SAFE public projects git trees. - safe/jmp/linux-2.6/blob - drivers/infiniband/ulp/ipoib/ipoib_main.c

   1 /*
   2  * Copyright (c) 2004 Topspin Communications.  All rights reserved.
   3  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
   4  * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
   5  *
   6  * This software is available to you under a choice of one of two
   7  * licenses.  You may choose to be licensed under the terms of the GNU
   8  * General Public License (GPL) Version 2, available from the file
   9  * COPYING in the main directory of this source tree, or the
  10  * OpenIB.org BSD license below:
  11  *
  12  *     Redistribution and use in source and binary forms, with or
  13  *     without modification, are permitted provided that the following
  14  *     conditions are met:
  15  *
  16  *      - Redistributions of source code must retain the above
  17  *        copyright notice, this list of conditions and the following
  18  *        disclaimer.
  19  *
  20  *      - Redistributions in binary form must reproduce the above
  21  *        copyright notice, this list of conditions and the following
  22  *        disclaimer in the documentation and/or other materials
  23  *        provided with the distribution.
  24  *
  25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  32  * SOFTWARE.
  33  *
  34  * $Id: ipoib_main.c 1377 2004-12-23 19:57:12Z roland $
  35  */
  36
  37 #include "ipoib.h"
  38
  39 #include <linux/module.h>
  40
  41 #include <linux/init.h>
  42 #include <linux/slab.h>
  43 #include <linux/vmalloc.h>
  44 #include <linux/kernel.h>
  45
  46 #include <linux/if_arp.h>       /* For ARPHRD_xxx */
  47
  48 #include <linux/ip.h>
  49 #include <linux/in.h>
  50
  51 #include <net/dst.h>
  52
  53 MODULE_AUTHOR("Roland Dreier");
  54 MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
  55 MODULE_LICENSE("Dual BSD/GPL");
  56
  57 int ipoib_sendq_size __read_mostly = IPOIB_TX_RING_SIZE;
  58 int ipoib_recvq_size __read_mostly = IPOIB_RX_RING_SIZE;
  59
  60 module_param_named(send_queue_size, ipoib_sendq_size, int, 0444);
  61 MODULE_PARM_DESC(send_queue_size, "Number of descriptors in send queue");
  62 module_param_named(recv_queue_size, ipoib_recvq_size, int, 0444);
  63 MODULE_PARM_DESC(recv_queue_size, "Number of descriptors in receive queue");
  64
  65 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
  66 int ipoib_debug_level;
  67
  68 module_param_named(debug_level, ipoib_debug_level, int, 0644);
  69 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
  70 #endif
  71
  72 struct ipoib_path_iter {
  73         struct net_device *dev;
  74         struct ipoib_path  path;
  75 };
  76
  77 static const u8 ipv4_bcast_addr[] = {
  78         0x00, 0xff, 0xff, 0xff,
  79         0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
  80         0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff
  81 };
  82
  83 struct workqueue_struct *ipoib_workqueue;
  84
  85 static void ipoib_add_one(struct ib_device *device);
  86 static void ipoib_remove_one(struct ib_device *device);
  87
  88 static struct ib_client ipoib_client = {
  89         .name   = "ipoib",
  90         .add    = ipoib_add_one,
  91         .remove = ipoib_remove_one
  92 };
  93
  94 int ipoib_open(struct net_device *dev)
  95 {
  96         struct ipoib_dev_priv *priv = netdev_priv(dev);
  97
  98         ipoib_dbg(priv, "bringing up interface\n");
  99
 100         set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 101
 102         if (ipoib_pkey_dev_delay_open(dev))
 103                 return 0;
 104
 105         if (ipoib_ib_dev_open(dev))
 106                 return -EINVAL;
 107
 108         if (ipoib_ib_dev_up(dev)) {
 109                 ipoib_ib_dev_stop(dev);
 110                 return -EINVAL;
 111         }
 112
 113         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 114                 struct ipoib_dev_priv *cpriv;
 115
 116                 /* Bring up any child interfaces too */
 117                 mutex_lock(&priv->vlan_mutex);
 118                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 119                         int flags;
 120
 121                         flags = cpriv->dev->flags;
 122                         if (flags & IFF_UP)
 123                                 continue;
 124
 125                         dev_change_flags(cpriv->dev, flags | IFF_UP);
 126                 }
 127                 mutex_unlock(&priv->vlan_mutex);
 128         }
 129
 130         netif_start_queue(dev);
 131
 132         return 0;
 133 }
 134
 135 static int ipoib_stop(struct net_device *dev)
 136 {
 137         struct ipoib_dev_priv *priv = netdev_priv(dev);
 138
 139         ipoib_dbg(priv, "stopping interface\n");
 140
 141         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 142
 143         netif_stop_queue(dev);
 144
 145         /*
 146          * Now flush workqueue to make sure a scheduled task doesn't
 147          * bring our internal state back up.
