Pileus Git - ~andy/linux/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
  35 #include "ipoib.h"
  36
  37 #include <linux/module.h>
  38
  39 #include <linux/init.h>
  40 #include <linux/slab.h>
  41 #include <linux/kernel.h>
  42 #include <linux/vmalloc.h>
  43
  44 #include <linux/if_arp.h>       /* For ARPHRD_xxx */
  45
  46 #include <linux/ip.h>
  47 #include <linux/in.h>
  48
  49 #include <net/dst.h>
  50
  51 MODULE_AUTHOR("Roland Dreier");
  52 MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
  53 MODULE_LICENSE("Dual BSD/GPL");
  54
  55 int ipoib_sendq_size __read_mostly = IPOIB_TX_RING_SIZE;
  56 int ipoib_recvq_size __read_mostly = IPOIB_RX_RING_SIZE;
  57
  58 module_param_named(send_queue_size, ipoib_sendq_size, int, 0444);
  59 MODULE_PARM_DESC(send_queue_size, "Number of descriptors in send queue");
  60 module_param_named(recv_queue_size, ipoib_recvq_size, int, 0444);
  61 MODULE_PARM_DESC(recv_queue_size, "Number of descriptors in receive queue");
  62
  63 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
  64 int ipoib_debug_level;
  65
  66 module_param_named(debug_level, ipoib_debug_level, int, 0644);
  67 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
  68 #endif
  69
  70 struct ipoib_path_iter {
  71         struct net_device *dev;
  72         struct ipoib_path  path;
  73 };
  74
  75 static const u8 ipv4_bcast_addr[] = {
  76         0x00, 0xff, 0xff, 0xff,
  77         0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
  78         0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff
  79 };
  80
  81 struct workqueue_struct *ipoib_workqueue;
  82
  83 struct ib_sa_client ipoib_sa_client;
  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                 goto err_disable;
 107
 108         if (ipoib_ib_dev_up(dev))
 109                 goto err_stop;
 110
 111         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 112                 struct ipoib_dev_priv *cpriv;
 113
 114                 /* Bring up any child interfaces too */
 115                 mutex_lock(&priv->vlan_mutex);
 116                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 117                         int flags;
 118
 119                         flags = cpriv->dev->flags;
 120                         if (flags & IFF_UP)
 121                                 continue;
 122
 123                         dev_change_flags(cpriv->dev, flags | IFF_UP);
 124                 }
 125                 mutex_unlock(&priv->vlan_mutex);
 126         }
 127
 128         netif_start_queue(dev);
 129
 130         return 0;
 131
 132 err_stop:
 133         ipoib_ib_dev_stop(dev, 1);
 134
 135 err_disable:
 136         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 137
 138         return -EINVAL;
 139 }
 140
 141 static int ipoib_stop(struct net_device *dev)
 142 {
 143         struct ipoib_dev_priv *priv = netdev_priv(dev);
 144
 145         ipoib_dbg(priv, "stopping interface\n");
 146
 147         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 148
 149         netif_stop_queue(dev);
 150
 151         ipoib_ib_dev_down(dev, 0);
 152         ipoib_ib_dev_stop(dev, 0);
 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         /* dev->mtu > 2K ==> connected mode */
 179         if (ipoib_cm_admin_enabled(dev)) {
 180                 if (new_mtu > ipoib_cm_max_mtu(dev))
 181                         return -EINVAL;
 182
 183                 if (new_mtu > priv->mcast_mtu)
 184                         ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
 185                                    priv->mcast_mtu);
 186
 187                 dev->mtu = new_mtu;
 188                 return 0;
 189         }
 190
 191         if (new_mtu > IPOIB_UD_MTU(priv->max_ib_mtu))
 192                 return -EINVAL;
 193
 194         priv->admin_mtu = new_mtu;
 195
 196         dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
 197
 198         return 0;
 199 }
 200
 201 static struct ipoib_path *__path_find(struct net_device *dev, void *gid)
 202 {
 203         struct ipoib_dev_priv *priv = netdev_priv(dev);
 204         struct rb_node *n = priv->path_tree.rb_node;
 205         struct ipoib_path *path;
 206         int ret;
 207
 208         while (n) {
 209                 path = rb_entry(n, struct ipoib_path, rb_node);
 210
 211                 ret = memcmp(gid, path->pathrec.dgid.raw,
 212                              sizeof (union ib_gid));
 213
 214                 if (ret < 0)
 215                         n = n->rb_left;
 216                 else if (ret > 0)
 217                         n = n->rb_right;
 218                 else
 219                         return path;
 220         }
 221
 222         return NULL;
 223 }
 224
 225 static int __path_add(struct net_device *dev, struct ipoib_path *path)
 226 {
 227         struct ipoib_dev_priv *priv = netdev_priv(dev);
 228         struct rb_node **n = &priv->path_tree.rb_node;
 229         struct rb_node *pn = NULL;
 230         struct ipoib_path *tpath;
 231         int ret;
 232
 233         while (*n) {
 234                 pn = *n;
 235                 tpath = rb_entry(pn, struct ipoib_path, rb_node);
 236
 237                 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
 238                              sizeof (union ib_gid));
 239                 if (ret < 0)
 240                         n = &pn->rb_left;
 241                 else if (ret > 0)
 242                         n = &pn->rb_right;
 243                 else
 244                         return -EEXIST;
 245         }
 246
 247         rb_link_node(&path->rb_node, pn, n);
 248         rb_insert_color(&path->rb_node, &priv->path_tree);
 249
 250         list_add_tail(&path->list, &priv->path_list);
 251
 252         return 0;
 253 }
 254
 255 static void path_free(struct net_device *dev, struct ipoib_path *path)
 256 {
 257         struct ipoib_dev_priv *priv = netdev_priv(dev);
 258         struct ipoib_neigh *neigh, *tn;
 259         struct sk_buff *skb;
 260         unsigned long flags;
 261
 262         while ((skb = __skb_dequeue(&path->queue)))
 263                 dev_kfree_skb_irq(skb);
 264
 265         spin_lock_irqsave(&priv->lock, flags);
 266
 267         list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
 268                 /*
 269                  * It's safe to call ipoib_put_ah() inside priv->lock
 270                  * here, because we know that path->ah will always
 271                  * hold one more reference, so ipoib_put_ah() will
 272                  * never do more than decrement the ref count.
