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                 int path_rate = ib_sa_rate_enum_to_int(pathrec->rate);
 306
 307                 if (path_rate > 0 && priv->local_rate > path_rate)
 308                         av.static_rate = (priv->local_rate - 1) / path_rate;
 309
 310                 ipoib_dbg(priv, "static_rate %d for local port %dX, path %dX\n",
 311                           av.static_rate, priv->local_rate,
 312                           ib_sa_rate_enum_to_int(pathrec->rate));
 313
 314                 ah = ipoib_create_ah(dev, priv->pd, &av);
 315         }
 316
 317         spin_lock_irqsave(&priv->lock, flags);
 318
 319         path->ah = ah;
 320
 321         if (ah) {
 322                 path->pathrec = *pathrec;
 323
 324                 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
 325                           ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
 326
 327                 while ((skb = __skb_dequeue(&path->queue)))
 328                         __skb_queue_tail(&skqueue, skb);
 329
 330                 list_for_each_entry(neigh, &path->neigh_list, list) {
 331                         kref_get(&path->ah->ref);
 332                         neigh->ah = path->ah;
 333
 334                         while ((skb = __skb_dequeue(&neigh->queue)))
 335                                 __skb_queue_tail(&skqueue, skb);
 336                 }
 337         } else
 338                 path->query = NULL;
 339
 340         complete(&path->done);
 341
 342         spin_unlock_irqrestore(&priv->lock, flags);
 343
 344         while ((skb = __skb_dequeue(&skqueue))) {
 345                 skb->dev = dev;
 346                 if (dev_queue_xmit(skb))
 347                         ipoib_warn(priv, "dev_queue_xmit failed "
 348                                    "to requeue packet\n");
 349         }
 350 }
 351
 352 static struct ipoib_path *path_rec_create(struct net_device *dev,
 353                                           union ib_gid *gid)
 354 {
 355         struct ipoib_dev_priv *priv = netdev_priv(dev);
 356         struct ipoib_path *path;
 357
 358         path = kmalloc(sizeof *path, GFP_ATOMIC);
 359         if (!path)
 360                 return NULL;
 361
 362         path->dev          = dev;
 363         path->pathrec.dlid = 0;
 364         path->ah           = NULL;
 365
 366         skb_queue_head_init(&path->queue);
 367
 368         INIT_LIST_HEAD(&path->neigh_list);
 369         path->query = NULL;
 370         init_completion(&path->done);
 371
 372         memcpy(path->pathrec.dgid.raw, gid->raw, sizeof (union ib_gid));
 373         path->pathrec.sgid      = priv->local_gid;
 374         path->pathrec.pkey      = cpu_to_be16(priv->pkey);
 375         path->pathrec.numb_path = 1;
 376
 377         return path;
 378 }
 379
 380 static int path_rec_start(struct net_device *dev,
 381                           struct ipoib_path *path)
 382 {
 383         struct ipoib_dev_priv *priv = netdev_priv(dev);
 384
 385         ipoib_dbg(priv, "Start path record lookup for " IPOIB_GID_FMT "\n",
 386                   IPOIB_GID_ARG(path->pathrec.dgid));
 387
 388         path->query_id =
 389                 ib_sa_path_rec_get(priv->ca, priv->port,
 390                                    &path->pathrec,
 391                                    IB_SA_PATH_REC_DGID          |
 392                                    IB_SA_PATH_REC_SGID          |
 393                                    IB_SA_PATH_REC_NUMB_PATH     |
 394                                    IB_SA_PATH_REC_PKEY,
 395                                    1000, GFP_ATOMIC,
 396                                    path_rec_completion,
 397                                    path, &path->query);
 398         if (path->query_id < 0) {
 399                 ipoib_warn(priv, "ib_sa_path_rec_get failed\n");
 400                 path->query = NULL;
 401                 return path->query_id;
 402         }
 403
 404         return 0;
 405 }
 406
 407 static void neigh_add_path(struct sk_buff *skb, struct net_device *dev)
 408 {
 409         struct ipoib_dev_priv *priv = netdev_priv(dev);
 410         struct ipoib_path *path;
 411         struct ipoib_neigh *neigh;
 412
 413         neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
 414         if (!neigh) {
 415                 ++priv->stats.tx_dropped;
 416                 dev_kfree_skb_any(skb);
 417                 return;
 418         }
 419
 420         skb_queue_head_init(&neigh->queue);
 421         neigh->neighbour = skb->dst->neighbour;
 422         *to_ipoib_neigh(skb->dst->neighbour) = neigh;
 423
 424         /*
 425          * We can only be called from ipoib_start_xmit, so we're
 426          * inside tx_lock -- no need to save/restore flags.
