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 struct ib_sa_client ipoib_sa_client;
  86
  87 static void ipoib_add_one(struct ib_device *device);
  88 static void ipoib_remove_one(struct ib_device *device);
  89
  90 static struct ib_client ipoib_client = {
  91         .name   = "ipoib",
  92         .add    = ipoib_add_one,
  93         .remove = ipoib_remove_one
  94 };
  95
  96 int ipoib_open(struct net_device *dev)
  97 {
  98         struct ipoib_dev_priv *priv = netdev_priv(dev);
  99
 100         ipoib_dbg(priv, "bringing up interface\n");
 101
 102         set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 103
 104         if (ipoib_pkey_dev_delay_open(dev))
 105                 return 0;
 106
 107         if (ipoib_ib_dev_open(dev))
 108                 return -EINVAL;
 109
 110         if (ipoib_ib_dev_up(dev)) {
 111                 ipoib_ib_dev_stop(dev);
 112                 return -EINVAL;
 113         }
 114
 115         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 116                 struct ipoib_dev_priv *cpriv;
 117
 118                 /* Bring up any child interfaces too */
 119                 mutex_lock(&priv->vlan_mutex);
 120                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 121                         int flags;
 122
 123                         flags = cpriv->dev->flags;
 124                         if (flags & IFF_UP)
 125                                 continue;
 126
 127                         dev_change_flags(cpriv->dev, flags | IFF_UP);
 128                 }
 129                 mutex_unlock(&priv->vlan_mutex);
 130         }
 131
 132         netif_start_queue(dev);
 133
 134         return 0;
 135 }
 136
 137 static int ipoib_stop(struct net_device *dev)
 138 {
 139         struct ipoib_dev_priv *priv = netdev_priv(dev);
 140
 141         ipoib_dbg(priv, "stopping interface\n");
 142
 143         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 144
 145         netif_stop_queue(dev);
 146
 147         /*
 148          * Now flush workqueue to make sure a scheduled task doesn't
 149          * bring our internal state back up.
 150          */
 151         flush_workqueue(ipoib_workqueue);
 152
 153         ipoib_ib_dev_down(dev, 1);
 154         ipoib_ib_dev_stop(dev);
 155
 156         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 157                 struct ipoib_dev_priv *cpriv;
 158
 159                 /* Bring down any child interfaces too */
 160                 mutex_lock(&priv->vlan_mutex);
 161                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 162                         int flags;
 163
 164                         flags = cpriv->dev->flags;
 165                         if (!(flags & IFF_UP))
 166                                 continue;
 167
 168                         dev_change_flags(cpriv->dev, flags & ~IFF_UP);
 169                 }
 170                 mutex_unlock(&priv->vlan_mutex);
 171         }
 172
 173         return 0;
 174 }
 175
 176 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
 177 {
 178         struct ipoib_dev_priv *priv = netdev_priv(dev);
 179
 180         if (new_mtu > IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN)
 181                 return -EINVAL;
 182
 183         priv->admin_mtu = new_mtu;
 184
 185         dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
 186
 187         return 0;
 188 }
 189
 190 static struct ipoib_path *__path_find(struct net_device *dev, void *gid)
 191 {
 192         struct ipoib_dev_priv *priv = netdev_priv(dev);
 193         struct rb_node *n = priv->path_tree.rb_node;
 194         struct ipoib_path *path;
 195         int ret;
 196
 197         while (n) {
 198                 path = rb_entry(n, struct ipoib_path, rb_node);
 199
 200                 ret = memcmp(gid, path->pathrec.dgid.raw,
 201                              sizeof (union ib_gid));
 202
 203                 if (ret < 0)
 204                         n = n->rb_left;
 205                 else if (ret > 0)
 206                         n = n->rb_right;
 207                 else
 208                         return path;
 209         }
 210
 211         return NULL;
 212 }
 213
 214 static int __path_add(struct net_device *dev, struct ipoib_path *path)
 215 {
 216         struct ipoib_dev_priv *priv = netdev_priv(dev);
 217         struct rb_node **n = &priv->path_tree.rb_node;
 218         struct rb_node *pn = NULL;
 219         struct ipoib_path *tpath;
 220         int ret;
 221
 222         while (*n) {
 223                 pn = *n;
 224                 tpath = rb_entry(pn, struct ipoib_path, rb_node);
 225
 226                 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
 227                              sizeof (union ib_gid));
 228                 if (ret < 0)
 229                         n = &pn->rb_left;
 230                 else if (ret > 0)
 231                         n = &pn->rb_right;
 232                 else
 233                         return -EEXIST;
 234         }
 235
 236         rb_link_node(&path->rb_node, pn, n);
 237         rb_insert_color(&path->rb_node, &priv->path_tree);
 238
 239         list_add_tail(&path->list, &priv->path_list);
 240
 241         return 0;
 242 }
 243
 244 static void path_free(struct net_device *dev, struct ipoib_path *path)
 245 {
 246         struct ipoib_dev_priv *priv = netdev_priv(dev);
 247         struct ipoib_neigh *neigh, *tn;
 248         struct sk_buff *skb;
 249         unsigned long flags;
 250
 251         while ((skb = __skb_dequeue(&path->queue)))
 252                 dev_kfree_skb_irq(skb);
 253
 254         spin_lock_irqsave(&priv->lock, flags);
 255
 256         list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
 257                 /*
 258                  * It's safe to call ipoib_put_ah() inside priv->lock
 259                  * here, because we know that path->ah will always
 260                  * hold one more reference, so ipoib_put_ah() will
 261                  * never do more than decrement the ref count.
