ah           = NULL;
 366
 367         skb_queue_head_init(&path->queue);
 368
 369         INIT_LIST_HEAD(&path->neigh_list);
 370         path->query = NULL;
 371         init_completion(&path->done);
 372
 373         memcpy(path->pathrec.dgid.raw, gid->raw, sizeof (union ib_gid));
 374         path->pathrec.sgid      = priv->local_gid;
 375         path->pathrec.pkey      = cpu_to_be16(priv->pkey);
 376         path->pathrec.numb_path = 1;
 377
 378         return path;
 379 }
 380
 381 static int path_rec_start(struct net_device *dev,
 382                           struct ipoib_path *path)
 383 {
 384         struct ipoib_dev_priv *priv = netdev_priv(dev);
 385
 386         ipoib_dbg(priv, "Start path record lookup for " IPOIB_GID_FMT "\n",
 387                   IPOIB_GID_ARG(path->pathrec.dgid));
 388
 389         path->query_id =
 390                 ib_sa_path_rec_get(priv->ca, priv->port,
 391                                    &path->pathrec,
 392                                    IB_SA_PATH_REC_DGID          |
 393                                    IB_SA_PATH_REC_SGID          |
 394                                    IB_SA_PATH_REC_NUMB_PATH     |
 395                                    IB_SA_PATH_REC_PKEY,
 396                                    1000, GFP_ATOMIC,
 397                                    path_rec_completion,
 398                                    path, &path->query);
 399         if (path->query_id < 0) {
 400                 ipoib_warn(priv, "ib_sa_path_rec_get failed\n");
 401                 path->query = NULL;
 402                 return path->query_id;
 403         }
 404
 405         return 0;
 406 }
 407
 408 static void neigh_add_path(struct sk_buff *skb, struct net_device *dev)
 409 {
 410         struct ipoib_dev_priv *priv = netdev_priv(dev);
 411         struct ipoib_path *path;
 412         struct ipoib_neigh *neigh;
 413
 414         neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
 415         if (!neigh) {
 416                 ++priv->stats.tx_dropped;
 417                 dev_kfree_skb_any(skb);
 418                 return;
 419         }
 420
 421         skb_queue_head_init(&neigh->queue);
 422         neigh->neighbour = skb->dst->neighbour;
 423         *to_ipoib_neigh(skb->dst->neighbour) = neigh;
 424
 425         /*
 426          * We can only be called from ipoib_start_xmit, so we're
 427          * inside tx_lock -- no need to save/restore flags.
 428          */
 429         spin_lock(&priv->lock);
 430
 431         path = __path_find(dev, (union ib_gid *) (skb->dst->neighbour->ha + 4));
 432         if (!path) {
 433                 path = path_rec_create(dev,
 434                                        (union ib_gid *) (skb->dst->neighbour->ha + 4));
 435                 if (!path)
 436                         goto err;
 437
 438                 __path_add(dev, path);
 439         }
 440
 441         list_add_tail(&neigh->list, &path->neigh_list);
 442
 443         if (path->pathrec.dlid) {
 444                 kref_get(&path->ah->ref);
 445                 neigh->ah = path->ah;
 446
 447                 ipoib_send(dev, skb, path->ah,
 448                            be32_to_cpup((__be32 *) skb->dst->neighbour->ha));
 449         } else {
 450                 neigh->ah  = NULL;
 451                 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 452                         __skb_queue_tail(&neigh->queue, skb);
 453                 } else {
 454                         ++priv->stats.tx_dropped;
 455                         dev_kfree_skb_any(skb);
 456                 }
 457
 458                 if (!path->query && path_rec_start(dev, path))
 459                         goto err;
 460         }
 461
 462         spin_unlock(&priv->lock);
 463         return;
 464
 465 err:
 466         *to_ipoib_neigh(skb->dst->neighbour) = NULL;