 427          */
 428         spin_lock(&priv->lock);
 429
 430         path = __path_find(dev, (union ib_gid *) (skb->dst->neighbour->ha + 4));
 431         if (!path) {
 432                 path = path_rec_create(dev,
 433                                        (union ib_gid *) (skb->dst->neighbour->ha + 4));
 434                 if (!path)
 435                         goto err;
 436
 437                 __path_add(dev, path);
 438         }
 439
 440         list_add_tail(&neigh->list, &path->neigh_list);
 441
 442         if (path->pathrec.dlid) {
 443                 kref_get(&path->ah->ref);
 444                 neigh->ah = path->ah;
 445
 446                 ipoib_send(dev, skb, path->ah,
 447                            be32_to_cpup((__be32 *) skb->dst->neighbour->ha));
 448         } else {
 449                 neigh->ah  = NULL;
 450                 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 451                         __skb_queue_tail(&neigh->queue, skb);
 452                 } else {
 453                         ++priv->stats.tx_dropped;
 454                         dev_kfree_skb_any(skb);
 455                 }
 456
 457                 if (!path->query && path_rec_start(dev, path))
 458                         goto err;
 459         }
 460
 461         spin_unlock(&priv->lock);
 462         return;
 463
 464 err:
 465         *to_ipoib_neigh(skb->dst->neighbour) = NULL;
 466         list_del(&neigh->list);
 467         neigh->neighbour->ops->destructor = NULL;
 468         kfree(neigh);
 469
 470         ++priv->stats.tx_dropped;
 471         dev_kfree_skb_any(skb);
 472
 473         spin_unlock(&priv->lock);
 474 }
 475
 476 static void path_lookup(struct sk_buff *skb, struct net_device *dev)
 477 {
 478         struct ipoib_dev_priv *priv = netdev_priv(skb->dev);
 479
 480         /* Look up path record for unicasts */
 481         if (skb->dst->neighbour->ha[4] != 0xff) {
 482                 neigh_add_path(skb, dev);
 483                 return;
 484         }
 485
 486         /* Add in the P_Key for multicasts */
 487         skb->dst->neighbour->ha[8] = (priv->pkey >> 8) & 0xff;
 488         skb->dst->neighbour->ha[9] = priv->pkey & 0xff;
 489         ipoib_mcast_send(dev, (union ib_gid *) (skb->dst->neighbour->ha + 4), skb);
 490 }
 491
 492 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
 493                              struct ipoib_pseudoheader *phdr)
 494 {
 495         struct ipoib_dev_priv *priv = netdev_priv(dev);
 496         struct ipoib_path *path;
 497
 498         /*
 499          * We can only be called from ipoib_start_xmit, so we're
 500          * inside tx_lock -- no need to save/restore flags.