 273                  */
 274                 if (neigh->ah)
 275                         ipoib_put_ah(neigh->ah);
 276
 277                 ipoib_neigh_free(dev, neigh);
 278         }
 279
 280         spin_unlock_irqrestore(&priv->lock, flags);
 281
 282         if (path->ah)
 283                 ipoib_put_ah(path->ah);
 284
 285         kfree(path);
 286 }
 287
 288 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 289
 290 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
 291 {
 292         struct ipoib_path_iter *iter;
 293
 294         iter = kmalloc(sizeof *iter, GFP_KERNEL);
 295         if (!iter)
 296                 return NULL;
 297
 298         iter->dev = dev;
 299         memset(iter->path.pathrec.dgid.raw, 0, 16);
 300
 301         if (ipoib_path_iter_next(iter)) {
 302                 kfree(iter);
 303                 return NULL;
 304         }
 305
 306         return iter;
 307 }
 308
 309 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
 310 {
 311         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
 312         struct rb_node *n;
 313         struct ipoib_path *path;
 314         int ret = 1;
 315
 316         spin_lock_irq(&priv->lock);
 317
 318         n = rb_first(&priv->path_tree);
 319
 320         while (n) {
 321                 path = rb_entry(n, struct ipoib_path, rb_node);
 322
 323                 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
 324                            sizeof (union ib_gid)) < 0) {
 325                         iter->path = *path;
 326                         ret = 0;
 327                         break;
 328                 }
 329
 330                 n = rb_next(n);
 331         }
 332
 333         spin_unlock_irq(&priv->lock);
 334
 335         return ret;
 336 }
 337
 338 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
 339                           struct ipoib_path *path)
 340 {
 341         *path = iter->path;
 342 }
 343
 344 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
 345
 346 void ipoib_mark_paths_invalid(struct net_device *dev)
 347 {
 348         struct ipoib_dev_priv *priv = netdev_priv(dev);
 349         struct ipoib_path *path, *tp;
 350
 351         spin_lock_irq(&priv->lock);
 352
 353         list_for_each_entry_safe(path, tp, &priv->path_list, list) {
 354                 ipoib_dbg(priv, "mark path LID 0x%04x GID %pI6 invalid\n",
 355                         be16_to_cpu(path->pathrec.dlid),
 356                         path->pathrec.dgid.raw);
 357                 path->valid =  0;
 358         }
 359
 360         spin_unlock_irq(&priv->lock);
 361 }
 362
 363 void ipoib_flush_paths(struct net_device *dev)
 364 {
 365         struct ipoib_dev_priv *priv = netdev_priv(dev);
 366         struct ipoib_path *path, *tp;
 367         LIST_HEAD(remove_list);
 368         unsigned long flags;
 369
 370         netif_tx_lock_bh(dev);
 371         spin_lock_irqsave(&priv->lock, flags);
 372
 373         list_splice_init(&priv->path_list, &remove_list);
 374
 375         list_for_each_entry(path, &remove_list, list)
 376                 rb_erase(&path->rb_node, &priv->path_tree);
 377
 378         list_for_each_entry_safe(path, tp, &remove_list, list) {
 379                 if (path->query)
 380                         ib_sa_cancel_query(path->query_id, path->query);
 381                 spin_unlock_irqrestore(&priv->lock, flags);
 382                 netif_tx_unlock_bh(dev);
 383                 wait_for_completion(&path->done);
 384                 path_free(dev, path);
 385                 netif_tx_lock_bh(dev);
 386                 spin_lock_irqsave(&priv->lock, flags);
 387         }
 388
 389         spin_unlock_irqrestore(&priv->lock, flags);
 390         netif_tx_unlock_bh(dev);
 391 }
 392
 393 static void path_rec_completion(int status,
 394                                 struct ib_sa_path_rec *pathrec,
 395                                 void *path_ptr)
 396 {
 397         struct ipoib_path *path = path_ptr;
 398         struct net_device *dev = path->dev;
 399         struct ipoib_dev_priv *priv = netdev_priv(dev);
 400         struct ipoib_ah *ah = NULL;
 401         struct ipoib_ah *old_ah = NULL;
 402         struct ipoib_neigh *neigh, *tn;
 403         struct sk_buff_head skqueue;
 404         struct sk_buff *skb;
 405         unsigned long flags;
 406
 407         if (!status)
 408                 ipoib_dbg(priv, "PathRec LID 0x%04x for GID %pI6\n",
 409                           be16_to_cpu(pathrec->dlid), pathrec->dgid.raw);
 410         else
 411                 ipoib_dbg(priv, "PathRec status %d for GID %pI6\n",
 412                           status, path->pathrec.dgid.raw);
 413
 414         skb_queue_head_init(&skqueue);
 415
 416         if (!status) {
 417                 struct ib_ah_attr av;
 418
 419                 if (!ib_init_ah_from_path(priv->ca, priv->port, pathrec, &av))
 420                         ah = ipoib_create_ah(dev, priv->pd, &av);
 421         }
 422
 423         spin_lock_irqsave(&priv->lock, flags);
 424
 425         if (ah) {
 426                 path->pathrec = *pathrec;
 427
 428                 old_ah   = path->ah;
 429                 path->ah = ah;
 430
 431                 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
 432                           ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
 433
 434                 while ((skb = __skb_dequeue(&path->queue)))
 435                         __skb_queue_tail(&skqueue, skb);
 436
 437                 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
 438                         if (neigh->ah) {
 439                                 WARN_ON(neigh->ah != old_ah);
 440                                 /*
 441                                  * Dropping the ah reference inside
 442                                  * priv->lock is safe here, because we
 443                                  * will hold one more reference from
 444                                  * the original value of path->ah (ie
 445                                  * old_ah).
