drivers/infiniband/ulp/ipoib/ipoib_main.c

   1 /*
   2  * Copyright (c) 2004 Topspin Communications.  All rights reserved.
   3  *
   4  * This software is available to you under a choice of one of two
   5  * licenses.  You may choose to be licensed under the terms of the GNU
   6  * General Public License (GPL) Version 2, available from the file
   7  * COPYING in the main directory of this source tree, or the
   8  * OpenIB.org BSD license below:
   9  *
  10  *     Redistribution and use in source and binary forms, with or
  11  *     without modification, are permitted provided that the following
  12  *     conditions are met:
  13  *
  14  *      - Redistributions of source code must retain the above
  15  *        copyright notice, this list of conditions and the following
  16  *        disclaimer.
  17  *
  18  *      - Redistributions in binary form must reproduce the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer in the documentation and/or other materials
  21  *        provided with the distribution.
  22  *
  23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30  * SOFTWARE.
  31  *
  32  * $Id: ipoib_main.c 1377 2004-12-23 19:57:12Z roland $
  33  */
  34
  35 #include "ipoib.h"
  36
  37 #include <linux/version.h>
  38 #include <linux/module.h>
  39
  40 #include <linux/init.h>
  41 #include <linux/slab.h>
  42 #include <linux/vmalloc.h>
  43
  44 #include <linux/if_arp.h>       /* For ARPHRD_xxx */
  45
  46 #include <linux/ip.h>
  47 #include <linux/in.h>
  48
  49 MODULE_AUTHOR("Roland Dreier");
  50 MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
  51 MODULE_LICENSE("Dual BSD/GPL");
  52
  53 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
  54 int ipoib_debug_level;
  55
  56 module_param_named(debug_level, ipoib_debug_level, int, 0644);
  57 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
  58 #endif
  59
  60 static const u8 ipv4_bcast_addr[] = {
  61         0x00, 0xff, 0xff, 0xff,
  62         0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
  63         0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff
  64 };
  65
  66 struct workqueue_struct *ipoib_workqueue;
  67
  68 static void ipoib_add_one(struct ib_device *device);
  69 static void ipoib_remove_one(struct ib_device *device);
  70
  71 static struct ib_client ipoib_client = {
  72         .name   = "ipoib",
  73         .add    = ipoib_add_one,
  74         .remove = ipoib_remove_one
  75 };
  76
  77 int ipoib_open(struct net_device *dev)
  78 {
  79         struct ipoib_dev_priv *priv = netdev_priv(dev);
  80
  81         ipoib_dbg(priv, "bringing up interface\n");
  82
  83         set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
  84
  85         if (ipoib_pkey_dev_delay_open(dev))
  86                 return 0;
  87
  88         if (ipoib_ib_dev_open(dev))
  89                 return -EINVAL;
  90
  91         if (ipoib_ib_dev_up(dev))
  92                 return -EINVAL;
  93
  94         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
  95                 struct ipoib_dev_priv *cpriv;
  96
  97                 /* Bring up any child interfaces too */
  98                 down(&priv->vlan_mutex);
  99                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 100                         int flags;
 101
 102                         flags = cpriv->dev->flags;
 103                         if (flags & IFF_UP)
 104                                 continue;
 105
 106                         dev_change_flags(cpriv->dev, flags | IFF_UP);
 107                 }
 108                 up(&priv->vlan_mutex);
 109         }
 110
 111         netif_start_queue(dev);
 112
 113         return 0;
 114 }
 115
 116 static int ipoib_stop(struct net_device *dev)
 117 {
 118         struct ipoib_dev_priv *priv = netdev_priv(dev);
 119
 120         ipoib_dbg(priv, "stopping interface\n");
 121
 122         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 123