 501          */
 502         spin_lock(&priv->lock);
 503
 504         path = __path_find(dev, (union ib_gid *) (phdr->hwaddr + 4));
 505         if (!path) {
 506                 path = path_rec_create(dev,
 507                                        (union ib_gid *) (phdr->hwaddr + 4));
 508                 if (path) {
 509                         /* put pseudoheader back on for next time */
 510                         skb_push(skb, sizeof *phdr);
 511                         __skb_queue_tail(&path->queue, skb);
 512
 513                         if (path_rec_start(dev, path)) {
 514                                 spin_unlock(&priv->lock);
 515                                 path_free(dev, path);
 516                                 return;
 517                         } else
 518                                 __path_add(dev, path);
 519                 } else {
 520                         ++priv->stats.tx_dropped;
 521                         dev_kfree_skb_any(skb);
 522                 }
 523
 524                 spin_unlock(&priv->lock);
 525                 return;
 526         }
 527
 528         if (path->pathrec.dlid) {
 529                 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
 530                           be16_to_cpu(path->pathrec.dlid));
 531
 532                 ipoib_send(dev, skb, path->ah,
 533                            be32_to_cpup((__be32 *) phdr->hwaddr));
 534         } else if ((path->query || !path_rec_start(dev, path)) &&
 535                    skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 536                 /* put pseudoheader back on for next time */
 537                 skb_push(skb, sizeof *phdr);
 538                 __skb_queue_tail(&path->queue, skb);
 539         } else {
 540                 ++priv->stats.tx_dropped;
 541                 dev_kfree_skb_any(skb);
 542         }
 543
 544         spin_unlock(&priv->lock);
 545 }
 546
 547 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
 548 {
 549         struct ipoib_dev_priv *priv = netdev_priv(dev);
 550         struct ipoib_neigh *neigh;
 551         unsigned long flags;
 552
 553         local_irq_save(flags);
 554         if (!spin_trylock(&priv->tx_lock)) {
 555                 local_irq_restore(flags);
 556                 return NETDEV_TX_LOCKED;
 557         }
 558
 559         /*
 560          * Check if our queue is stopped.  Since we have the LLTX bit
 561          * set, we can't rely on netif_stop_queue() preventing our
 562          * xmit function from being called with a full queue.
 563          */
 564         if (unlikely(netif_queue_stopped(dev))) {
 565                 spin_unlock_irqrestore(&priv->tx_lock, flags);
 566                 return NETDEV_TX_BUSY;
 567         }
 568
 569         if (skb->dst && skb->dst->neighbour) {
 570                 if (unlikely(!*to_ipoib_neigh(skb->dst->neighbour))) {
 571                         path_lookup(skb, dev);
 572                         goto out;
 573                 }
 574
 575                 neigh = *to_ipoib_neigh(skb->dst->neighbour);
 576
 577                 if (likely(neigh->ah)) {
 578                         ipoib_send(dev, skb, neigh->ah,
 579                                    be32_to_cpup((__be32 *) skb->dst->neighbour->ha));
 580                         goto out;
 581                 }
 582
 583                 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 584                         spin_lock(&priv->lock);
 585                         __skb_queue_tail(&neigh->queue, skb);
 586                         spin_unlock(&priv->lock);
 587                 } else {
 588                         ++priv->stats.tx_dropped;
 589                         dev_kfree_skb_any(skb);
 590                 }
 591         } else {
 592                 struct ipoib_pseudoheader *phdr =
 593                         (struct ipoib_pseudoheader *) skb->data;
 594                 skb_pull(skb, sizeof *phdr);
 595
 596                 if (phdr->hwaddr[4] == 0xff) {
 597                         /* Add in the P_Key for multicast*/
 598                         phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
