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