 124         netif_stop_queue(dev);
 125
 126         ipoib_ib_dev_down(dev);
 127         ipoib_ib_dev_stop(dev);
 128
 129         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 130                 struct ipoib_dev_priv *cpriv;
 131
 132                 /* Bring down any child interfaces too */
 133                 down(&priv->vlan_mutex);
 134                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 135                         int flags;
 136
 137                         flags = cpriv->dev->flags;
 138                         if (!(flags & IFF_UP))
 139                                 continue;
 140
 141                         dev_change_flags(cpriv->dev, flags & ~IFF_UP);
 142                 }
 143                 up(&priv->vlan_mutex);
 144         }
 145
 146         return 0;
 147 }
 148
 149 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
 150 {
 151         struct ipoib_dev_priv *priv = netdev_priv(dev);
 152
 153         if (new_mtu > IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN)
 154                 return -EINVAL;
 155
 156         priv->admin_mtu = new_mtu;
 157
 158         dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
 159
 160         return 0;
 161 }
 162
 163 static struct ipoib_path *__path_find(struct net_device *dev,
 164                                       union ib_gid *gid)
 165 {
 166         struct ipoib_dev_priv *priv = netdev_priv(dev);
 167         struct rb_node *n = priv->path_tree.rb_node;
 168         struct ipoib_path *path;
 169         int ret;
 170
 171         while (n) {
 172                 path = rb_entry(n, struct ipoib_path, rb_node);
 173
 174                 ret = memcmp(gid->raw, path->pathrec.dgid.raw,
 175                              sizeof (union ib_gid));
 176
 177                 if (ret < 0)
 178                         n = n->rb_left;
 179                 else if (ret > 0)
 180                         n = n->rb_right;
 181                 else
 182                         return path;
 183         }
 184
 185         return NULL;
 186 }
 187
 188 static int __path_add(struct net_device *dev, struct ipoib_path *path)
 189 {
 190         struct ipoib_dev_priv *priv = netdev_priv(dev);
 191         struct rb_node **n = &priv->path_tree.rb_node;
 192         struct rb_node *pn = NULL;
 193         struct ipoib_path *tpath;
 194         int ret;
 195
 196         while (*n) {
 197                 pn = *n;
 198                 tpath = rb_entry(pn, struct ipoib_path, rb_node);
 199
 200                 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
 201                              sizeof (union ib_gid));
 202                 if (ret < 0)
 203                         n = &pn->rb_left;
 204                 else if (ret > 0)
 205                         n = &pn->rb_right;
 206                 else
 207                         return -EEXIST;
 208         }
 209
 210         rb_link_node(&path->rb_node, pn, n);
 211         rb_insert_color(&path->rb_node, &priv->path_tree);
 212
 213         list_add_tail(&path->list, &priv->path_list);
 214
 215         return 0;
 216 }
 217
 218 static void path_free(struct net_device *dev, struct ipoib_path *path)
 219 {
 220         struct ipoib_dev_priv *priv = netdev_priv(dev);
 221         struct ipoib_neigh *neigh, *tn;
 222         struct sk_buff *skb;
 223         unsigned long flags;
 224
 225         while ((skb = __skb_dequeue(&path->queue)))
 226                 dev_kfree_skb_irq(skb);
 227
 228         spin_lock_irqsave(&priv->lock, flags);
 229
 230         list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
 231                 /*
 232                  * It's safe to call ipoib_put_ah() inside priv->lock
 233                  * here, because we know that path->ah will always
 234                  * hold one more reference, so ipoib_put_ah() will
 235                  * never do more than decrement the ref count.