 148          */
 149         flush_workqueue(ipoib_workqueue);
 150
 151         ipoib_ib_dev_down(dev, 1);
 152         ipoib_ib_dev_stop(dev);
 153
 154         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 155                 struct ipoib_dev_priv *cpriv;
 156
 157                 /* Bring down any child interfaces too */
 158                 mutex_lock(&priv->vlan_mutex);
 159                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 160                         int flags;
 161
 162                         flags = cpriv->dev->flags;
 163                         if (!(flags & IFF_UP))
 164                                 continue;
 165
 166                         dev_change_flags(cpriv->dev, flags & ~IFF_UP);
 167                 }
 168                 mutex_unlock(&priv->vlan_mutex);
 169         }
 170
 171         return 0;
 172 }
 173
 174 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
 175 {
 176         struct ipoib_dev_priv *priv = netdev_priv(dev);
 177
 178         if (new_mtu > IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN)
 179                 return -EINVAL;
 180
 181         priv->admin_mtu = new_mtu;
 182
 183         dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
 184
 185         return 0;
 186 }
 187
 188 static struct ipoib_path *__path_find(struct net_device *dev,
 189                                       union ib_gid *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->raw, 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         unsigned long flags;
 340
 341         spin_lock_irqsave(&priv->lock, flags);
 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_irqrestore(&priv->lock, flags);
 353                 wait_for_completion(&path->done);
 354                 path_free(dev, path);
 355                 spin_lock_irqsave(&priv->lock, flags);
 356         }
 357         spin_unlock_irqrestore(&priv->lock, flags);
 358 }
 359
 360 static void path_rec_completion(int status,
 361                                 struct ib_sa_path_rec *pathrec,
 362                                 void *path_ptr)
 363 {
 364         struct ipoib_path *path = path_ptr;
 365         struct net_device *dev = path->dev;
 366         struct ipoib_dev_priv *priv = netdev_priv(dev);
 367         struct ipoib_ah *ah = NULL;
 368         struct ipoib_neigh *neigh;
 369         struct sk_buff_head skqueue;
 370         struct sk_buff *skb;
 371         unsigned long flags;
 372
 373         if (pathrec)
 374                 ipoib_dbg(priv, "PathRec LID 0x%04x for GID " IPOIB_GID_FMT "\n",
 375                           be16_to_cpu(pathrec->dlid), IPOIB_GID_ARG(pathrec->dgid));
 376         else
 377                 ipoib_dbg(priv, "PathRec status %d for GID " IPOIB_GID_FMT "\n",
 378                           status, IPOIB_GID_ARG(path->pathrec.dgid));
 379
 380         skb_queue_head_init(&skqueue);
 381
 382         if (!status) {
 383                 struct ib_ah_attr av = {
 384                         .dlid          = be16_to_cpu(pathrec->dlid),
 385                         .sl            = pathrec->sl,
 386                         .port_num      = priv->port,
 387                         .static_rate   = pathrec->rate
 388                 };
 389
 390                 ah = ipoib_create_ah(dev, priv->pd, &av);
 391         }
 392
 393         spin_lock_irqsave(&priv->lock, flags);
 394
 395         path->ah = ah;
 396
 397         if (ah) {
 398                 path->pathrec = *pathrec;
 399
 400                 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
 401                           ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
 402
 403                 while ((skb = __skb_dequeue(&path->queue)))
 404                         __skb_queue_tail(&skqueue, skb);
 405
 406                 list_for_each_entry(neigh, &path->neigh_list, list) {
 407                         kref_get(&path->ah->ref);
 408                         neigh->ah = path->ah;
 409
 410                         while ((skb = __skb_dequeue(&neigh->queue)))
 411                                 __skb_queue_tail(&skqueue, skb);
 412                 }
 413         }
 414
 415         path->query = NULL;
 416         complete(&path->done);
 417
 418         spin_unlock_irqrestore(&priv->lock, flags);
