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 static int lro;
  64 module_param(lro, bool, 0444);
  65 MODULE_PARM_DESC(lro,  "Enable LRO (Large Receive Offload)");
  66
  67 static int lro_max_aggr = IPOIB_LRO_MAX_AGGR;
  68 module_param(lro_max_aggr, int, 0644);
  69 MODULE_PARM_DESC(lro_max_aggr, "LRO: Max packets to be aggregated "
  70                 "(default = 64)");
  71
  72 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
  73 int ipoib_debug_level;
  74
  75 module_param_named(debug_level, ipoib_debug_level, int, 0644);
  76 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
  77 #endif
  78
  79 struct ipoib_path_iter {
  80         struct net_device *dev;
  81         struct ipoib_path  path;
  82 };
  83
  84 static const u8 ipv4_bcast_addr[] = {
  85         0x00, 0xff, 0xff, 0xff,
  86         0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
  87         0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff
  88 };
  89
  90 struct workqueue_struct *ipoib_workqueue;
  91
  92 struct ib_sa_client ipoib_sa_client;
  93
  94 static void ipoib_add_one(struct ib_device *device);
  95 static void ipoib_remove_one(struct ib_device *device);
  96
  97 static struct ib_client ipoib_client = {
  98         .name   = "ipoib",
  99         .add    = ipoib_add_one,
 100         .remove = ipoib_remove_one
 101 };
 102
 103 int ipoib_open(struct net_device *dev)
 104 {
 105         struct ipoib_dev_priv *priv = netdev_priv(dev);
 106
 107         ipoib_dbg(priv, "bringing up interface\n");
 108
 109         set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 110
 111         if (ipoib_pkey_dev_delay_open(dev))
 112                 return 0;
 113
 114         napi_enable(&priv->napi);
 115
 116         if (ipoib_ib_dev_open(dev)) {
 117                 napi_disable(&priv->napi);
 118                 return -EINVAL;
 119         }
 120
 121         if (ipoib_ib_dev_up(dev)) {
 122                 ipoib_ib_dev_stop(dev, 1);
 123                 napi_disable(&priv->napi);
 124                 return -EINVAL;
 125         }
 126
 127         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 128                 struct ipoib_dev_priv *cpriv;
 129
 130                 /* Bring up any child interfaces too */
 131                 mutex_lock(&priv->vlan_mutex);
 132                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 133                         int flags;
 134
 135                         flags = cpriv->dev->flags;
 136                         if (flags & IFF_UP)
 137                                 continue;
 138
 139                         dev_change_flags(cpriv->dev, flags | IFF_UP);
 140                 }
 141                 mutex_unlock(&priv->vlan_mutex);
 142         }
 143
 144         netif_start_queue(dev);
 145
 146         return 0;
 147 }
 148
 149 static int ipoib_stop(struct net_device *dev)
 150 {
 151         struct ipoib_dev_priv *priv = netdev_priv(dev);
 152
 153         ipoib_dbg(priv, "stopping interface\n");
 154
 155         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 156         napi_disable(&priv->napi);
 157
 158         netif_stop_queue(dev);
 159
 160         ipoib_ib_dev_down(dev, 0);
 161         ipoib_ib_dev_stop(dev, 0);
 162
 163         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 164                 struct ipoib_dev_priv *cpriv;
 165
 166                 /* Bring down any child interfaces too */
 167                 mutex_lock(&priv->vlan_mutex);
 168                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 169                         int flags;
 170
 171                         flags = cpriv->dev->flags;
 172                         if (!(flags & IFF_UP))
 173                                 continue;
 174
 175                         dev_change_flags(cpriv->dev, flags & ~IFF_UP);
 176                 }
 177                 mutex_unlock(&priv->vlan_mutex);
 178         }
 179
 180         return 0;
 181 }
 182
 183 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
 184 {
 185         struct ipoib_dev_priv *priv = netdev_priv(dev);
 186
 187         /* dev->mtu > 2K ==> connected mode */
 188         if (ipoib_cm_admin_enabled(dev)) {
 189                 if (new_mtu > ipoib_cm_max_mtu(dev))
 190                         return -EINVAL;
 191
 192                 if (new_mtu > priv->mcast_mtu)
 193                         ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
 194                                    priv->mcast_mtu);
 195
 196                 dev->mtu = new_mtu;
 197                 return 0;
 198         }
 199
 200         if (new_mtu > IPOIB_UD_MTU(priv->max_ib_mtu))
 201                 return -EINVAL;
 202
 203         priv->admin_mtu = new_mtu;
 204
 205         dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
 206
 207         return 0;
 208 }
 209
 210 static struct ipoib_path *__path_find(struct net_device *dev, void *gid)
 211 {
 212         struct ipoib_dev_priv *priv = netdev_priv(dev);
 213         struct rb_node *n = priv->path_tree.rb_node;
 214         struct ipoib_path *path;
 215         int ret;
 216
 217         while (n) {
 218                 path = rb_entry(n, struct ipoib_path, rb_node);
 219
 220                 ret = memcmp(gid, path->pathrec.dgid.raw,
 221                              sizeof (union ib_gid));
 222
 223                 if (ret < 0)
 224                         n = n->rb_left;
 225                 else if (ret > 0)
 226                         n = n->rb_right;
 227                 else
 228                         return path;
 229         }
 230
 231         return NULL;
 232 }
 233
 234 static int __path_add(struct net_device *dev, struct ipoib_path *path)
 235 {
 236         struct ipoib_dev_priv *priv = netdev_priv(dev);
 237         struct rb_node **n = &priv->path_tree.rb_node;
 238         struct rb_node *pn = NULL;
 239         struct ipoib_path *tpath;
 240         int ret;
 241
 242         while (*n) {
 243                 pn = *n;
 244                 tpath = rb_entry(pn, struct ipoib_path, rb_node);
 245
 246                 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
 247                              sizeof (union ib_gid));
 248                 if (ret < 0)
 249                         n = &pn->rb_left;
 250                 else if (ret > 0)
 251                         n = &pn->rb_right;
 252                 else
 253                         return -EEXIST;
 254         }
 255
 256         rb_link_node(&path->rb_node, pn, n);
 257         rb_insert_color(&path->rb_node, &priv->path_tree);
 258
 259         list_add_tail(&path->list, &priv->path_list);
 260
 261         return 0;
 262 }
 263
 264 static void path_free(struct net_device *dev, struct ipoib_path *path)
 265 {
 266         struct ipoib_dev_priv *priv = netdev_priv(dev);
 267         struct ipoib_neigh *neigh, *tn;
 268         struct sk_buff *skb;
 269         unsigned long flags;
 270
 271         while ((skb = __skb_dequeue(&path->queue)))
 272                 dev_kfree_skb_irq(skb);
 273
 274         spin_lock_irqsave(&priv->lock, flags);
 275
 276         list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
 277                 /*
 278                  * It's safe to call ipoib_put_ah() inside priv->lock
 279                  * here, because we know that path->ah will always
 280                  * hold one more reference, so ipoib_put_ah() will
 281                  * never do more than decrement the ref count.