 236                  */
 237                 if (neigh->ah)
 238                         ipoib_put_ah(neigh->ah);
 239                 *to_ipoib_neigh(neigh->neighbour) = NULL;
 240                 neigh->neighbour->ops->destructor = NULL;
 241                 kfree(neigh);
 242         }
 243
 244         spin_unlock_irqrestore(&priv->lock, flags);
 245
 246         if (path->ah)
 247                 ipoib_put_ah(path->ah);
 248
 249         kfree(path);
 250 }
 251
 252 void ipoib_flush_paths(struct net_device *dev)
 253 {
 254         struct ipoib_dev_priv *priv = netdev_priv(dev);
 255         struct ipoib_path *path, *tp;
 256         LIST_HEAD(remove_list);
 257         unsigned long flags;
 258
 259         spin_lock_irqsave(&priv->lock, flags);
 260
 261         list_splice(&priv->path_list, &remove_list);
 262         INIT_LIST_HEAD(&priv->path_list);
 263
 264         list_for_each_entry(path, &remove_list, list)
 265                 rb_erase(&path->rb_node, &priv->path_tree);
 266
 267         spin_unlock_irqrestore(&priv->lock, flags);
 268
 269         list_for_each_entry_safe(path, tp, &remove_list, list) {
 270                 if (path->query)
 271                         ib_sa_cancel_query(path->query_id, path->query);
 272                 wait_for_completion(&path->done);
 273                 path_free(dev, path);
 274         }
 275 }
 276
 277 static void path_rec_completion(int status,
 278                                 struct ib_sa_path_rec *pathrec,
 279                                 void *path_ptr)
 280 {
 281         struct ipoib_path *path = path_ptr;
 282         struct net_device *dev = path->dev;
 283         struct ipoib_dev_priv *priv = netdev_priv(dev);
 284         struct ipoib_ah *ah = NULL;
 285         struct ipoib_neigh *neigh;
 286         struct sk_buff_head skqueue;
 287         struct sk_buff *skb;
 288         unsigned long flags;
 289
 290         if (pathrec)
 291                 ipoib_dbg(priv, "PathRec LID 0x%04x for GID " IPOIB_GID_FMT "\n",
 292                           be16_to_cpu(pathrec->dlid), IPOIB_GID_ARG(pathrec->dgid));
 293         else
 294                 ipoib_dbg(priv, "PathRec status %d for GID " IPOIB_GID_FMT "\n",
 295                           status, IPOIB_GID_ARG(path->pathrec.dgid));
 296
 297         skb_queue_head_init(&skqueue);
 298
 299         if (!status) {
 300                 struct ib_ah_attr av = {
 301                         .dlid          = be16_to_cpu(pathrec->dlid),
 302                         .sl            = pathrec->sl,
 303                         .port_num      = priv->port
 304                 };
 305
 306                 if (ib_sa_rate_enum_to_int(pathrec->rate) > 0)
 307                         av.static_rate = (2 * priv->local_rate -
 308                                           ib_sa_rate_enum_to_int(pathrec->rate) - 1) /
 309                                 (priv->local_rate ? priv->local_rate : 1);
 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 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                         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 reply */
 605
 606                         if (be16_to_cpup((u16 *) skb->data) != ETH_P_ARP) {
 607                                 ipoib_warn(priv, "Unicast, no %s: type %04x, QPN %06x "
 608                                            IPOIB_GID_FMT "\n",
 609                                            skb->dst ? "neigh" : "dst",
 610                                            be16_to_cpup((u16 *) skb->data),
 611                                            be32_to_cpup((u32 *) phdr->hwaddr),
 612                                            IPOIB_GID_ARG(*(union ib_gid *) (phdr->hwaddr + 4)));
 613                                 dev_kfree_skb_any(skb);
 614                                 ++priv->stats.tx_dropped;
 615                                 goto out;
 616                         }
 617
 618                         unicast_arp_send(skb, dev, phdr);
 619                 }
 620         }
 621
 622 out:
 623         spin_unlock_irqrestore(&priv->tx_lock, flags);
 624
 625         return NETDEV_TX_OK;
 626 }
 627
 628 static struct net_device_stats *ipoib_get_stats(struct net_device *dev)
 629 {
 630         struct ipoib_dev_priv *priv = netdev_priv(dev);
 631
 632         return &priv->stats;
 633 }
 634
 635 static void ipoib_timeout(struct net_device *dev)
 636 {
 637         struct ipoib_dev_priv *priv = netdev_priv(dev);
 638
 639         ipoib_warn(priv, "transmit timeout: latency %ld\n",
 640                    jiffies - dev->trans_start);
 641         /* XXX reset QP, etc. */
 642 }
 643
 644 static int ipoib_hard_header(struct sk_buff *skb,
 645                              struct net_device *dev,
 646                              unsigned short type,
 647                              void *daddr, void *saddr, unsigned len)
 648 {
 649         struct ipoib_header *header;
 650
 651         header = (struct ipoib_header *) skb_push(skb, sizeof *header);
 652
 653         header->proto = htons(type);
 654         header->reserved = 0;
 655
 656         /*
 657          * If we don't have a neighbour structure, stuff the
 658          * destination address onto the front of the skb so we can
 659          * figure out where to send the packet later.