 599                         phdr->hwaddr[9] = priv->pkey & 0xff;
 600
 601                         ipoib_mcast_send(dev, (union ib_gid *) (phdr->hwaddr + 4), skb);
 602                 } else {
 603                         /* unicast GID -- should be ARP reply */
 604
 605                         if (be16_to_cpup((u16 *) skb->data) != ETH_P_ARP) {
 606                                 ipoib_warn(priv, "Unicast, no %s: type %04x, QPN %06x "
 607                                            IPOIB_GID_FMT "\n",
 608                                            skb->dst ? "neigh" : "dst",
 609                                            be16_to_cpup((u16 *) skb->data),
 610                                            be32_to_cpup((u32 *) phdr->hwaddr),
 611                                            IPOIB_GID_ARG(*(union ib_gid *) (phdr->hwaddr + 4)));
 612                                 dev_kfree_skb_any(skb);
 613                                 ++priv->stats.tx_dropped;
 614                                 goto out;
 615                         }
 616
 617                         unicast_arp_send(skb, dev, phdr);
 618                 }
 619         }
 620
 621 out:
 622         spin_unlock_irqrestore(&priv->tx_lock, flags);
 623
 624         return NETDEV_TX_OK;
 625 }
 626
 627 static struct net_device_stats *ipoib_get_stats(struct net_device *dev)
 628 {
 629         struct ipoib_dev_priv *priv = netdev_priv(dev);
 630
 631         return &priv->stats;
 632 }
 633
 634 static void ipoib_timeout(struct net_device *dev)
 635 {
 636         struct ipoib_dev_priv *priv = netdev_priv(dev);
 637
 638         ipoib_warn(priv, "transmit timeout: latency %ld\n",
 639                    jiffies - dev->trans_start);
 640         /* XXX reset QP, etc. */
 641 }
 642
 643 static int ipoib_hard_header(struct sk_buff *skb,
 644                              struct net_device *dev,
 645                              unsigned short type,
 646                              void *daddr, void *saddr, unsigned len)
 647 {
 648         struct ipoib_header *header;
 649
 650         header = (struct ipoib_header *) skb_push(skb, sizeof *header);
 651
 652         header->proto = htons(type);
 653         header->reserved = 0;
 654
 655         /*
 656          * If we don't have a neighbour structure, stuff the
 657          * destination address onto the front of the skb so we can
 658          * figure out where to send the packet later.
 659          */
 660         if (!skb->dst || !skb->dst->neighbour) {
 661                 struct ipoib_pseudoheader *phdr =
 662                         (struct ipoib_pseudoheader *) skb_push(skb, sizeof *phdr);
 663                 memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
 664         }
 665
 666         return 0;
 667 }
 668
 669 static void ipoib_set_mcast_list(struct net_device *dev)
 670 {
 671         struct ipoib_dev_priv *priv = netdev_priv(dev);
 672
 673         schedule_work(&priv->restart_task);
 674 }
 675
 676 static void ipoib_neigh_destructor(struct neighbour *n)
 677 {
 678         struct ipoib_neigh *neigh;
 679         struct ipoib_dev_priv *priv = netdev_priv(n->dev);
 680         unsigned long flags;
 681         struct ipoib_ah *ah = NULL;
 682
 683         ipoib_dbg(priv,
 684                   "neigh_destructor for %06x " IPOIB_GID_FMT "\n",
 685                   be32_to_cpup((__be32 *) n->ha),
 686                   IPOIB_GID_ARG(*((union ib_gid *) (n->ha + 4))));
 687
 688         spin_lock_irqsave(&priv->lock, flags);
 689
 690         neigh = *to_ipoib_neigh(n);
 691         if (neigh) {
 692                 if (neigh->ah)
 693                         ah = neigh->ah;
 694                 list_del(&neigh->list);
 695                 *to_ipoib_neigh(n) = NULL;
 696                 kfree(neigh);
 697         }
 698
 699         spin_unlock_irqrestore(&priv->lock, flags);
 700
 701         if (ah)
 702                 ipoib_put_ah(ah);
 703 }
 704
 705 static int ipoib_neigh_setup(struct neighbour *neigh)
 706 {
 707         /*
 708          * Is this kosher?  I can't find anybody in the kernel that
 709          * sets neigh->destructor, so we should be able to set it here
 710          * without trouble.