 282                  */
 283                 if (neigh->ah)
 284                         ipoib_put_ah(neigh->ah);
 285
 286                 ipoib_neigh_free(dev, neigh);
 287         }
 288
 289         spin_unlock_irqrestore(&priv->lock, flags);
 290
 291         if (path->ah)
 292                 ipoib_put_ah(path->ah);
 293
 294         kfree(path);
 295 }
 296
 297 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 298
 299 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
 300 {
 301         struct ipoib_path_iter *iter;
 302
 303         iter = kmalloc(sizeof *iter, GFP_KERNEL);
 304         if (!iter)
 305                 return NULL;
 306
 307         iter->dev = dev;
 308         memset(iter->path.pathrec.dgid.raw, 0, 16);
 309
 310         if (ipoib_path_iter_next(iter)) {
 311                 kfree(iter);
 312                 return NULL;
 313         }
 314
 315         return iter;
 316 }
 317
 318 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
 319 {
 320         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
 321         struct rb_node *n;
 322         struct ipoib_path *path;
 323         int ret = 1;
 324
 325         spin_lock_irq(&priv->lock);
 326
 327         n = rb_first(&priv->path_tree);
 328
 329         while (n) {
 330                 path = rb_entry(n, struct ipoib_path, rb_node);
 331
 332                 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
 333                            sizeof (union ib_gid)) < 0) {
 334                         iter->path = *path;
 335                         ret = 0;
 336                         break;
 337                 }
 338
 339                 n = rb_next(n);
 340         }
 341
 342         spin_unlock_irq(&priv->lock);
 343
 344         return ret;
 345 }
 346
 347 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
 348                           struct ipoib_path *path)
 349 {
 350         *path = iter->path;
 351 }
 352
 353 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
 354
 355 void ipoib_mark_paths_invalid(struct net_device *dev)
 356 {
 357         struct ipoib_dev_priv *priv = netdev_priv(dev);
 358         struct ipoib_path *path, *tp;
 359
 360         spin_lock_irq(&priv->lock);
 361
 362         list_for_each_entry_safe(path, tp, &priv->path_list, list) {
 363                 ipoib_dbg(priv, "mark path LID 0x%04x GID " IPOIB_GID_FMT " invalid\n",
 364                         be16_to_cpu(path->pathrec.dlid),
 365                         IPOIB_GID_ARG(path->pathrec.dgid));
 366                 path->valid =  0;
 367         }
 368
 369         spin_unlock_irq(&priv->lock);
 370 }
 371
 372 void ipoib_flush_paths(struct net_device *dev)
 373 {
 374         struct ipoib_dev_priv *priv = netdev_priv(dev);
 375         struct ipoib_path *path, *tp;
 376         LIST_HEAD(remove_list);
 377         unsigned long flags;
 378
 379         netif_tx_lock_bh(dev);
 380         spin_lock_irqsave(&priv->lock, flags);
 381
 382         list_splice_init(&priv->path_list, &remove_list);
 383
 384         list_for_each_entry(path, &remove_list, list)
 385                 rb_erase(&path->rb_node, &priv->path_tree);
 386
 387         list_for_each_entry_safe(path, tp, &remove_list, list) {
 388                 if (path->query)
 389                         ib_sa_cancel_query(path->query_id, path->query);
 390                 spin_unlock_irqrestore(&priv->lock, flags);
 391                 netif_tx_unlock_bh(dev);
 392                 wait_for_completion(&path->done);
 393                 path_free(dev, path);
 394                 netif_tx_lock_bh(dev);
 395                 spin_lock_irqsave(&priv->lock, flags);
 396         }
 397
 398         spin_unlock_irqrestore(&priv->lock, flags);
 399         netif_tx_unlock_bh(dev);
 400 }
 401
 402 static void path_rec_completion(int status,
 403                                 struct ib_sa_path_rec *pathrec,
 404                                 void *path_ptr)
 405 {
 406         struct ipoib_path *path = path_ptr;
 407         struct net_device *dev = path->dev;
 408         struct ipoib_dev_priv *priv = netdev_priv(dev);
 409         struct ipoib_ah *ah = NULL;
 410         struct ipoib_ah *old_ah = NULL;
 411         struct ipoib_neigh *neigh, *tn;
 412         struct sk_buff_head skqueue;
 413         struct sk_buff *skb;
 414         unsigned long flags;
 415
 416         if (!status)
 417                 ipoib_dbg(priv, "PathRec LID 0x%04x for GID " IPOIB_GID_FMT "\n",
 418                           be16_to_cpu(pathrec->dlid), IPOIB_GID_ARG(pathrec->dgid));
 419         else
 420                 ipoib_dbg(priv, "PathRec status %d for GID " IPOIB_GID_FMT "\n",
 421                           status, IPOIB_GID_ARG(path->pathrec.dgid));
 422
 423         skb_queue_head_init(&skqueue);
 424
 425         if (!status) {
 426                 struct ib_ah_attr av;
 427
 428                 if (!ib_init_ah_from_path(priv->ca, priv->port, pathrec, &av))
 429                         ah = ipoib_create_ah(dev, priv->pd, &av);
 430         }
 431
 432         spin_lock_irqsave(&priv->lock, flags);
 433
 434         if (ah) {
 435                 path->pathrec = *pathrec;
 436
 437                 old_ah   = path->ah;
 438                 path->ah = ah;
 439
 440                 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
 441                           ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
 442
 443                 while ((skb = __skb_dequeue(&path->queue)))
 444                         __skb_queue_tail(&skqueue, skb);
 445
 446                 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
 447                         if (neigh->ah) {
 448                                 WARN_ON(neigh->ah != old_ah);
 449                                 /*
 450                                  * Dropping the ah reference inside
 451                                  * priv->lock is safe here, because we
 452                                  * will hold one more reference from
 453                                  * the original value of path->ah (ie
 454                                  * old_ah).
