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