drivers/infiniband/ulp/ipoib/ipoib_main.c

   1 /*
   2  * Copyright (c) 2004 Topspin Communications.  All rights reserved.
   3  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
   4  * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
   5  *
   6  * This software is available to you under a choice of one of two
   7  * licenses.  You may choose to be licensed under the terms of the GNU
   8  * General Public License (GPL) Version 2, available from the file
   9  * COPYING in the main directory of this source tree, or the
  10  * OpenIB.org BSD license below:
  11  *
  12  *     Redistribution and use in source and binary forms, with or
  13  *     without modification, are permitted provided that the following
  14  *     conditions are met:
  15  *
  16  *      - Redistributions of source code must retain the above
  17  *        copyright notice, this list of conditions and the following
  18  *        disclaimer.
  19  *
  20  *      - Redistributions in binary form must reproduce the above
  21  *        copyright notice, this list of conditions and the following
  22  *        disclaimer in the documentation and/or other materials
  23  *        provided with the distribution.
  24  *
  25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  32  * SOFTWARE.
  33  *
  34  * $Id: ipoib_main.c 1377 2004-12-23 19:57:12Z roland $
  35  */
  36
  37 #include "ipoib.h"
  38
  39 #include <linux/module.h>
  40
  41 #include <linux/init.h>
  42 #include <linux/slab.h>
  43 #include <linux/vmalloc.h>
  44
  45 #include <linux/if_arp.h>       /* For ARPHRD_xxx */
  46
  47 #include <linux/ip.h>
  48 #include <linux/in.h>
  49
  50 #include <net/dst.h>
  51
  52 MODULE_AUTHOR("Roland Dreier");
  53 MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
  54 MODULE_LICENSE("Dual BSD/GPL");
  55
  56 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
  57 int ipoib_debug_level;
  58
  59 module_param_named(debug_level, ipoib_debug_level, int, 0644);
  60 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
  61 #endif
  62
  63 struct ipoib_path_iter {
  64         struct net_device *dev;
  65         struct ipoib_path  path;
  66 };
  67
  68 static const u8 ipv4_bcast_addr[] = {
  69         0x00, 0xff, 0xff, 0xff,
  70         0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
  71         0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff
  72 };
  73
  74 struct workqueue_struct *ipoib_workqueue;
  75
  76 static void ipoib_add_one(struct ib_device *device);
  77 static void ipoib_remove_one(struct ib_device *device);
  78
  79 static struct ib_client ipoib_client = {
  80         .name   = "ipoib",
  81         .add    = ipoib_add_one,
  82         .remove = ipoib_remove_one
  83 };
  84
  85 int ipoib_open(struct net_device *dev)
  86 {
  87         struct ipoib_dev_priv *priv = netdev_priv(dev);
  88
  89         ipoib_dbg(priv, "bringing up interface\n");
  90
  91         set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
  92
  93         if (ipoib_pkey_dev_delay_open(dev))
  94                 return 0;
  95
  96         if (ipoib_ib_dev_open(dev))
  97                 return -EINVAL;
  98
  99         if (ipoib_ib_dev_up(dev)) {
 100                 ipoib_ib_dev_stop(dev);
 101                 return -EINVAL;
 102         }
 103
 104         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 105                 struct ipoib_dev_priv *cpriv;
 106
 107                 /* Bring up any child interfaces too */
 108                 mutex_lock(&priv->vlan_mutex);
 109                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 110                         int flags;
 111
 112                         flags = cpriv->dev->flags;
 113                         if (flags & IFF_UP)
 114                                 continue;
 115
 116                         dev_change_flags(cpriv->dev, flags | IFF_UP);
 117                 }
 118                 mutex_unlock(&priv->vlan_mutex);
 119         }
 120
 121         netif_start_queue(dev);
 122
 123         return 0;
 124 }
 125
 126 static int ipoib_stop(struct net_device *dev)
 127 {
 128         struct ipoib_dev_priv *priv = netdev_priv(dev);
 129
 130         ipoib_dbg(priv, "stopping interface\n");
 131
 132         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 133
 134         netif_stop_queue(dev);
 135
 136         /*
 137          * Now flush workqueue to make sure a scheduled task doesn't
 138          * bring our internal state back up.