 455                                  */
 456                                 ipoib_put_ah(neigh->ah);
 457                         }
 458                         kref_get(&path->ah->ref);
 459                         neigh->ah = path->ah;
 460                         memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
 461                                sizeof(union ib_gid));
 462
 463                         if (ipoib_cm_enabled(dev, neigh->neighbour)) {
 464                                 if (!ipoib_cm_get(neigh))
 465                                         ipoib_cm_set(neigh, ipoib_cm_create_tx(dev,
 466                                                                                path,
 467                                                                                neigh));
 468                                 if (!ipoib_cm_get(neigh)) {
 469                                         list_del(&neigh->list);
 470                                         if (neigh->ah)
 471                                                 ipoib_put_ah(neigh->ah);
 472                                         ipoib_neigh_free(dev, neigh);
 473                                         continue;
 474                                 }
 475                         }
 476
 477                         while ((skb = __skb_dequeue(&neigh->queue)))
 478                                 __skb_queue_tail(&skqueue, skb);
 479                 }
 480                 path->valid = 1;
 481         }
 482
 483         path->query = NULL;
 484         complete(&path->done);
 485
 486         spin_unlock_irqrestore(&priv->lock, flags);
 487
 488         if (old_ah)
 489                 ipoib_put_ah(old_ah);
 490
 491         while ((skb = __skb_dequeue(&skqueue))) {
 492                 skb->dev = dev;
 493                 if (dev_queue_xmit(skb))
 494                         ipoib_warn(priv, "dev_queue_xmit failed "
 495                                    "to requeue packet\n");
 496         }
 497 }
 498
 499 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
 500 {
 501         struct ipoib_dev_priv *priv = netdev_priv(dev);
 502         struct ipoib_path *path;
 503
 504         if (!priv->broadcast)
 505                 return NULL;
 506
 507         path = kzalloc(sizeof *path, GFP_ATOMIC);
 508         if (!path)
 509                 return NULL;
 510
 511         path->dev = dev;
 512
 513         skb_queue_head_init(&path->queue);
 514
 515         INIT_LIST_HEAD(&path->neigh_list);
 516
 517         memcpy(path->pathrec.dgid.raw, gid, sizeof (union ib_gid));
 518         path->pathrec.sgid          = priv->local_gid;
 519         path->pathrec.pkey          = cpu_to_be16(priv->pkey);
 520         path->pathrec.numb_path     = 1;
 521         path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
 522
 523         return path;
 524 }
 525
 526 static int path_rec_start(struct net_device *dev,
 527                           struct ipoib_path *path)
 528 {
 529         struct ipoib_dev_priv *priv = netdev_priv(dev);
 530
 531         ipoib_dbg(priv, "Start path record lookup for " IPOIB_GID_FMT "\n",
 532                   IPOIB_GID_ARG(path->pathrec.dgid));
 533
 534         init_completion(&path->done);
 535
 536         path->query_id =
 537                 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
 538                                    &path->pathrec,
 539                                    IB_SA_PATH_REC_DGID          |
 540                                    IB_SA_PATH_REC_SGID          |
 541                                    IB_SA_PATH_REC_NUMB_PATH     |
 542                                    IB_SA_PATH_REC_TRAFFIC_CLASS |
 543                                    IB_SA_PATH_REC_PKEY,
 544                                    1000, GFP_ATOMIC,
 545                                    path_rec_completion,
 546                                    path, &path->query);
 547         if (path->query_id < 0) {
 548                 ipoib_warn(priv, "ib_sa_path_rec_get failed: %d\n", path->query_id);
 549                 path->query = NULL;
 550                 complete(&path->done);
 551                 return path->query_id;
 552         }
 553
 554         return 0;
 555 }
 556
 557 static void neigh_add_path(struct sk_buff *skb, struct net_device *dev)
 558 {
 559         struct ipoib_dev_priv *priv = netdev_priv(dev);
 560         struct ipoib_path *path;
 561         struct ipoib_neigh *neigh;
 562         unsigned long flags;
 563
 564         neigh = ipoib_neigh_alloc(skb->dst->neighbour, skb->dev);
 565         if (!neigh) {
 566                 ++dev->stats.tx_dropped;
 567                 dev_kfree_skb_any(skb);
 568                 return;
 569         }
 570
 571         spin_lock_irqsave(&priv->lock, flags);
 572
 573         path = __path_find(dev, skb->dst->neighbour->ha + 4);
 574         if (!path) {
 575                 path = path_rec_create(dev, skb->dst->neighbour->ha + 4);
 576                 if (!path)
 577                         goto err_path;
 578
 579                 __path_add(dev, path);
 580         }
 581
 582         list_add_tail(&neigh->list, &path->neigh_list);
 583
 584         if (path->ah) {
 585                 kref_get(&path->ah->ref);
 586                 neigh->ah = path->ah;
 587                 memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
 588                        sizeof(union ib_gid));
 589
 590                 if (ipoib_cm_enabled(dev, neigh->neighbour)) {
 591                         if (!ipoib_cm_get(neigh))
 592                                 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
 593                         if (!ipoib_cm_get(neigh)) {
 594                                 list_del(&neigh->list);
