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         napi_enable(&priv->napi);
 110         set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 111
 112         if (ipoib_pkey_dev_delay_open(dev))
 113                 return 0;
 114
 115         if (ipoib_ib_dev_open(dev)) {
 116                 napi_disable(&priv->napi);
 117                 return -EINVAL;
 118         }
 119
 120         if (ipoib_ib_dev_up(dev)) {
 121                 ipoib_ib_dev_stop(dev, 1);
 122                 napi_disable(&priv->napi);
 123                 return -EINVAL;
 124         }
 125
 126         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 127                 struct ipoib_dev_priv *cpriv;
 128
 129                 /* Bring up any child interfaces too */
 130                 mutex_lock(&priv->vlan_mutex);
 131                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 132                         int flags;
 133
 134                         flags = cpriv->dev->flags;
 135                         if (flags & IFF_UP)
 136                                 continue;
 137
 138                         dev_change_flags(cpriv->dev, flags | IFF_UP);
 139                 }
 140                 mutex_unlock(&priv->vlan_mutex);
 141         }
 142
 143         netif_start_queue(dev);
 144
 145         return 0;
 146 }
 147
 148 static int ipoib_stop(struct net_device *dev)
 149 {
 150         struct ipoib_dev_priv *priv = netdev_priv(dev);
 151
 152         ipoib_dbg(priv, "stopping interface\n");
 153
 154         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 155         napi_disable(&priv->napi);
 156
 157         netif_stop_queue(dev);
 158
 159         ipoib_ib_dev_down(dev, 0);
 160         ipoib_ib_dev_stop(dev, 0);
 161
 162         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 163                 struct ipoib_dev_priv *cpriv;
 164
 165                 /* Bring down any child interfaces too */
 166                 mutex_lock(&priv->vlan_mutex);
 167                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 168                         int flags;
 169
 170                         flags = cpriv->dev->flags;
 171                         if (!(flags & IFF_UP))
 172                                 continue;
 173
 174                         dev_change_flags(cpriv->dev, flags & ~IFF_UP);
 175                 }
 176                 mutex_unlock(&priv->vlan_mutex);
 177         }
 178
 179         return 0;
 180 }
 181
 182 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
 183 {
 184         struct ipoib_dev_priv *priv = netdev_priv(dev);
 185
 186         /* dev->mtu > 2K ==> connected mode */
 187         if (ipoib_cm_admin_enabled(dev)) {
 188                 if (new_mtu > ipoib_cm_max_mtu(dev))
 189                         return -EINVAL;
 190
 191                 if (new_mtu > priv->mcast_mtu)
 192                         ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
 193                                    priv->mcast_mtu);
 194
 195                 dev->mtu = new_mtu;
 196                 return 0;
 197         }
 198
 199         if (new_mtu > IPOIB_UD_MTU(priv->max_ib_mtu))
 200                 return -EINVAL;
 201
 202         priv->admin_mtu = new_mtu;
 203
 204         dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
 205
 206         return 0;
 207 }
 208
 209 static struct ipoib_path *__path_find(struct net_device *dev, void *gid)
 210 {
 211         struct ipoib_dev_priv *priv = netdev_priv(dev);
 212         struct rb_node *n = priv->path_tree.rb_node;
 213         struct ipoib_path *path;
 214         int ret;
 215
 216         while (n) {
 217                 path = rb_entry(n, struct ipoib_path, rb_node);
 218
 219                 ret = memcmp(gid, path->pathrec.dgid.raw,
 220                              sizeof (union ib_gid));
 221
 222                 if (ret < 0)
 223                         n = n->rb_left;
 224                 else if (ret > 0)
 225                         n = n->rb_right;
 226                 else
 227                         return path;
 228         }
 229
 230         return NULL;
 231 }
 232
 233 static int __path_add(struct net_device *dev, struct ipoib_path *path)
 234 {
 235         struct ipoib_dev_priv *priv = netdev_priv(dev);
 236         struct rb_node **n = &priv->path_tree.rb_node;
 237         struct rb_node *pn = NULL;
 238         struct ipoib_path *tpath;
 239         int ret;
 240
 241         while (*n) {
 242                 pn = *n;
 243                 tpath = rb_entry(pn, struct ipoib_path, rb_node);
 244
 245                 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
 246                              sizeof (union ib_gid));
 247                 if (ret < 0)
 248                         n = &pn->rb_left;
 249                 else if (ret > 0)
 250                         n = &pn->rb_right;
 251                 else
 252                         return -EEXIST;
 253         }
 254
 255         rb_link_node(&path->rb_node, pn, n);
 256         rb_insert_color(&path->rb_node, &priv->path_tree);
 257
 258         list_add_tail(&path->list, &priv->path_list);
 259
 260         return 0;
 261 }
 262
 263 static void path_free(struct net_device *dev, struct ipoib_path *path)
 264 {
 265         struct ipoib_dev_priv *priv = netdev_priv(dev);
 266         struct ipoib_neigh *neigh, *tn;
 267         struct sk_buff *skb;
 268         unsigned long flags;
 269
 270         while ((skb = __skb_dequeue(&path->queue)))
 271                 dev_kfree_skb_irq(skb);
 272
 273         spin_lock_irqsave(&priv->lock, flags);
 274
 275         list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
 276                 /*
 277                  * It's safe to call ipoib_put_ah() inside priv->lock
 278                  * here, because we know that path->ah will always
 279                  * hold one more reference, so ipoib_put_ah() will
 280                  * never do more than decrement the ref count.