 660          */
 661         if (!skb->dst || !skb->dst->neighbour) {
 662                 struct ipoib_pseudoheader *phdr =
 663                         (struct ipoib_pseudoheader *) skb_push(skb, sizeof *phdr);
 664                 memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
 665         }
 666
 667         return 0;
 668 }
 669
 670 static void ipoib_set_mcast_list(struct net_device *dev)
 671 {
 672         struct ipoib_dev_priv *priv = netdev_priv(dev);
 673
 674         schedule_work(&priv->restart_task);
 675 }
 676
 677 static void ipoib_neigh_destructor(struct neighbour *n)
 678 {
 679         struct ipoib_neigh *neigh;
 680         struct ipoib_dev_priv *priv = netdev_priv(n->dev);
 681         unsigned long flags;
 682         struct ipoib_ah *ah = NULL;
 683
 684         ipoib_dbg(priv,
 685                   "neigh_destructor for %06x " IPOIB_GID_FMT "\n",
 686                   be32_to_cpup((__be32 *) n->ha),
 687                   IPOIB_GID_ARG(*((union ib_gid *) (n->ha + 4))));
 688
 689         spin_lock_irqsave(&priv->lock, flags);
 690
 691         neigh = *to_ipoib_neigh(n);
 692         if (neigh) {
 693                 if (neigh->ah)
 694                         ah = neigh->ah;
 695                 list_del(&neigh->list);
 696                 *to_ipoib_neigh(n) = NULL;
 697                 kfree(neigh);
 698         }
 699
 700         spin_unlock_irqrestore(&priv->lock, flags);
 701
 702         if (ah)
 703                 ipoib_put_ah(ah);
 704 }
 705
 706 static int ipoib_neigh_setup(struct neighbour *neigh)
 707 {
 708         /*
 709          * Is this kosher?  I can't find anybody in the kernel that
 710          * sets neigh->destructor, so we should be able to set it here
 711          * without trouble.
 712          */
 713         neigh->ops->destructor = ipoib_neigh_destructor;
 714
 715         return 0;
 716 }
 717
 718 static int ipoib_neigh_setup_dev(struct net_device *dev, struct neigh_parms *parms)
 719 {
 720         parms->neigh_setup = ipoib_neigh_setup;
 721
 722         return 0;
 723 }
 724
 725 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
 726 {
 727         struct ipoib_dev_priv *priv = netdev_priv(dev);
 728
 729         /* Allocate RX/TX "rings" to hold queued skbs */
 730
 731         priv->rx_ring = kmalloc(IPOIB_RX_RING_SIZE * sizeof (struct ipoib_buf),
 732                                 GFP_KERNEL);
 733         if (!priv->rx_ring) {
 734                 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
 735                        ca->name, IPOIB_RX_RING_SIZE);
 736                 goto out;
 737         }
 738         memset(priv->rx_ring, 0,
 739                IPOIB_RX_RING_SIZE * sizeof (struct ipoib_buf));
 740
 741         priv->tx_ring = kmalloc(IPOIB_TX_RING_SIZE * sizeof (struct ipoib_buf),
 742                                 GFP_KERNEL);
 743         if (!priv->tx_ring) {
 744                 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
 745                        ca->name, IPOIB_TX_RING_SIZE);
 746                 goto out_rx_ring_cleanup;
 747         }
 748         memset(priv->tx_ring, 0,
 749                IPOIB_TX_RING_SIZE * sizeof (struct ipoib_buf));
 750
 751         /* priv->tx_head & tx_tail are already 0 */
 752
 753         if (ipoib_ib_dev_init(dev, ca, port))
 754                 goto out_tx_ring_cleanup;
 755
 756         return 0;
 757
 758 out_tx_ring_cleanup:
 759         kfree(priv->tx_ring);
 760
 761 out_rx_ring_cleanup:
 762         kfree(priv->rx_ring);
 763
 764 out:
 765         return -ENOMEM;
 766 }
 767
 768 void ipoib_dev_cleanup(struct net_device *dev)
 769 {
 770         struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
 771
 772         ipoib_delete_debug_file(dev);
 773
 774         /* Delete any child interfaces first */