 419
 420         while ((skb = __skb_dequeue(&skqueue))) {
 421                 skb->dev = dev;
 422                 if (dev_queue_xmit(skb))
 423                         ipoib_warn(priv, "dev_queue_xmit failed "
 424                                    "to requeue packet\n");
 425         }
 426 }
 427
 428 static struct ipoib_path *path_rec_create(struct net_device *dev,
 429                                           union ib_gid *gid)
 430 {
 431         struct ipoib_dev_priv *priv = netdev_priv(dev);
 432         struct ipoib_path *path;
 433
 434         path = kzalloc(sizeof *path, GFP_ATOMIC);
 435         if (!path)
 436                 return NULL;
 437
 438         path->dev = dev;
 439
 440         skb_queue_head_init(&path->queue);
 441
 442         INIT_LIST_HEAD(&path->neigh_list);
 443
 444         memcpy(path->pathrec.dgid.raw, gid->raw, sizeof (union ib_gid));
 445         path->pathrec.sgid      = priv->local_gid;
 446         path->pathrec.pkey      = cpu_to_be16(priv->pkey);
 447         path->pathrec.numb_path = 1;
 448
 449         return path;
 450 }
 451
 452 static int path_rec_start(struct net_device *dev,
 453                           struct ipoib_path *path)
 454 {
 455         struct ipoib_dev_priv *priv = netdev_priv(dev);
 456
 457         ipoib_dbg(priv, "Start path record lookup for " IPOIB_GID_FMT "\n",
 458                   IPOIB_GID_ARG(path->pathrec.dgid));
 459
 460         init_completion(&path->done);
 461
 462         path->query_id =
 463                 ib_sa_path_rec_get(priv->ca, priv->port,
 464                                    &path->pathrec,
 465                                    IB_SA_PATH_REC_DGID          |
 466                                    IB_SA_PATH_REC_SGID          |
 467                                    IB_SA_PATH_REC_NUMB_PATH     |
 468                                    IB_SA_PATH_REC_PKEY,
 469                                    1000, GFP_ATOMIC,
 470                                    path_rec_completion,
 471                                    path, &path->query);
 472         if (path->query_id < 0) {
 473                 ipoib_warn(priv, "ib_sa_path_rec_get failed\n");
 474                 path->query = NULL;
 475                 return path->query_id;
 476         }
 477
 478         return 0;
 479 }
 480
 481 static void neigh_add_path(struct sk_buff *skb, struct net_device *dev)
 482 {
 483         struct ipoib_dev_priv *priv = netdev_priv(dev);
 484         struct ipoib_path *path;
 485         struct ipoib_neigh *neigh;
 486
 487         neigh = ipoib_neigh_alloc(skb->dst->neighbour);
 488         if (!neigh) {
 489                 ++priv->stats.tx_dropped;
 490                 dev_kfree_skb_any(skb);
 491                 return;
 492         }
 493
 494         skb_queue_head_init(&neigh->queue);
 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_path;
 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_list;
 526         }
 527
 528         spin_unlock(&priv->lock);
 529         return;
 530
 531 err_list:
 532         list_del(&neigh->list);
 533
 534 err_path:
 535         ipoib_neigh_free(neigh);
 536         ++priv->stats.tx_dropped;
 537         dev_kfree_skb_any(skb);
 538
 539         spin_unlock(&priv->lock);
 540 }
 541
 542 static void ipoib_path_lookup(struct sk_buff *skb, struct net_device *dev)
 543 {
 544         struct ipoib_dev_priv *priv = netdev_priv(skb->dev);
 545
 546         /* Look up path record for unicasts */
 547         if (skb->dst->neighbour->ha[4] != 0xff) {
 548                 neigh_add_path(skb, dev);
 549                 return;
 550         }
 551
 552         /* Add in the P_Key for multicasts */
 553         skb->dst->neighbour->ha[8] = (priv->pkey >> 8) & 0xff;
 554         skb->dst->neighbour->ha[9] = priv->pkey & 0xff;
 555         ipoib_mcast_send(dev, (union ib_gid *) (skb->dst->neighbour->ha + 4), skb);
 556 }
 557
 558 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
 559                              struct ipoib_pseudoheader *phdr)
 560 {
 561         struct ipoib_dev_priv *priv = netdev_priv(dev);
 562         struct ipoib_path *path;
 563
 564         /*
 565          * We can only be called from ipoib_start_xmit, so we're
 566          * inside tx_lock -- no need to save/restore flags.