 262                  */
 263                 if (neigh->ah)
 264                         ipoib_put_ah(neigh->ah);
 265
 266                 ipoib_neigh_free(neigh);
 267         }
 268
 269         spin_unlock_irqrestore(&priv->lock, flags);
 270
 271         if (path->ah)
 272                 ipoib_put_ah(path->ah);
 273
 274         kfree(path);
 275 }
 276
 277 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 278
 279 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
 280 {
 281         struct ipoib_path_iter *iter;
 282
 283         iter = kmalloc(sizeof *iter, GFP_KERNEL);
 284         if (!iter)
 285                 return NULL;
 286
 287         iter->dev = dev;
 288         memset(iter->path.pathrec.dgid.raw, 0, 16);
 289
 290         if (ipoib_path_iter_next(iter)) {
 291                 kfree(iter);
 292                 return NULL;
 293         }
 294
 295         return iter;
 296 }
 297
 298 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
 299 {
 300         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
 301         struct rb_node *n;
 302         struct ipoib_path *path;
 303         int ret = 1;
 304
 305         spin_lock_irq(&priv->lock);
 306
 307         n = rb_first(&priv->path_tree);
 308
 309         while (n) {
 310                 path = rb_entry(n, struct ipoib_path, rb_node);
 311
 312                 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
 313                            sizeof (union ib_gid)) < 0) {
 314                         iter->path = *path;
 315                         ret = 0;
 316                         break;
 317                 }
 318
 319                 n = rb_next(n);
 320         }
 321
 322         spin_unlock_irq(&priv->lock);
 323
 324         return ret;
 325 }
 326
 327 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
 328                           struct ipoib_path *path)
 329 {
 330         *path = iter->path;
 331 }
 332
 333 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
 334
 335 void ipoib_flush_paths(struct net_device *dev)
 336 {
 337         struct ipoib_dev_priv *priv = netdev_priv(dev);
 338         struct ipoib_path *path, *tp;
 339         LIST_HEAD(remove_list);
 340
 341         spin_lock_irq(&priv->tx_lock);
 342         spin_lock(&priv->lock);
 343
 344         list_splice(&priv->path_list, &remove_list);
 345         INIT_LIST_HEAD(&priv->path_list);
 346
 347         list_for_each_entry(path, &remove_list, list)
 348                 rb_erase(&path->rb_node, &priv->path_tree);
 349
 350         list_for_each_entry_safe(path, tp, &remove_list, list) {
 351                 if (path->query)
 352                         ib_sa_cancel_query(path->query_id, path->query);
 353                 spin_unlock(&priv->lock);
 354                 spin_unlock_irq(&priv->tx_lock);
 355                 wait_for_completion(&path->done);
 356                 path_free(dev, path);
 357                 spin_lock_irq(&priv->tx_lock);
 358                 spin_lock(&priv->lock);
 359         }
 360         spin_unlock(&priv->lock);
 361         spin_unlock_irq(&priv->tx_lock);
 362 }
 363
 364 static void path_rec_completion(int status,
 365                                 struct ib_sa_path_rec *pathrec,
 366                                 void *path_ptr)
 367 {
 368         struct ipoib_path *path = path_ptr;
 369         struct net_device *dev = path->dev;
 370         struct ipoib_dev_priv *priv = netdev_priv(dev);
 371         struct ipoib_ah *ah = NULL;
 372         struct ipoib_neigh *neigh;
 373         struct sk_buff_head skqueue;
 374         struct sk_buff *skb;
 375         unsigned long flags;
 376
 377         if (pathrec)
 378                 ipoib_dbg(priv, "PathRec LID 0x%04x for GID " IPOIB_GID_FMT "\n",
 379                           be16_to_cpu(pathrec->dlid), IPOIB_GID_ARG(pathrec->dgid));
 380         else
 381                 ipoib_dbg(priv, "PathRec status %d for GID " IPOIB_GID_FMT "\n",
 382                           status, IPOIB_GID_ARG(path->pathrec.dgid));
 383
 384         skb_queue_head_init(&skqueue);
 385
 386         if (!status) {
 387                 struct ib_ah_attr av = {
 388                         .dlid          = be16_to_cpu(pathrec->dlid),
 389                         .sl            = pathrec->sl,