 467         list_del(&neigh->list);
 468         neigh->neighbour->ops->destructor = NULL;
 469         kfree(neigh);
 470
 471         ++priv->stats.tx_dropped;
 472         dev_kfree_skb_any(skb);
 473
 474         spin_unlock(&priv->lock);
 475 }
 476
 477 static void ipoib_path_lookup(struct sk_buff *skb, struct net_device *dev)
 478 {
 479         struct ipoib_dev_priv *priv = netdev_priv(skb->dev);
 480
 481         /* Look up path record for unicasts */
 482         if (skb->dst->neighbour->ha[4] != 0xff) {
 483                 neigh_add_path(skb, dev);
 484                 return;
 485         }
 486
 487         /* Add in the P_Key for multicasts */
 488         skb->dst->neighbour->ha[8] = (priv->pkey >> 8) & 0xff;
 489         skb->dst->neighbour->ha[9] = priv->pkey & 0xff;
 490         ipoib_mcast_send(dev, (union ib_gid *) (skb->dst->neighbour->ha + 4), skb);
 491 }
 492
 493 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
 494                              struct ipoib_pseudoheader *phdr)
 495 {
 496         struct ipoib_dev_priv *priv = netdev_priv(dev);
 497         struct ipoib_path *path;
 498
 499         /*
 500          * We can only be called from ipoib_start_xmit, so we're
 501          * inside tx_lock -- no need to save/restore flags.
 502          */
 503         spin_lock(&priv->lock);
 504
 505         path = __path_find(dev, (union ib_gid *) (phdr->hwaddr + 4));
 506         if (!path) {
 507                 path = path_rec_create(dev,
 508                                        (union ib_gid *) (phdr->hwaddr + 4));
 509                 if (path) {
 510                         /* put pseudoheader back on for next time */
 511                         skb_push(skb, sizeof *phdr);
 512                         __skb_queue_tail(&path->queue, skb);
 513
 514                         if (path_rec_start(dev, path)) {
 515                                 spin_unlock(&priv->lock);
 516                                 path_free(dev, path);
 517                                 return;
 518                         } else
 519                                 __path_add(dev, path);
 520                 } else {
 521                         ++priv->stats.tx_dropped;
 522                         dev_kfree_skb_any(skb);
 523                 }
 524
 525                 spin_unlock(&priv->lock);
 526                 return;
 527         }
 528
 529         if (path->pathrec.dlid) {
 530                 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
 531                           be16_to_cpu(path->pathrec.dlid));
 532
 533                 ipoib_send(dev, skb, path->ah,
 534                            be32_to_cpup((__be32 *) phdr->hwaddr));
 535         } else if ((path->query || !path_rec_start(dev, path)) &&
 536                    skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 537                 /* put pseudoheader back on for next time */
 538                 skb_push(skb, sizeof *phdr);
 539                 __skb_queue_tail(&path->queue, skb);
 540         } else {
 541                 ++priv->stats.tx_dropped;
 542                 dev_kfree_skb_any(skb);
 543         }
 544
 545         spin_unlock(&priv->lock);
 546 }
 547
 548 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
 549 {
 550         struct ipoib_dev_priv *priv = netdev_priv(dev);
 551         struct ipoib_neigh *neigh;
 552         unsigned long flags;
 553
 554         local_irq_save(flags);
 555         if (!spin_trylock(&priv->tx_lock)) {
 556                 local_irq_restore(flags);
 557                 return NETDEV_TX_LOCKED;
 558         }
 559
 560         /*
 561          * Check if our queue is stopped.  Since we have the LLTX bit
 562          * set, we can't rely on netif_stop_queue() preventing our
 563          * xmit function from being called with a full queue.
