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

   1 /*
   2  * Copyright (c) 2004 Topspin Communications.  All rights reserved.
   3  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
   4  * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
   5  *
   6  * This software is available to you under a choice of one of two
   7  * licenses.  You may choose to be licensed under the terms of the GNU
   8  * General Public License (GPL) Version 2, available from the file
   9  * COPYING in the main directory of this source tree, or the
  10  * OpenIB.org BSD license below:
  11  *
  12  *     Redistribution and use in source and binary forms, with or
  13  *     without modification, are permitted provided that the following
  14  *     conditions are met:
  15  *
  16  *      - Redistributions of source code must retain the above
  17  *        copyright notice, this list of conditions and the following
  18  *        disclaimer.
  19  *
  20  *      - Redistributions in binary form must reproduce the above
  21  *        copyright notice, this list of conditions and the following
  22  *        disclaimer in the documentation and/or other materials
  23  *        provided with the distribution.
  24  *
  25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  32  * SOFTWARE.
  33  *
  34  * $Id: ipoib_main.c 1377 2004-12-23 19:57:12Z roland $
  35  */
  36
  37 #include "ipoib.h"
  38
  39 #include <linux/module.h>
  40
  41 #include <linux/init.h>
  42 #include <linux/slab.h>
  43 #include <linux/kernel.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 int ipoib_sendq_size __read_mostly = IPOIB_TX_RING_SIZE;
  57 int ipoib_recvq_size __read_mostly = IPOIB_RX_RING_SIZE;
  58
  59 module_param_named(send_queue_size, ipoib_sendq_size, int, 0444);
  60 MODULE_PARM_DESC(send_queue_size, "Number of descriptors in send queue");
  61 module_param_named(recv_queue_size, ipoib_recvq_size, int, 0444);
  62 MODULE_PARM_DESC(recv_queue_size, "Number of descriptors in receive queue");
  63
  64 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
  65 int ipoib_debug_level;
  66
  67 module_param_named(debug_level, ipoib_debug_level, int, 0644);
  68 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
  69 #endif
  70
  71 struct ipoib_path_iter {
  72         struct net_device *dev;
  73         struct ipoib_path  path;
  74 };
  75
  76 static const u8 ipv4_bcast_addr[] = {
  77         0x00, 0xff, 0xff, 0xff,
  78         0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
  79         0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff
  80 };
  81
  82 struct workqueue_struct *ipoib_workqueue;
  83
  84 struct ib_sa_client ipoib_sa_client;
  85
  86 static void ipoib_add_one(struct ib_device *device);
  87 static void ipoib_remove_one(struct ib_device *device);
  88
  89 static struct ib_client ipoib_client = {
  90         .name   = "ipoib",
  91         .add    = ipoib_add_one,
  92         .remove = ipoib_remove_one
  93 };
  94
  95 int ipoib_open(struct net_device *dev)
  96 {
  97         struct ipoib_dev_priv *priv = netdev_priv(dev);
  98
  99         ipoib_dbg(priv, "bringing up interface\n");
 100
 101         set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 102
 103         if (ipoib_pkey_dev_delay_open(dev))
 104                 return 0;
 105
 106         if (ipoib_ib_dev_open(dev))
 107                 return -EINVAL;
 108
 109         if (ipoib_ib_dev_up(dev)) {
 110                 ipoib_ib_dev_stop(dev);
 111                 return -EINVAL;
 112         }
 113
 114         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 115                 struct ipoib_dev_priv *cpriv;
 116
 117                 /* Bring up any child interfaces too */
 118                 mutex_lock(&priv->vlan_mutex);
 119                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 120                         int flags;
 121
 122                         flags = cpriv->dev->flags;
 123                         if (flags & IFF_UP)
 124                                 continue;
 125
 126                         dev_change_flags(cpriv->dev, flags | IFF_UP);
 127                 }
 128                 mutex_unlock(&priv->vlan_mutex);
 129         }
 130
 131         netif_start_queue(dev);
 132
 133         return 0;
 134 }
 135
 136 static int ipoib_stop(struct net_device *dev)
 137 {
 138         struct ipoib_dev_priv *priv = netdev_priv(dev);
 139
 140         ipoib_dbg(priv, "stopping interface\n");
 141
 142         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 143
 144         netif_stop_queue(dev);
 145
 146         /*
 147          * Now flush workqueue to make sure a scheduled task doesn't
 148          * bring our internal state back up.