 567          */
 568         spin_lock(&priv->lock);
 569
 570         path = __path_find(dev, (union ib_gid *) (phdr->hwaddr + 4));
 571         if (!path) {
 572                 path = path_rec_create(dev,
 573                                        (union ib_gid *) (phdr->hwaddr + 4));
 574                 if (path) {
 575                         /* put pseudoheader back on for next time */
 576                         skb_push(skb, sizeof *phdr);
 577                         __skb_queue_tail(&path->queue, skb);
 578
 579                         if (path_rec_start(dev, path)) {
 580                                 spin_unlock(&priv->lock);
 581                                 path_free(dev, path);
 582                                 return;
 583                         } else
 584                                 __path_add(dev, path);
 585                 } else {
 586                         ++priv->stats.tx_dropped;
 587                         dev_kfree_skb_any(skb);
 588                 }
 589
 590                 spin_unlock(&priv->lock);
 591                 return;
 592         }
 593
 594         if (path->ah) {
 595                 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
 596                           be16_to_cpu(path->pathrec.dlid));
 597
 598                 ipoib_send(dev, skb, path->ah,
 599                            be32_to_cpup((__be32 *) phdr->hwaddr));
 600         } else if ((path->query || !path_rec_start(dev, path)) &&
 601                    skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 602                 /* put pseudoheader back on for next time */
 603                 skb_push(skb, sizeof *phdr);
 604                 __skb_queue_tail(&path->queue, skb);
 605         } else {
 606                 ++priv->stats.tx_dropped;
 607                 dev_kfree_skb_any(skb);
 608         }
 609
 610         spin_unlock(&priv->lock);
 611 }
 612
 613 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
 614 {
 615         struct ipoib_dev_priv *priv = netdev_priv(dev);
 616         struct ipoib_neigh *neigh;
 617         unsigned long flags;
 618
 619         if (!spin_trylock_irqsave(&priv->tx_lock, flags))
 620                 return NETDEV_TX_LOCKED;
 621
 622         /*
 623          * Check if our queue is stopped.  Since we have the LLTX bit
 624          * set, we can't rely on netif_stop_queue() preventing our
 625          * xmit function from being called with a full queue.
 626          */
 627         if (unlikely(netif_queue_stopped(dev))) {
 628                 spin_unlock_irqrestore(&priv->tx_lock, flags);
 629                 return NETDEV_TX_BUSY;
 630         }
 631
 632         if (skb->dst && skb->dst->neighbour) {
 633                 if (unlikely(!*to_ipoib_neigh(skb->dst->neighbour))) {
 634                         ipoib_path_lookup(skb, dev);
 635                         goto out;
 636                 }
 637
 638                 neigh = *to_ipoib_neigh(skb->dst->neighbour);
 639
 640                 if (likely(neigh->ah)) {
 641                         ipoib_send(dev, skb, neigh->ah,
 642                                    be32_to_cpup((__be32 *) skb->dst->neighbour->ha));
 643                         goto out;
 644                 }
 645
 646                 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 647                         spin_lock(&priv->lock);
 648                         __skb_queue_tail(&neigh->queue, skb);
 649                         spin_unlock(&priv->lock);
 650                 } else {
 651                         ++priv->stats.tx_dropped;
 652                         dev_kfree_skb_any(skb);
 653                 }
 654         } else {
 655                 struct ipoib_pseudoheader *phdr =
 656                         (struct ipoib_pseudoheader *) skb->data;
 657                 skb_pull(skb, sizeof *phdr);
 658
 659                 if (phdr->hwaddr[4] == 0xff) {
 660                         /* Add in the P_Key for multicast*/
 661                         phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
 662                         phdr->hwaddr[9] = priv->pkey & 0xff;
 663
 664                         ipoib_mcast_send(dev, (union ib_gid *) (phdr->hwaddr + 4), skb);