 446                                  */
 447                                 ipoib_put_ah(neigh->ah);
 448                         }
 449                         kref_get(&path->ah->ref);
 450                         neigh->ah = path->ah;
 451                         memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
 452                                sizeof(union ib_gid));
 453
 454                         if (ipoib_cm_enabled(dev, neigh->neighbour)) {
 455                                 if (!ipoib_cm_get(neigh))
 456                                         ipoib_cm_set(neigh, ipoib_cm_create_tx(dev,
 457                                                                                path,
 458                                                                                neigh));
 459                                 if (!ipoib_cm_get(neigh)) {
 460                                         list_del(&neigh->list);
 461                                         if (neigh->ah)
 462                                                 ipoib_put_ah(neigh->ah);
 463                                         ipoib_neigh_free(dev, neigh);
 464                                         continue;
 465                                 }
 466                         }
 467
 468                         while ((skb = __skb_dequeue(&neigh->queue)))
 469                                 __skb_queue_tail(&skqueue, skb);
 470                 }
 471                 path->valid = 1;
 472         }
 473
 474         path->query = NULL;
 475         complete(&path->done);
 476
 477         spin_unlock_irqrestore(&priv->lock, flags);
 478
 479         if (old_ah)
 480                 ipoib_put_ah(old_ah);
 481
 482         while ((skb = __skb_dequeue(&skqueue))) {
 483                 skb->dev = dev;
 484                 if (dev_queue_xmit(skb))
 485                         ipoib_warn(priv, "dev_queue_xmit failed "
 486                                    "to requeue packet\n");
 487         }
 488 }
 489
 490 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
 491 {
 492         struct ipoib_dev_priv *priv = netdev_priv(dev);
 493         struct ipoib_path *path;
 494
 495         if (!priv->broadcast)
 496                 return NULL;
 497
 498         path = kzalloc(sizeof *path, GFP_ATOMIC);
 499         if (!path)
 500                 return NULL;
 501
 502         path->dev = dev;
 503
 504         skb_queue_head_init(&path->queue);
 505
 506         INIT_LIST_HEAD(&path->neigh_list);
 507
 508         memcpy(path->pathrec.dgid.raw, gid, sizeof (union ib_gid));
 509         path->pathrec.sgid          = priv->local_gid;
 510         path->pathrec.pkey          = cpu_to_be16(priv->pkey);
 511         path->pathrec.numb_path     = 1;
 512         path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
 513
 514         return path;
 515 }
 516
 517 static int path_rec_start(struct net_device *dev,
 518                           struct ipoib_path *path)
 519 {
 520         struct ipoib_dev_priv *priv = netdev_priv(dev);
 521
 522         ipoib_dbg(priv, "Start path record lookup for %pI6\n",
 523                   path->pathrec.dgid.raw);
 524
 525         init_completion(&path->done);
 526
 527         path->query_id =
 528                 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
 529                                    &path->pathrec,
 530                                    IB_SA_PATH_REC_DGID          |
 531                                    IB_SA_PATH_REC_SGID          |
 532                                    IB_SA_PATH_REC_NUMB_PATH     |
 533                                    IB_SA_PATH_REC_TRAFFIC_CLASS |
 534                                    IB_SA_PATH_REC_PKEY,
 535                                    1000, GFP_ATOMIC,
 536                                    path_rec_completion,
 537                                    path, &path->query);
 538         if (path->query_id < 0) {
 539                 ipoib_warn(priv, "ib_sa_path_rec_get failed: %d\n", path->query_id);
 540                 path->query = NULL;
 541                 complete(&path->done);
 542                 return path->query_id;
 543         }
 544
 545         return 0;
 546 }
 547
 548 static void neigh_add_path(struct sk_buff *skb, struct net_device *dev)
 549 {
 550         struct ipoib_dev_priv *priv = netdev_priv(dev);
 551         struct ipoib_path *path;
 552         struct ipoib_neigh *neigh;
 553         unsigned long flags;
 554
 555         neigh = ipoib_neigh_alloc(skb_dst(skb)->neighbour, skb->dev);
 556         if (!neigh) {
 557                 ++dev->stats.tx_dropped;
 558                 dev_kfree_skb_any(skb);
 559                 return;
 560         }
 561
 562         spin_lock_irqsave(&priv->lock, flags);
 563
 564         path = __path_find(dev, skb_dst(skb)->neighbour->ha + 4);
 565         if (!path) {
 566                 path = path_rec_create(dev, skb_dst(skb)->neighbour->ha + 4);
 567                 if (!path)
 568                         goto err_path;
 569
 570                 __path_add(dev, path);
 571         }
 572
 573         list_add_tail(&neigh->list, &path->neigh_list);
 574
 575         if (path->ah) {
 576                 kref_get(&path->ah->ref);
 577                 neigh->ah = path->ah;
 578                 memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
 579                        sizeof(union ib_gid));
 580
 581                 if (ipoib_cm_enabled(dev, neigh->neighbour)) {
 582                         if (!ipoib_cm_get(neigh))
 583                                 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
 584                         if (!ipoib_cm_get(neigh)) {
 585                                 list_del(&neigh->list);
 586                                 if (neigh->ah)
 587                                         ipoib_put_ah(neigh->ah);