 149          */
 150         flush_workqueue(ipoib_workqueue);
 151
 152         ipoib_ib_dev_down(dev, 1);
 153         ipoib_ib_dev_stop(dev);
 154
 155         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 156                 struct ipoib_dev_priv *cpriv;
 157
 158                 /* Bring down any child interfaces too */
 159                 mutex_lock(&priv->vlan_mutex);
 160                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 161                         int flags;
 162
 163                         flags = cpriv->dev->flags;
 164                         if (!(flags & IFF_UP))
 165                                 continue;
 166
 167                         dev_change_flags(cpriv->dev, flags & ~IFF_UP);
 168                 }
 169                 mutex_unlock(&priv->vlan_mutex);
 170         }
 171
 172         return 0;
 173 }
 174
 175 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
 176 {
 177         struct ipoib_dev_priv *priv = netdev_priv(dev);
 178
 179         if (new_mtu > IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN)
 180                 return -EINVAL;
 181
 182         priv->admin_mtu = new_mtu;
 183
 184         dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
 185
 186         return 0;
 187 }
 188
 189 static struct ipoib_path *__path_find(struct net_device *dev, void *gid)
 190 {
 191         struct ipoib_dev_priv *priv = netdev_priv(dev);
 192         struct rb_node *n = priv->path_tree.rb_node;
 193         struct ipoib_path *path;
 194         int ret;
 195
 196         while (n) {
 197                 path = rb_entry(n, struct ipoib_path, rb_node);
 198
 199                 ret = memcmp(gid, path->pathrec.dgid.raw,
 200                              sizeof (union ib_gid));
 201
 202                 if (ret < 0)
 203                         n = n->rb_left;
 204                 else if (ret > 0)
 205                         n = n->rb_right;
 206                 else
 207                         return path;
 208         }
 209
 210         return NULL;
 211 }
 212
 213 static int __path_add(struct net_device *dev, struct ipoib_path *path)
 214 {
 215         struct ipoib_dev_priv *priv = netdev_priv(dev);
 216         struct rb_node **n = &priv->path_tree.rb_node;
 217         struct rb_node *pn = NULL;
 218         struct ipoib_path *tpath;
 219         int ret;
 220
 221         while (*n) {
 222                 pn = *n;
 223                 tpath = rb_entry(pn, struct ipoib_path, rb_node);
 224
 225                 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
 226                              sizeof (union ib_gid));
 227                 if (ret < 0)
 228                         n = &pn->rb_left;
 229                 else if (ret > 0)
 230                         n = &pn->rb_right;
 231                 else
 232                         return -EEXIST;
 233         }
 234
 235         rb_link_node(&path->rb_node, pn, n);
 236         rb_insert_color(&path->rb_node, &priv->path_tree);
 237
 238         list_add_tail(&path->list, &priv->path_list);
 239
 240         return 0;
 241 }
 242
 243 static void path_free(struct net_device *dev, struct ipoib_path *path)
 244 {
 245         struct ipoib_dev_priv *priv = netdev_priv(dev);
 246         struct ipoib_neigh *neigh, *tn;
 247         struct sk_buff *skb;
 248         unsigned long flags;
 249
 250         while ((skb = __skb_dequeue(&path->queue)))
 251                 dev_kfree_skb_irq(skb);
 252
 253         spin_lock_irqsave(&priv->lock, flags);
 254
 255         list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
 256                 /*
 257                  * It's safe to call ipoib_put_ah() inside priv->lock
 258                  * here, because we know that path->ah will always
 259                  * hold one more reference, so ipoib_put_ah() will
 260                  * never do more than decrement the ref count.
 261                  */
 262                 if (neigh->ah)
 263                         ipoib_put_ah(neigh->ah);
 264
 265                 ipoib_neigh_free(neigh);
 266         }
 267
 268         spin_unlock_irqrestore(&priv->lock, flags);
 269
 270         if (path->ah)
 271                 ipoib_put_ah(path->ah);
 272
 273         kfree(path);
 274 }
 275
 276 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 277
 278 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
 279 {
 280         struct ipoib_path_iter *iter;
 281
 282         iter = kmalloc(sizeof *iter, GFP_KERNEL);
 283         if (!iter)
 284                 return NULL;
 285
 286         iter->dev = dev;
 287         memset(iter->path.pathrec.dgid.raw, 0, 16);
 288
 289         if (ipoib_path_iter_next(iter)) {
 290                 kfree(iter);
 291                 return NULL;
 292         }
 293
 294         return iter;
 295 }
 296
 297 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
 298 {
 299         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
 300         struct rb_node *n;
 301         struct ipoib_path *path;
 302         int ret = 1;
 303
 304         spin_lock_irq(&priv->lock);
 305
 306         n = rb_first(&priv->path_tree);
 307
 308         while (n) {
 309                 path = rb_entry(n, struct ipoib_path, rb_node);
 310
 311                 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
 312                            sizeof (union ib_gid)) < 0) {
 313                         iter->path = *path;
 314                         ret = 0;
 315                         break;
 316                 }
 317
 318                 n = rb_next(n);
 319         }
 320
 321         spin_unlock_irq(&priv->lock);
 322
 323         return ret;
 324 }
 325
 326 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
 327                           struct ipoib_path *path)
 328 {
 329         *path = iter->path;
 330 }
 331
 332 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
 333
 334 void ipoib_flush_paths(struct net_device *dev)
 335 {
 336         struct ipoib_dev_priv *priv = netdev_priv(dev);
 337         struct ipoib_path *path, *tp;
 338         LIST_HEAD(remove_list);
 339
 340         spin_lock_irq(&priv->tx_lock);
 341         spin_lock(&priv->lock);
 342