 139          */
 140         flush_workqueue(ipoib_workqueue);
 141
 142         ipoib_ib_dev_down(dev, 1);
 143         ipoib_ib_dev_stop(dev);
 144
 145         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 146                 struct ipoib_dev_priv *cpriv;
 147
 148                 /* Bring down any child interfaces too */
 149                 mutex_lock(&priv->vlan_mutex);
 150                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 151                         int flags;
 152
 153                         flags = cpriv->dev->flags;
 154                         if (!(flags & IFF_UP))
 155                                 continue;
 156
 157                         dev_change_flags(cpriv->dev, flags & ~IFF_UP);
 158                 }
 159                 mutex_unlock(&priv->vlan_mutex);
 160         }
 161
 162         return 0;
 163 }
 164
 165 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
 166 {
 167         struct ipoib_dev_priv *priv = netdev_priv(dev);
 168
 169         if (new_mtu > IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN)
 170                 return -EINVAL;
 171
 172         priv->admin_mtu = new_mtu;
 173
 174         dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
 175
 176         return 0;
 177 }
 178
 179 static struct ipoib_path *__path_find(struct net_device *dev,
 180                                       union ib_gid *gid)
 181 {
 182         struct ipoib_dev_priv *priv = netdev_priv(dev);
 183         struct rb_node *n = priv->path_tree.rb_node;
 184         struct ipoib_path *path;
 185         int ret;
 186
 187         while (n) {
 188                 path = rb_entry(n, struct ipoib_path, rb_node);
 189
 190                 ret = memcmp(gid->raw, path->pathrec.dgid.raw,
 191                              sizeof (union ib_gid));
 192
 193                 if (ret < 0)
 194                         n = n->rb_left;
 195                 else if (ret > 0)
 196                         n = n->rb_right;
 197                 else
 198                         return path;
 199         }
 200
 201         return NULL;
 202 }
 203
 204 static int __path_add(struct net_device *dev, struct ipoib_path *path)
 205 {
 206         struct ipoib_dev_priv *priv = netdev_priv(dev);
 207         struct rb_node **n = &priv->path_tree.rb_node;
 208         struct rb_node *pn = NULL;
 209         struct ipoib_path *tpath;
 210         int ret;
 211
 212         while (*n) {
 213                 pn = *n;
 214                 tpath = rb_entry(pn, struct ipoib_path, rb_node);
 215
 216                 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
 217                              sizeof (union ib_gid));
 218                 if (ret < 0)
 219                         n = &pn->rb_left;
 220                 else if (ret > 0)
 221                         n = &pn->rb_right;
 222                 else
 223                         return -EEXIST;
 224         }
 225
 226         rb_link_node(&path->rb_node, pn, n);
 227         rb_insert_color(&path->rb_node, &priv->path_tree);
 228
 229         list_add_tail(&path->list, &priv->path_list);
 230
 231         return 0;
 232 }
 233
 234 static void path_free(struct net_device *dev, struct ipoib_path *path)
 235 {
 236         struct ipoib_dev_priv *priv = netdev_priv(dev);
 237         struct ipoib_neigh *neigh, *tn;
 238         struct sk_buff *skb;
 239         unsigned long flags;
 240
 241         while ((skb = __skb_dequeue(&path->queue)))
 242                 dev_kfree_skb_irq(skb);
 243
 244         spin_lock_irqsave(&priv->lock, flags);
 245
 246         list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
 247                 /*
 248                  * It's safe to call ipoib_put_ah() inside priv->lock
 249                  * here, because we know that path->ah will always
 250                  * hold one more reference, so ipoib_put_ah() will
 251                  * never do more than decrement the ref count.
 252                  */
 253                 if (neigh->ah)
 254                         ipoib_put_ah(neigh->ah);
 255                 *to_ipoib_neigh(neigh->neighbour) = NULL;
 256                 neigh->neighbour->ops->destructor = NULL;
 257                 kfree(neigh);
 258         }
 259
 260         spin_unlock_irqrestore(&priv->lock, flags);
 261
 262         if (path->ah)
 263                 ipoib_put_ah(path->ah);
 264
 265         kfree(path);
 266 }
 267
 268 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 269
 270 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
 271 {
 272         struct ipoib_path_iter *iter;
 273
 274         iter = kmalloc(sizeof *iter, GFP_KERNEL);
 275         if (!iter)
 276                 return NULL;
 277
 278         iter->dev = dev;
 279         memset(iter->path.pathrec.dgid.raw, 0, 16);
 280
 281         if (ipoib_path_iter_next(iter)) {
 282                 kfree(iter);
 283                 return NULL;
 284         }
 285
 286         return iter;
 287 }
 288
 289 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
 290 {
 291         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
 292         struct rb_node *n;
 293         struct ipoib_path *path;
 294         int ret = 1;
 295
 296         spin_lock_irq(&priv->lock);
 297
 298         n = rb_first(&priv->path_tree);
 299
 300         while (n) {
 301                 path = rb_entry(n, struct ipoib_path, rb_node);
 302
 303                 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
 304                            sizeof (union ib_gid)) < 0) {
 305                         iter->path = *path;
 306                         ret = 0;
 307                         break;
 308                 }
 309
 310                 n = rb_next(n);
 311         }
 312
 313         spin_unlock_irq(&priv->lock);
 314
 315         return ret;
 316 }
 317
 318 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
 319                           struct ipoib_path *path)
 320 {
 321         *path = iter->path;
 322 }
 323
 324 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
 325
 326 void ipoib_flush_paths(struct net_device *dev)
 327 {
 328         struct ipoib_dev_priv *priv = netdev_priv(dev);
 329         struct ipoib_path *path, *tp;
 330         LIST_HEAD(remove_list);
 331         unsigned long flags;
 332
 333         spin_lock_irqsave(&priv->lock, flags);
 334
 335         list_splice(&priv->path_list, &remove_list);