 588                                 ipoib_neigh_free(dev, neigh);
 589                                 goto err_drop;
 590                         }
 591                         if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE)
 592                                 __skb_queue_tail(&neigh->queue, skb);
 593                         else {
 594                                 ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
 595                                            skb_queue_len(&neigh->queue));
 596                                 goto err_drop;
 597                         }
 598                 } else {
 599                         spin_unlock_irqrestore(&priv->lock, flags);
 600                         ipoib_send(dev, skb, path->ah, IPOIB_QPN(skb_dst(skb)->neighbour->ha));
 601                         return;
 602                 }
 603         } else {
 604                 neigh->ah  = NULL;
 605
 606                 if (!path->query && path_rec_start(dev, path))
 607                         goto err_list;
 608
 609                 __skb_queue_tail(&neigh->queue, skb);
 610         }
 611
 612         spin_unlock_irqrestore(&priv->lock, flags);
 613         return;
 614
 615 err_list:
 616         list_del(&neigh->list);
 617
 618 err_path:
 619         ipoib_neigh_free(dev, neigh);
 620 err_drop:
 621         ++dev->stats.tx_dropped;
 622         dev_kfree_skb_any(skb);
 623
 624         spin_unlock_irqrestore(&priv->lock, flags);
 625 }
 626
 627 static void ipoib_path_lookup(struct sk_buff *skb, struct net_device *dev)
 628 {
 629         struct ipoib_dev_priv *priv = netdev_priv(skb->dev);
 630
 631         /* Look up path record for unicasts */
 632         if (skb_dst(skb)->neighbour->ha[4] != 0xff) {
 633                 neigh_add_path(skb, dev);
 634                 return;
 635         }
 636
 637         /* Add in the P_Key for multicasts */
 638         skb_dst(skb)->neighbour->ha[8] = (priv->pkey >> 8) & 0xff;
 639         skb_dst(skb)->neighbour->ha[9] = priv->pkey & 0xff;
 640         ipoib_mcast_send(dev, skb_dst(skb)->neighbour->ha + 4, skb);
 641 }
 642
 643 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
 644                              struct ipoib_pseudoheader *phdr)
 645 {
 646         struct ipoib_dev_priv *priv = netdev_priv(dev);
 647         struct ipoib_path *path;
 648         unsigned long flags;
 649
 650         spin_lock_irqsave(&priv->lock, flags);
 651
 652         path = __path_find(dev, phdr->hwaddr + 4);
 653         if (!path || !path->valid) {
 654                 int new_path = 0;
 655
 656                 if (!path) {
 657                         path = path_rec_create(dev, phdr->hwaddr + 4);
 658                         new_path = 1;
 659                 }
 660                 if (path) {
 661                         /* put pseudoheader back on for next time */
 662                         skb_push(skb, sizeof *phdr);
 663                         __skb_queue_tail(&path->queue, skb);
 664
 665                         if (!path->query && path_rec_start(dev, path)) {
 666                                 spin_unlock_irqrestore(&priv->lock, flags);
 667                                 if (new_path)
 668                                         path_free(dev, path);
 669                                 return;
 670                         } else
 671                                 __path_add(dev, path);
 672                 } else {
 673                         ++dev->stats.tx_dropped;
 674                         dev_kfree_skb_any(skb);
 675                 }
 676
 677                 spin_unlock_irqrestore(&priv->lock, flags);
 678                 return;
 679         }
 680
 681         if (path->ah) {
 682                 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
 683                           be16_to_cpu(path->pathrec.dlid));
 684
 685                 spin_unlock_irqrestore(&priv->lock, flags);
 686                 ipoib_send(dev, skb, path->ah, IPOIB_QPN(phdr->hwaddr));
 687                 return;
 688         } else if ((path->query || !path_rec_start(dev, path)) &&
 689                    skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 690                 /* put pseudoheader back on for next time */
 691                 skb_push(skb, sizeof *phdr);
 692                 __skb_queue_tail(&path->queue, skb);
 693         } else {
 694                 ++dev->stats.tx_dropped;
 695                 dev_kfree_skb_any(skb);
 696         }
 697
 698         spin_unlock_irqrestore(&priv->lock, flags);
 699 }
 700
 701 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
 702 {
 703         struct ipoib_dev_priv *priv = netdev_priv(dev);
 704         struct ipoib_neigh *neigh;
 705         unsigned long flags;
 706
 707         if (likely(skb_dst(skb) && skb_dst(skb)->neighbour)) {
 708                 if (unlikely(!*to_ipoib_neigh(skb_dst(skb)->neighbour))) {
 709                         ipoib_path_lookup(skb, dev);
 710                         return NETDEV_TX_OK;
 711                 }
 712
 713                 neigh = *to_ipoib_neigh(skb_dst(skb)->neighbour);
 714
 715                 if (unlikely((memcmp(&neigh->dgid.raw,
 716                                      skb_dst(skb)->neighbour->ha + 4,
 717                                      sizeof(union ib_gid))) ||
 718                              (neigh->dev != dev))) {
 719                         spin_lock_irqsave(&priv->lock, flags);
 720                         /*
 721                          * It's safe to call ipoib_put_ah() inside
 722                          * priv->lock here, because we know that
 723                          * path->ah will always hold one more reference,
 724                          * so ipoib_put_ah() will never do more than
 725                          * decrement the ref count.