 343         list_splice(&priv->path_list, &remove_list);
 344         INIT_LIST_HEAD(&priv->path_list);
 345
 346         list_for_each_entry(path, &remove_list, list)
 347                 rb_erase(&path->rb_node, &priv->path_tree);
 348
 349         list_for_each_entry_safe(path, tp, &remove_list, list) {
 350                 if (path->query)
 351                         ib_sa_cancel_query(path->query_id, path->query);
 352                 spin_unlock(&priv->lock);
 353                 spin_unlock_irq(&priv->tx_lock);
 354                 wait_for_completion(&path->done);
 355                 path_free(dev, path);
 356                 spin_lock_irq(&priv->tx_lock);
 357                 spin_lock(&priv->lock);
 358         }
 359         spin_unlock(&priv->lock);
 360         spin_unlock_irq(&priv->tx_lock);
 361 }
 362
 363 static void path_rec_completion(int status,
 364                                 struct ib_sa_path_rec *pathrec,
 365                                 void *path_ptr)
 366 {
 367         struct ipoib_path *path = path_ptr;
 368         struct net_device *dev = path->dev;
 369         struct ipoib_dev_priv *priv = netdev_priv(dev);
 370         struct ipoib_ah *ah = NULL;
 371         struct ipoib_neigh *neigh;
 372         struct sk_buff_head skqueue;
 373         struct sk_buff *skb;
 374         unsigned long flags;
 375
 376         if (pathrec)
 377                 ipoib_dbg(priv, "PathRec LID 0x%04x for GID " IPOIB_GID_FMT "\n",
 378                           be16_to_cpu(pathrec->dlid), IPOIB_GID_ARG(pathrec->dgid));
 379         else
 380                 ipoib_dbg(priv, "PathRec status %d for GID " IPOIB_GID_FMT "\n",
 381                           status, IPOIB_GID_ARG(path->pathrec.dgid));
 382
 383         skb_queue_head_init(&skqueue);
 384
 385         if (!status) {
 386                 struct ib_ah_attr av = {
 387                         .dlid          = be16_to_cpu(pathrec->dlid),
 388                         .sl            = pathrec->sl,
 389                         .port_num      = priv->port,
 390                         .static_rate   = pathrec->rate
 391                 };
 392
 393                 ah = ipoib_create_ah(dev, priv->pd, &av);
 394         }
 395
 396         spin_lock_irqsave(&priv->lock, flags);
 397
 398         path->ah = ah;
 399
 400         if (ah) {
 401                 path->pathrec = *pathrec;
 402
 403                 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
 404                           ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
 405
 406                 while ((skb = __skb_dequeue(&path->queue)))
 407                         __skb_queue_tail(&skqueue, skb);
 408
 409                 list_for_each_entry(neigh, &path->neigh_list, list) {
 410                         kref_get(&path->ah->ref);
 411                         neigh->ah = path->ah;
 412                         memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
 413                                sizeof(union ib_gid));
 414
 415                         while ((skb = __skb_dequeue(&neigh->queue)))
 416                                 __skb_queue_tail(&skqueue, skb);
 417                 }
 418         }
 419
 420         path->query = NULL;
 421         complete(&path->done);
 422
 423         spin_unlock_irqrestore(&priv->lock, flags);
 424
 425         while ((skb = __skb_dequeue(&skqueue))) {
 426                 skb->dev = dev;
 427                 if (dev_queue_xmit(skb))
 428                         ipoib_warn(priv, "dev_queue_xmit failed "
 429                                    "to requeue packet\n");
 430         }
 431 }
 432
 433 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
 434 {
 435         struct ipoib_dev_priv *priv = netdev_priv(dev);
 436         struct ipoib_path *path;
 437
 438         path = kzalloc(sizeof *path, GFP_ATOMIC);
 439         if (!path)
 440                 return NULL;
 441
 442         path->dev = dev;
 443
 444         skb_queue_head_init(&path->queue);
 445
 446         INIT_LIST_HEAD(&path->neigh_list);
 447
 448         memcpy(path->pathrec.dgid.raw, gid, sizeof (union ib_gid));
 449         path->pathrec.sgid      = priv->local_gid;
 450         path->pathrec.pkey      = cpu_to_be16(priv->pkey);
 451         path->pathrec.numb_path = 1;
 452
 453         return path;
 454 }
 455
 456 static int path_rec_start(struct net_device *dev,
 457                           struct ipoib_path *path)
 458 {
 459         struct ipoib_dev_priv *priv = netdev_priv(dev);
 460
 461         ipoib_dbg(priv, "Start path record lookup for " IPOIB_GID_FMT "\n",
 462                   IPOIB_GID_ARG(path->pathrec.dgid));
 463
 464         init_completion(&path->done);
 465
 466         path->query_id =
 467                 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
 468                                    &path->pathrec,
 469                                    IB_SA_PATH_REC_DGID          |
 470                                    IB_SA_PATH_REC_SGID          |
 471                                    IB_SA_PATH_REC_NUMB_PATH     |
 472                                    IB_SA_PATH_REC_PKEY,
 473                                    1000, GFP_ATOMIC,
 474                                    path_rec_completion,
 475                                    path, &path->query);
 476         if (path->query_id < 0) {
 477                 ipoib_warn(priv, "ib_sa_path_rec_get failed\n");
 478                 path->query = NULL;
 479                 return path->query_id;
 480         }
 481
 482         return 0;
 483 }
 484
 485 static void neigh_add_path(struct sk_buff *skb, struct net_device *dev)
 486 {
 487         struct ipoib_dev_priv *priv = netdev_priv(dev);
 488         struct ipoib_path *path;
 489         struct ipoib_neigh *neigh;
 490
 491         neigh = ipoib_neigh_alloc(skb->dst->neighbour);
 492         if (!neigh) {
 493                 ++priv->stats.tx_dropped;
 494                 dev_kfree_skb_any(skb);
 495                 return;
 496         }
 497
 498         skb_queue_head_init(&neigh->queue);
 499
 500         /*
 501          * We can only be called from ipoib_start_xmit, so we're
 502          * inside tx_lock -- no need to save/restore flags.