 564          */
 565         if (unlikely(netif_queue_stopped(dev))) {
 566                 spin_unlock_irqrestore(&priv->tx_lock, flags);
 567                 return NETDEV_TX_BUSY;
 568         }
 569
 570         if (skb->dst && skb->dst->neighbour) {
 571                 if (unlikely(!*to_ipoib_neigh(skb->dst->neighbour))) {
 572                         ipoib_path_lookup(skb, dev);
 573                         goto out;
 574                 }
 575
 576                 neigh = *to_ipoib_neigh(skb->dst->neighbour);
 577
 578                 if (likely(neigh->ah)) {
 579                         ipoib_send(dev, skb, neigh->ah,
 580                                    be32_to_cpup((__be32 *) skb->dst->neighbour->ha));
 581                         goto out;
 582                 }
 583
 584                 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 585                         spin_lock(&priv->lock);
 586                         __skb_queue_tail(&neigh->queue, skb);
 587                         spin_unlock(&priv->lock);
 588                 } else {
 589                         ++priv->stats.tx_dropped;
 590                         dev_kfree_skb_any(skb);
 591                 }
 592         } else {
 593                 struct ipoib_pseudoheader *phdr =
 594                         (struct ipoib_pseudoheader *) skb->data;
 595                 skb_pull(skb, sizeof *phdr);
 596
 597                 if (phdr->hwaddr[4] == 0xff) {
 598                         /* Add in the P_Key for multicast*/
 599                         phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
 600                         phdr->hwaddr[9] = priv->pkey & 0xff;
 601
 602                         ipoib_mcast_send(dev, (union ib_gid *) (phdr->hwaddr + 4), skb);
 603                 } else {
 604                         /* unicast GID -- should be ARP or RARP reply */
 605
 606                         if ((be16_to_cpup((__be16 *) skb->data) != ETH_P_ARP) &&
 607                             (be16_to_cpup((__be16 *) skb->data) != ETH_P_RARP)) {
 608                                 ipoib_warn(priv, "Unicast, no %s: type %04x, QPN %06x "
 609                                            IPOIB_GID_FMT "\n",
 610                                            skb->dst ? "neigh" : "dst",
 611                                            be16_to_cpup((__be16 *) skb->data),
 612                                            be32_to_cpup((__be32 *) phdr->hwaddr),
 613                                            IPOIB_GID_ARG(*(union ib_gid *) (phdr->hwaddr + 4)));
 614                                 dev_kfree_skb_any(skb);
 615                                 ++priv->stats.tx_dropped;
 616                                 goto out;
 617                         }
 618
 619                         unicast_arp_send(skb, dev, phdr);
 620                 }
 621         }
 622
 623 out:
 624         spin_unlock_irqrestore(&priv->tx_lock, flags);
 625
 626         return NETDEV_TX_OK;
 627 }
 628
 629 static struct net_device_stats *ipoib_get_stats(struct net_device *dev)
 630 {
 631         struct ipoib_dev_priv *priv = netdev_priv(dev);
 632
 633         return &priv->stats;
 634 }
 635
 636 static void ipoib_timeout(struct net_device *dev)
 637 {
 638         struct ipoib_dev_priv *priv = netdev_priv(dev);
 639
 640         ipoib_warn(priv, "transmit timeout: latency %ld\n",
 641                    jiffies - dev->trans_start);
 642         /* XXX reset QP, etc. */
 643 }
 644
 645 static int ipoib_hard_header(struct sk_buff *skb,
 646                              struct net_device *dev,
 647                              unsigned short type,
 648                              void *daddr, void *saddr, unsigned len)
 649 {
 650         struct ipoib_header *header;
 651
 652         header = (struct ipoib_header *) skb_push(skb, sizeof *header);
 653
 654         header->proto = htons(type);
 655         header->reserved = 0;
 656
 657         /*
 658          * If we don't have a neighbour structure, stuff the
 659          * destination address onto the front of the skb so we can
 660          * figure out where to send the packet later.