 665                 } else {
 666                         /* unicast GID -- should be ARP or RARP reply */
 667
 668                         if ((be16_to_cpup((__be16 *) skb->data) != ETH_P_ARP) &&
 669                             (be16_to_cpup((__be16 *) skb->data) != ETH_P_RARP)) {
 670                                 ipoib_warn(priv, "Unicast, no %s: type %04x, QPN %06x "
 671                                            IPOIB_GID_FMT "\n",
 672                                            skb->dst ? "neigh" : "dst",
 673                                            be16_to_cpup((__be16 *) skb->data),
 674                                            be32_to_cpup((__be32 *) phdr->hwaddr),
 675                                            IPOIB_GID_ARG(*(union ib_gid *) (phdr->hwaddr + 4)));
 676                                 dev_kfree_skb_any(skb);
 677                                 ++priv->stats.tx_dropped;
 678                                 goto out;
 679                         }
 680
 681                         unicast_arp_send(skb, dev, phdr);
 682                 }
 683         }
 684
 685 out:
 686         spin_unlock_irqrestore(&priv->tx_lock, flags);
 687
 688         return NETDEV_TX_OK;
 689 }
 690
 691 static struct net_device_stats *ipoib_get_stats(struct net_device *dev)
 692 {
 693         struct ipoib_dev_priv *priv = netdev_priv(dev);
 694
 695         return &priv->stats;
 696 }
 697
 698 static void ipoib_timeout(struct net_device *dev)
 699 {
 700         struct ipoib_dev_priv *priv = netdev_priv(dev);
 701
 702         ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
 703                    jiffies_to_msecs(jiffies - dev->trans_start));
 704         ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
 705                    netif_queue_stopped(dev),
 706                    priv->tx_head, priv->tx_tail);
 707         /* XXX reset QP, etc. */
 708 }
 709
 710 static int ipoib_hard_header(struct sk_buff *skb,
 711                              struct net_device *dev,
 712                              unsigned short type,
 713                              void *daddr, void *saddr, unsigned len)
 714 {
 715         struct ipoib_header *header;
 716
 717         header = (struct ipoib_header *) skb_push(skb, sizeof *header);
 718
 719         header->proto = htons(type);
 720         header->reserved = 0;
 721
 722         /*
 723          * If we don't have a neighbour structure, stuff the
 724          * destination address onto the front of the skb so we can
 725          * figure out where to send the packet later.
 726          */
 727         if ((!skb->dst || !skb->dst->neighbour) && daddr) {
 728                 struct ipoib_pseudoheader *phdr =
 729                         (struct ipoib_pseudoheader *) skb_push(skb, sizeof *phdr);
 730                 memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
 731         }
 732
 733         return 0;
 734 }
 735
 736 static void ipoib_set_mcast_list(struct net_device *dev)
 737 {
 738         struct ipoib_dev_priv *priv = netdev_priv(dev);
 739
 740         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
 741                 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
 742                 return;
 743         }
 744
 745         queue_work(ipoib_workqueue, &priv->restart_task);
 746 }
 747
 748 static void ipoib_neigh_destructor(struct neighbour *n)
 749 {
 750         struct ipoib_neigh *neigh;
 751         struct ipoib_dev_priv *priv = netdev_priv(n->dev);
 752         unsigned long flags;
 753         struct ipoib_ah *ah = NULL;
 754
 755         ipoib_dbg(priv,
 756                   "neigh_destructor for %06x " IPOIB_GID_FMT "\n",
 757                   be32_to_cpup((__be32 *) n->ha),
 758                   IPOIB_GID_ARG(*((union ib_gid *) (n->ha + 4))));
 759
 760         spin_lock_irqsave(&priv->lock, flags);
 761
 762         neigh = *to_ipoib_neigh(n);
 763         if (neigh) {
 764                 if (neigh->ah)
 765                         ah = neigh->ah;
 766                 list_del(&neigh->list);
 767                 ipoib_neigh_free(neigh);
 768         }
 769
 770         spin_unlock_irqrestore(&priv->lock, flags);