 775         list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
 776                 unregister_netdev(cpriv->dev);
 777                 ipoib_dev_cleanup(cpriv->dev);
 778                 free_netdev(cpriv->dev);
 779         }
 780
 781         ipoib_ib_dev_cleanup(dev);
 782
 783         if (priv->rx_ring) {
 784                 kfree(priv->rx_ring);
 785                 priv->rx_ring = NULL;
 786         }
 787
 788         if (priv->tx_ring) {
 789                 kfree(priv->tx_ring);
 790                 priv->tx_ring = NULL;
 791         }
 792 }
 793
 794 static void ipoib_setup(struct net_device *dev)
 795 {
 796         struct ipoib_dev_priv *priv = netdev_priv(dev);
 797
 798         dev->open                = ipoib_open;
 799         dev->stop                = ipoib_stop;
 800         dev->change_mtu          = ipoib_change_mtu;
 801         dev->hard_start_xmit     = ipoib_start_xmit;
 802         dev->get_stats           = ipoib_get_stats;
 803         dev->tx_timeout          = ipoib_timeout;
 804         dev->hard_header         = ipoib_hard_header;
 805         dev->set_multicast_list  = ipoib_set_mcast_list;
 806         dev->neigh_setup         = ipoib_neigh_setup_dev;
 807
 808         dev->watchdog_timeo      = HZ;
 809
 810         dev->rebuild_header      = NULL;
 811         dev->set_mac_address     = NULL;
 812         dev->header_cache_update = NULL;
 813
 814         dev->flags              |= IFF_BROADCAST | IFF_MULTICAST;
 815
 816         /*
 817          * We add in INFINIBAND_ALEN to allow for the destination
 818          * address "pseudoheader" for skbs without neighbour struct.
 819          */
 820         dev->hard_header_len     = IPOIB_ENCAP_LEN + INFINIBAND_ALEN;
 821         dev->addr_len            = INFINIBAND_ALEN;
 822         dev->type                = ARPHRD_INFINIBAND;
 823         dev->tx_queue_len        = IPOIB_TX_RING_SIZE * 2;
 824         dev->features            = NETIF_F_VLAN_CHALLENGED | NETIF_F_LLTX;
 825
 826         /* MTU will be reset when mcast join happens */
 827         dev->mtu                 = IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN;
 828         priv->mcast_mtu          = priv->admin_mtu = dev->mtu;
 829
 830         memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
 831
 832         netif_carrier_off(dev);
 833
 834         SET_MODULE_OWNER(dev);
 835
 836         priv->dev = dev;
 837
 838         spin_lock_init(&priv->lock);
 839         spin_lock_init(&priv->tx_lock);
 840
 841         init_MUTEX(&priv->mcast_mutex);
 842         init_MUTEX(&priv->vlan_mutex);
 843
 844         INIT_LIST_HEAD(&priv->path_list);
 845         INIT_LIST_HEAD(&priv->child_intfs);
 846         INIT_LIST_HEAD(&priv->dead_ahs);
 847         INIT_LIST_HEAD(&priv->multicast_list);
 848
 849         INIT_WORK(&priv->pkey_task,    ipoib_pkey_poll,          priv->dev);
 850         INIT_WORK(&priv->mcast_task,   ipoib_mcast_join_task,    priv->dev);
 851         INIT_WORK(&priv->flush_task,   ipoib_ib_dev_flush,       priv->dev);
 852         INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task, priv->dev);
 853         INIT_WORK(&priv->ah_reap_task, ipoib_reap_ah,            priv->dev);
 854 }
 855
 856 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
 857 {
 858         struct net_device *dev;
 859
 860         dev = alloc_netdev((int) sizeof (struct ipoib_dev_priv), name,
 861                            ipoib_setup);
 862         if (!dev)
 863                 return NULL;
 864
 865         return netdev_priv(dev);
 866 }
 867
 868 static ssize_t show_pkey(struct class_device *cdev, char *buf)
 869 {
 870         struct ipoib_dev_priv *priv =
 871                 netdev_priv(container_of(cdev, struct net_device, class_dev));
 872
 873         return sprintf(buf, "0x%04x\n", priv->pkey);
 874 }