 595                                 if (neigh->ah)
 596                                         ipoib_put_ah(neigh->ah);
 597                                 ipoib_neigh_free(dev, neigh);
 598                                 goto err_drop;
 599                         }
 600                         if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE)
 601                                 __skb_queue_tail(&neigh->queue, skb);
 602                         else {
 603                                 ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
 604                                            skb_queue_len(&neigh->queue));
 605                                 goto err_drop;
 606                         }
 607                 } else
 608                         ipoib_send(dev, skb, path->ah, IPOIB_QPN(skb->dst->neighbour->ha));
 609         } else {
 610                 neigh->ah  = NULL;
 611
 612                 if (!path->query && path_rec_start(dev, path))
 613                         goto err_list;
 614
 615                 __skb_queue_tail(&neigh->queue, skb);
 616         }
 617
 618         spin_unlock_irqrestore(&priv->lock, flags);
 619         return;
 620
 621 err_list:
 622         list_del(&neigh->list);
 623
 624 err_path:
 625         ipoib_neigh_free(dev, neigh);
 626 err_drop:
 627         ++dev->stats.tx_dropped;
 628         dev_kfree_skb_any(skb);
 629
 630         spin_unlock_irqrestore(&priv->lock, flags);
 631 }
 632
 633 static void ipoib_path_lookup(struct sk_buff *skb, struct net_device *dev)
 634 {
 635         struct ipoib_dev_priv *priv = netdev_priv(skb->dev);
 636
 637         /* Look up path record for unicasts */
 638         if (skb->dst->neighbour->ha[4] != 0xff) {
 639                 neigh_add_path(skb, dev);
 640                 return;
 641         }
 642
 643         /* Add in the P_Key for multicasts */
 644         skb->dst->neighbour->ha[8] = (priv->pkey >> 8) & 0xff;
 645         skb->dst->neighbour->ha[9] = priv->pkey & 0xff;
 646         ipoib_mcast_send(dev, skb->dst->neighbour->ha + 4, skb);
 647 }
 648
 649 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
 650                              struct ipoib_pseudoheader *phdr)
 651 {
 652         struct ipoib_dev_priv *priv = netdev_priv(dev);
 653         struct ipoib_path *path;
 654         unsigned long flags;
 655
 656         spin_lock_irqsave(&priv->lock, flags);
 657
 658         path = __path_find(dev, phdr->hwaddr + 4);
 659         if (!path || !path->valid) {
 660                 if (!path)
 661                         path = path_rec_create(dev, phdr->hwaddr + 4);
 662                 if (path) {
 663                         /* put pseudoheader back on for next time */
 664                         skb_push(skb, sizeof *phdr);
 665                         __skb_queue_tail(&path->queue, skb);
 666
 667                         if (!path->query && path_rec_start(dev, path)) {
 668                                 spin_unlock_irqrestore(&priv->lock, flags);
 669                                 path_free(dev, path);
 670                                 return;
 671                         } else
 672                                 __path_add(dev, path);
 673                 } else {
 674                         ++dev->stats.tx_dropped;
 675                         dev_kfree_skb_any(skb);
 676                 }
 677
 678                 spin_unlock_irqrestore(&priv->lock, flags);
 679                 return;
 680         }
 681
 682         if (path->ah) {
 683                 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
 684                           be16_to_cpu(path->pathrec.dlid));
 685
 686                 ipoib_send(dev, skb, path->ah, IPOIB_QPN(phdr->hwaddr));
 687         } else if ((path->query || !path_rec_start(dev, path)) &&
 688                    skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 689                 /* put pseudoheader back on for next time */
 690                 skb_push(skb, sizeof *phdr);
 691                 __skb_queue_tail(&path->queue, skb);
 692         } else {
 693                 ++dev->stats.tx_dropped;
 694                 dev_kfree_skb_any(skb);
 695         }
 696
 697         spin_unlock_irqrestore(&priv->lock, flags);
 698 }
 699
 700 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
 701 {
 702         struct ipoib_dev_priv *priv = netdev_priv(dev);
 703         struct ipoib_neigh *neigh;
 704         unsigned long flags;
 705
 706         if (likely(skb->dst && skb->dst->neighbour)) {
 707                 if (unlikely(!*to_ipoib_neigh(skb->dst->neighbour))) {
 708                         ipoib_path_lookup(skb, dev);
 709                         return NETDEV_TX_OK;
 710                 }
 711
 712                 neigh = *to_ipoib_neigh(skb->dst->neighbour);
 713
 714                 if (neigh->ah)
 715                         if (unlikely((memcmp(&neigh->dgid.raw,
 716                                             skb->dst->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                                 ipoib_put_ah(neigh->ah);
 728                                 list_del(&neigh->list);
 729                                 ipoib_neigh_free(dev, neigh);
 730                                 spin_unlock_irqrestore(&priv->lock, flags);
 731                                 ipoib_path_lookup(skb, dev);
 732                                 return NETDEV_TX_OK;
 733                         }
 734
 735                 if (ipoib_cm_get(neigh)) {
 736                         if (ipoib_cm_up(neigh)) {
 737                                 ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
 738                                 return NETDEV_TX_OK;