 336         INIT_LIST_HEAD(&priv->path_list);
 337
 338         list_for_each_entry(path, &remove_list, list)
 339                 rb_erase(&path->rb_node, &priv->path_tree);
 340
 341         spin_unlock_irqrestore(&priv->lock, flags);
 342
 343         list_for_each_entry_safe(path, tp, &remove_list, list) {
 344                 if (path->query)
 345                         ib_sa_cancel_query(path->query_id, path->query);
 346                 wait_for_completion(&path->done);
 347                 path_free(dev, path);
 348         }
 349 }
 350
 351 static void path_rec_completion(int status,
 352                                 struct ib_sa_path_rec *pathrec,
 353                                 void *path_ptr)
 354 {
 355         struct ipoib_path *path = path_ptr;
 356         struct net_device *dev = path->dev;
 357         struct ipoib_dev_priv *priv = netdev_priv(dev);
 358         struct ipoib_ah *ah = NULL;
 359         struct ipoib_neigh *neigh;
 360         struct sk_buff_head skqueue;
 361         struct sk_buff *skb;
 362         unsigned long flags;
 363
 364         if (pathrec)
 365                 ipoib_dbg(priv, "PathRec LID 0x%04x for GID " IPOIB_GID_FMT "\n",
 366                           be16_to_cpu(pathrec->dlid), IPOIB_GID_ARG(pathrec->dgid));
 367         else
 368                 ipoib_dbg(priv, "PathRec status %d for GID " IPOIB_GID_FMT "\n",
 369                           status, IPOIB_GID_ARG(path->pathrec.dgid));
 370
 371         skb_queue_head_init(&skqueue);
 372
 373         if (!status) {
 374                 struct ib_ah_attr av = {
 375                         .dlid          = be16_to_cpu(pathrec->dlid),
 376                         .sl            = pathrec->sl,
 377                         .port_num      = priv->port
 378                 };
 379                 int path_rate = ib_sa_rate_enum_to_int(pathrec->rate);
 380
 381                 if (path_rate > 0 && priv->local_rate > path_rate)
 382                         av.static_rate = (priv->local_rate - 1) / path_rate;
 383
 384                 ipoib_dbg(priv, "static_rate %d for local port %dX, path %dX\n",
 385                           av.static_rate, priv->local_rate,
 386                           ib_sa_rate_enum_to_int(pathrec->rate));
 387
 388                 ah = ipoib_create_ah(dev, priv->pd, &av);
 389         }
 390
 391         spin_lock_irqsave(&priv->lock, flags);
 392
 393         path->ah = ah;
 394
 395         if (ah) {
 396                 path->pathrec = *pathrec;
 397
 398                 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
 399                           ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
 400
 401                 while ((skb = __skb_dequeue(&path->queue)))
 402                         __skb_queue_tail(&skqueue, skb);
 403
 404                 list_for_each_entry(neigh, &path->neigh_list, list) {
 405                         kref_get(&path->ah->ref);
 406                         neigh->ah = path->ah;
 407
 408                         while ((skb = __skb_dequeue(&neigh->queue)))
 409                                 __skb_queue_tail(&skqueue, skb);
 410                 }
 411         }
 412
 413         path->query = NULL;
 414         complete(&path->done);
 415
 416         spin_unlock_irqrestore(&priv->lock, flags);
 417
 418         while ((skb = __skb_dequeue(&skqueue))) {
 419                 skb->dev = dev;
 420                 if (dev_queue_xmit(skb))
 421                         ipoib_warn(priv, "dev_queue_xmit failed "
 422                                    "to requeue packet\n");
 423         }
 424 }
 425
 426 static struct ipoib_path *path_rec_create(struct net_device *dev,
 427                                           union ib_gid *gid)
 428 {
 429         struct ipoib_dev_priv *priv = netdev_priv(dev);
 430         struct ipoib_path *path;
 431
 432         path = kzalloc(sizeof *path, GFP_ATOMIC);
 433         if (!path)
 434                 return NULL;
 435
 436         path->dev = dev;
 437
 438         skb_queue_head_init(&path->queue);
 439
 440         INIT_LIST_HEAD(&path->neigh_list);
 441
 442         memcpy(path->pathrec.dgid.raw, gid->raw, sizeof (union ib_gid));
 443         path->pathrec.sgid      = priv->local_gid;
 444         path->pathrec.pkey      = cpu_to_be16(priv->pkey);
 445         path->pathrec.numb_path = 1;
 446
 447         return path;
 448 }
 449
 450 static int path_rec_start(struct net_device *dev,
 451                           struct ipoib_path *path)
 452 {
 453         struct ipoib_dev_priv *priv = netdev_priv(dev);
 454
 455         ipoib_dbg(priv, "Start path record lookup for " IPOIB_GID_FMT "\n",
 456                   IPOIB_GID_ARG(path->pathrec.dgid));
 457
 458         init_completion(&path->done);
 459
 460         path->query_id =
 461                 ib_sa_path_rec_get(priv->ca, priv->port,
 462                                    &path->pathrec,
 463                                    IB_SA_PATH_REC_DGID          |
 464                                    IB_SA_PATH_REC_SGID          |
 465                                    IB_SA_PATH_REC_NUMB_PATH     |
 466                                    IB_SA_PATH_REC_PKEY,
 467                                    1000, GFP_ATOMIC,
 468                                    path_rec_completion,
 469                                    path, &path->query);
 470         if (path->query_id < 0) {
 471                 ipoib_warn(priv, "ib_sa_path_rec_get failed\n");
 472                 path->query = NULL;
 473                 return path->query_id;
 474         }
 475
 476         return 0;
 477 }
 478
 479 static void neigh_add_path(struct sk_buff *skb, struct net_device *dev)
 480 {
 481         struct ipoib_dev_priv *priv = netdev_priv(dev);
 482         struct ipoib_path *path;
 483         struct ipoib_neigh *neigh;
 484
 485         neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
 486         if (!neigh) {
 487                 ++priv->stats.tx_dropped;
 488                 dev_kfree_skb_any(skb);
 489                 return;
 490         }
 491
 492         skb_queue_head_init(&neigh->queue);
 493         neigh->neighbour = skb->dst->neighbour;
 494         *to_ipoib_neigh(skb->dst->neighbour) = neigh;
 495
 496         /*
 497          * We can only be called from ipoib_start_xmit, so we're
 498          * inside tx_lock -- no need to save/restore flags.