 771
 772         if (ah)
 773                 ipoib_put_ah(ah);
 774 }
 775
 776 struct ipoib_neigh *ipoib_neigh_alloc(struct neighbour *neighbour)
 777 {
 778         struct ipoib_neigh *neigh;
 779
 780         neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
 781         if (!neigh)
 782                 return NULL;
 783
 784         neigh->neighbour = neighbour;
 785         *to_ipoib_neigh(neighbour) = neigh;
 786
 787         return neigh;
 788 }
 789
 790 void ipoib_neigh_free(struct ipoib_neigh *neigh)
 791 {
 792         *to_ipoib_neigh(neigh->neighbour) = NULL;
 793         kfree(neigh);
 794 }
 795
 796 static int ipoib_neigh_setup_dev(struct net_device *dev, struct neigh_parms *parms)
 797 {
 798         parms->neigh_destructor = ipoib_neigh_destructor;
 799
 800         return 0;
 801 }
 802
 803 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
 804 {
 805         struct ipoib_dev_priv *priv = netdev_priv(dev);
 806
 807         /* Allocate RX/TX "rings" to hold queued skbs */
 808         priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
 809                                 GFP_KERNEL);
 810         if (!priv->rx_ring) {
 811                 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
 812                        ca->name, ipoib_recvq_size);
 813                 goto out;
 814         }
 815
 816         priv->tx_ring = kzalloc(ipoib_sendq_size * sizeof *priv->tx_ring,
 817                                 GFP_KERNEL);
 818         if (!priv->tx_ring) {
 819                 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
 820                        ca->name, ipoib_sendq_size);
 821                 goto out_rx_ring_cleanup;
 822         }
 823
 824         /* priv->tx_head & tx_tail are already 0 */
 825
 826         if (ipoib_ib_dev_init(dev, ca, port))
 827                 goto out_tx_ring_cleanup;
 828
 829         return 0;
 830
 831 out_tx_ring_cleanup:
 832         kfree(priv->tx_ring);
 833
 834 out_rx_ring_cleanup:
 835         kfree(priv->rx_ring);
 836
 837 out:
 838         return -ENOMEM;
 839 }
 840
 841 void ipoib_dev_cleanup(struct net_device *dev)
 842 {
 843         struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
 844
 845         ipoib_delete_debug_files(dev);
 846
 847         /* Delete any child interfaces first */
 848         list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
 849                 unregister_netdev(cpriv->dev);
 850                 ipoib_dev_cleanup(cpriv->dev);
 851                 free_netdev(cpriv->dev);
 852         }
 853
 854         ipoib_ib_dev_cleanup(dev);
 855
 856         kfree(priv->rx_ring);
 857         kfree(priv->tx_ring);
 858
 859         priv->rx_ring = NULL;
 860         priv->tx_ring = NULL;
 861 }
 862
 863 static void ipoib_setup(struct net_device *dev)
 864 {
 865         struct ipoib_dev_priv *priv = netdev_priv(dev);
 866
 867         dev->open                = ipoib_open;
 868         dev->stop                = ipoib_stop;
 869         dev->change_mtu          = ipoib_change_mtu;
 870         dev->hard_start_xmit     = ipoib_start_xmit;
 871         dev->get_stats           = ipoib_get_stats;
 872         dev->tx_timeout          = ipoib_timeout;
 873         dev->hard_header         = ipoib_hard_header;
 874         dev->set_multicast_list  = ipoib_set_mcast_list;
 875         dev->neigh_setup         = ipoib_neigh_setup_dev;
 876
 877         dev->watchdog_timeo      = HZ;
 878
 879         dev->flags              |= IFF_BROADCAST | IFF_MULTICAST;
 880
 881         /*
 882          * We add in INFINIBAND_ALEN to allow for the destination
 883          * address "pseudoheader" for skbs without neighbour struct.
 884          */
 885         dev->hard_header_len     = IPOIB_ENCAP_LEN + INFINIBAND_ALEN;
 886         dev->addr_len            = INFINIBAND_ALEN;
 887         dev->type                = ARPHRD_INFINIBAND;
 888         dev->tx_queue_len        = ipoib_sendq_size * 2;
 889         dev->features            = NETIF_F_VLAN_CHALLENGED | NETIF_F_LLTX;
 890
 891         /* MTU will be reset when mcast join happens */
 892         dev->mtu                 = IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN;