 726                          */
 727                         if (neigh->ah)
 728                                 ipoib_put_ah(neigh->ah);
 729                         list_del(&neigh->list);
 730                         ipoib_neigh_free(dev, neigh);
 731                         spin_unlock_irqrestore(&priv->lock, flags);
 732                         ipoib_path_lookup(skb, dev);
 733                         return NETDEV_TX_OK;
 734                 }
 735
 736                 if (ipoib_cm_get(neigh)) {
 737                         if (ipoib_cm_up(neigh)) {
 738                                 ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
 739                                 return NETDEV_TX_OK;
 740                         }
 741                 } else if (neigh->ah) {
 742                         ipoib_send(dev, skb, neigh->ah, IPOIB_QPN(skb_dst(skb)->neighbour->ha));
 743                         return NETDEV_TX_OK;
 744                 }
 745
 746                 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 747                         spin_lock_irqsave(&priv->lock, flags);
 748                         __skb_queue_tail(&neigh->queue, skb);
 749                         spin_unlock_irqrestore(&priv->lock, flags);
 750                 } else {
 751                         ++dev->stats.tx_dropped;
 752                         dev_kfree_skb_any(skb);
 753                 }
 754         } else {
 755                 struct ipoib_pseudoheader *phdr =
 756                         (struct ipoib_pseudoheader *) skb->data;
 757                 skb_pull(skb, sizeof *phdr);
 758
 759                 if (phdr->hwaddr[4] == 0xff) {
 760                         /* Add in the P_Key for multicast*/
 761                         phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
 762                         phdr->hwaddr[9] = priv->pkey & 0xff;
 763
 764                         ipoib_mcast_send(dev, phdr->hwaddr + 4, skb);
 765                 } else {
 766                         /* unicast GID -- should be ARP or RARP reply */
 767
 768                         if ((be16_to_cpup((__be16 *) skb->data) != ETH_P_ARP) &&
 769                             (be16_to_cpup((__be16 *) skb->data) != ETH_P_RARP)) {
 770                                 ipoib_warn(priv, "Unicast, no %s: type %04x, QPN %06x %pI6\n",
 771                                            skb_dst(skb) ? "neigh" : "dst",
 772                                            be16_to_cpup((__be16 *) skb->data),
 773                                            IPOIB_QPN(phdr->hwaddr),
 774                                            phdr->hwaddr + 4);
 775                                 dev_kfree_skb_any(skb);
 776                                 ++dev->stats.tx_dropped;
 777                                 return NETDEV_TX_OK;
 778                         }
 779
 780                         unicast_arp_send(skb, dev, phdr);
 781                 }
 782         }
 783
 784         return NETDEV_TX_OK;
 785 }
 786
 787 static void ipoib_timeout(struct net_device *dev)
 788 {
 789         struct ipoib_dev_priv *priv = netdev_priv(dev);
 790
 791         ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
 792                    jiffies_to_msecs(jiffies - dev->trans_start));
 793         ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
 794                    netif_queue_stopped(dev),
 795                    priv->tx_head, priv->tx_tail);
 796         /* XXX reset QP, etc. */
 797 }
 798
 799 static int ipoib_hard_header(struct sk_buff *skb,
 800                              struct net_device *dev,
 801                              unsigned short type,
 802                              const void *daddr, const void *saddr, unsigned len)
 803 {
 804         struct ipoib_header *header;
 805
 806         header = (struct ipoib_header *) skb_push(skb, sizeof *header);
 807
 808         header->proto = htons(type);
 809         header->reserved = 0;
 810
 811         /*
 812          * If we don't have a neighbour structure, stuff the
 813          * destination address onto the front of the skb so we can
 814          * figure out where to send the packet later.
 815          */
 816         if ((!skb_dst(skb) || !skb_dst(skb)->neighbour) && daddr) {
 817                 struct ipoib_pseudoheader *phdr =
 818                         (struct ipoib_pseudoheader *) skb_push(skb, sizeof *phdr);
 819                 memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
 820         }
 821
 822         return 0;
 823 }
 824
 825 static void ipoib_set_mcast_list(struct net_device *dev)
 826 {
 827         struct ipoib_dev_priv *priv = netdev_priv(dev);
 828
 829         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
 830                 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
 831                 return;
 832         }
 833
 834         queue_work(ipoib_workqueue, &priv->restart_task);
 835 }
 836
 837 static void ipoib_neigh_cleanup(struct neighbour *n)
 838 {
 839         struct ipoib_neigh *neigh;
 840         struct ipoib_dev_priv *priv = netdev_priv(n->dev);
 841         unsigned long flags;
 842         struct ipoib_ah *ah = NULL;
 843
 844         neigh = *to_ipoib_neigh(n);
 845         if (neigh)
 846                 priv = netdev_priv(neigh->dev);
 847         else
 848                 return;
 849         ipoib_dbg(priv,
 850                   "neigh_cleanup for %06x %pI6\n",
 851                   IPOIB_QPN(n->ha),
 852                   n->ha + 4);
 853
 854         spin_lock_irqsave(&priv->lock, flags);
 855
 856         if (neigh->ah)
 857                 ah = neigh->ah;
 858         list_del(&neigh->list);
 859         ipoib_neigh_free(n->dev, neigh);
 860
 861         spin_unlock_irqrestore(&priv->lock, flags);
 862
 863         if (ah)
 864                 ipoib_put_ah(ah);
 865 }
 866
 867 struct ipoib_neigh *ipoib_neigh_alloc(struct neighbour *neighbour,
 868                                       struct net_device *dev)
 869 {
 870         struct ipoib_neigh *neigh;
 871
 872         neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
 873         if (!neigh)
 874                 return NULL;
 875
 876         neigh->neighbour = neighbour;
 877         neigh->dev = dev;
 878         memset(&neigh->dgid.raw, 0, sizeof (union ib_gid));
 879         *to_ipoib_neigh(neighbour) = neigh;
 880         skb_queue_head_init(&neigh->queue);
 881         ipoib_cm_set(neigh, NULL);
 882
 883         return neigh;
 884 }
 885
 886 void ipoib_neigh_free(struct net_device *dev, struct ipoib_neigh *neigh)
 887 {
 888         struct sk_buff *skb;
 889         *to_ipoib_neigh(neigh->neighbour) = NULL;
 890         while ((skb = __skb_dequeue(&neigh->queue))) {
 891                 ++dev->stats.tx_dropped;
 892                 dev_kfree_skb_any(skb);
 893         }
 894         if (ipoib_cm_get(neigh))
 895                 ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
 896         kfree(neigh);
 897 }
 898
 899 static int ipoib_neigh_setup_dev(struct net_device *dev, struct neigh_parms *parms)
 900 {
 901         parms->neigh_cleanup = ipoib_neigh_cleanup;
 902
 903         return 0;