 281                  */
 282                 if (neigh->ah)
 283                         ipoib_put_ah(neigh->ah);
 284
 285                 ipoib_neigh_free(dev, neigh);
 286         }
 287
 288         spin_unlock_irqrestore(&priv->lock, flags);
 289
 290         if (path->ah)
 291                 ipoib_put_ah(path->ah);
 292
 293         kfree(path);
 294 }
 295
 296 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 297
 298 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
 299 {
 300         struct ipoib_path_iter *iter;
 301
 302         iter = kmalloc(sizeof *iter, GFP_KERNEL);
 303         if (!iter)
 304                 return NULL;
 305
 306         iter->dev = dev;
 307         memset(iter->path.pathrec.dgid.raw, 0, 16);
 308
 309         if (ipoib_path_iter_next(iter)) {
 310                 kfree(iter);
 311                 return NULL;
 312         }
 313
 314         return iter;
 315 }
 316
 317 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
 318 {
 319         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
 320         struct rb_node *n;
 321         struct ipoib_path *path;
 322         int ret = 1;
 323
 324         spin_lock_irq(&priv->lock);
 325
 326         n = rb_first(&priv->path_tree);
 327
 328         while (n) {
 329                 path = rb_entry(n, struct ipoib_path, rb_node);
 330
 331                 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
 332                            sizeof (union ib_gid)) < 0) {
 333                         iter->path = *path;
 334                         ret = 0;
 335                         break;
 336                 }
 337
 338                 n = rb_next(n);
 339         }
 340
 341         spin_unlock_irq(&priv->lock);
 342
 343         return ret;
 344 }
 345
 346 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
 347                           struct ipoib_path *path)
 348 {
 349         *path = iter->path;
 350 }
 351
 352 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
 353
 354 void ipoib_mark_paths_invalid(struct net_device *dev)
 355 {
 356         struct ipoib_dev_priv *priv = netdev_priv(dev);
 357         struct ipoib_path *path, *tp;
 358
 359         spin_lock_irq(&priv->lock);
 360
 361         list_for_each_entry_safe(path, tp, &priv->path_list, list) {
 362                 ipoib_dbg(priv, "mark path LID 0x%04x GID " IPOIB_GID_FMT " invalid\n",
 363                         be16_to_cpu(path->pathrec.dlid),
 364                         IPOIB_GID_ARG(path->pathrec.dgid));
 365                 path->valid =  0;
 366         }
 367
 368         spin_unlock_irq(&priv->lock);
 369 }
 370
 371 void ipoib_flush_paths(struct net_device *dev)
 372 {
 373         struct ipoib_dev_priv *priv = netdev_priv(dev);
 374         struct ipoib_path *path, *tp;
 375         LIST_HEAD(remove_list);
 376         unsigned long flags;
 377
 378         netif_tx_lock_bh(dev);
 379         spin_lock_irqsave(&priv->lock, flags);
 380
 381         list_splice_init(&priv->path_list, &remove_list);
 382
 383         list_for_each_entry(path, &remove_list, list)
 384                 rb_erase(&path->rb_node, &priv->path_tree);
 385
 386         list_for_each_entry_safe(path, tp, &remove_list, list) {
 387                 if (path->query)
 388                         ib_sa_cancel_query(path->query_id, path->query);
 389                 spin_unlock_irqrestore(&priv->lock, flags);
 390                 netif_tx_unlock_bh(dev);
 391                 wait_for_completion(&path->done);
 392                 path_free(dev, path);
 393                 netif_tx_lock_bh(dev);
 394                 spin_lock_irqsave(&priv->lock, flags);
 395         }
 396
 397         spin_unlock_irqrestore(&priv->lock, flags);
 398         netif_tx_unlock_bh(dev);
 399 }
 400
 401 static void path_rec_completion(int status,
 402                                 struct ib_sa_path_rec *pathrec,
 403                                 void *path_ptr)
 404 {
 405         struct ipoib_path *path = path_ptr;
 406         struct net_device *dev = path->dev;
 407         struct ipoib_dev_priv *priv = netdev_priv(dev);
 408         struct ipoib_ah *ah = NULL;
 409         struct ipoib_ah *old_ah = NULL;
 410         struct ipoib_neigh *neigh, *tn;
 411         struct sk_buff_head skqueue;
 412         struct sk_buff *skb;
 413         unsigned long flags;
 414
 415         if (!status)
 416                 ipoib_dbg(priv, "PathRec LID 0x%04x for GID " IPOIB_GID_FMT "\n",
 417                           be16_to_cpu(pathrec->dlid), IPOIB_GID_ARG(pathrec->dgid));
 418         else
 419                 ipoib_dbg(priv, "PathRec status %d for GID " IPOIB_GID_FMT "\n",
 420                           status, IPOIB_GID_ARG(path->pathrec.dgid));
 421
 422         skb_queue_head_init(&skqueue);
 423
 424         if (!status) {
 425                 struct ib_ah_attr av;
 426
 427                 if (!ib_init_ah_from_path(priv->ca, priv->port, pathrec, &av))
 428                         ah = ipoib_create_ah(dev, priv->pd, &av);
 429         }