 390                         .port_num      = priv->port,
 391                         .static_rate   = pathrec->rate
 392                 };
 393
 394                 ah = ipoib_create_ah(dev, priv->pd, &av);
 395         }
 396
 397         spin_lock_irqsave(&priv->lock, flags);
 398
 399         path->ah = ah;
 400
 401         if (ah) {
 402                 path->pathrec = *pathrec;
 403
 404                 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
 405                           ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
 406
 407                 while ((skb = __skb_dequeue(&path->queue)))
 408                         __skb_queue_tail(&skqueue, skb);
 409
 410                 list_for_each_entry(neigh, &path->neigh_list, list) {
 411                         kref_get(&path->ah->ref);
 412                         neigh->ah = path->ah;
 413                         memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
 414                                sizeof(union ib_gid));
 415
 416                         while ((skb = __skb_dequeue(&neigh->queue)))
 417                                 __skb_queue_tail(&skqueue, skb);
 418                 }
 419         }
 420
 421         path->query = NULL;
 422         complete(&path->done);
 423
 424         spin_unlock_irqrestore(&priv->lock, flags);
 425
 426         while ((skb = __skb_dequeue(&skqueue))) {
 427                 skb->dev = dev;
 428                 if (dev_queue_xmit(skb))
 429                         ipoib_warn(priv, "dev_queue_xmit failed "
 430                                    "to requeue packet\n");
 431         }
 432 }
 433
 434 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
 435 {
 436         struct ipoib_dev_priv *priv = netdev_priv(dev);
 437         struct ipoib_path *path;
 438
 439         path = kzalloc(sizeof *path, GFP_ATOMIC);
 440         if (!path)
 441                 return NULL;
 442
 443         path->dev = dev;
 444
 445         skb_queue_head_init(&path->queue);
 446
 447         INIT_LIST_HEAD(&path->neigh_list);
 448
 449         memcpy(path->pathrec.dgid.raw, gid, sizeof (union ib_gid));
 450         path->pathrec.sgid      = priv->local_gid;
 451         path->pathrec.pkey      = cpu_to_be16(priv->pkey);
 452         path->pathrec.numb_path = 1;
 453
 454         return path;
 455 }
 456
 457 static int path_rec_start(struct net_device *dev,
 458                           struct ipoib_path *path)
 459 {
 460         struct ipoib_dev_priv *priv = netdev_priv(dev);
 461
 462         ipoib_dbg(priv, "Start path record lookup for " IPOIB_GID_FMT "\n",
 463                   IPOIB_GID_ARG(path->pathrec.dgid));
 464
 465         init_completion(&path->done);
 466
 467         path->query_id =
 468                 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
 469                                    &path->pathrec,
 470                                    IB_SA_PATH_REC_DGID          |
 471                                    IB_SA_PATH_REC_SGID          |
 472                                    IB_SA_PATH_REC_NUMB_PATH     |
 473                                    IB_SA_PATH_REC_PKEY,
 474                                    1000, GFP_ATOMIC,
 475                                    path_rec_completion,
 476                                    path, &path->query);
 477         if (path->query_id < 0) {
 478                 ipoib_warn(priv, "ib_sa_path_rec_get failed\n");
 479                 path->query = NULL;
 480                 return path->query_id;
 481         }
 482
 483         return 0;
 484 }
 485
 486 static void neigh_add_path(struct sk_buff *skb, struct net_device *dev)
 487 {
 488         struct ipoib_dev_priv *priv = netdev_priv(dev);
 489         struct ipoib_path *path;
 490         struct ipoib_neigh *neigh;
 491
 492         neigh = ipoib_neigh_alloc(skb->dst->neighbour);
 493         if (!neigh) {
 494                 ++priv->stats.tx_dropped;
 495                 dev_kfree_skb_any(skb);
 496                 return;
 497         }
 498
 499         skb_queue_head_init(&neigh->queue);
 500
 501         /*
 502          * We can only be called from ipoib_start_xmit, so we're
 503          * inside tx_lock -- no need to save/restore flags.