 499          */
 500         spin_lock(&priv->lock);
 501
 502         path = __path_find(dev, (union ib_gid *) (skb->dst->neighbour->ha + 4));
 503         if (!path) {
 504                 path = path_rec_create(dev,
 505                                        (union ib_gid *) (skb->dst->neighbour->ha + 4));
 506                 if (!path)
 507                         goto err;
 508
 509                 __path_add(dev, path);
 510         }
 511
 512         list_add_tail(&neigh->list, &path->neigh_list);
 513
 514         if (path->ah) {
 515                 kref_get(&path->ah->ref);
 516                 neigh->ah = path->ah;
 517
 518                 ipoib_send(dev, skb, path->ah,
 519                            be32_to_cpup((__be32 *) skb->dst->neighbour->ha));
 520         } else {
 521                 neigh->ah  = NULL;
 522                 __skb_queue_tail(&neigh->queue, skb);
 523
 524                 if (!path->query && path_rec_start(dev, path))
 525                         goto err;
 526         }
 527
 528         spin_unlock(&priv->lock);
 529         return;
 530
 531 err:
 532         *to_ipoib_neigh(skb->dst->neighbour) = NULL;
 533         list_del(&neigh->list);
 534         neigh->neighbour->ops->destructor = NULL;
 535         kfree(neigh);
 536
 537         ++priv->stats.tx_dropped;
 538         dev_kfree_skb_any(skb);
 539
 540         spin_unlock(&priv->lock);
 541 }
 542
 543 static void ipoib_path_lookup(struct sk_buff *skb, struct net_device *dev)
 544 {
 545         struct ipoib_dev_priv *priv = netdev_priv(skb->dev);
 546
 547         /* Look up path record for unicasts */
 548         if (skb->dst->neighbour->ha[4] != 0xff) {
 549                 neigh_add_path(skb, dev);
 550                 return;
 551         }
 552
 553         /* Add in the P_Key for multicasts */
 554         skb->dst->neighbour->ha[8] = (priv->pkey >> 8) & 0xff;
 555         skb->dst->neighbour->ha[9] = priv->pkey & 0xff;
 556         ipoib_mcast_send(dev, (union ib_gid *) (skb->dst->neighbour->ha + 4), skb);
 557 }
 558
 559 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
 560                              struct ipoib_pseudoheader *phdr)
 561 {
 562         struct ipoib_dev_priv *priv = netdev_priv(dev);
 563         struct ipoib_path *path;
 564
 565         /*
 566          * We can only be called from ipoib_start_xmit, so we're
 567          * inside tx_lock -- no need to save/restore flags.
 568          */
 569         spin_lock(&priv->lock);
 570
 571         path = __path_find(dev, (union ib_gid *) (phdr->hwaddr + 4));
 572         if (!path) {
 573                 path = path_rec_create(dev,
 574                                        (union ib_gid *) (phdr->hwaddr + 4));
 575                 if (path) {
 576                         /* put pseudoheader back on for next time */
 577                         skb_push(skb, sizeof *phdr);
 578                         __skb_queue_tail(&path->queue, skb);
 579
 580                         if (path_rec_start(dev, path)) {
 581                                 spin_unlock(&priv->lock);
 582                                 path_free(dev, path);
 583                                 return;
 584                         } else
 585                                 __path_add(dev, path);
 586                 } else {
 587                         ++priv->stats.tx_dropped;
 588                         dev_kfree_skb_any(skb);
 589                 }
 590
 591                 spin_unlock(&priv->lock);
 592                 return;
 593         }
 594
 595         if (path->ah) {
 596                 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
 597                           be16_to_cpu(path->pathrec.dlid));
 598
 599                 ipoib_send(dev, skb, path->ah,
 600                            be32_to_cpup((__be32 *) phdr->hwaddr));
 601         } else if ((path->query || !path_rec_start(dev, path)) &&
 602                    skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 603                 /* put pseudoheader back on for next time */
 604                 skb_push(skb, sizeof *phdr);
 605                 __skb_queue_tail(&path->queue, skb);
 606         } else {
 607                 ++priv->stats.tx_dropped;
 608                 dev_kfree_skb_any(skb);
 609         }
 610
 611         spin_unlock(&priv->lock);
 612 }
 613
 614 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
 615 {
 616         struct ipoib_dev_priv *priv = netdev_priv(dev);
 617         struct ipoib_neigh *neigh;
 618         unsigned long flags;
 619
 620         if (!spin_trylock_irqsave(&priv->tx_lock, flags))
 621                 return NETDEV_TX_LOCKED;
 622
 623         /*
 624          * Check if our queue is stopped.  Since we have the LLTX bit
 625          * set, we can't rely on netif_stop_queue() preventing our
 626          * xmit function from being called with a full queue.