 661          */
 662         if (!skb->dst || !skb->dst->neighbour) {
 663                 struct ipoib_pseudoheader *phdr =
 664                         (struct ipoib_pseudoheader *) skb_push(skb, sizeof *phdr);
 665                 memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
 666         }
 667
 668         return 0;
 669 }
 670
 671 static void ipoib_set_mcast_list(struct net_device *dev)
 672 {
 673         struct ipoib_dev_priv *priv = netdev_priv(dev);
 674
 675         queue_work(ipoib_workqueue, &priv->restart_task);
 676 }
 677
 678 static void ipoib_neigh_destructor(struct neighbour *n)
 679 {
 680         struct ipoib_neigh *neigh;
 681         struct ipoib_dev_priv *priv = netdev_priv(n->dev);
 682         unsigned long flags;
 683         struct ipoib_ah *ah = NULL;
 684
 685         ipoib_dbg(priv,
 686                   "neigh_destructor for %06x " IPOIB_GID_FMT "\n",
 687                   be32_to_cpup((__be32 *) n->ha),
 688                   IPOIB_GID_ARG(*((union ib_gid *) (n->ha + 4))));
 689
 690         spin_lock_irqsave(&priv->lock, flags);
 691
 692         neigh = *to_ipoib_neigh(n);
 693         if (neigh) {
 694                 if (neigh->ah)
 695                         ah = neigh->ah;
 696                 list_del(&neigh->list);
 697                 *to_ipoib_neigh(n) = NULL;
 698                 kfree(neigh);
 699         }
 700
 701         spin_unlock_irqrestore(&priv->lock, flags);
 702
 703         if (ah)
 704                 ipoib_put_ah(ah);
 705 }
 706
 707 static int ipoib_neigh_setup(struct neighbour *neigh)
 708 {
 709         /*
 710          * Is this kosher?  I can't find anybody in the kernel that
 711          * sets neigh->destructor, so we should be able to set it here
 712          * without trouble.
 713          */
 714         neigh->ops->destructor = ipoib_neigh_destructor;
 715
 716         return 0;
 717 }
 718
 719 static int ipoib_neigh_setup_dev(struct net_device *dev, struct neigh_parms *parms)
 720 {
 721         parms->neigh_setup = ipoib_neigh_setup;
 722
 723         return 0;
 724 }
 725
 726 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
 727 {
 728         struct ipoib_dev_priv *priv = netdev_priv(dev);
 729
 730         /* Allocate RX/TX "rings" to hold queued skbs */
 731
 732         priv->rx_ring = kmalloc(IPOIB_RX_RING_SIZE * sizeof (struct ipoib_buf),
 733                                 GFP_KERNEL);
 734         if (!priv->rx_ring) {
 735                 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
 736                        ca->name, IPOIB_RX_RING_SIZE);
 737                 goto out;
 738         }
 739         memset(priv->rx_ring, 0,
 740                IPOIB_RX_RING_SIZE * sizeof (struct ipoib_buf));
 741
 742         priv->tx_ring = kmalloc(IPOIB_TX_RING_SIZE * sizeof (struct ipoib_buf),
 743                                 GFP_KERNEL);
 744         if (!priv->tx_ring) {
 745                 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
 746                        ca->name, IPOIB_TX_RING_SIZE);
 747                 goto out_rx_ring_cleanup;
 748         }
 749         memset(priv->tx_ring, 0,
 750                IPOIB_TX_RING_SIZE * sizeof (struct ipoib_buf));
 751
 752         /* priv->tx_head & tx_tail are already 0 */
 753
 754         if (ipoib_ib_dev_init(dev, ca, port))
 755                 goto out_tx_ring_cleanup;
 756
 757         return 0;
 758
 759 out_tx_ring_cleanup:
 760         kfree(priv->tx_ring);
 761
 762 out_rx_ring_cleanup:
 763         kfree(priv->rx_ring);
 764
 765 out:
 766         return -ENOMEM;
 767 }
 768
 769 void ipoib_dev_cleanup(struct net_device *dev)
 770 {
 771         struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
 772
 773         ipoib_delete_debug_file(dev);
 774
 775         /* Delete any child interfaces first */
 776         list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
 777                 unregister_netdev(cpriv->dev);
 778                 ipoib_dev_cleanup(cpriv->dev);
 779                 free_netdev(cpriv->dev);