 739                         }
 740                 } else if (neigh->ah) {
 741                         ipoib_send(dev, skb, neigh->ah, IPOIB_QPN(skb->dst->neighbour->ha));
 742                         return NETDEV_TX_OK;
 743                 }
 744
 745                 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 746                         spin_lock_irqsave(&priv->lock, flags);
 747                         __skb_queue_tail(&neigh->queue, skb);
 748                         spin_unlock_irqrestore(&priv->lock, flags);
 749                 } else {
 750                         ++dev->stats.tx_dropped;
 751                         dev_kfree_skb_any(skb);
 752                 }
 753         } else {
 754                 struct ipoib_pseudoheader *phdr =
 755                         (struct ipoib_pseudoheader *) skb->data;
 756                 skb_pull(skb, sizeof *phdr);
 757
 758                 if (phdr->hwaddr[4] == 0xff) {
 759                         /* Add in the P_Key for multicast*/
 760                         phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
 761                         phdr->hwaddr[9] = priv->pkey & 0xff;
 762
 763                         ipoib_mcast_send(dev, phdr->hwaddr + 4, skb);
 764                 } else {
 765                         /* unicast GID -- should be ARP or RARP reply */
 766
 767                         if ((be16_to_cpup((__be16 *) skb->data) != ETH_P_ARP) &&
 768                             (be16_to_cpup((__be16 *) skb->data) != ETH_P_RARP)) {
 769                                 ipoib_warn(priv, "Unicast, no %s: type %04x, QPN %06x "
 770                                            IPOIB_GID_FMT "\n",
 771                                            skb->dst ? "neigh" : "dst",
 772                                            be16_to_cpup((__be16 *) skb->data),
 773                                            IPOIB_QPN(phdr->hwaddr),
 774                                            IPOIB_GID_RAW_ARG(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->dst->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 " IPOIB_GID_FMT "\n",
 851                   IPOIB_QPN(n->ha),
 852                   IPOIB_GID_RAW_ARG(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         *to_ipoib_neigh(neighbour) = neigh;
 879         skb_queue_head_init(&neigh->queue);
 880         ipoib_cm_set(neigh, NULL);
 881
 882         return neigh;
 883 }
 884
 885 void ipoib_neigh_free(struct net_device *dev, struct ipoib_neigh *neigh)
 886 {
 887         struct sk_buff *skb;
 888         *to_ipoib_neigh(neigh->neighbour) = NULL;
 889         while ((skb = __skb_dequeue(&neigh->queue))) {
 890                 ++dev->stats.tx_dropped;
 891                 dev_kfree_skb_any(skb);
 892         }
 893         if (ipoib_cm_get(neigh))
 894                 ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
 895         kfree(neigh);
 896 }
 897
 898 static int ipoib_neigh_setup_dev(struct net_device *dev, struct neigh_parms *parms)
 899 {
 900         parms->neigh_cleanup = ipoib_neigh_cleanup;
 901
 902         return 0;
 903 }
 904
 905 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
 906 {
 907         struct ipoib_dev_priv *priv = netdev_priv(dev);
 908
 909         /* Allocate RX/TX "rings" to hold queued skbs */
 910         priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
 911                                 GFP_KERNEL);
 912         if (!priv->rx_ring) {
 913                 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
 914                        ca->name, ipoib_recvq_size);
 915                 goto out;
 916         }
 917
 918         priv->tx_ring = vmalloc(ipoib_sendq_size * sizeof *priv->tx_ring);
 919         if (!priv->tx_ring) {
 920                 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
 921                        ca->name, ipoib_sendq_size);
 922                 goto out_rx_ring_cleanup;
 923         }
 924         memset(priv->tx_ring, 0, ipoib_sendq_size * sizeof *priv->tx_ring);
 925
 926         /* priv->tx_head, tx_tail & tx_outstanding are already 0 */
 927
 928         if (ipoib_ib_dev_init(dev, ca, port))
 929                 goto out_tx_ring_cleanup;
 930
 931         return 0;
 932
 933 out_tx_ring_cleanup:
 934         vfree(priv->tx_ring);
 935
 936 out_rx_ring_cleanup:
 937         kfree(priv->rx_ring);
 938
 939 out:
 940         return -ENOMEM;
 941 }
 942
 943 void ipoib_dev_cleanup(struct net_device *dev)
 944 {
 945         struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
 946
 947         ipoib_delete_debug_files(dev);
 948
 949         /* Delete any child interfaces first */
 950         list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
 951                 unregister_netdev(cpriv->dev);
 952                 ipoib_dev_cleanup(cpriv->dev);
 953                 free_netdev(cpriv->dev);
 954         }
 955
 956         ipoib_ib_dev_cleanup(dev);
 957
 958         kfree(priv->rx_ring);
 959         vfree(priv->tx_ring);
 960
 961         priv->rx_ring = NULL;
 962         priv->tx_ring = NULL;
 963 }
 964
 965 static const struct header_ops ipoib_header_ops = {
 966         .create = ipoib_hard_header,
 967 };
 968
 969 static int get_skb_hdr(struct sk_buff *skb, void **iphdr,
 970                        void **tcph, u64 *hdr_flags, void *priv)
 971 {
 972         unsigned int ip_len;
 973         struct iphdr *iph;
 974
 975         if (unlikely(skb->protocol != htons(ETH_P_IP)))
 976                 return -1;
 977
 978         /*
 979          * In the future we may add an else clause that verifies the
 980          * checksum and allows devices which do not calculate checksum
 981          * to use LRO.