 504          */
 505         spin_lock(&priv->lock);
 506
 507         path = __path_find(dev, skb->dst->neighbour->ha + 4);
 508         if (!path) {
 509                 path = path_rec_create(dev, skb->dst->neighbour->ha + 4);
 510                 if (!path)
 511                         goto err_path;
 512
 513                 __path_add(dev, path);
 514         }
 515
 516         list_add_tail(&neigh->list, &path->neigh_list);
 517
 518         if (path->ah) {
 519                 kref_get(&path->ah->ref);
 520                 neigh->ah = path->ah;
 521                 memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
 522                        sizeof(union ib_gid));
 523
 524                 ipoib_send(dev, skb, path->ah,
 525                            be32_to_cpup((__be32 *) skb->dst->neighbour->ha));
 526         } else {
 527                 neigh->ah  = NULL;
 528                 __skb_queue_tail(&neigh->queue, skb);
 529
 530                 if (!path->query && path_rec_start(dev, path))
 531                         goto err_list;
 532         }
 533
 534         spin_unlock(&priv->lock);
 535         return;
 536
 537 err_list:
 538         list_del(&neigh->list);
 539
 540 err_path:
 541         ipoib_neigh_free(neigh);
 542         ++priv->stats.tx_dropped;
 543         dev_kfree_skb_any(skb);
 544
 545         spin_unlock(&priv->lock);
 546 }
 547
 548 static void ipoib_path_lookup(struct sk_buff *skb, struct net_device *dev)
 549 {
 550         struct ipoib_dev_priv *priv = netdev_priv(skb->dev);
 551
 552         /* Look up path record for unicasts */
 553         if (skb->dst->neighbour->ha[4] != 0xff) {
 554                 neigh_add_path(skb, dev);
 555                 return;
 556         }
 557
 558         /* Add in the P_Key for multicasts */
 559         skb->dst->neighbour->ha[8] = (priv->pkey >> 8) & 0xff;
 560         skb->dst->neighbour->ha[9] = priv->pkey & 0xff;
 561         ipoib_mcast_send(dev, skb->dst->neighbour->ha + 4, skb);
 562 }
 563
 564 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
 565                              struct ipoib_pseudoheader *phdr)
 566 {
 567         struct ipoib_dev_priv *priv = netdev_priv(dev);
 568         struct ipoib_path *path;
 569
 570         /*
 571          * We can only be called from ipoib_start_xmit, so we're
 572          * inside tx_lock -- no need to save/restore flags.
 573          */
 574         spin_lock(&priv->lock);
 575
 576         path = __path_find(dev, phdr->hwaddr + 4);
 577         if (!path) {
 578                 path = path_rec_create(dev, phdr->hwaddr + 4);
 579                 if (path) {
 580                         /* put pseudoheader back on for next time */
 581                         skb_push(skb, sizeof *phdr);
 582                         __skb_queue_tail(&path->queue, skb);
 583
 584                         if (path_rec_start(dev, path)) {
 585                                 spin_unlock(&priv->lock);
 586                                 path_free(dev, path);
 587                                 return;
 588                         } else
 589                                 __path_add(dev, path);
 590                 } else {
 591                         ++priv->stats.tx_dropped;
 592                         dev_kfree_skb_any(skb);
 593                 }
 594
 595                 spin_unlock(&priv->lock);
 596                 return;
 597         }
 598
 599         if (path->ah) {
 600                 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
 601                           be16_to_cpu(path->pathrec.dlid));
 602
 603                 ipoib_send(dev, skb, path->ah,
 604                            be32_to_cpup((__be32 *) phdr->hwaddr));
 605         } else if ((path->query || !path_rec_start(dev, path)) &&
 606                    skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 607                 /* put pseudoheader back on for next time */
 608                 skb_push(skb, sizeof *phdr);
 609                 __skb_queue_tail(&path->queue, skb);
 610         } else {
 611                 ++priv->stats.tx_dropped;
 612                 dev_kfree_skb_any(skb);
 613         }
 614
 615         spin_unlock(&priv->lock);
 616 }
 617
 618 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
 619 {
 620         struct ipoib_dev_priv *priv = netdev_priv(dev);
 621         struct ipoib_neigh *neigh;
 622         unsigned long flags;
 623
 624         if (!spin_trylock_irqsave(&priv->tx_lock, flags))
 625                 return NETDEV_TX_LOCKED;
 626
 627         /*
 628          * Check if our queue is stopped.  Since we have the LLTX bit
 629          * set, we can't rely on netif_stop_queue() preventing our
 630          * xmit function from being called with a full queue.
 631          */
 632         if (unlikely(netif_queue_stopped(dev))) {
 633                 spin_unlock_irqrestore(&priv->tx_lock, flags);
 634                 return NETDEV_TX_BUSY;
 635         }
 636
 637         if (skb->dst && skb->dst->neighbour) {
 638                 if (unlikely(!*to_ipoib_neigh(skb->dst->neighbour))) {
 639                         ipoib_path_lookup(skb, dev);
 640                         goto out;
 641                 }
 642
 643                 neigh = *to_ipoib_neigh(skb->dst->neighbour);
 644
 645                 if (likely(neigh->ah)) {
 646                         if (unlikely(memcmp(&neigh->dgid.raw,
 647                                             skb->dst->neighbour->ha + 4,
 648                                             sizeof(union ib_gid)))) {
 649                                 spin_lock(&priv->lock);
 650                                 /*
 651                                  * It's safe to call ipoib_put_ah() inside
 652                                  * priv->lock here, because we know that
 653                                  * path->ah will always hold one more reference,
 654                                  * so ipoib_put_ah() will never do more than
 655                                  * decrement the ref count.