 904 }
 905
 906 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
 907 {
 908         struct ipoib_dev_priv *priv = netdev_priv(dev);
 909
 910         /* Allocate RX/TX "rings" to hold queued skbs */
 911         priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
 912                                 GFP_KERNEL);
 913         if (!priv->rx_ring) {
 914                 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
 915                        ca->name, ipoib_recvq_size);
 916                 goto out;
 917         }
 918
 919         priv->tx_ring = vmalloc(ipoib_sendq_size * sizeof *priv->tx_ring);
 920         if (!priv->tx_ring) {
 921                 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
 922                        ca->name, ipoib_sendq_size);
 923                 goto out_rx_ring_cleanup;
 924         }
 925         memset(priv->tx_ring, 0, ipoib_sendq_size * sizeof *priv->tx_ring);
 926
 927         /* priv->tx_head, tx_tail & tx_outstanding are already 0 */
 928
 929         if (ipoib_ib_dev_init(dev, ca, port))
 930                 goto out_tx_ring_cleanup;
 931
 932         return 0;
 933
 934 out_tx_ring_cleanup:
 935         vfree(priv->tx_ring);
 936
 937 out_rx_ring_cleanup:
 938         kfree(priv->rx_ring);
 939
 940 out:
 941         return -ENOMEM;
 942 }
 943
 944 void ipoib_dev_cleanup(struct net_device *dev)
 945 {
 946         struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
 947
 948         ipoib_delete_debug_files(dev);
 949
 950         /* Delete any child interfaces first */
 951         list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
 952                 unregister_netdev(cpriv->dev);
 953                 ipoib_dev_cleanup(cpriv->dev);
 954                 free_netdev(cpriv->dev);
 955         }
 956
 957         ipoib_ib_dev_cleanup(dev);
 958
 959         kfree(priv->rx_ring);
 960         vfree(priv->tx_ring);
 961
 962         priv->rx_ring = NULL;
 963         priv->tx_ring = NULL;
 964 }
 965
 966 static const struct header_ops ipoib_header_ops = {
 967         .create = ipoib_hard_header,
 968 };
 969
 970 static const struct net_device_ops ipoib_netdev_ops = {
 971         .ndo_open                = ipoib_open,
 972         .ndo_stop                = ipoib_stop,
 973         .ndo_change_mtu          = ipoib_change_mtu,
 974         .ndo_start_xmit          = ipoib_start_xmit,
 975         .ndo_tx_timeout          = ipoib_timeout,
 976         .ndo_set_multicast_list  = ipoib_set_mcast_list,
 977         .ndo_neigh_setup         = ipoib_neigh_setup_dev,
 978 };
 979
 980 static void ipoib_setup(struct net_device *dev)
 981 {
 982         struct ipoib_dev_priv *priv = netdev_priv(dev);
 983
 984         dev->netdev_ops          = &ipoib_netdev_ops;
 985         dev->header_ops          = &ipoib_header_ops;
 986
 987         ipoib_set_ethtool_ops(dev);
 988
 989         netif_napi_add(dev, &priv->napi, ipoib_poll, 100);
 990
 991         dev->watchdog_timeo      = HZ;
 992
 993         dev->flags              |= IFF_BROADCAST | IFF_MULTICAST;
 994
 995         /*
 996          * We add in INFINIBAND_ALEN to allow for the destination
 997          * address "pseudoheader" for skbs without neighbour struct.
 998          */
 999         dev->hard_header_len     = IPOIB_ENCAP_LEN + INFINIBAND_ALEN;
1000         dev->addr_len            = INFINIBAND_ALEN;
1001         dev->type                = ARPHRD_INFINIBAND;
1002         dev->tx_queue_len        = ipoib_sendq_size * 2;
1003         dev->features            = (NETIF_F_VLAN_CHALLENGED     |
1004                                     NETIF_F_HIGHDMA);
1005         dev->priv_flags         &= ~IFF_XMIT_DST_RELEASE;
1006
1007         memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
1008
1009         netif_carrier_off(dev);
1010
1011         priv->dev = dev;
1012
1013         spin_lock_init(&priv->lock);
1014
1015         mutex_init(&priv->vlan_mutex);
1016
1017         INIT_LIST_HEAD(&priv->path_list);
1018         INIT_LIST_HEAD(&priv->child_intfs);
1019         INIT_LIST_HEAD(&priv->dead_ahs);
1020         INIT_LIST_HEAD(&priv->multicast_list);
1021
1022         INIT_DELAYED_WORK(&priv->pkey_poll_task, ipoib_pkey_poll);
1023         INIT_DELAYED_WORK(&priv->mcast_task,   ipoib_mcast_join_task);
1024         INIT_WORK(&priv->carrier_on_task, ipoib_mcast_carrier_on_task);
1025         INIT_WORK(&priv->flush_light,   ipoib_ib_dev_flush_light);
1026         INIT_WORK(&priv->flush_normal,   ipoib_ib_dev_flush_normal);
1027         INIT_WORK(&priv->flush_heavy,   ipoib_ib_dev_flush_heavy);
1028         INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
1029         INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
1030 }
1031
1032 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
1033 {
1034         struct net_device *dev;
1035
1036         dev = alloc_netdev((int) sizeof (struct ipoib_dev_priv), name,
1037                            ipoib_setup);
1038         if (!dev)
1039                 return NULL;
1040
1041         return netdev_priv(dev);
1042 }
1043
1044 static ssize_t show_pkey(struct device *dev,
1045                          struct device_attribute *attr, char *buf)
1046 {
1047         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1048
1049         return sprintf(buf, "0x%04x\n", priv->pkey);
1050 }
1051 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1052
1053 static ssize_t show_umcast(struct device *dev,
1054                            struct device_attribute *attr, char *buf)
1055 {
1056         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1057
1058         return sprintf(buf, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
1059 }
1060
1061 static ssize_t set_umcast(struct device *dev,
1062                           struct device_attribute *attr,
1063                           const char *buf, size_t count)
1064 {
1065         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1066         unsigned long umcast_val = simple_strtoul(buf, NULL, 0);
1067
1068         if (umcast_val > 0) {
1069                 set_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1070                 ipoib_warn(priv, "ignoring multicast groups joined directly "
1071                                 "by userspace\n");
1072         } else
1073                 clear_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1074
1075         return count;
1076 }
1077 static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast);
1078
1079 int ipoib_add_umcast_attr(struct net_device *dev)
1080 {
1081         return device_create_file(&dev->dev, &dev_attr_umcast);
1082 }
1083
1084 static ssize_t create_child(struct device *dev,
1085                             struct device_attribute *attr,
1086                             const char *buf, size_t count)
1087 {
1088         int pkey;
1089         int ret;
1090
1091         if (sscanf(buf, "%i", &pkey) != 1)
1092                 return -EINVAL;
1093
1094         if (pkey < 0 || pkey > 0xffff)
1095                 return -EINVAL;
1096
1097         /*
1098          * Set the full membership bit, so that we join the right
1099          * broadcast group, etc.