 503          */
 504         spin_lock(&priv->lock);
 505
 506         path = __path_find(dev, skb->dst->neighbour->ha + 4);
 507         if (!path) {
 508                 path = path_rec_create(dev, skb->dst->neighbour->ha + 4);
 509                 if (!path)
 510                         goto err_path;
 511
 512                 __path_add(dev, path);
 513         }
 514
 515         list_add_tail(&neigh->list, &path->neigh_list);
 516
 517         if (path->ah) {
 518                 kref_get(&path->ah->ref);
 519                 neigh->ah = path->ah;
 520                 memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
 521                        sizeof(union ib_gid));
 522
 523                 ipoib_send(dev, skb, path->ah,
 524                            be32_to_cpup((__be32 *) skb->dst->neighbour->ha));
 525         } else {
 526                 neigh->ah  = NULL;
 527                 __skb_queue_tail(&neigh->queue, skb);
 528
 529                 if (!path->query && path_rec_start(dev, path))
 530                         goto err_list;
 531         }
 532
 533         spin_unlock(&priv->lock);
 534         return;
 535
 536 err_list:
 537         list_del(&neigh->list);
 538
 539 err_path:
 540         ipoib_neigh_free(neigh);
 541         ++priv->stats.tx_dropped;
 542         dev_kfree_skb_any(skb);
 543
 544         spin_unlock(&priv->lock);
 545 }
 546
 547 static void ipoib_path_lookup(struct sk_buff *skb, struct net_device *dev)
 548 {
 549         struct ipoib_dev_priv *priv = netdev_priv(skb->dev);
 550
 551         /* Look up path record for unicasts */
 552         if (skb->dst->neighbour->ha[4] != 0xff) {
 553                 neigh_add_path(skb, dev);
 554                 return;
 555         }
 556
 557         /* Add in the P_Key for multicasts */
 558         skb->dst->neighbour->ha[8] = (priv->pkey >> 8) & 0xff;
 559         skb->dst->neighbour->ha[9] = priv->pkey & 0xff;
 560         ipoib_mcast_send(dev, skb->dst->neighbour->ha + 4, skb);
 561 }
 562
 563 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
 564                              struct ipoib_pseudoheader *phdr)
 565 {
 566         struct ipoib_dev_priv *priv = netdev_priv(dev);
 567         struct ipoib_path *path;
 568
 569         /*
 570          * We can only be called from ipoib_start_xmit, so we're
 571          * inside tx_lock -- no need to save/restore flags.
 572          */
 573         spin_lock(&priv->lock);
 574
 575         path = __path_find(dev, phdr->hwaddr + 4);
 576         if (!path) {
 577                 path = path_rec_create(dev, phdr->hwaddr + 4);
 578                 if (path) {
 579                         /* put pseudoheader back on for next time */
 580                         skb_push(skb, sizeof *phdr);
 581                         __skb_queue_tail(&path->queue, skb);
 582
 583                         if (path_rec_start(dev, path)) {
 584                                 spin_unlock(&priv->lock);
 585                                 path_free(dev, path);
 586                                 return;
 587                         } else
 588                                 __path_add(dev, path);
 589                 } else {
 590                         ++priv->stats.tx_dropped;
 591                         dev_kfree_skb_any(skb);
 592                 }
 593
 594                 spin_unlock(&priv->lock);
 595                 return;
 596         }
 597
 598         if (path->ah) {
 599                 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
 600                           be16_to_cpu(path->pathrec.dlid));
 601
 602                 ipoib_send(dev, skb, path->ah,
 603                            be32_to_cpup((__be32 *) phdr->hwaddr));
 604         } else if ((path->query || !path_rec_start(dev, path)) &&
 605                    skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 606                 /* put pseudoheader back on for next time */
 607                 skb_push(skb, sizeof *phdr);
 608                 __skb_queue_tail(&path->queue, skb);
 609         } else {
 610                 ++priv->stats.tx_dropped;
 611                 dev_kfree_skb_any(skb);
 612         }
 613
 614         spin_unlock(&priv->lock);
 615 }
 616
 617 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
 618 {
 619         struct ipoib_dev_priv *priv = netdev_priv(dev);
 620         struct ipoib_neigh *neigh;
 621         unsigned long flags;
 622
 623         if (unlikely(!spin_trylock_irqsave(&priv->tx_lock, flags)))
 624                 return NETDEV_TX_LOCKED;
 625
 626         /*
 627          * Check if our queue is stopped.  Since we have the LLTX bit
 628          * set, we can't rely on netif_stop_queue() preventing our
 629          * xmit function from being called with a full queue.