 982          */
 983         if (unlikely(skb->ip_summed != CHECKSUM_UNNECESSARY))
 984                 return -1;
 985
 986         /* Check for non-TCP packet */
 987         skb_reset_network_header(skb);
 988         iph = ip_hdr(skb);
 989         if (iph->protocol != IPPROTO_TCP)
 990                 return -1;
 991
 992         ip_len = ip_hdrlen(skb);
 993         skb_set_transport_header(skb, ip_len);
 994         *tcph = tcp_hdr(skb);
 995
 996         /* check if IP header and TCP header are complete */
 997         if (ntohs(iph->tot_len) < ip_len + tcp_hdrlen(skb))
 998                 return -1;
 999
1000         *hdr_flags = LRO_IPV4 | LRO_TCP;
1001         *iphdr = iph;
1002
1003         return 0;
1004 }
1005
1006 static void ipoib_lro_setup(struct ipoib_dev_priv *priv)
1007 {
1008         priv->lro.lro_mgr.max_aggr       = lro_max_aggr;
1009         priv->lro.lro_mgr.max_desc       = IPOIB_MAX_LRO_DESCRIPTORS;
1010         priv->lro.lro_mgr.lro_arr        = priv->lro.lro_desc;
1011         priv->lro.lro_mgr.get_skb_header = get_skb_hdr;
1012         priv->lro.lro_mgr.features       = LRO_F_NAPI;
1013         priv->lro.lro_mgr.dev            = priv->dev;
1014         priv->lro.lro_mgr.ip_summed_aggr = CHECKSUM_UNNECESSARY;
1015 }
1016
1017 static void ipoib_setup(struct net_device *dev)
1018 {
1019         struct ipoib_dev_priv *priv = netdev_priv(dev);
1020
1021         dev->open                = ipoib_open;
1022         dev->stop                = ipoib_stop;
1023         dev->change_mtu          = ipoib_change_mtu;
1024         dev->hard_start_xmit     = ipoib_start_xmit;
1025         dev->tx_timeout          = ipoib_timeout;
1026         dev->header_ops          = &ipoib_header_ops;
1027         dev->set_multicast_list  = ipoib_set_mcast_list;
1028         dev->neigh_setup         = ipoib_neigh_setup_dev;
1029
1030         ipoib_set_ethtool_ops(dev);
1031
1032         netif_napi_add(dev, &priv->napi, ipoib_poll, 100);
1033
1034         dev->watchdog_timeo      = HZ;
1035
1036         dev->flags              |= IFF_BROADCAST | IFF_MULTICAST;
1037
1038         /*
1039          * We add in INFINIBAND_ALEN to allow for the destination
1040          * address "pseudoheader" for skbs without neighbour struct.
1041          */
1042         dev->hard_header_len     = IPOIB_ENCAP_LEN + INFINIBAND_ALEN;
1043         dev->addr_len            = INFINIBAND_ALEN;
1044         dev->type                = ARPHRD_INFINIBAND;
1045         dev->tx_queue_len        = ipoib_sendq_size * 2;
1046         dev->features            = (NETIF_F_VLAN_CHALLENGED     |
1047                                     NETIF_F_HIGHDMA);
1048
1049         memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
1050
1051         netif_carrier_off(dev);
1052
1053         priv->dev = dev;
1054
1055         ipoib_lro_setup(priv);
1056
1057         spin_lock_init(&priv->lock);
1058
1059         mutex_init(&priv->vlan_mutex);
1060
1061         INIT_LIST_HEAD(&priv->path_list);
1062         INIT_LIST_HEAD(&priv->child_intfs);
1063         INIT_LIST_HEAD(&priv->dead_ahs);
1064         INIT_LIST_HEAD(&priv->multicast_list);
1065
1066         INIT_DELAYED_WORK(&priv->pkey_poll_task, ipoib_pkey_poll);
1067         INIT_DELAYED_WORK(&priv->mcast_task,   ipoib_mcast_join_task);
1068         INIT_WORK(&priv->carrier_on_task, ipoib_mcast_carrier_on_task);
1069         INIT_WORK(&priv->flush_light,   ipoib_ib_dev_flush_light);
1070         INIT_WORK(&priv->flush_normal,   ipoib_ib_dev_flush_normal);
1071         INIT_WORK(&priv->flush_heavy,   ipoib_ib_dev_flush_heavy);
1072         INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
1073         INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
1074 }
1075
1076 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
1077 {
1078         struct net_device *dev;
1079
1080         dev = alloc_netdev((int) sizeof (struct ipoib_dev_priv), name,
1081                            ipoib_setup);
1082         if (!dev)
1083                 return NULL;
1084
1085         return netdev_priv(dev);
1086 }
1087
1088 static ssize_t show_pkey(struct device *dev,
1089                          struct device_attribute *attr, char *buf)
1090 {
1091         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1092
1093         return sprintf(buf, "0x%04x\n", priv->pkey);
1094 }
1095 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1096
1097 static ssize_t show_umcast(struct device *dev,
1098                            struct device_attribute *attr, char *buf)
1099 {
1100         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1101
1102         return sprintf(buf, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
1103 }
1104
1105 static ssize_t set_umcast(struct device *dev,
1106                           struct device_attribute *attr,
1107                           const char *buf, size_t count)
1108 {
1109         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1110         unsigned long umcast_val = simple_strtoul(buf, NULL, 0);
1111
1112         if (umcast_val > 0) {
1113                 set_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1114                 ipoib_warn(priv, "ignoring multicast groups joined directly "
1115                                 "by userspace\n");
1116         } else
1117                 clear_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1118
1119         return count;
1120 }
1121 static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast);
1122
1123 int ipoib_add_umcast_attr(struct net_device *dev)
1124 {
1125         return device_create_file(&dev->dev, &dev_attr_umcast);
1126 }
1127
1128 static ssize_t create_child(struct device *dev,
1129                             struct device_attribute *attr,
1130                             const char *buf, size_t count)
1131 {
1132         int pkey;
1133         int ret;
1134
1135         if (sscanf(buf, "%i", &pkey) != 1)
1136                 return -EINVAL;
1137
1138         if (pkey < 0 || pkey > 0xffff)
1139                 return -EINVAL;
1140
1141         /*
1142          * Set the full membership bit, so that we join the right
1143          * broadcast group, etc.