 711          */
 712         neigh->ops->destructor = ipoib_neigh_destructor;
 713
 714         return 0;
 715 }
 716
 717 static int ipoib_neigh_setup_dev(struct net_device *dev, struct neigh_parms *parms)
 718 {
 719         parms->neigh_setup = ipoib_neigh_setup;
 720
 721         return 0;
 722 }
 723
 724 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
 725 {
 726         struct ipoib_dev_priv *priv = netdev_priv(dev);
 727
 728         /* Allocate RX/TX "rings" to hold queued skbs */
 729
 730         priv->rx_ring = kmalloc(IPOIB_RX_RING_SIZE * sizeof (struct ipoib_buf),
 731                                 GFP_KERNEL);
 732         if (!priv->rx_ring) {
 733                 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
 734                        ca->name, IPOIB_RX_RING_SIZE);
 735                 goto out;
 736         }
 737         memset(priv->rx_ring, 0,
 738                IPOIB_RX_RING_SIZE * sizeof (struct ipoib_buf));
 739
 740         priv->tx_ring = kmalloc(IPOIB_TX_RING_SIZE * sizeof (struct ipoib_buf),
 741                                 GFP_KERNEL);
 742         if (!priv->tx_ring) {
 743                 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
 744                        ca->name, IPOIB_TX_RING_SIZE);
 745                 goto out_rx_ring_cleanup;
 746         }
 747         memset(priv->tx_ring, 0,
 748                IPOIB_TX_RING_SIZE * sizeof (struct ipoib_buf));
 749
 750         /* priv->tx_head & tx_tail are already 0 */
 751
 752         if (ipoib_ib_dev_init(dev, ca, port))
 753                 goto out_tx_ring_cleanup;
 754
 755         return 0;
 756
 757 out_tx_ring_cleanup:
 758         kfree(priv->tx_ring);
 759
 760 out_rx_ring_cleanup:
 761         kfree(priv->rx_ring);
 762
 763 out:
 764         return -ENOMEM;
 765 }
 766
 767 void ipoib_dev_cleanup(struct net_device *dev)
 768 {
 769         struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
 770
 771         ipoib_delete_debug_file(dev);
 772
 773         /* Delete any child interfaces first */
 774         list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
 775                 unregister_netdev(cpriv->dev);
 776                 ipoib_dev_cleanup(cpriv->dev);
 777                 free_netdev(cpriv->dev);
 778         }
 779
 780         ipoib_ib_dev_cleanup(dev);
 781
 782         if (priv->rx_ring) {
 783                 kfree(priv->rx_ring);
 784                 priv->rx_ring = NULL;
 785         }
 786
 787         if (priv->tx_ring) {
 788                 kfree(priv->tx_ring);
 789                 priv->tx_ring = NULL;
 790         }
 791 }
 792
 793 static void ipoib_setup(struct net_device *dev)
 794 {
 795         struct ipoib_dev_priv *priv = netdev_priv(dev);
 796
 797         dev->open                = ipoib_open;
 798         dev->stop                = ipoib_stop;
 799         dev->change_mtu          = ipoib_change_mtu;
 800         dev->hard_start_xmit     = ipoib_start_xmit;
 801         dev->get_stats           = ipoib_get_stats;
 802         dev->tx_timeout          = ipoib_timeout;
 803         dev->hard_header         = ipoib_hard_header;
 804         dev->set_multicast_list  = ipoib_set_mcast_list;
 805         dev->neigh_setup         = ipoib_neigh_setup_dev;
 806
 807         dev->watchdog_timeo      = HZ;
 808
 809         dev->rebuild_header      = NULL;
 810         dev->set_mac_address     = NULL;
 811         dev->header_cache_update = NULL;
 812
 813         dev->flags              |= IFF_BROADCAST | IFF_MULTICAST;
 814
 815         /*
 816          * We add in INFINIBAND_ALEN to allow for the destination
 817          * address "pseudoheader" for skbs without neighbour struct.