 780         }
 781
 782         ipoib_ib_dev_cleanup(dev);
 783
 784         kfree(priv->rx_ring);
 785         kfree(priv->tx_ring);
 786
 787         priv->rx_ring = NULL;
 788         priv->tx_ring = NULL;
 789 }
 790
 791 static void ipoib_setup(struct net_device *dev)
 792 {
 793         struct ipoib_dev_priv *priv = netdev_priv(dev);
 794
 795         dev->open                = ipoib_open;
 796         dev->stop                = ipoib_stop;
 797         dev->change_mtu          = ipoib_change_mtu;
 798         dev->hard_start_xmit     = ipoib_start_xmit;
 799         dev->get_stats           = ipoib_get_stats;
 800         dev->tx_timeout          = ipoib_timeout;
 801         dev->hard_header         = ipoib_hard_header;
 802         dev->set_multicast_list  = ipoib_set_mcast_list;
 803         dev->neigh_setup         = ipoib_neigh_setup_dev;
 804
 805         dev->watchdog_timeo      = HZ;
 806
 807         dev->rebuild_header      = NULL;
 808         dev->set_mac_address     = NULL;
 809         dev->header_cache_update = NULL;
 810
 811         dev->flags              |= IFF_BROADCAST | IFF_MULTICAST;
 812
 813         /*
 814          * We add in INFINIBAND_ALEN to allow for the destination
 815          * address "pseudoheader" for skbs without neighbour struct.
 816          */
 817         dev->hard_header_len     = IPOIB_ENCAP_LEN + INFINIBAND_ALEN;
 818         dev->addr_len            = INFINIBAND_ALEN;
 819         dev->type                = ARPHRD_INFINIBAND;
 820         dev->tx_queue_len        = IPOIB_TX_RING_SIZE * 2;
 821         dev->features            = NETIF_F_VLAN_CHALLENGED | NETIF_F_LLTX;
 822
 823         /* MTU will be reset when mcast join happens */
 824         dev->mtu                 = IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN;
 825         priv->mcast_mtu          = priv->admin_mtu = dev->mtu;
 826
 827         memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
 828
 829         netif_carrier_off(dev);
 830
 831         SET_MODULE_OWNER(dev);
 832
 833         priv->dev = dev;
 834
 835         spin_lock_init(&priv->lock);
 836         spin_lock_init(&priv->tx_lock);
 837
 838         init_MUTEX(&priv->mcast_mutex);
 839         init_MUTEX(&priv->vlan_mutex);
 840
 841         INIT_LIST_HEAD(&priv->path_list);
 842         INIT_LIST_HEAD(&priv->child_intfs);
 843         INIT_LIST_HEAD(&priv->dead_ahs);
 844         INIT_LIST_HEAD(&priv->multicast_list);
 845
 846         INIT_WORK(&priv->pkey_task,    ipoib_pkey_poll,          priv->dev);
 847         INIT_WORK(&priv->mcast_task,   ipoib_mcast_join_task,    priv->dev);
 848         INIT_WORK(&priv->flush_task,   ipoib_ib_dev_flush,       priv->dev);
 849         INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task, priv->dev);
 850         INIT_WORK(&priv->ah_reap_task, ipoib_reap_ah,            priv->dev);
 851 }
 852
 853 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
 854 {
 855         struct net_device *dev;
 856
 857         dev = alloc_netdev((int) sizeof (struct ipoib_dev_priv), name,
 858                            ipoib_setup);
 859         if (!dev)
 860                 return NULL;
 861
 862         return netdev_priv(dev);
 863 }
 864
 865 static ssize_t show_pkey(struct class_device *cdev, char *buf)
 866 {
 867         struct ipoib_dev_priv *priv =
 868                 netdev_priv(container_of(cdev, struct net_device, class_dev));
 869
 870         return sprintf(buf, "0x%04x\n", priv->pkey);
 871 }
 872 static CLASS_DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
 873
 874 static ssize_t create_child(struct class_device *cdev,
 875                             const char *buf, size_t count)
 876 {
 877         int pkey;
 878         int ret;
 879
 880         if (sscanf(buf, "%i", &pkey) != 1)
 881                 return -EINVAL;
 882
 883         if (pkey < 0 || pkey > 0xffff)
 884                 return -EINVAL;
 885
 886         /*
 887          * Set the full membership bit, so that we join the right
 888          * broadcast group, etc.