 656                                  */
 657                                 ipoib_put_ah(neigh->ah);
 658                                 list_del(&neigh->list);
 659                                 ipoib_neigh_free(neigh);
 660                                 spin_unlock(&priv->lock);
 661                                 ipoib_path_lookup(skb, dev);
 662                                 goto out;
 663                         }
 664
 665                         ipoib_send(dev, skb, neigh->ah,
 666                                    be32_to_cpup((__be32 *) skb->dst->neighbour->ha));
 667                         goto out;
 668                 }
 669
 670                 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 671                         spin_lock(&priv->lock);
 672                         __skb_queue_tail(&neigh->queue, skb);
 673                         spin_unlock(&priv->lock);
 674                 } else {
 675                         ++priv->stats.tx_dropped;
 676                         dev_kfree_skb_any(skb);
 677                 }
 678         } else {
 679                 struct ipoib_pseudoheader *phdr =
 680                         (struct ipoib_pseudoheader *) skb->data;
 681                 skb_pull(skb, sizeof *phdr);
 682
 683                 if (phdr->hwaddr[4] == 0xff) {
 684                         /* Add in the P_Key for multicast*/
 685                         phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
 686                         phdr->hwaddr[9] = priv->pkey & 0xff;
 687
 688                         ipoib_mcast_send(dev, phdr->hwaddr + 4, skb);
 689                 } else {
 690                         /* unicast GID -- should be ARP or RARP reply */
 691
 692                         if ((be16_to_cpup((__be16 *) skb->data) != ETH_P_ARP) &&
 693                             (be16_to_cpup((__be16 *) skb->data) != ETH_P_RARP)) {
 694                                 ipoib_warn(priv, "Unicast, no %s: type %04x, QPN %06x "
 695                                            IPOIB_GID_FMT "\n",
 696                                            skb->dst ? "neigh" : "dst",
 697                                            be16_to_cpup((__be16 *) skb->data),
 698                                            be32_to_cpup((__be32 *) phdr->hwaddr),
 699                                            IPOIB_GID_RAW_ARG(phdr->hwaddr + 4));
 700                                 dev_kfree_skb_any(skb);
 701                                 ++priv->stats.tx_dropped;
 702                                 goto out;
 703                         }
 704
 705                         unicast_arp_send(skb, dev, phdr);
 706                 }
 707         }
 708
 709 out:
 710         spin_unlock_irqrestore(&priv->tx_lock, flags);
 711
 712         return NETDEV_TX_OK;
 713 }
 714
 715 static struct net_device_stats *ipoib_get_stats(struct net_device *dev)
 716 {
 717         struct ipoib_dev_priv *priv = netdev_priv(dev);
 718
 719         return &priv->stats;
 720 }
 721
 722 static void ipoib_timeout(struct net_device *dev)
 723 {
 724         struct ipoib_dev_priv *priv = netdev_priv(dev);
 725
 726         ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
 727                    jiffies_to_msecs(jiffies - dev->trans_start));
 728         ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
 729                    netif_queue_stopped(dev),
 730                    priv->tx_head, priv->tx_tail);
 731         /* XXX reset QP, etc. */
 732 }
 733
 734 static int ipoib_hard_header(struct sk_buff *skb,
 735                              struct net_device *dev,
 736                              unsigned short type,
 737                              void *daddr, void *saddr, unsigned len)
 738 {
 739         struct ipoib_header *header;
 740
 741         header = (struct ipoib_header *) skb_push(skb, sizeof *header);
 742
 743         header->proto = htons(type);
 744         header->reserved = 0;
 745
 746         /*
 747          * If we don't have a neighbour structure, stuff the
 748          * destination address onto the front of the skb so we can
 749          * figure out where to send the packet later.