 627          */
 628         if (unlikely(netif_queue_stopped(dev))) {
 629                 spin_unlock_irqrestore(&priv->tx_lock, flags);
 630                 return NETDEV_TX_BUSY;
 631         }
 632
 633         if (skb->dst && skb->dst->neighbour) {
 634                 if (unlikely(!*to_ipoib_neigh(skb->dst->neighbour))) {
 635                         ipoib_path_lookup(skb, dev);
 636                         goto out;
 637                 }
 638
 639                 neigh = *to_ipoib_neigh(skb->dst->neighbour);
 640
 641                 if (likely(neigh->ah)) {
 642                         ipoib_send(dev, skb, neigh->ah,
 643                                    be32_to_cpup((__be32 *) skb->dst->neighbour->ha));
 644                         goto out;
 645                 }
 646
 647                 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 648                         spin_lock(&priv->lock);
 649                         __skb_queue_tail(&neigh->queue, skb);
 650                         spin_unlock(&priv->lock);
 651                 } else {
 652                         ++priv->stats.tx_dropped;
 653                         dev_kfree_skb_any(skb);
 654                 }
 655         } else {
 656                 struct ipoib_pseudoheader *phdr =
 657                         (struct ipoib_pseudoheader *) skb->data;
 658                 skb_pull(skb, sizeof *phdr);
 659
 660                 if (phdr->hwaddr[4] == 0xff) {
 661                         /* Add in the P_Key for multicast*/
 662                         phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
 663                         phdr->hwaddr[9] = priv->pkey & 0xff;
 664
 665                         ipoib_mcast_send(dev, (union ib_gid *) (phdr->hwaddr + 4), skb);
 666                 } else {
 667                         /* unicast GID -- should be ARP or RARP reply */
 668
 669                         if ((be16_to_cpup((__be16 *) skb->data) != ETH_P_ARP) &&
 670                             (be16_to_cpup((__be16 *) skb->data) != ETH_P_RARP)) {
 671                                 ipoib_warn(priv, "Unicast, no %s: type %04x, QPN %06x "
 672                                            IPOIB_GID_FMT "\n",
 673                                            skb->dst ? "neigh" : "dst",
 674                                            be16_to_cpup((__be16 *) skb->data),
 675                                            be32_to_cpup((__be32 *) phdr->hwaddr),
 676                                            IPOIB_GID_ARG(*(union ib_gid *) (phdr->hwaddr + 4)));
 677                                 dev_kfree_skb_any(skb);
 678                                 ++priv->stats.tx_dropped;
 679                                 goto out;
 680                         }
 681
 682                         unicast_arp_send(skb, dev, phdr);
 683                 }
 684         }
 685
 686 out:
 687         spin_unlock_irqrestore(&priv->tx_lock, flags);
 688
 689         return NETDEV_TX_OK;
 690 }
 691
 692 static struct net_device_stats *ipoib_get_stats(struct net_device *dev)
 693 {
 694         struct ipoib_dev_priv *priv = netdev_priv(dev);
 695
 696         return &priv->stats;
 697 }
 698
 699 static void ipoib_timeout(struct net_device *dev)
 700 {
 701         struct ipoib_dev_priv *priv = netdev_priv(dev);
 702
 703         ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
 704                    jiffies_to_msecs(jiffies - dev->trans_start));
 705         ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
 706                    netif_queue_stopped(dev),
 707                    priv->tx_head, priv->tx_tail);
 708         /* XXX reset QP, etc. */
 709 }
 710
 711 static int ipoib_hard_header(struct sk_buff *skb,
 712                              struct net_device *dev,
 713                              unsigned short type,
 714                              void *daddr, void *saddr, unsigned len)
 715 {
 716         struct ipoib_header *header;
 717
 718         header = (struct ipoib_header *) skb_push(skb, sizeof *header);
 719
 720         header->proto = htons(type);
 721         header->reserved = 0;
 722
 723         /*
 724          * If we don't have a neighbour structure, stuff the
 725          * destination address onto the front of the skb so we can
 726          * figure out where to send the packet later.