1144          */
1145         pkey |= 0x8000;
1146
1147         ret = ipoib_vlan_add(to_net_dev(dev), pkey);
1148
1149         return ret ? ret : count;
1150 }
1151 static DEVICE_ATTR(create_child, S_IWUGO, NULL, create_child);
1152
1153 static ssize_t delete_child(struct device *dev,
1154                             struct device_attribute *attr,
1155                             const char *buf, size_t count)
1156 {
1157         int pkey;
1158         int ret;
1159
1160         if (sscanf(buf, "%i", &pkey) != 1)
1161                 return -EINVAL;
1162
1163         if (pkey < 0 || pkey > 0xffff)
1164                 return -EINVAL;
1165
1166         ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
1167
1168         return ret ? ret : count;
1169
1170 }
1171 static DEVICE_ATTR(delete_child, S_IWUGO, NULL, delete_child);
1172
1173 int ipoib_add_pkey_attr(struct net_device *dev)
1174 {
1175         return device_create_file(&dev->dev, &dev_attr_pkey);
1176 }
1177
1178 int ipoib_set_dev_features(struct ipoib_dev_priv *priv, struct ib_device *hca)
1179 {
1180         struct ib_device_attr *device_attr;
1181         int result = -ENOMEM;
1182
1183         device_attr = kmalloc(sizeof *device_attr, GFP_KERNEL);
1184         if (!device_attr) {
1185                 printk(KERN_WARNING "%s: allocation of %zu bytes failed\n",
1186                        hca->name, sizeof *device_attr);
1187                 return result;
1188         }
1189
1190         result = ib_query_device(hca, device_attr);
1191         if (result) {
1192                 printk(KERN_WARNING "%s: ib_query_device failed (ret = %d)\n",
1193                        hca->name, result);
1194                 kfree(device_attr);
1195                 return result;
1196         }
1197         priv->hca_caps = device_attr->device_cap_flags;
1198
1199         kfree(device_attr);
1200
1201         if (priv->hca_caps & IB_DEVICE_UD_IP_CSUM) {
1202                 set_bit(IPOIB_FLAG_CSUM, &priv->flags);
1203                 priv->dev->features |= NETIF_F_SG | NETIF_F_IP_CSUM;
1204         }
1205
1206         if (lro)
1207                 priv->dev->features |= NETIF_F_LRO;
1208
1209         if (priv->dev->features & NETIF_F_SG && priv->hca_caps & IB_DEVICE_UD_TSO)
1210                 priv->dev->features |= NETIF_F_TSO;
1211
1212         return 0;
1213 }
1214
1215
1216 static struct net_device *ipoib_add_port(const char *format,
1217                                          struct ib_device *hca, u8 port)
1218 {
1219         struct ipoib_dev_priv *priv;
1220         struct ib_port_attr attr;
1221         int result = -ENOMEM;
1222
1223         priv = ipoib_intf_alloc(format);
1224         if (!priv)
1225                 goto alloc_mem_failed;
1226
1227         SET_NETDEV_DEV(priv->dev, hca->dma_device);
1228
1229         if (!ib_query_port(hca, port, &attr))
1230                 priv->max_ib_mtu = ib_mtu_enum_to_int(attr.max_mtu);
1231         else {
1232                 printk(KERN_WARNING "%s: ib_query_port %d failed\n",
1233                        hca->name, port);
1234                 goto device_init_failed;
1235         }
1236
1237         /* MTU will be reset when mcast join happens */
1238         priv->dev->mtu  = IPOIB_UD_MTU(priv->max_ib_mtu);
1239         priv->mcast_mtu  = priv->admin_mtu = priv->dev->mtu;
1240
1241         result = ib_query_pkey(hca, port, 0, &priv->pkey);
1242         if (result) {
1243                 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
1244                        hca->name, port, result);
1245                 goto device_init_failed;
1246         }
1247
1248         if (ipoib_set_dev_features(priv, hca))
1249                 goto device_init_failed;
1250
1251         /*
1252          * Set the full membership bit, so that we join the right
1253          * broadcast group, etc.