 750          */
 751         if ((!skb->dst || !skb->dst->neighbour) && daddr) {
 752                 struct ipoib_pseudoheader *phdr =
 753                         (struct ipoib_pseudoheader *) skb_push(skb, sizeof *phdr);
 754                 memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
 755         }
 756
 757         return 0;
 758 }
 759
 760 static void ipoib_set_mcast_list(struct net_device *dev)
 761 {
 762         struct ipoib_dev_priv *priv = netdev_priv(dev);
 763
 764         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
 765                 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
 766                 return;
 767         }
 768
 769         queue_work(ipoib_workqueue, &priv->restart_task);
 770 }
 771
 772 static void ipoib_neigh_destructor(struct neighbour *n)
 773 {
 774         struct ipoib_neigh *neigh;
 775         struct ipoib_dev_priv *priv = netdev_priv(n->dev);
 776         unsigned long flags;
 777         struct ipoib_ah *ah = NULL;
 778
 779         ipoib_dbg(priv,
 780                   "neigh_destructor for %06x " IPOIB_GID_FMT "\n",
 781                   be32_to_cpup((__be32 *) n->ha),
 782                   IPOIB_GID_RAW_ARG(n->ha + 4));
 783
 784         spin_lock_irqsave(&priv->lock, flags);
 785
 786         neigh = *to_ipoib_neigh(n);
 787         if (neigh) {
 788                 if (neigh->ah)
 789                         ah = neigh->ah;
 790                 list_del(&neigh->list);
 791                 ipoib_neigh_free(neigh);
 792         }
 793
 794         spin_unlock_irqrestore(&priv->lock, flags);
 795
 796         if (ah)
 797                 ipoib_put_ah(ah);
 798 }
 799
 800 struct ipoib_neigh *ipoib_neigh_alloc(struct neighbour *neighbour)
 801 {
 802         struct ipoib_neigh *neigh;
 803
 804         neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
 805         if (!neigh)
 806                 return NULL;
 807
 808         neigh->neighbour = neighbour;
 809         *to_ipoib_neigh(neighbour) = neigh;
 810
 811         return neigh;
 812 }
 813
 814 void ipoib_neigh_free(struct ipoib_neigh *neigh)
 815 {
 816         *to_ipoib_neigh(neigh->neighbour) = NULL;
 817         kfree(neigh);
 818 }
 819
 820 static int ipoib_neigh_setup_dev(struct net_device *dev, struct neigh_parms *parms)
 821 {
 822         parms->neigh_destructor = ipoib_neigh_destructor;
 823
 824         return 0;
 825 }
 826
 827 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
 828 {
 829         struct ipoib_dev_priv *priv = netdev_priv(dev);
 830
 831         /* Allocate RX/TX "rings" to hold queued skbs */
 832         priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
 833                                 GFP_KERNEL);
 834         if (!priv->rx_ring) {
 835                 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
 836                        ca->name, ipoib_recvq_size);
 837                 goto out;
 838         }
 839
 840         priv->tx_ring = kzalloc(ipoib_sendq_size * sizeof *priv->tx_ring,
 841                                 GFP_KERNEL);
 842         if (!priv->tx_ring) {
 843                 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
 844                        ca->name, ipoib_sendq_size);
 845                 goto out_rx_ring_cleanup;
 846         }
 847
 848         /* priv->tx_head & tx_tail are already 0 */
 849
 850         if (ipoib_ib_dev_init(dev, ca, port))
 851                 goto out_tx_ring_cleanup;
 852
 853         return 0;
 854
 855 out_tx_ring_cleanup:
 856         kfree(priv->tx_ring);
 857
 858 out_rx_ring_cleanup:
 859         kfree(priv->rx_ring);
 860
 861 out:
 862         return -ENOMEM;
 863 }
 864
 865 void ipoib_dev_cleanup(struct net_device *dev)
 866 {
 867         struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
 868
 869         ipoib_delete_debug_files(dev);
 870
 871         /* Delete any child interfaces first */
 872         list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
 873                 unregister_netdev(cpriv->dev);
 874                 ipoib_dev_cleanup(cpriv->dev);
 875                 free_netdev(cpriv->dev);
 876         }
 877
 878         ipoib_ib_dev_cleanup(dev);
 879
 880         kfree(priv->rx_ring);
 881         kfree(priv->tx_ring);
 882
 883         priv->rx_ring = NULL;
 884         priv->tx_ring = NULL;
 885 }
 886
 887 static void ipoib_setup(struct net_device *dev)
 888 {
 889         struct ipoib_dev_priv *priv = netdev_priv(dev);
 890
 891         dev->open                = ipoib_open;
 892         dev->stop                = ipoib_stop;
 893         dev->change_mtu          = ipoib_change_mtu;
 894         dev->hard_start_xmit     = ipoib_start_xmit;
 895         dev->get_stats           = ipoib_get_stats;
 896         dev->tx_timeout          = ipoib_timeout;
 897         dev->hard_header         = ipoib_hard_header;
 898         dev->set_multicast_list  = ipoib_set_mcast_list;
 899         dev->neigh_setup         = ipoib_neigh_setup_dev;
 900
 901         dev->watchdog_timeo      = HZ;
 902
 903         dev->flags              |= IFF_BROADCAST | IFF_MULTICAST;
 904
 905         /*
 906          * We add in INFINIBAND_ALEN to allow for the destination
 907          * address "pseudoheader" for skbs without neighbour struct.