 727          */
 728         if (!skb->dst || !skb->dst->neighbour) {
 729                 struct ipoib_pseudoheader *phdr =
 730                         (struct ipoib_pseudoheader *) skb_push(skb, sizeof *phdr);
 731                 memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
 732         }
 733
 734         return 0;
 735 }
 736
 737 static void ipoib_set_mcast_list(struct net_device *dev)
 738 {
 739         struct ipoib_dev_priv *priv = netdev_priv(dev);
 740
 741         queue_work(ipoib_workqueue, &priv->restart_task);
 742 }
 743
 744 static void ipoib_neigh_destructor(struct neighbour *n)
 745 {
 746         struct ipoib_neigh *neigh;
 747         struct ipoib_dev_priv *priv = netdev_priv(n->dev);
 748         unsigned long flags;
 749         struct ipoib_ah *ah = NULL;
 750
 751         ipoib_dbg(priv,
 752                   "neigh_destructor for %06x " IPOIB_GID_FMT "\n",
 753                   be32_to_cpup((__be32 *) n->ha),
 754                   IPOIB_GID_ARG(*((union ib_gid *) (n->ha + 4))));
 755
 756         spin_lock_irqsave(&priv->lock, flags);
 757
 758         neigh = *to_ipoib_neigh(n);
 759         if (neigh) {
 760                 if (neigh->ah)
 761                         ah = neigh->ah;
 762                 list_del(&neigh->list);
 763                 *to_ipoib_neigh(n) = NULL;
 764                 kfree(neigh);
 765         }
 766
 767         spin_unlock_irqrestore(&priv->lock, flags);
 768
 769         if (ah)
 770                 ipoib_put_ah(ah);
 771 }
 772
 773 static int ipoib_neigh_setup(struct neighbour *neigh)
 774 {
 775         /*
 776          * Is this kosher?  I can't find anybody in the kernel that
 777          * sets neigh->destructor, so we should be able to set it here
 778          * without trouble.
 779          */
 780         neigh->ops->destructor = ipoib_neigh_destructor;
 781
 782         return 0;
 783 }
 784
 785 static int ipoib_neigh_setup_dev(struct net_device *dev, struct neigh_parms *parms)
 786 {
 787         parms->neigh_setup = ipoib_neigh_setup;
 788
 789         return 0;
 790 }
 791
 792 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
 793 {
 794         struct ipoib_dev_priv *priv = netdev_priv(dev);
 795
 796         /* Allocate RX/TX "rings" to hold queued skbs */
 797
 798         priv->rx_ring = kzalloc(IPOIB_RX_RING_SIZE * sizeof (struct ipoib_rx_buf),
 799                                 GFP_KERNEL);
 800         if (!priv->rx_ring) {
 801                 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
 802                        ca->name, IPOIB_RX_RING_SIZE);
 803                 goto out;
 804         }
 805
 806         priv->tx_ring = kzalloc(IPOIB_TX_RING_SIZE * sizeof (struct ipoib_tx_buf),
 807                                 GFP_KERNEL);
 808         if (!priv->tx_ring) {
 809                 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
 810                        ca->name, IPOIB_TX_RING_SIZE);
 811                 goto out_rx_ring_cleanup;
 812         }
 813
 814         /* priv->tx_head & tx_tail are already 0 */
 815
 816         if (ipoib_ib_dev_init(dev, ca, port))
 817                 goto out_tx_ring_cleanup;
 818
 819         return 0;
 820
 821 out_tx_ring_cleanup:
 822         kfree(priv->tx_ring);
 823
 824 out_rx_ring_cleanup:
 825         kfree(priv->rx_ring);
 826
 827 out:
 828         return -ENOMEM;
 829 }
 830
 831 void ipoib_dev_cleanup(struct net_device *dev)
 832 {
 833         struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
 834
 835         ipoib_delete_debug_files(dev);
 836
 837         /* Delete any child interfaces first */
 838         list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
 839                 unregister_netdev(cpriv->dev);
 840                 ipoib_dev_cleanup(cpriv->dev);
 841                 free_netdev(cpriv->dev);
 842         }
 843
 844         ipoib_ib_dev_cleanup(dev);
 845
 846         kfree(priv->rx_ring);
 847         kfree(priv->tx_ring);
 848
 849         priv->rx_ring = NULL;
 850         priv->tx_ring = NULL;
 851 }
 852
 853 static void ipoib_setup(struct net_device *dev)
 854 {
 855         struct ipoib_dev_priv *priv = netdev_priv(dev);
 856
 857         dev->open                = ipoib_open;
 858         dev->stop                = ipoib_stop;
 859         dev->change_mtu          = ipoib_change_mtu;
 860         dev->hard_start_xmit     = ipoib_start_xmit;
 861         dev->get_stats           = ipoib_get_stats;
 862         dev->tx_timeout          = ipoib_timeout;
 863         dev->hard_header         = ipoib_hard_header;
 864         dev->set_multicast_list  = ipoib_set_mcast_list;
 865         dev->neigh_setup         = ipoib_neigh_setup_dev;
 866
 867         dev->watchdog_timeo      = HZ;
 868
 869         dev->flags              |= IFF_BROADCAST | IFF_MULTICAST;
 870
 871         /*
 872          * We add in INFINIBAND_ALEN to allow for the destination
 873          * address "pseudoheader" for skbs without neighbour struct.