 818          */
 819         dev->hard_header_len     = IPOIB_ENCAP_LEN + INFINIBAND_ALEN;
 820         dev->addr_len            = INFINIBAND_ALEN;
 821         dev->type                = ARPHRD_INFINIBAND;
 822         dev->tx_queue_len        = IPOIB_TX_RING_SIZE * 2;
 823         dev->features            = NETIF_F_VLAN_CHALLENGED | NETIF_F_LLTX;
 824
 825         /* MTU will be reset when mcast join happens */
 826         dev->mtu                 = IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN;
 827         priv->mcast_mtu          = priv->admin_mtu = dev->mtu;
 828
 829         memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
 830
 831         netif_carrier_off(dev);
 832
 833         SET_MODULE_OWNER(dev);
 834
 835         priv->dev = dev;
 836
 837         spin_lock_init(&priv->lock);
 838         spin_lock_init(&priv->tx_lock);
 839
 840         init_MUTEX(&priv->mcast_mutex);
 841         init_MUTEX(&priv->vlan_mutex);
 842
 843         INIT_LIST_HEAD(&priv->path_list);
 844         INIT_LIST_HEAD(&priv->child_intfs);
 845         INIT_LIST_HEAD(&priv->dead_ahs);
 846         INIT_LIST_HEAD(&priv->multicast_list);
 847
 848         INIT_WORK(&priv->pkey_task,    ipoib_pkey_poll,          priv->dev);
 849         INIT_WORK(&priv->mcast_task,   ipoib_mcast_join_task,    priv->dev);
 850         INIT_WORK(&priv->flush_task,   ipoib_ib_dev_flush,       priv->dev);
 851         INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task, priv->dev);
 852         INIT_WORK(&priv->ah_reap_task, ipoib_reap_ah,            priv->dev);
 853 }
 854
 855 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
 856 {
 857         struct net_device *dev;
 858
 859         dev = alloc_netdev((int) sizeof (struct ipoib_dev_priv), name,
 860                            ipoib_setup);
 861         if (!dev)
 862                 return NULL;
 863
 864         return netdev_priv(dev);
 865 }
 866
 867 static ssize_t show_pkey(struct class_device *cdev, char *buf)
 868 {
 869         struct ipoib_dev_priv *priv =
 870                 netdev_priv(container_of(cdev, struct net_device, class_dev));
 871
 872         return sprintf(buf, "0x%04x\n", priv->pkey);
 873 }
 874 static CLASS_DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
 875
 876 static ssize_t create_child(struct class_device *cdev,
 877                             const char *buf, size_t count)
 878 {
 879         int pkey;
 880         int ret;
 881
 882         if (sscanf(buf, "%i", &pkey) != 1)
 883                 return -EINVAL;
 884
 885         if (pkey < 0 || pkey > 0xffff)
 886                 return -EINVAL;
 887
 888         ret = ipoib_vlan_add(container_of(cdev, struct net_device, class_dev),
 889                              pkey);
 890
 891         return ret ? ret : count;
 892 }
 893 static CLASS_DEVICE_ATTR(create_child, S_IWUGO, NULL, create_child);
 894
 895 static ssize_t delete_child(struct class_device *cdev,
 896                             const char *buf, size_t count)
 897 {
 898         int pkey;
 899         int ret;
 900
 901         if (sscanf(buf, "%i", &pkey) != 1)
 902                 return -EINVAL;
 903
 904         if (pkey < 0 || pkey > 0xffff)
 905                 return -EINVAL;
 906
 907         ret = ipoib_vlan_delete(container_of(cdev, struct net_device, class_dev),
 908                                 pkey);
 909
 910         return ret ? ret : count;
 911
 912 }
 913 static CLASS_DEVICE_ATTR(delete_child, S_IWUGO, NULL, delete_child);
 914
 915 int ipoib_add_pkey_attr(struct net_device *dev)
 916 {
 917         return class_device_create_file(&dev->class_dev,
 918                                         &class_device_attr_pkey);
 919 }
 920
 921 static struct net_device *ipoib_add_port(const char *format,
 922                                          struct ib_device *hca, u8 port)
 923 {
 924         struct ipoib_dev_priv *priv;
 925         int result = -ENOMEM;
 926
 927         priv = ipoib_intf_alloc(format);
 928         if (!priv)
 929                 goto alloc_mem_failed;
 930
 931         SET_NETDEV_DEV(priv->dev, hca->dma_device);
 932
 933         result = ib_query_pkey(hca, port, 0, &priv->pkey);