 889          */
 890         pkey |= 0x8000;
 891
 892         ret = ipoib_vlan_add(container_of(cdev, struct net_device, class_dev),
 893                              pkey);
 894
 895         return ret ? ret : count;
 896 }
 897 static CLASS_DEVICE_ATTR(create_child, S_IWUGO, NULL, create_child);
 898
 899 static ssize_t delete_child(struct class_device *cdev,
 900                             const char *buf, size_t count)
 901 {
 902         int pkey;
 903         int ret;
 904
 905         if (sscanf(buf, "%i", &pkey) != 1)
 906                 return -EINVAL;
 907
 908         if (pkey < 0 || pkey > 0xffff)
 909                 return -EINVAL;
 910
 911         ret = ipoib_vlan_delete(container_of(cdev, struct net_device, class_dev),
 912                                 pkey);
 913
 914         return ret ? ret : count;
 915
 916 }
 917 static CLASS_DEVICE_ATTR(delete_child, S_IWUGO, NULL, delete_child);
 918
 919 int ipoib_add_pkey_attr(struct net_device *dev)
 920 {
 921         return class_device_create_file(&dev->class_dev,
 922                                         &class_device_attr_pkey);
 923 }
 924
 925 static struct net_device *ipoib_add_port(const char *format,
 926                                          struct ib_device *hca, u8 port)
 927 {
 928         struct ipoib_dev_priv *priv;
 929         int result = -ENOMEM;
 930
 931         priv = ipoib_intf_alloc(format);
 932         if (!priv)
 933                 goto alloc_mem_failed;
 934
 935         SET_NETDEV_DEV(priv->dev, hca->dma_device);
 936
 937         result = ib_query_pkey(hca, port, 0, &priv->pkey);
 938         if (result) {
 939                 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
 940                        hca->name, port, result);
 941                 goto alloc_mem_failed;
 942         }
 943
 944         /*
 945          * Set the full membership bit, so that we join the right
 946          * broadcast group, etc.
 947          */
 948         priv->pkey |= 0x8000;
 949
 950         priv->dev->broadcast[8] = priv->pkey >> 8;
 951         priv->dev->broadcast[9] = priv->pkey & 0xff;
 952
 953         result = ib_query_gid(hca, port, 0, &priv->local_gid);
 954         if (result) {
 955                 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
 956                        hca->name, port, result);
 957                 goto alloc_mem_failed;
 958         } else
 959                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
 960
 961
 962         result = ipoib_dev_init(priv->dev, hca, port);
 963         if (result < 0) {
 964                 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
 965                        hca->name, port, result);
 966                 goto device_init_failed;
 967         }
 968
 969         INIT_IB_EVENT_HANDLER(&priv->event_handler,
 970                               priv->ca, ipoib_event);
 971         result = ib_register_event_handler(&priv->event_handler);
 972         if (result < 0) {
 973                 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
 974                        "port %d (ret = %d)\n",
 975                        hca->name, port, result);
 976                 goto event_failed;
 977         }
 978
 979         result = register_netdev(priv->dev);
 980         if (result) {
 981                 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
 982                        hca->name, port, result);
 983                 goto register_failed;
 984         }
 985
 986         if (ipoib_create_debug_file(priv->dev))
 987                 goto debug_failed;
 988
 989         if (ipoib_add_pkey_attr(priv->dev))
 990                 goto sysfs_failed;
 991         if (class_device_create_file(&priv->dev->class_dev,