 874          */
 875         dev->hard_header_len     = IPOIB_ENCAP_LEN + INFINIBAND_ALEN;
 876         dev->addr_len            = INFINIBAND_ALEN;
 877         dev->type                = ARPHRD_INFINIBAND;
 878         dev->tx_queue_len        = IPOIB_TX_RING_SIZE * 2;
 879         dev->features            = NETIF_F_VLAN_CHALLENGED | NETIF_F_LLTX;
 880
 881         /* MTU will be reset when mcast join happens */
 882         dev->mtu                 = IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN;
 883         priv->mcast_mtu          = priv->admin_mtu = dev->mtu;
 884
 885         memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
 886
 887         netif_carrier_off(dev);
 888
 889         SET_MODULE_OWNER(dev);
 890
 891         priv->dev = dev;
 892
 893         spin_lock_init(&priv->lock);
 894         spin_lock_init(&priv->tx_lock);
 895
 896         mutex_init(&priv->mcast_mutex);
 897         mutex_init(&priv->vlan_mutex);
 898
 899         INIT_LIST_HEAD(&priv->path_list);
 900         INIT_LIST_HEAD(&priv->child_intfs);
 901         INIT_LIST_HEAD(&priv->dead_ahs);
 902         INIT_LIST_HEAD(&priv->multicast_list);
 903
 904         INIT_WORK(&priv->pkey_task,    ipoib_pkey_poll,          priv->dev);
 905         INIT_WORK(&priv->mcast_task,   ipoib_mcast_join_task,    priv->dev);
 906         INIT_WORK(&priv->flush_task,   ipoib_ib_dev_flush,       priv->dev);
 907         INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task, priv->dev);
 908         INIT_WORK(&priv->ah_reap_task, ipoib_reap_ah,            priv->dev);
 909 }
 910
 911 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
 912 {
 913         struct net_device *dev;
 914
 915         dev = alloc_netdev((int) sizeof (struct ipoib_dev_priv), name,
 916                            ipoib_setup);
 917         if (!dev)
 918                 return NULL;
 919
 920         return netdev_priv(dev);
 921 }
 922
 923 static ssize_t show_pkey(struct class_device *cdev, char *buf)
 924 {
 925         struct ipoib_dev_priv *priv =
 926                 netdev_priv(container_of(cdev, struct net_device, class_dev));
 927
 928         return sprintf(buf, "0x%04x\n", priv->pkey);
 929 }
 930 static CLASS_DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
 931
 932 static ssize_t create_child(struct class_device *cdev,
 933                             const char *buf, size_t count)
 934 {
 935         int pkey;
 936         int ret;
 937
 938         if (sscanf(buf, "%i", &pkey) != 1)
 939                 return -EINVAL;
 940
 941         if (pkey < 0 || pkey > 0xffff)
 942                 return -EINVAL;
 943
 944         /*
 945          * Set the full membership bit, so that we join the right
 946          * broadcast group, etc.
 947          */
 948         pkey |= 0x8000;
 949
 950         ret = ipoib_vlan_add(container_of(cdev, struct net_device, class_dev),
 951                              pkey);
 952
 953         return ret ? ret : count;
 954 }
 955 static CLASS_DEVICE_ATTR(create_child, S_IWUGO, NULL, create_child);
 956
 957 static ssize_t delete_child(struct class_device *cdev,
 958                             const char *buf, size_t count)
 959 {
 960         int pkey;
 961         int ret;
 962
 963         if (sscanf(buf, "%i", &pkey) != 1)
 964                 return -EINVAL;
 965
 966         if (pkey < 0 || pkey > 0xffff)
 967                 return -EINVAL;
 968
 969         ret = ipoib_vlan_delete(container_of(cdev, struct net_device, class_dev),
 970                                 pkey);
 971
 972         return ret ? ret : count;
 973
 974 }
 975 static CLASS_DEVICE_ATTR(delete_child, S_IWUGO, NULL, delete_child);
 976
 977 int ipoib_add_pkey_attr(struct net_device *dev)
 978 {
 979         return class_device_create_file(&dev->class_dev,
 980                                         &class_device_attr_pkey);
 981 }
 982
 983 static struct net_device *ipoib_add_port(const char *format,
 984                                          struct ib_device *hca, u8 port)
 985 {
 986         struct ipoib_dev_priv *priv;
 987         int result = -ENOMEM;
 988
 989         priv = ipoib_intf_alloc(format);
 990         if (!priv)
 991                 goto alloc_mem_failed;
 992
 993         SET_NETDEV_DEV(priv->dev, hca->dma_device);
 994
 995         result = ib_query_pkey(hca, port, 0, &priv->pkey);
 996         if (result) {
 997                 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
 998                        hca->name, port, result);
 999                 goto alloc_mem_failed;
1000         }
1001
1002         /*
1003          * Set the full membership bit, so that we join the right
1004          * broadcast group, etc.