 934         if (result) {
 935                 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
 936                        hca->name, port, result);
 937                 goto alloc_mem_failed;
 938         }
 939
 940         priv->dev->broadcast[8] = priv->pkey >> 8;
 941         priv->dev->broadcast[9] = priv->pkey & 0xff;
 942
 943         result = ib_query_gid(hca, port, 0, &priv->local_gid);
 944         if (result) {
 945                 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
 946                        hca->name, port, result);
 947                 goto alloc_mem_failed;
 948         } else
 949                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
 950
 951
 952         result = ipoib_dev_init(priv->dev, hca, port);
 953         if (result < 0) {
 954                 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
 955                        hca->name, port, result);
 956                 goto device_init_failed;
 957         }
 958
 959         INIT_IB_EVENT_HANDLER(&priv->event_handler,
 960                               priv->ca, ipoib_event);
 961         result = ib_register_event_handler(&priv->event_handler);
 962         if (result < 0) {
 963                 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
 964                        "port %d (ret = %d)\n",
 965                        hca->name, port, result);
 966                 goto event_failed;
 967         }
 968
 969         result = register_netdev(priv->dev);
 970         if (result) {
 971                 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
 972                        hca->name, port, result);
 973                 goto register_failed;
 974         }
 975
 976         if (ipoib_create_debug_file(priv->dev))
 977                 goto debug_failed;
 978
 979         if (ipoib_add_pkey_attr(priv->dev))
 980                 goto sysfs_failed;
 981         if (class_device_create_file(&priv->dev->class_dev,
 982                                      &class_device_attr_create_child))
 983                 goto sysfs_failed;
 984         if (class_device_create_file(&priv->dev->class_dev,
 985                                      &class_device_attr_delete_child))
 986                 goto sysfs_failed;
 987
 988         return priv->dev;
 989
 990 sysfs_failed:
 991         ipoib_delete_debug_file(priv->dev);
 992
 993 debug_failed:
 994         unregister_netdev(priv->dev);
 995
 996 register_failed:
 997         ib_unregister_event_handler(&priv->event_handler);
 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
1051                 unregister_netdev(priv->dev);
1052                 ipoib_dev_cleanup(priv->dev);
1053                 free_netdev(priv->dev);
1054         }
1055 }
1056
1057 static int __init ipoib_init_module(void)
1058 {
1059         int ret;
1060
1061         ret = ipoib_register_debugfs();
1062         if (ret)
1063                 return ret;
1064
1065         /*
1066          * We create our own workqueue mainly because we want to be
1067          * able to flush it when devices are being removed.  We can't
1068          * use schedule_work()/flush_scheduled_work() because both
1069          * unregister_netdev() and linkwatch_event take the rtnl lock,
1070          * so flush_scheduled_work() can deadlock during device
1071          * removal.
1072          */
1073         ipoib_workqueue = create_singlethread_workqueue("ipoib");
1074         if (!ipoib_workqueue) {
1075                 ret = -ENOMEM;
1076                 goto err_fs;
1077         }
1078
1079         ret = ib_register_client(&ipoib_client);
1080         if (ret)
1081                 goto err_wq;
1082
1083         return 0;
1084
1085 err_wq:
1086         destroy_workqueue(ipoib_workqueue);
1087
1088 err_fs:
1089         ipoib_unregister_debugfs();
1090
1091         return ret;
1092 }
1093
1094 static void __exit ipoib_cleanup_module(void)
1095 {
1096         ib_unregister_client(&ipoib_client);
1097         ipoib_unregister_debugfs();
1098         destroy_workqueue(ipoib_workqueue);
1099 }
1100
1101 module_init(ipoib_init_module);
1102 module_exit(ipoib_cleanup_module);