 875 static CLASS_DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
 876
 877 static ssize_t create_child(struct class_device *cdev,
 878                             const char *buf, size_t count)
 879 {
 880         int pkey;
 881         int ret;
 882
 883         if (sscanf(buf, "%i", &pkey) != 1)
 884                 return -EINVAL;
 885
 886         if (pkey < 0 || pkey > 0xffff)
 887                 return -EINVAL;
 888
 889         ret = ipoib_vlan_add(container_of(cdev, struct net_device, class_dev),
 890                              pkey);
 891
 892         return ret ? ret : count;
 893 }
 894 static CLASS_DEVICE_ATTR(create_child, S_IWUGO, NULL, create_child);
 895
 896 static ssize_t delete_child(struct class_device *cdev,
 897                             const char *buf, size_t count)
 898 {
 899         int pkey;
 900         int ret;
 901
 902         if (sscanf(buf, "%i", &pkey) != 1)
 903                 return -EINVAL;
 904
 905         if (pkey < 0 || pkey > 0xffff)
 906                 return -EINVAL;
 907
 908         ret = ipoib_vlan_delete(container_of(cdev, struct net_device, class_dev),
 909                                 pkey);
 910
 911         return ret ? ret : count;
 912
 913 }
 914 static CLASS_DEVICE_ATTR(delete_child, S_IWUGO, NULL, delete_child);
 915
 916 int ipoib_add_pkey_attr(struct net_device *dev)
 917 {
 918         return class_device_create_file(&dev->class_dev,
 919                                         &class_device_attr_pkey);
 920 }
 921
 922 static struct net_device *ipoib_add_port(const char *format,
 923                                          struct ib_device *hca, u8 port)
 924 {
 925         struct ipoib_dev_priv *priv;
 926         int result = -ENOMEM;
 927
 928         priv = ipoib_intf_alloc(format);
 929         if (!priv)
 930                 goto alloc_mem_failed;
 931
 932         SET_NETDEV_DEV(priv->dev, hca->dma_device);
 933
 934         result = ib_query_pkey(hca, port, 0, &priv->pkey);
 935         if (result) {
 936                 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
 937                        hca->name, port, result);
 938                 goto alloc_mem_failed;
 939         }
 940
 941         priv->dev->broadcast[8] = priv->pkey >> 8;
 942         priv->dev->broadcast[9] = priv->pkey & 0xff;
 943
 944         result = ib_query_gid(hca, port, 0, &priv->local_gid);
 945         if (result) {
 946                 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
 947                        hca->name, port, result);
 948                 goto alloc_mem_failed;
 949         } else
 950                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
 951
 952
 953         result = ipoib_dev_init(priv->dev, hca, port);
 954         if (result < 0) {
 955                 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
 956                        hca->name, port, result);
 957                 goto device_init_failed;
 958         }
 959
 960         INIT_IB_EVENT_HANDLER(&priv->event_handler,
 961                               priv->ca, ipoib_event);
 962         result = ib_register_event_handler(&priv->event_handler);
 963         if (result < 0) {
 964                 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
 965                        "port %d (ret = %d)\n",
 966                        hca->name, port, result);
 967                 goto event_failed;
 968         }
 969
 970         result = register_netdev(priv->dev);
 971         if (result) {
 972                 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
 973                        hca->name, port, result);
 974                 goto register_failed;
 975         }
 976
 977         if (ipoib_create_debug_file(priv->dev))
 978                 goto debug_failed;
 979
 980         if (ipoib_add_pkey_attr(priv->dev))