 630          */
 631         if (unlikely(netif_queue_stopped(dev))) {
 632                 spin_unlock_irqrestore(&priv->tx_lock, flags);
 633                 return NETDEV_TX_BUSY;
 634         }
 635
 636         if (likely(skb->dst && skb->dst->neighbour)) {
 637                 if (unlikely(!*to_ipoib_neigh(skb->dst->neighbour))) {
 638                         ipoib_path_lookup(skb, dev);
 639                         goto out;
 640                 }
 641
 642                 neigh = *to_ipoib_neigh(skb->dst->neighbour);
 643
 644                 if (likely(neigh->ah)) {
 645                         if (unlikely(memcmp(&neigh->dgid.raw,
 646                                             skb->dst->neighbour->ha + 4,
 647                                             sizeof(union ib_gid)))) {
 648                                 spin_lock(&priv->lock);
 649                                 /*
 650                                  * It's safe to call ipoib_put_ah() inside
 651                                  * priv->lock here, because we know that
 652                                  * path->ah will always hold one more reference,
 653                                  * so ipoib_put_ah() will never do more than
 654                                  * decrement the ref count.
 655                                  */
 656                                 ipoib_put_ah(neigh->ah);
 657                                 list_del(&neigh->list);
 658                                 ipoib_neigh_free(neigh);
 659                                 spin_unlock(&priv->lock);
 660                                 ipoib_path_lookup(skb, dev);
 661                                 goto out;
 662                         }
 663
 664                         ipoib_send(dev, skb, neigh->ah,
 665                                    be32_to_cpup((__be32 *) skb->dst->neighbour->ha));
 666                         goto out;
 667                 }
 668
 669                 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 670                         spin_lock(&priv->lock);
 671                         __skb_queue_tail(&neigh->queue, skb);
 672                         spin_unlock(&priv->lock);
 673                 } else {
 674                         ++priv->stats.tx_dropped;
 675                         dev_kfree_skb_any(skb);
 676                 }
 677         } else {
 678                 struct ipoib_pseudoheader *phdr =
 679                         (struct ipoib_pseudoheader *) skb->data;
 680                 skb_pull(skb, sizeof *phdr);
 681
 682                 if (phdr->hwaddr[4] == 0xff) {
 683                         /* Add in the P_Key for multicast*/
 684                         phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
 685                         phdr->hwaddr[9] = priv->pkey & 0xff;
 686
 687                         ipoib_mcast_send(dev, phdr->hwaddr + 4, skb);
 688                 } else {
 689                         /* unicast GID -- should be ARP or RARP reply */
 690
 691                         if ((be16_to_cpup((__be16 *) skb->data) != ETH_P_ARP) &&
 692                             (be16_to_cpup((__be16 *) skb->data) != ETH_P_RARP)) {
 693                                 ipoib_warn(priv, "Unicast, no %s: type %04x, QPN %06x "
 694                                            IPOIB_GID_FMT "\n",
 695                                            skb->dst ? "neigh" : "dst",
 696                                            be16_to_cpup((__be16 *) skb->data),
 697                                            be32_to_cpup((__be32 *) phdr->hwaddr),
 698                                            IPOIB_GID_RAW_ARG(phdr->hwaddr + 4));
 699                                 dev_kfree_skb_any(skb);
 700                                 ++priv->stats.tx_dropped;
 701                                 goto out;
 702                         }
 703
 704                         unicast_arp_send(skb, dev, phdr);
 705                 }
 706         }
 707
 708 out:
 709         spin_unlock_irqrestore(&priv->tx_lock, flags);
 710
 711         return NETDEV_TX_OK;
 712 }
 713
 714 static struct net_device_stats *ipoib_get_stats(struct net_device *dev)
 715 {
 716         struct ipoib_dev_priv *priv = netdev_priv(dev);
 717
 718         return &priv->stats;
 719 }
 720
 721 static void ipoib_timeout(struct net_device *dev)
 722 {
 723         struct ipoib_dev_priv *priv = netdev_priv(dev);
 724
 725         ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
 726                    jiffies_to_msecs(jiffies - dev->trans_start));
 727         ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
 728                    netif_queue_stopped(dev),
 729                    priv->tx_head, priv->tx_tail);
 730         /* XXX reset QP, etc. */
 731 }
 732
 733 static int ipoib_hard_header(struct sk_buff *skb,
 734                              struct net_device *dev,
 735                              unsigned short type,
 736                              void *daddr, void *saddr, unsigned len)
 737 {
 738         struct ipoib_header *header;
 739
 740         header = (struct ipoib_header *) skb_push(skb, sizeof *header);
 741
 742         header->proto = htons(type);
 743         header->reserved = 0;
 744
 745         /*
 746          * If we don't have a neighbour structure, stuff the
 747          * destination address onto the front of the skb so we can
 748          * figure out where to send the packet later.