 893         priv->mcast_mtu          = priv->admin_mtu = dev->mtu;
 894
 895         memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
 896
 897         netif_carrier_off(dev);
 898
 899         SET_MODULE_OWNER(dev);
 900
 901         priv->dev = dev;
 902
 903         spin_lock_init(&priv->lock);
 904         spin_lock_init(&priv->tx_lock);
 905
 906         mutex_init(&priv->mcast_mutex);
 907         mutex_init(&priv->vlan_mutex);
 908
 909         INIT_LIST_HEAD(&priv->path_list);
 910         INIT_LIST_HEAD(&priv->child_intfs);
 911         INIT_LIST_HEAD(&priv->dead_ahs);
 912         INIT_LIST_HEAD(&priv->multicast_list);
 913
 914         INIT_WORK(&priv->pkey_task,    ipoib_pkey_poll,          priv->dev);
 915         INIT_WORK(&priv->mcast_task,   ipoib_mcast_join_task,    priv->dev);
 916         INIT_WORK(&priv->flush_task,   ipoib_ib_dev_flush,       priv->dev);
 917         INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task, priv->dev);
 918         INIT_WORK(&priv->ah_reap_task, ipoib_reap_ah,            priv->dev);
 919 }
 920
 921 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
 922 {
 923         struct net_device *dev;
 924
 925         dev = alloc_netdev((int) sizeof (struct ipoib_dev_priv), name,
 926                            ipoib_setup);
 927         if (!dev)
 928                 return NULL;
 929
 930         return netdev_priv(dev);
 931 }
 932
 933 static ssize_t show_pkey(struct class_device *cdev, char *buf)
 934 {
 935         struct ipoib_dev_priv *priv =
 936                 netdev_priv(container_of(cdev, struct net_device, class_dev));
 937
 938         return sprintf(buf, "0x%04x\n", priv->pkey);
 939 }
 940 static CLASS_DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
 941
 942 static ssize_t create_child(struct class_device *cdev,
 943                             const char *buf, size_t count)
 944 {
 945         int pkey;
 946         int ret;
 947
 948         if (sscanf(buf, "%i", &pkey) != 1)
 949                 return -EINVAL;
 950
 951         if (pkey < 0 || pkey > 0xffff)
 952                 return -EINVAL;
 953
 954         /*
 955          * Set the full membership bit, so that we join the right
 956          * broadcast group, etc.
 957          */
 958         pkey |= 0x8000;
 959
 960         ret = ipoib_vlan_add(container_of(cdev, struct net_device, class_dev),
 961                              pkey);
 962
 963         return ret ? ret : count;
 964 }
 965 static CLASS_DEVICE_ATTR(create_child, S_IWUGO, NULL, create_child);
 966
 967 static ssize_t delete_child(struct class_device *cdev,
 968                             const char *buf, size_t count)
 969 {
 970         int pkey;
 971         int ret;
 972
 973         if (sscanf(buf, "%i", &pkey) != 1)
 974                 return -EINVAL;
 975
 976         if (pkey < 0 || pkey > 0xffff)
 977                 return -EINVAL;
 978
 979         ret = ipoib_vlan_delete(container_of(cdev, struct net_device, class_dev),
 980                                 pkey);
 981
 982         return ret ? ret : count;
 983
 984 }
 985 static CLASS_DEVICE_ATTR(delete_child, S_IWUGO, NULL, delete_child);
 986
 987 int ipoib_add_pkey_attr(struct net_device *dev)
 988 {
 989         return class_device_create_file(&dev->class_dev,
 990                                         &class_device_attr_pkey);
 991 }
 992
 993 static struct net_device *ipoib_add_port(const char *format,
 994                                          struct ib_device *hca, u8 port)
 995 {
 996         struct ipoib_dev_priv *priv;
 997         int result = -ENOMEM;
 998
 999         priv = ipoib_intf_alloc(format);
1000         if (!priv)
1001                 goto alloc_mem_failed;
1002
1003         SET_NETDEV_DEV(priv->dev, hca->dma_device);
1004
1005         result = ib_query_pkey(hca, port, 0, &priv->pkey);
1006         if (result) {
1007                 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
1008                        hca->name, port, result);
1009                 goto alloc_mem_failed;
1010         }
1011
1012         /*
1013          * Set the full membership bit, so that we join the right
1014          * broadcast group, etc.