 981                 goto sysfs_failed;
 982         if (class_device_create_file(&priv->dev->class_dev,
 983                                      &class_device_attr_create_child))
 984                 goto sysfs_failed;
 985         if (class_device_create_file(&priv->dev->class_dev,
 986                                      &class_device_attr_delete_child))
 987                 goto sysfs_failed;
 988
 989         return priv->dev;
 990
 991 sysfs_failed:
 992         ipoib_delete_debug_file(priv->dev);
 993
 994 debug_failed:
 995         unregister_netdev(priv->dev);
 996
 997 register_failed:
 998         ib_unregister_event_handler(&priv->event_handler);
 999
1000 event_failed:
1001         ipoib_dev_cleanup(priv->dev);
1002
1003 device_init_failed:
1004         free_netdev(priv->dev);
1005
1006 alloc_mem_failed:
1007         return ERR_PTR(result);
1008 }
1009
1010 static void ipoib_add_one(struct ib_device *device)
1011 {
1012         struct list_head *dev_list;
1013         struct net_device *dev;
1014         struct ipoib_dev_priv *priv;
1015         int s, e, p;
1016
1017         dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1018         if (!dev_list)
1019                 return;
1020
1021         INIT_LIST_HEAD(dev_list);
1022
1023         if (device->node_type == IB_NODE_SWITCH) {
1024                 s = 0;
1025                 e = 0;
1026         } else {
1027                 s = 1;
1028                 e = device->phys_port_cnt;
1029         }
1030
1031         for (p = s; p <= e; ++p) {
1032                 dev = ipoib_add_port("ib%d", device, p);
1033                 if (!IS_ERR(dev)) {
1034                         priv = netdev_priv(dev);
1035                         list_add_tail(&priv->list, dev_list);
1036                 }
1037         }
1038
1039         ib_set_client_data(device, &ipoib_client, dev_list);
1040 }
1041
1042 static void ipoib_remove_one(struct ib_device *device)
1043 {
1044         struct ipoib_dev_priv *priv, *tmp;
1045         struct list_head *dev_list;
1046
1047         dev_list = ib_get_client_data(device, &ipoib_client);
1048
1049         list_for_each_entry_safe(priv, tmp, dev_list, list) {
1050                 ib_unregister_event_handler(&priv->event_handler);
1051
1052                 unregister_netdev(priv->dev);
1053                 ipoib_dev_cleanup(priv->dev);
1054                 free_netdev(priv->dev);
1055         }
1056 }
1057
1058 static int __init ipoib_init_module(void)
1059 {
1060         int ret;
1061
1062         ret = ipoib_register_debugfs();
1063         if (ret)
1064                 return ret;
1065
1066         /*
1067          * We create our own workqueue mainly because we want to be
1068          * able to flush it when devices are being removed.  We can't
1069          * use schedule_work()/flush_scheduled_work() because both
1070          * unregister_netdev() and linkwatch_event take the rtnl lock,
1071          * so flush_scheduled_work() can deadlock during device
1072          * removal.
1073          */
1074         ipoib_workqueue = create_singlethread_workqueue("ipoib");
1075         if (!ipoib_workqueue) {
1076                 ret = -ENOMEM;
1077                 goto err_fs;
1078         }
1079
1080         ret = ib_register_client(&ipoib_client);
1081         if (ret)
1082                 goto err_wq;
1083
1084         return 0;
1085
1086 err_fs:
1087         ipoib_unregister_debugfs();
1088
1089 err_wq:
1090         destroy_workqueue(ipoib_workqueue);
1091
1092         return ret;
1093 }
1094
1095 static void __exit ipoib_cleanup_module(void)
1096 {
1097         ipoib_unregister_debugfs();
1098         ib_unregister_client(&ipoib_client);
1099         destroy_workqueue(ipoib_workqueue);
1100 }
1101
1102 module_init(ipoib_init_module);
1103 module_exit(ipoib_cleanup_module);