 749          */
 750         if ((!skb->dst || !skb->dst->neighbour) && daddr) {
 751                 struct ipoib_pseudoheader *phdr =
 752                         (struct ipoib_pseudoheader *) skb_push(skb, sizeof *phdr);
 753                 memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
 754         }
 755
 756         return 0;
 757 }
 758
 759 static void ipoib_set_mcast_list(struct net_device *dev)
 760 {
 761         struct ipoib_dev_priv *priv = netdev_priv(dev);
 762
 763         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
 764                 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
 765                 return;
 766         }
 767
 768         queue_work(ipoib_workqueue, &priv->restart_task);
 769 }
 770
 771 static void ipoib_neigh_destructor(struct neighbour *n)
 772 {
 773         struct ipoib_neigh *neigh;
 774         struct ipoib_dev_priv *priv = netdev_priv(n->dev);
 775         unsigned long flags;
 776         struct ipoib_ah *ah = NULL;
 777
 778         ipoib_dbg(priv,
 779                   "neigh_destructor for %06x " IPOIB_GID_FMT "\n",
 780                   be32_to_cpup((__be32 *) n->ha),
 781                   IPOIB_GID_RAW_ARG(n->ha + 4));
 782
 783         spin_lock_irqsave(&priv->lock, flags);
 784
 785         neigh = *to_ipoib_neigh(n);
 786         if (neigh) {
 787                 if (neigh->ah)
 788                         ah = neigh->ah;
 789                 list_del(&neigh->list);
 790                 ipoib_neigh_free(neigh);
 791         }
 792
 793         spin_unlock_irqrestore(&priv->lock, flags);
 794
 795         if (ah)
 796                 ipoib_put_ah(ah);
 797 }
 798
 799 struct ipoib_neigh *ipoib_neigh_alloc(struct neighbour *neighbour)
 800 {
 801         struct ipoib_neigh *neigh;
 802
 803         neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
 804         if (!neigh)
 805                 return NULL;
 806
 807         neigh->neighbour = neighbour;
 808         *to_ipoib_neigh(neighbour) = neigh;
 809
 810         return neigh;
 811 }
 812
 813 void ipoib_neigh_free(struct ipoib_neigh *neigh)
 814 {
 815         *to_ipoib_neigh(neigh->neighbour) = NULL;
 816         kfree(neigh);
 817 }
 818
 819 static int ipoib_neigh_setup_dev(struct net_device *dev, struct neigh_parms *parms)
 820 {
 821         parms->neigh_destructor = ipoib_neigh_destructor;
 822
 823         return 0;
 824 }
 825
 826 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
 827 {
 828         struct ipoib_dev_priv *priv = netdev_priv(dev);
 829
 830         /* Allocate RX/TX "rings" to hold queued skbs */
 831         priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
 832                                 GFP_KERNEL);
 833         if (!priv->rx_ring) {
 834                 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
 835                        ca->name, ipoib_recvq_size);
 836                 goto out;
 837         }
 838
 839         priv->tx_ring = kzalloc(ipoib_sendq_size * sizeof *priv->tx_ring,
 840                                 GFP_KERNEL);
 841         if (!priv->tx_ring) {
 842                 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
 843                        ca->name, ipoib_sendq_size);
 844                 goto out_rx_ring_cleanup;
 845         }
 846
 847         /* priv->tx_head & tx_tail are already 0 */
 848
 849         if (ipoib_ib_dev_init(dev, ca, port))
 850                 goto out_tx_ring_cleanup;
 851
 852         return 0;
 853
 854 out_tx_ring_cleanup:
 855         kfree(priv->tx_ring);
 856
 857 out_rx_ring_cleanup:
 858         kfree(priv->rx_ring);
 859
 860 out:
 861         return -ENOMEM;
 862 }
 863
 864 void ipoib_dev_cleanup(struct net_device *dev)
 865 {
 866         struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
 867
 868         ipoib_delete_debug_files(dev);
 869
 870         /* Delete any child interfaces first */
 871         list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
 872                 unregister_netdev(cpriv->dev);
 873                 ipoib_dev_cleanup(cpriv->dev);
 874                 free_netdev(cpriv->dev);
 875         }
 876
 877         ipoib_ib_dev_cleanup(dev);
 878
 879         kfree(priv->rx_ring);
 880         kfree(priv->tx_ring);
 881
 882         priv->rx_ring = NULL;
 883         priv->tx_ring = NULL;
 884 }
 885
 886 static void ipoib_setup(struct net_device *dev)
 887 {
 888         struct ipoib_dev_priv *priv = netdev_priv(dev);
 889
 890         dev->open                = ipoib_open;
 891         dev->stop                = ipoib_stop;
 892         dev->change_mtu          = ipoib_change_mtu;
 893         dev->hard_start_xmit     = ipoib_start_xmit;
 894         dev->get_stats           = ipoib_get_stats;
 895         dev->tx_timeout          = ipoib_timeout;
 896         dev->hard_header         = ipoib_hard_header;
 897         dev->set_multicast_list  = ipoib_set_mcast_list;
 898         dev->neigh_setup         = ipoib_neigh_setup_dev;
 899
 900         dev->watchdog_timeo      = HZ;
 901
 902         dev->flags              |= IFF_BROADCAST | IFF_MULTICAST;
 903
 904         /*
 905          * We add in INFINIBAND_ALEN to allow for the destination
 906          * address "pseudoheader" for skbs without neighbour struct.