1005          */
1006         priv->pkey |= 0x8000;
1007
1008         priv->dev->broadcast[8] = priv->pkey >> 8;
1009         priv->dev->broadcast[9] = priv->pkey & 0xff;
1010
1011         result = ib_query_gid(hca, port, 0, &priv->local_gid);
1012         if (result) {
1013                 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1014                        hca->name, port, result);
1015                 goto alloc_mem_failed;
1016         } else
1017                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1018
1019
1020         result = ipoib_dev_init(priv->dev, hca, port);
1021         if (result < 0) {
1022                 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1023                        hca->name, port, result);
1024                 goto device_init_failed;
1025         }
1026
1027         INIT_IB_EVENT_HANDLER(&priv->event_handler,
1028                               priv->ca, ipoib_event);
1029         result = ib_register_event_handler(&priv->event_handler);
1030         if (result < 0) {
1031                 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1032                        "port %d (ret = %d)\n",
1033                        hca->name, port, result);
1034                 goto event_failed;
1035         }
1036
1037         result = register_netdev(priv->dev);
1038         if (result) {
1039                 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1040                        hca->name, port, result);
1041                 goto register_failed;
1042         }
1043
1044         ipoib_create_debug_files(priv->dev);
1045
1046         if (ipoib_add_pkey_attr(priv->dev))
1047                 goto sysfs_failed;
1048         if (class_device_create_file(&priv->dev->class_dev,
1049                                      &class_device_attr_create_child))
1050                 goto sysfs_failed;
1051         if (class_device_create_file(&priv->dev->class_dev,
1052                                      &class_device_attr_delete_child))
1053                 goto sysfs_failed;
1054
1055         return priv->dev;
1056
1057 sysfs_failed:
1058         ipoib_delete_debug_files(priv->dev);
1059         unregister_netdev(priv->dev);
1060
1061 register_failed:
1062         ib_unregister_event_handler(&priv->event_handler);
1063         flush_scheduled_work();
1064
1065 event_failed:
1066         ipoib_dev_cleanup(priv->dev);
1067
1068 device_init_failed:
1069         free_netdev(priv->dev);
1070
1071 alloc_mem_failed:
1072         return ERR_PTR(result);
1073 }
1074
1075 static void ipoib_add_one(struct ib_device *device)
1076 {
1077         struct list_head *dev_list;
1078         struct net_device *dev;
1079         struct ipoib_dev_priv *priv;
1080         int s, e, p;
1081
1082         dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1083         if (!dev_list)
1084                 return;
1085
1086         INIT_LIST_HEAD(dev_list);
1087
1088         if (device->node_type == IB_NODE_SWITCH) {
1089                 s = 0;
1090                 e = 0;
1091         } else {
1092                 s = 1;
1093                 e = device->phys_port_cnt;
1094         }
1095
1096         for (p = s; p <= e; ++p) {
1097                 dev = ipoib_add_port("ib%d", device, p);
1098                 if (!IS_ERR(dev)) {
1099                         priv = netdev_priv(dev);
1100                         list_add_tail(&priv->list, dev_list);
1101                 }
1102         }
1103
1104         ib_set_client_data(device, &ipoib_client, dev_list);
1105 }
1106
1107 static void ipoib_remove_one(struct ib_device *device)
1108 {
1109         struct ipoib_dev_priv *priv, *tmp;
1110         struct list_head *dev_list;
1111
1112         dev_list = ib_get_client_data(device, &ipoib_client);
1113
1114         list_for_each_entry_safe(priv, tmp, dev_list, list) {
1115                 ib_unregister_event_handler(&priv->event_handler);
1116                 flush_scheduled_work();
1117
1118                 unregister_netdev(priv->dev);
1119                 ipoib_dev_cleanup(priv->dev);
1120                 free_netdev(priv->dev);
1121         }
1122
1123         kfree(dev_list);
1124 }
1125
1126 static int __init ipoib_init_module(void)
1127 {
1128         int ret;
1129
1130         ret = ipoib_register_debugfs();
1131         if (ret)
1132                 return ret;
1133
1134         /*
1135          * We create our own workqueue mainly because we want to be
1136          * able to flush it when devices are being removed.  We can't
1137          * use schedule_work()/flush_scheduled_work() because both
1138          * unregister_netdev() and linkwatch_event take the rtnl lock,
1139          * so flush_scheduled_work() can deadlock during device
1140          * removal.
1141          */
1142         ipoib_workqueue = create_singlethread_workqueue("ipoib");
1143         if (!ipoib_workqueue) {
1144                 ret = -ENOMEM;
1145                 goto err_fs;
1146         }
1147
1148         ret = ib_register_client(&ipoib_client);
1149         if (ret)
1150                 goto err_wq;
1151
1152         return 0;
1153
1154 err_wq:
1155         destroy_workqueue(ipoib_workqueue);
1156
1157 err_fs:
1158         ipoib_unregister_debugfs();
1159
1160         return ret;
1161 }
1162
1163 static void __exit ipoib_cleanup_module(void)
1164 {
1165         ib_unregister_client(&ipoib_client);
1166         ipoib_unregister_debugfs();
1167         destroy_workqueue(ipoib_workqueue);
1168 }
1169
1170 module_init(ipoib_init_module);
1171 module_exit(ipoib_cleanup_module);