 908          */
 909         dev->hard_header_len     = IPOIB_ENCAP_LEN + INFINIBAND_ALEN;
 910         dev->addr_len            = INFINIBAND_ALEN;
 911         dev->type                = ARPHRD_INFINIBAND;
 912         dev->tx_queue_len        = ipoib_sendq_size * 2;
 913         dev->features            = NETIF_F_VLAN_CHALLENGED | NETIF_F_LLTX;
 914
 915         /* MTU will be reset when mcast join happens */
 916         dev->mtu                 = IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN;
 917         priv->mcast_mtu          = priv->admin_mtu = dev->mtu;
 918
 919         memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
 920
 921         netif_carrier_off(dev);
 922
 923         SET_MODULE_OWNER(dev);
 924
 925         priv->dev = dev;
 926
 927         spin_lock_init(&priv->lock);
 928         spin_lock_init(&priv->tx_lock);
 929
 930         mutex_init(&priv->mcast_mutex);
 931         mutex_init(&priv->vlan_mutex);
 932
 933         INIT_LIST_HEAD(&priv->path_list);
 934         INIT_LIST_HEAD(&priv->child_intfs);
 935         INIT_LIST_HEAD(&priv->dead_ahs);
 936         INIT_LIST_HEAD(&priv->multicast_list);
 937
 938         INIT_WORK(&priv->pkey_task,    ipoib_pkey_poll,          priv->dev);
 939         INIT_WORK(&priv->mcast_task,   ipoib_mcast_join_task,    priv->dev);
 940         INIT_WORK(&priv->flush_task,   ipoib_ib_dev_flush,       priv->dev);
 941         INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task, priv->dev);
 942         INIT_WORK(&priv->ah_reap_task, ipoib_reap_ah,            priv->dev);
 943 }
 944
 945 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
 946 {
 947         struct net_device *dev;
 948
 949         dev = alloc_netdev((int) sizeof (struct ipoib_dev_priv), name,
 950                            ipoib_setup);
 951         if (!dev)
 952                 return NULL;
 953
 954         return netdev_priv(dev);
 955 }
 956
 957 static ssize_t show_pkey(struct class_device *cdev, char *buf)
 958 {
 959         struct ipoib_dev_priv *priv =
 960                 netdev_priv(container_of(cdev, struct net_device, class_dev));
 961
 962         return sprintf(buf, "0x%04x\n", priv->pkey);
 963 }
 964 static CLASS_DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
 965
 966 static ssize_t create_child(struct class_device *cdev,
 967                             const char *buf, size_t count)
 968 {
 969         int pkey;
 970         int ret;
 971
 972         if (sscanf(buf, "%i", &pkey) != 1)
 973                 return -EINVAL;
 974
 975         if (pkey < 0 || pkey > 0xffff)
 976                 return -EINVAL;
 977
 978         /*
 979          * Set the full membership bit, so that we join the right
 980          * broadcast group, etc.
 981          */
 982         pkey |= 0x8000;
 983
 984         ret = ipoib_vlan_add(container_of(cdev, struct net_device, class_dev),
 985                              pkey);
 986
 987         return ret ? ret : count;
 988 }
 989 static CLASS_DEVICE_ATTR(create_child, S_IWUGO, NULL, create_child);
 990
 991 static ssize_t delete_child(struct class_device *cdev,
 992                             const char *buf, size_t count)
 993 {
 994         int pkey;
 995         int ret;
 996
 997         if (sscanf(buf, "%i", &pkey) != 1)
 998                 return -EINVAL;
 999
1000         if (pkey < 0 || pkey > 0xffff)
1001                 return -EINVAL;
1002
1003         ret = ipoib_vlan_delete(container_of(cdev, struct net_device, class_dev),
1004                                 pkey);
1005
1006         return ret ? ret : count;
1007
1008 }
1009 static CLASS_DEVICE_ATTR(delete_child, S_IWUGO, NULL, delete_child);
1010
1011 int ipoib_add_pkey_attr(struct net_device *dev)
1012 {
1013         return class_device_create_file(&dev->class_dev,
1014                                         &class_device_attr_pkey);
1015 }
1016
1017 static struct net_device *ipoib_add_port(const char *format,
1018                                          struct ib_device *hca, u8 port)
1019 {
1020         struct ipoib_dev_priv *priv;
1021         int result = -ENOMEM;
1022
1023         priv = ipoib_intf_alloc(format);
1024         if (!priv)
1025                 goto alloc_mem_failed;
1026
1027         SET_NETDEV_DEV(priv->dev, hca->dma_device);
1028
1029         result = ib_query_pkey(hca, port, 0, &priv->pkey);
1030         if (result) {
1031                 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
1032                        hca->name, port, result);
1033                 goto alloc_mem_failed;
1034         }
1035
1036         /*
1037          * Set the full membership bit, so that we join the right
1038          * broadcast group, etc.