1100          */
1101         pkey |= 0x8000;
1102
1103         ret = ipoib_vlan_add(to_net_dev(dev), pkey);
1104
1105         return ret ? ret : count;
1106 }
1107 static DEVICE_ATTR(create_child, S_IWUSR, NULL, create_child);
1108
1109 static ssize_t delete_child(struct device *dev,
1110                             struct device_attribute *attr,
1111                             const char *buf, size_t count)
1112 {
1113         int pkey;
1114         int ret;
1115
1116         if (sscanf(buf, "%i", &pkey) != 1)
1117                 return -EINVAL;
1118
1119         if (pkey < 0 || pkey > 0xffff)
1120                 return -EINVAL;
1121
1122         ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
1123
1124         return ret ? ret : count;
1125
1126 }
1127 static DEVICE_ATTR(delete_child, S_IWUSR, NULL, delete_child);
1128
1129 int ipoib_add_pkey_attr(struct net_device *dev)
1130 {
1131         return device_create_file(&dev->dev, &dev_attr_pkey);
1132 }
1133
1134 int ipoib_set_dev_features(struct ipoib_dev_priv *priv, struct ib_device *hca)
1135 {
1136         struct ib_device_attr *device_attr;
1137         int result = -ENOMEM;
1138
1139         device_attr = kmalloc(sizeof *device_attr, GFP_KERNEL);
1140         if (!device_attr) {
1141                 printk(KERN_WARNING "%s: allocation of %zu bytes failed\n",
1142                        hca->name, sizeof *device_attr);
1143                 return result;
1144         }
1145
1146         result = ib_query_device(hca, device_attr);
1147         if (result) {
1148                 printk(KERN_WARNING "%s: ib_query_device failed (ret = %d)\n",
1149                        hca->name, result);
1150                 kfree(device_attr);
1151                 return result;
1152         }
1153         priv->hca_caps = device_attr->device_cap_flags;
1154
1155         kfree(device_attr);
1156
1157         if (priv->hca_caps & IB_DEVICE_UD_IP_CSUM) {
1158                 set_bit(IPOIB_FLAG_CSUM, &priv->flags);
1159                 priv->dev->features |= NETIF_F_SG | NETIF_F_IP_CSUM;
1160         }
1161
1162         priv->dev->features |= NETIF_F_GRO;
1163
1164         if (priv->dev->features & NETIF_F_SG && priv->hca_caps & IB_DEVICE_UD_TSO)
1165                 priv->dev->features |= NETIF_F_TSO;
1166
1167         return 0;
1168 }
1169
1170
1171 static struct net_device *ipoib_add_port(const char *format,
1172                                          struct ib_device *hca, u8 port)
1173 {
1174         struct ipoib_dev_priv *priv;
1175         struct ib_port_attr attr;
1176         int result = -ENOMEM;
1177
1178         priv = ipoib_intf_alloc(format);
1179         if (!priv)
1180                 goto alloc_mem_failed;
1181
1182         SET_NETDEV_DEV(priv->dev, hca->dma_device);
1183         priv->dev->dev_id = port - 1;
1184
1185         if (!ib_query_port(hca, port, &attr))
1186                 priv->max_ib_mtu = ib_mtu_enum_to_int(attr.max_mtu);
1187         else {
1188                 printk(KERN_WARNING "%s: ib_query_port %d failed\n",
1189                        hca->name, port);
1190                 goto device_init_failed;
1191         }
1192
1193         /* MTU will be reset when mcast join happens */
1194         priv->dev->mtu  = IPOIB_UD_MTU(priv->max_ib_mtu);
1195         priv->mcast_mtu  = priv->admin_mtu = priv->dev->mtu;
1196
1197         result = ib_query_pkey(hca, port, 0, &priv->pkey);
1198         if (result) {
1199                 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
1200                        hca->name, port, result);
1201                 goto device_init_failed;
1202         }
1203
1204         if (ipoib_set_dev_features(priv, hca))
1205                 goto device_init_failed;
1206
1207         /*
1208          * Set the full membership bit, so that we join the right
1209          * broadcast group, etc.