 907          */
 908         dev->hard_header_len     = IPOIB_ENCAP_LEN + INFINIBAND_ALEN;
 909         dev->addr_len            = INFINIBAND_ALEN;
 910         dev->type                = ARPHRD_INFINIBAND;
 911         dev->tx_queue_len        = ipoib_sendq_size * 2;
 912         dev->features            = NETIF_F_VLAN_CHALLENGED | NETIF_F_LLTX;
 913
 914         /* MTU will be reset when mcast join happens */
 915         dev->mtu                 = IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN;
 916         priv->mcast_mtu          = priv->admin_mtu = dev->mtu;
 917
 918         memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
 919
 920         netif_carrier_off(dev);
 921
 922         SET_MODULE_OWNER(dev);
 923
 924         priv->dev = dev;
 925
 926         spin_lock_init(&priv->lock);
 927         spin_lock_init(&priv->tx_lock);
 928
 929         mutex_init(&priv->mcast_mutex);
 930         mutex_init(&priv->vlan_mutex);
 931
 932         INIT_LIST_HEAD(&priv->path_list);
 933         INIT_LIST_HEAD(&priv->child_intfs);
 934         INIT_LIST_HEAD(&priv->dead_ahs);
 935         INIT_LIST_HEAD(&priv->multicast_list);
 936
 937         INIT_WORK(&priv->pkey_task,    ipoib_pkey_poll,          priv->dev);
 938         INIT_WORK(&priv->mcast_task,   ipoib_mcast_join_task,    priv->dev);
 939         INIT_WORK(&priv->flush_task,   ipoib_ib_dev_flush,       priv->dev);
 940         INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task, priv->dev);
 941         INIT_WORK(&priv->ah_reap_task, ipoib_reap_ah,            priv->dev);
 942 }
 943
 944 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
 945 {
 946         struct net_device *dev;
 947
 948         dev = alloc_netdev((int) sizeof (struct ipoib_dev_priv), name,
 949                            ipoib_setup);
 950         if (!dev)
 951                 return NULL;
 952
 953         return netdev_priv(dev);
 954 }
 955
 956 static ssize_t show_pkey(struct class_device *cdev, char *buf)
 957 {
 958         struct ipoib_dev_priv *priv =
 959                 netdev_priv(container_of(cdev, struct net_device, class_dev));
 960
 961         return sprintf(buf, "0x%04x\n", priv->pkey);
 962 }
 963 static CLASS_DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
 964
 965 static ssize_t create_child(struct class_device *cdev,
 966                             const char *buf, size_t count)
 967 {
 968         int pkey;
 969         int ret;
 970
 971         if (sscanf(buf, "%i", &pkey) != 1)
 972                 return -EINVAL;
 973
 974         if (pkey < 0 || pkey > 0xffff)
 975                 return -EINVAL;
 976
 977         /*
 978          * Set the full membership bit, so that we join the right
 979          * broadcast group, etc.
 980          */
 981         pkey |= 0x8000;
 982
 983         ret = ipoib_vlan_add(container_of(cdev, struct net_device, class_dev),
 984                              pkey);
 985
 986         return ret ? ret : count;
 987 }
 988 static CLASS_DEVICE_ATTR(create_child, S_IWUGO, NULL, create_child);
 989
 990 static ssize_t delete_child(struct class_device *cdev,
 991                             const char *buf, size_t count)
 992 {
 993         int pkey;
 994         int ret;
 995
 996         if (sscanf(buf, "%i", &pkey) != 1)
 997                 return -EINVAL;
 998
 999         if (pkey < 0 || pkey > 0xffff)
1000                 return -EINVAL;
1001
1002         ret = ipoib_vlan_delete(container_of(cdev, struct net_device, class_dev),
1003                                 pkey);
1004
1005         return ret ? ret : count;
1006
1007 }
1008 static CLASS_DEVICE_ATTR(delete_child, S_IWUGO, NULL, delete_child);
1009
1010 int ipoib_add_pkey_attr(struct net_device *dev)
1011 {
1012         return class_device_create_file(&dev->class_dev,
1013                                         &class_device_attr_pkey);
1014 }
1015
1016 static struct net_device *ipoib_add_port(const char *format,
1017                                          struct ib_device *hca, u8 port)
1018 {
1019         struct ipoib_dev_priv *priv;
1020         int result = -ENOMEM;
1021
1022         priv = ipoib_intf_alloc(format);
1023         if (!priv)
1024                 goto alloc_mem_failed;
1025
1026         SET_NETDEV_DEV(priv->dev, hca->dma_device);
1027
1028         result = ib_query_pkey(hca, port, 0, &priv->pkey);
1029         if (result) {
1030                 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
1031                        hca->name, port, result);
1032                 goto alloc_mem_failed;
1033         }
1034
1035         /*
1036          * Set the full membership bit, so that we join the right
1037          * broadcast group, etc.