1039          */
1040         priv->pkey |= 0x8000;
1041
1042         priv->dev->broadcast[8] = priv->pkey >> 8;
1043         priv->dev->broadcast[9] = priv->pkey & 0xff;
1044
1045         result = ib_query_gid(hca, port, 0, &priv->local_gid);
1046         if (result) {
1047                 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1048                        hca->name, port, result);
1049                 goto alloc_mem_failed;
1050         } else
1051                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1052
1053
1054         result = ipoib_dev_init(priv->dev, hca, port);
1055         if (result < 0) {
1056                 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1057                        hca->name, port, result);
1058                 goto device_init_failed;
1059         }
1060
1061         INIT_IB_EVENT_HANDLER(&priv->event_handler,
1062                               priv->ca, ipoib_event);
1063         result = ib_register_event_handler(&priv->event_handler);
1064         if (result < 0) {
1065                 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1066                        "port %d (ret = %d)\n",
1067                        hca->name, port, result);
1068                 goto event_failed;
1069         }
1070
1071         result = register_netdev(priv->dev);
1072         if (result) {
1073                 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1074                        hca->name, port, result);
1075                 goto register_failed;
1076         }
1077
1078         ipoib_create_debug_files(priv->dev);
1079
1080         if (ipoib_add_pkey_attr(priv->dev))
1081                 goto sysfs_failed;
1082         if (class_device_create_file(&priv->dev->class_dev,
1083                                      &class_device_attr_create_child))
1084                 goto sysfs_failed;
1085         if (class_device_create_file(&priv->dev->class_dev,
1086                                      &class_device_attr_delete_child))
1087                 goto sysfs_failed;
1088
1089         return priv->dev;
1090
1091 sysfs_failed:
1092         ipoib_delete_debug_files(priv->dev);
1093         unregister_netdev(priv->dev);
1094
1095 register_failed:
1096         ib_unregister_event_handler(&priv->event_handler);
1097         flush_scheduled_work();
1098
1099 event_failed:
1100         ipoib_dev_cleanup(priv->dev);
1101
1102 device_init_failed:
1103         free_netdev(priv->dev);
1104
1105 alloc_mem_failed:
1106         return ERR_PTR(result);
1107 }
1108
1109 static void ipoib_add_one(struct ib_device *device)
1110 {
1111         struct list_head *dev_list;
1112         struct net_device *dev;
1113         struct ipoib_dev_priv *priv;
1114         int s, e, p;
1115
1116         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1117                 return;
1118
1119         dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1120         if (!dev_list)
1121                 return;
1122
1123         INIT_LIST_HEAD(dev_list);
1124
1125         if (device->node_type == RDMA_NODE_IB_SWITCH) {
1126                 s = 0;
1127                 e = 0;
1128         } else {
1129                 s = 1;
1130                 e = device->phys_port_cnt;
1131         }
1132
1133         for (p = s; p <= e; ++p) {
1134                 dev = ipoib_add_port("ib%d", device, p);
1135                 if (!IS_ERR(dev)) {
1136                         priv = netdev_priv(dev);
1137                         list_add_tail(&priv->list, dev_list);
1138                 }
1139         }
1140
1141         ib_set_client_data(device, &ipoib_client, dev_list);
1142 }
1143
1144 static void ipoib_remove_one(struct ib_device *device)
1145 {
1146         struct ipoib_dev_priv *priv, *tmp;
1147         struct list_head *dev_list;
1148
1149         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1150                 return;
1151
1152         dev_list = ib_get_client_data(device, &ipoib_client);
1153
1154         list_for_each_entry_safe(priv, tmp, dev_list, list) {
1155                 ib_unregister_event_handler(&priv->event_handler);
1156                 flush_scheduled_work();
1157
1158                 unregister_netdev(priv->dev);
1159                 ipoib_dev_cleanup(priv->dev);
1160                 free_netdev(priv->dev);
1161         }
1162
1163         kfree(dev_list);
1164 }
1165
1166 static int __init ipoib_init_module(void)
1167 {
1168         int ret;
1169
1170         ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1171         ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1172         ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1173
1174         ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1175         ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1176         ipoib_sendq_size = max(ipoib_sendq_size, IPOIB_MIN_QUEUE_SIZE);
1177
1178         ret = ipoib_register_debugfs();
1179         if (ret)
1180                 return ret;
1181
1182         /*
1183          * We create our own workqueue mainly because we want to be
1184          * able to flush it when devices are being removed.  We can't
1185          * use schedule_work()/flush_scheduled_work() because both
1186          * unregister_netdev() and linkwatch_event take the rtnl lock,
1187          * so flush_scheduled_work() can deadlock during device
1188          * removal.
1189          */
1190         ipoib_workqueue = create_singlethread_workqueue("ipoib");
1191         if (!ipoib_workqueue) {
1192                 ret = -ENOMEM;
1193                 goto err_fs;
1194         }
1195
1196         ib_sa_register_client(&ipoib_sa_client);
1197
1198         ret = ib_register_client(&ipoib_client);
1199         if (ret)
1200                 goto err_sa;
1201
1202         return 0;
1203
1204 err_sa:
1205         ib_sa_unregister_client(&ipoib_sa_client);
1206         destroy_workqueue(ipoib_workqueue);
1207
1208 err_fs:
1209         ipoib_unregister_debugfs();
1210
1211         return ret;
1212 }
1213
1214 static void __exit ipoib_cleanup_module(void)
1215 {
1216         ib_unregister_client(&ipoib_client);
1217         ib_sa_unregister_client(&ipoib_sa_client);
1218         ipoib_unregister_debugfs();
1219         destroy_workqueue(ipoib_workqueue);
1220 }
1221
1222 module_init(ipoib_init_module);
1223 module_exit(ipoib_cleanup_module);