1210          */
1211         priv->pkey |= 0x8000;
1212
1213         priv->dev->broadcast[8] = priv->pkey >> 8;
1214         priv->dev->broadcast[9] = priv->pkey & 0xff;
1215
1216         result = ib_query_gid(hca, port, 0, &priv->local_gid);
1217         if (result) {
1218                 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1219                        hca->name, port, result);
1220                 goto device_init_failed;
1221         } else
1222                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1223
1224         result = ipoib_dev_init(priv->dev, hca, port);
1225         if (result < 0) {
1226                 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1227                        hca->name, port, result);
1228                 goto device_init_failed;
1229         }
1230
1231         INIT_IB_EVENT_HANDLER(&priv->event_handler,
1232                               priv->ca, ipoib_event);
1233         result = ib_register_event_handler(&priv->event_handler);
1234         if (result < 0) {
1235                 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1236                        "port %d (ret = %d)\n",
1237                        hca->name, port, result);
1238                 goto event_failed;
1239         }
1240
1241         result = register_netdev(priv->dev);
1242         if (result) {
1243                 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1244                        hca->name, port, result);
1245                 goto register_failed;
1246         }
1247
1248         ipoib_create_debug_files(priv->dev);
1249
1250         if (ipoib_cm_add_mode_attr(priv->dev))
1251                 goto sysfs_failed;
1252         if (ipoib_add_pkey_attr(priv->dev))
1253                 goto sysfs_failed;
1254         if (ipoib_add_umcast_attr(priv->dev))
1255                 goto sysfs_failed;
1256         if (device_create_file(&priv->dev->dev, &dev_attr_create_child))
1257                 goto sysfs_failed;
1258         if (device_create_file(&priv->dev->dev, &dev_attr_delete_child))
1259                 goto sysfs_failed;
1260
1261         return priv->dev;
1262
1263 sysfs_failed:
1264         ipoib_delete_debug_files(priv->dev);
1265         unregister_netdev(priv->dev);
1266
1267 register_failed:
1268         ib_unregister_event_handler(&priv->event_handler);
1269         flush_workqueue(ipoib_workqueue);
1270
1271 event_failed:
1272         ipoib_dev_cleanup(priv->dev);
1273
1274 device_init_failed:
1275         free_netdev(priv->dev);
1276
1277 alloc_mem_failed:
1278         return ERR_PTR(result);
1279 }
1280
1281 static void ipoib_add_one(struct ib_device *device)
1282 {
1283         struct list_head *dev_list;
1284         struct net_device *dev;
1285         struct ipoib_dev_priv *priv;
1286         int s, e, p;
1287
1288         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1289                 return;
1290
1291         dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1292         if (!dev_list)
1293                 return;
1294
1295         INIT_LIST_HEAD(dev_list);
1296
1297         if (device->node_type == RDMA_NODE_IB_SWITCH) {
1298                 s = 0;
1299                 e = 0;
1300         } else {
1301                 s = 1;
1302                 e = device->phys_port_cnt;
1303         }
1304
1305         for (p = s; p <= e; ++p) {
1306                 if (rdma_port_get_link_layer(device, p) != IB_LINK_LAYER_INFINIBAND)
1307                         continue;
1308                 dev = ipoib_add_port("ib%d", device, p);
1309                 if (!IS_ERR(dev)) {
1310                         priv = netdev_priv(dev);
1311                         list_add_tail(&priv->list, dev_list);
1312                 }
1313         }
1314
1315         ib_set_client_data(device, &ipoib_client, dev_list);
1316 }
1317
1318 static void ipoib_remove_one(struct ib_device *device)
1319 {
1320         struct ipoib_dev_priv *priv, *tmp;
1321         struct list_head *dev_list;
1322
1323         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1324                 return;
1325
1326         dev_list = ib_get_client_data(device, &ipoib_client);
1327
1328         list_for_each_entry_safe(priv, tmp, dev_list, list) {
1329                 ib_unregister_event_handler(&priv->event_handler);
1330
1331                 rtnl_lock();
1332                 dev_change_flags(priv->dev, priv->dev->flags & ~IFF_UP);
1333                 rtnl_unlock();
1334
1335                 flush_workqueue(ipoib_workqueue);
1336
1337                 unregister_netdev(priv->dev);
1338                 ipoib_dev_cleanup(priv->dev);
1339                 free_netdev(priv->dev);
1340         }
1341
1342         kfree(dev_list);
1343 }
1344
1345 static int __init ipoib_init_module(void)
1346 {
1347         int ret;
1348
1349         ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1350         ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1351         ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1352
1353         ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1354         ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1355         ipoib_sendq_size = max3(ipoib_sendq_size, 2 * MAX_SEND_CQE, IPOIB_MIN_QUEUE_SIZE);
1356 #ifdef CONFIG_INFINIBAND_IPOIB_CM
1357         ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
1358 #endif
1359
1360         /*
1361          * When copying small received packets, we only copy from the
1362          * linear data part of the SKB, so we rely on this condition.
1363          */
1364         BUILD_BUG_ON(IPOIB_CM_COPYBREAK > IPOIB_CM_HEAD_SIZE);
1365
1366         ret = ipoib_register_debugfs();
1367         if (ret)
1368                 return ret;
1369
1370         /*
1371          * We create our own workqueue mainly because we want to be
1372          * able to flush it when devices are being removed.  We can't
1373          * use schedule_work()/flush_scheduled_work() because both
1374          * unregister_netdev() and linkwatch_event take the rtnl lock,
1375          * so flush_scheduled_work() can deadlock during device
1376          * removal.
1377          */
1378         ipoib_workqueue = create_singlethread_workqueue("ipoib");
1379         if (!ipoib_workqueue) {
1380                 ret = -ENOMEM;
1381                 goto err_fs;
1382         }
1383
1384         ib_sa_register_client(&ipoib_sa_client);
1385
1386         ret = ib_register_client(&ipoib_client);
1387         if (ret)
1388                 goto err_sa;
1389
1390         return 0;
1391
1392 err_sa:
1393         ib_sa_unregister_client(&ipoib_sa_client);
1394         destroy_workqueue(ipoib_workqueue);
1395
1396 err_fs:
1397         ipoib_unregister_debugfs();
1398
1399         return ret;
1400 }
1401
1402 static void __exit ipoib_cleanup_module(void)
1403 {
1404         ib_unregister_client(&ipoib_client);
1405         ib_sa_unregister_client(&ipoib_sa_client);
1406         ipoib_unregister_debugfs();
1407         destroy_workqueue(ipoib_workqueue);
1408 }
1409
1410 module_init(ipoib_init_module);
1411 module_exit(ipoib_cleanup_module);