1015          */
1016         priv->pkey |= 0x8000;
1017
1018         priv->dev->broadcast[8] = priv->pkey >> 8;
1019         priv->dev->broadcast[9] = priv->pkey & 0xff;
1020
1021         result = ib_query_gid(hca, port, 0, &priv->local_gid);
1022         if (result) {
1023                 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1024                        hca->name, port, result);
1025                 goto alloc_mem_failed;
1026         } else
1027                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1028
1029
1030         result = ipoib_dev_init(priv->dev, hca, port);
1031         if (result < 0) {
1032                 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1033                        hca->name, port, result);
1034                 goto device_init_failed;
1035         }
1036
1037         INIT_IB_EVENT_HANDLER(&priv->event_handler,
1038                               priv->ca, ipoib_event);
1039         result = ib_register_event_handler(&priv->event_handler);
1040         if (result < 0) {
1041                 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1042                        "port %d (ret = %d)\n",
1043                        hca->name, port, result);
1044                 goto event_failed;
1045         }
1046
1047         result = register_netdev(priv->dev);
1048         if (result) {
1049                 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1050                        hca->name, port, result);
1051                 goto register_failed;
1052         }
1053
1054         ipoib_create_debug_files(priv->dev);
1055
1056         if (ipoib_add_pkey_attr(priv->dev))
1057                 goto sysfs_failed;
1058         if (class_device_create_file(&priv->dev->class_dev,
1059                                      &class_device_attr_create_child))
1060                 goto sysfs_failed;
1061         if (class_device_create_file(&priv->dev->class_dev,
1062                                      &class_device_attr_delete_child))
1063                 goto sysfs_failed;
1064
1065         return priv->dev;
1066
1067 sysfs_failed:
1068         ipoib_delete_debug_files(priv->dev);
1069         unregister_netdev(priv->dev);
1070
1071 register_failed:
1072         ib_unregister_event_handler(&priv->event_handler);
1073         flush_scheduled_work();
1074
1075 event_failed:
1076         ipoib_dev_cleanup(priv->dev);
1077
1078 device_init_failed:
1079         free_netdev(priv->dev);
1080
1081 alloc_mem_failed:
1082         return ERR_PTR(result);
1083 }
1084
1085 static void ipoib_add_one(struct ib_device *device)
1086 {
1087         struct list_head *dev_list;
1088         struct net_device *dev;
1089         struct ipoib_dev_priv *priv;
1090         int s, e, p;
1091
1092         dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1093         if (!dev_list)
1094                 return;
1095
1096         INIT_LIST_HEAD(dev_list);
1097
1098         if (device->node_type == IB_NODE_SWITCH) {
1099                 s = 0;
1100                 e = 0;
1101         } else {
1102                 s = 1;
1103                 e = device->phys_port_cnt;
1104         }
1105
1106         for (p = s; p <= e; ++p) {
1107                 dev = ipoib_add_port("ib%d", device, p);
1108                 if (!IS_ERR(dev)) {
1109                         priv = netdev_priv(dev);
1110                         list_add_tail(&priv->list, dev_list);
1111                 }
1112         }
1113
1114         ib_set_client_data(device, &ipoib_client, dev_list);
1115 }
1116
1117 static void ipoib_remove_one(struct ib_device *device)
1118 {
1119         struct ipoib_dev_priv *priv, *tmp;
1120         struct list_head *dev_list;
1121
1122         dev_list = ib_get_client_data(device, &ipoib_client);
1123
1124         list_for_each_entry_safe(priv, tmp, dev_list, list) {
1125                 ib_unregister_event_handler(&priv->event_handler);
1126                 flush_scheduled_work();
1127
1128                 unregister_netdev(priv->dev);
1129                 ipoib_dev_cleanup(priv->dev);
1130                 free_netdev(priv->dev);
1131         }
1132
1133         kfree(dev_list);
1134 }
1135
1136 static int __init ipoib_init_module(void)
1137 {
1138         int ret;
1139
1140         ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1141         ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1142         ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1143
1144         ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1145         ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1146         ipoib_sendq_size = max(ipoib_sendq_size, IPOIB_MIN_QUEUE_SIZE);
1147
1148         ret = ipoib_register_debugfs();
1149         if (ret)
1150                 return ret;
1151
1152         /*
1153          * We create our own workqueue mainly because we want to be
1154          * able to flush it when devices are being removed.  We can't
1155          * use schedule_work()/flush_scheduled_work() because both
1156          * unregister_netdev() and linkwatch_event take the rtnl lock,
1157          * so flush_scheduled_work() can deadlock during device
1158          * removal.
1159          */
1160         ipoib_workqueue = create_singlethread_workqueue("ipoib");
1161         if (!ipoib_workqueue) {
1162                 ret = -ENOMEM;
1163                 goto err_fs;
1164         }
1165
1166         ret = ib_register_client(&ipoib_client);
1167         if (ret)
1168                 goto err_wq;
1169
1170         return 0;
1171
1172 err_wq:
1173         destroy_workqueue(ipoib_workqueue);
1174
1175 err_fs:
1176         ipoib_unregister_debugfs();
1177
1178         return ret;
1179 }
1180
1181 static void __exit ipoib_cleanup_module(void)
1182 {
1183         ib_unregister_client(&ipoib_client);
1184         ipoib_unregister_debugfs();
1185         destroy_workqueue(ipoib_workqueue);
1186 }
1187
1188 module_init(ipoib_init_module);
1189 module_exit(ipoib_cleanup_module);