1254          */
1255         priv->pkey |= 0x8000;
1256
1257         priv->dev->broadcast[8] = priv->pkey >> 8;
1258         priv->dev->broadcast[9] = priv->pkey & 0xff;
1259
1260         result = ib_query_gid(hca, port, 0, &priv->local_gid);
1261         if (result) {
1262                 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1263                        hca->name, port, result);
1264                 goto device_init_failed;
1265         } else
1266                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1267
1268         result = ipoib_dev_init(priv->dev, hca, port);
1269         if (result < 0) {
1270                 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1271                        hca->name, port, result);
1272                 goto device_init_failed;
1273         }
1274
1275         INIT_IB_EVENT_HANDLER(&priv->event_handler,
1276                               priv->ca, ipoib_event);
1277         result = ib_register_event_handler(&priv->event_handler);
1278         if (result < 0) {
1279                 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1280                        "port %d (ret = %d)\n",
1281                        hca->name, port, result);
1282                 goto event_failed;
1283         }
1284
1285         result = register_netdev(priv->dev);
1286         if (result) {
1287                 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1288                        hca->name, port, result);
1289                 goto register_failed;
1290         }
1291
1292         ipoib_create_debug_files(priv->dev);
1293
1294         if (ipoib_cm_add_mode_attr(priv->dev))
1295                 goto sysfs_failed;
1296         if (ipoib_add_pkey_attr(priv->dev))
1297                 goto sysfs_failed;
1298         if (ipoib_add_umcast_attr(priv->dev))
1299                 goto sysfs_failed;
1300         if (device_create_file(&priv->dev->dev, &dev_attr_create_child))
1301                 goto sysfs_failed;
1302         if (device_create_file(&priv->dev->dev, &dev_attr_delete_child))
1303                 goto sysfs_failed;
1304
1305         return priv->dev;
1306
1307 sysfs_failed:
1308         ipoib_delete_debug_files(priv->dev);
1309         unregister_netdev(priv->dev);
1310
1311 register_failed:
1312         ib_unregister_event_handler(&priv->event_handler);
1313         flush_workqueue(ipoib_workqueue);
1314
1315 event_failed:
1316         ipoib_dev_cleanup(priv->dev);
1317
1318 device_init_failed:
1319         free_netdev(priv->dev);
1320
1321 alloc_mem_failed:
1322         return ERR_PTR(result);
1323 }
1324
1325 static void ipoib_add_one(struct ib_device *device)
1326 {
1327         struct list_head *dev_list;
1328         struct net_device *dev;
1329         struct ipoib_dev_priv *priv;
1330         int s, e, p;
1331
1332         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1333                 return;
1334
1335         dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1336         if (!dev_list)
1337                 return;
1338
1339         INIT_LIST_HEAD(dev_list);
1340
1341         if (device->node_type == RDMA_NODE_IB_SWITCH) {
1342                 s = 0;
1343                 e = 0;
1344         } else {
1345                 s = 1;
1346                 e = device->phys_port_cnt;
1347         }
1348
1349         for (p = s; p <= e; ++p) {
1350                 dev = ipoib_add_port("ib%d", device, p);
1351                 if (!IS_ERR(dev)) {
1352                         priv = netdev_priv(dev);
1353                         list_add_tail(&priv->list, dev_list);
1354                 }
1355         }
1356
1357         ib_set_client_data(device, &ipoib_client, dev_list);
1358 }
1359
1360 static void ipoib_remove_one(struct ib_device *device)
1361 {
1362         struct ipoib_dev_priv *priv, *tmp;
1363         struct list_head *dev_list;
1364
1365         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1366                 return;
1367
1368         dev_list = ib_get_client_data(device, &ipoib_client);
1369
1370         list_for_each_entry_safe(priv, tmp, dev_list, list) {
1371                 ib_unregister_event_handler(&priv->event_handler);
1372
1373                 rtnl_lock();
1374                 dev_change_flags(priv->dev, priv->dev->flags & ~IFF_UP);
1375                 rtnl_unlock();
1376
1377                 flush_workqueue(ipoib_workqueue);
1378
1379                 unregister_netdev(priv->dev);
1380                 ipoib_dev_cleanup(priv->dev);
1381                 free_netdev(priv->dev);
1382         }
1383
1384         kfree(dev_list);
1385 }
1386
1387 static int __init ipoib_init_module(void)
1388 {
1389         int ret;
1390
1391         ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1392         ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1393         ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1394
1395         ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1396         ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1397         ipoib_sendq_size = max(ipoib_sendq_size, max(2 * MAX_SEND_CQE,
1398                                                      IPOIB_MIN_QUEUE_SIZE));
1399 #ifdef CONFIG_INFINIBAND_IPOIB_CM
1400         ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
1401 #endif
1402
1403         /*
1404          * When copying small received packets, we only copy from the
1405          * linear data part of the SKB, so we rely on this condition.
1406          */
1407         BUILD_BUG_ON(IPOIB_CM_COPYBREAK > IPOIB_CM_HEAD_SIZE);
1408
1409         ret = ipoib_register_debugfs();
1410         if (ret)
1411                 return ret;
1412
1413         /*
1414          * We create our own workqueue mainly because we want to be
1415          * able to flush it when devices are being removed.  We can't
1416          * use schedule_work()/flush_scheduled_work() because both
1417          * unregister_netdev() and linkwatch_event take the rtnl lock,
1418          * so flush_scheduled_work() can deadlock during device
1419          * removal.
1420          */
1421         ipoib_workqueue = create_singlethread_workqueue("ipoib");
1422         if (!ipoib_workqueue) {
1423                 ret = -ENOMEM;
1424                 goto err_fs;
1425         }
1426
1427         ib_sa_register_client(&ipoib_sa_client);
1428
1429         ret = ib_register_client(&ipoib_client);
1430         if (ret)
1431                 goto err_sa;
1432
1433         return 0;
1434
1435 err_sa:
1436         ib_sa_unregister_client(&ipoib_sa_client);
1437         destroy_workqueue(ipoib_workqueue);
1438
1439 err_fs:
1440         ipoib_unregister_debugfs();
1441
1442         return ret;
1443 }
1444
1445 static void __exit ipoib_cleanup_module(void)
1446 {
1447         ib_unregister_client(&ipoib_client);
1448         ib_sa_unregister_client(&ipoib_sa_client);
1449         ipoib_unregister_debugfs();
1450         destroy_workqueue(ipoib_workqueue);
1451 }
1452
1453 module_init(ipoib_init_module);
1454 module_exit(ipoib_cleanup_module);