1038          */
1039         priv->pkey |= 0x8000;
1040
1041         priv->dev->broadcast[8] = priv->pkey >> 8;
1042         priv->dev->broadcast[9] = priv->pkey & 0xff;
1043
1044         result = ib_query_gid(hca, port, 0, &priv->local_gid);
1045         if (result) {
1046                 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1047                        hca->name, port, result);
1048                 goto alloc_mem_failed;
1049         } else
1050                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1051
1052
1053         result = ipoib_dev_init(priv->dev, hca, port);
1054         if (result < 0) {
1055                 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1056                        hca->name, port, result);
1057                 goto device_init_failed;
1058         }
1059
1060         INIT_IB_EVENT_HANDLER(&priv->event_handler,
1061                               priv->ca, ipoib_event);
1062         result = ib_register_event_handler(&priv->event_handler);
1063         if (result < 0) {
1064                 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1065                        "port %d (ret = %d)\n",
1066                        hca->name, port, result);
1067                 goto event_failed;
1068         }
1069
1070         result = register_netdev(priv->dev);
1071         if (result) {
1072                 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1073                        hca->name, port, result);
1074                 goto register_failed;
1075         }
1076
1077         ipoib_create_debug_files(priv->dev);
1078
1079         if (ipoib_add_pkey_attr(priv->dev))
1080                 goto sysfs_failed;
1081         if (class_device_create_file(&priv->dev->class_dev,
1082                                      &class_device_attr_create_child))
1083                 goto sysfs_failed;
1084         if (class_device_create_file(&priv->dev->class_dev,
1085                                      &class_device_attr_delete_child))
1086                 goto sysfs_failed;
1087
1088         return priv->dev;
1089
1090 sysfs_failed:
1091         ipoib_delete_debug_files(priv->dev);
1092         unregister_netdev(priv->dev);
1093
1094 register_failed:
1095         ib_unregister_event_handler(&priv->event_handler);
1096         flush_scheduled_work();
1097
1098 event_failed:
1099         ipoib_dev_cleanup(priv->dev);
1100
1101 device_init_failed:
1102         free_netdev(priv->dev);
1103
1104 alloc_mem_failed:
1105         return ERR_PTR(result);
1106 }
1107
1108 static void ipoib_add_one(struct ib_device *device)
1109 {
1110         struct list_head *dev_list;
1111         struct net_device *dev;
1112         struct ipoib_dev_priv *priv;
1113         int s, e, p;
1114
1115         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1116                 return;
1117
1118         dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1119         if (!dev_list)
1120                 return;
1121
1122         INIT_LIST_HEAD(dev_list);
1123
1124         if (device->node_type == RDMA_NODE_IB_SWITCH) {
1125                 s = 0;
1126                 e = 0;
1127         } else {
1128                 s = 1;
1129                 e = device->phys_port_cnt;
1130         }
1131
1132         for (p = s; p <= e; ++p) {
1133                 dev = ipoib_add_port("ib%d", device, p);
1134                 if (!IS_ERR(dev)) {
1135                         priv = netdev_priv(dev);
1136                         list_add_tail(&priv->list, dev_list);
1137                 }
1138         }
1139
1140         ib_set_client_data(device, &ipoib_client, dev_list);
1141 }
1142
1143 static void ipoib_remove_one(struct ib_device *device)
1144 {
1145         struct ipoib_dev_priv *priv, *tmp;
1146         struct list_head *dev_list;
1147
1148         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1149                 return;
1150
1151         dev_list = ib_get_client_data(device, &ipoib_client);
1152
1153         list_for_each_entry_safe(priv, tmp, dev_list, list) {
1154                 ib_unregister_event_handler(&priv->event_handler);
1155                 flush_scheduled_work();
1156
1157                 unregister_netdev(priv->dev);
1158                 ipoib_dev_cleanup(priv->dev);
1159                 free_netdev(priv->dev);
1160         }
1161
1162         kfree(dev_list);
1163 }
1164
1165 static int __init ipoib_init_module(void)
1166 {
1167         int ret;
1168
1169         ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1170         ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1171         ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1172
1173         ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1174         ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1175         ipoib_sendq_size = max(ipoib_sendq_size, IPOIB_MIN_QUEUE_SIZE);
1176
1177         ret = ipoib_register_debugfs();
1178         if (ret)
1179                 return ret;
1180
1181         /*
1182          * We create our own workqueue mainly because we want to be
1183          * able to flush it when devices are being removed.  We can't
1184          * use schedule_work()/flush_scheduled_work() because both
1185          * unregister_netdev() and linkwatch_event take the rtnl lock,
1186          * so flush_scheduled_work() can deadlock during device
1187          * removal.
1188          */
1189         ipoib_workqueue = create_singlethread_workqueue("ipoib");
1190         if (!ipoib_workqueue) {
1191                 ret = -ENOMEM;
1192                 goto err_fs;
1193         }
1194
1195         ib_sa_register_client(&ipoib_sa_client);
1196
1197         ret = ib_register_client(&ipoib_client);
1198         if (ret)
1199                 goto err_sa;
1200
1201         return 0;
1202
1203 err_sa:
1204         ib_sa_unregister_client(&ipoib_sa_client);
1205         destroy_workqueue(ipoib_workqueue);
1206
1207 err_fs:
1208         ipoib_unregister_debugfs();
1209
1210         return ret;
1211 }
1212
1213 static void __exit ipoib_cleanup_module(void)
1214 {
1215         ib_unregister_client(&ipoib_client);
1216         ib_sa_unregister_client(&ipoib_sa_client);
1217         ipoib_unregister_debugfs();
1218         destroy_workqueue(ipoib_workqueue);
1219 }
1220
1221 module_init(ipoib_init_module);
1222 module_exit(ipoib_cleanup_module);