 992                                      &class_device_attr_create_child))
 993                 goto sysfs_failed;
 994         if (class_device_create_file(&priv->dev->class_dev,
 995                                      &class_device_attr_delete_child))
 996                 goto sysfs_failed;
 997
 998         return priv->dev;
 999
1000 sysfs_failed:
1001         ipoib_delete_debug_file(priv->dev);
1002
1003 debug_failed:
1004         unregister_netdev(priv->dev);
1005
1006 register_failed:
1007         ib_unregister_event_handler(&priv->event_handler);
1008         flush_scheduled_work();
1009
1010 event_failed:
1011         ipoib_dev_cleanup(priv->dev);
1012
1013 device_init_failed:
1014         free_netdev(priv->dev);
1015
1016 alloc_mem_failed:
1017         return ERR_PTR(result);
1018 }
1019
1020 static void ipoib_add_one(struct ib_device *device)
1021 {
1022         struct list_head *dev_list;
1023         struct net_device *dev;
1024         struct ipoib_dev_priv *priv;
1025         int s, e, p;
1026
1027         dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1028         if (!dev_list)
1029                 return;
1030
1031         INIT_LIST_HEAD(dev_list);
1032
1033         if (device->node_type == IB_NODE_SWITCH) {
1034                 s = 0;
1035                 e = 0;
1036         } else {
1037                 s = 1;
1038                 e = device->phys_port_cnt;
1039         }
1040
1041         for (p = s; p <= e; ++p) {
1042                 dev = ipoib_add_port("ib%d", device, p);
1043                 if (!IS_ERR(dev)) {
1044                         priv = netdev_priv(dev);
1045                         list_add_tail(&priv->list, dev_list);
1046                 }
1047         }
1048
1049         ib_set_client_data(device, &ipoib_client, dev_list);
1050 }
1051
1052 static void ipoib_remove_one(struct ib_device *device)
1053 {
1054         struct ipoib_dev_priv *priv, *tmp;
1055         struct list_head *dev_list;
1056
1057         dev_list = ib_get_client_data(device, &ipoib_client);
1058
1059         list_for_each_entry_safe(priv, tmp, dev_list, list) {
1060                 ib_unregister_event_handler(&priv->event_handler);
1061                 flush_scheduled_work();
1062
1063                 unregister_netdev(priv->dev);
1064                 ipoib_dev_cleanup(priv->dev);
1065                 free_netdev(priv->dev);
1066         }
1067
1068         kfree(dev_list);
1069 }
1070
1071 static int __init ipoib_init_module(void)
1072 {
1073         int ret;
1074
1075         ret = ipoib_register_debugfs();
1076         if (ret)
1077                 return ret;
1078
1079         /*
1080          * We create our own workqueue mainly because we want to be
1081          * able to flush it when devices are being removed.  We can't
1082          * use schedule_work()/flush_scheduled_work() because both
1083          * unregister_netdev() and linkwatch_event take the rtnl lock,
1084          * so flush_scheduled_work() can deadlock during device
1085          * removal.
1086          */
1087         ipoib_workqueue = create_singlethread_workqueue("ipoib");
1088         if (!ipoib_workqueue) {
1089                 ret = -ENOMEM;
1090                 goto err_fs;
1091         }
1092
1093         ret = ib_register_client(&ipoib_client);
1094         if (ret)
1095                 goto err_wq;
1096
1097         return 0;
1098
1099 err_wq:
1100         destroy_workqueue(ipoib_workqueue);
1101
1102 err_fs:
1103         ipoib_unregister_debugfs();
1104
1105         return ret;
1106 }
1107
1108 static void __exit ipoib_cleanup_module(void)
1109 {
1110         ib_unregister_client(&ipoib_client);
1111         ipoib_unregister_debugfs();
1112         destroy_workqueue(ipoib_workqueue);
1113 }
1114
1115 module_init(ipoib_init_module);
1116 module_exit(ipoib_cleanup_module);