 430
 431         spin_lock_irqsave(&priv->lock, flags);
 432
 433         if (ah) {
 434                 path->pathrec = *pathrec;
 435
 436                 old_ah   = path->ah;
 437                 path->ah = ah;
 438
 439                 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
 440                           ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
 441
 442                 while ((skb = __skb_dequeue(&path->queue)))
 443                         __skb_queue_tail(&skqueue, skb);
 444
 445                 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
 446                         if (neigh->ah) {
 447                                 WARN_ON(neigh->ah != old_ah);
 448                                 /*
 449                                  * Dropping the ah reference inside
 450                                  * priv->lock is safe here, because we
 451                                  * will hold one more reference from
 452                                  * the original value of path->ah (ie
 453                                  * old_ah).
 454                                  */
 455                                 ipoib_put_ah(neigh->ah);
 456                         }
 457                         kref_get(&path->ah->ref);
 458                         neigh->ah = path->ah;
 459                         memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
 460                                sizeof(union ib_gid));
 461
 462                         if (ipoib_cm_enabled(dev, neigh->neighbour)) {
 463                                 if (!ipoib_cm_get(neigh))
 464                                         ipoib_cm_set(neigh, ipoib_cm_create_tx(dev,
 465                                                                                path,
 466                                                                                neigh));
 467                                 if (!ipoib_cm_get(neigh)) {
 468                                         list_del(&neigh->list);
 469                                         if (neigh->ah)
 470                                                 ipoib_put_ah(neigh->ah);
 471                                         ipoib_neigh_free(dev, neigh);
 472                                         continue;
 473                                 }
 474                         }
 475
 476                         while ((skb = __skb_dequeue(&neigh->queue)))
 477                                 __skb_queue_tail(&skqueue, skb);
 478                 }
 479                 path->valid = 1;
 480         }
 481
 482         path->query = NULL;
 483         complete(&path->done);
 484
 485         spin_unlock_irqrestore(&priv->lock, flags);
 486
 487         if (old_ah)
 488                 ipoib_put_ah(old_ah);
 489
 490         while ((skb = __skb_dequeue(&skqueue))) {
 491                 skb->dev = dev;
 492                 if (dev_queue_xmit(skb))
 493                         ipoib_warn(priv, "dev_queue_xmit failed "
 494                                    "to requeue packet\n");
 495         }
 496 }
 497
 498 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
 499 {
 500         struct ipoib_dev_priv *priv = netdev_priv(dev);
 501         struct ipoib_path *path;
 502
 503         if (!priv->broadcast)
 504                 return NULL;
 505
 506         path = kzalloc(sizeof *path, GFP_ATOMIC);
 507         if (!path)
 508                 return NULL;
 509
 510         path->dev = dev;
 511
 512         skb_queue_head_init(&path->queue);
 513
 514         INIT_LIST_HEAD(&path->neigh_list);
 515
 516         memcpy(path->pathrec.dgid.raw, gid, sizeof (union ib_gid));
 517         path->pathrec.sgid          = priv->local_gid;
 518         path->pathrec.pkey          = cpu_to_be16(priv->pkey);
 519         path->pathrec.numb_path     = 1;
 520         path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
 521
 522         return path;
 523 }
 524
 525 static int path_rec_start(struct net_device *dev,
 526                           struct ipoib_path *path)
 527 {
 528         struct ipoib_dev_priv *priv = netdev_priv(dev);
 529
 530         ipoib_dbg(priv, "Start path record lookup for " IPOIB_GID_FMT "\n",
 531                   IPOIB_GID_ARG(path->pathrec.dgid));
 532
 533         init_completion(&path->done);
 534
 535         path->query_id =
 536                 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
 537                                    &path->pathrec,
 538                                    IB_SA_PATH_REC_DGID          |
 539                                    IB_SA_PATH_REC_SGID          |
 540                                    IB_SA_PATH_REC_NUMB_PATH     |
 541                                    IB_SA_PATH_REC_TRAFFIC_CLASS |
 542                                    IB_SA_PATH_REC_PKEY,
 543                                    1000, GFP_ATOMIC,
 544                                    path_rec_completion,
 545                                    path, &path->query);
 546         if (path->query_id < 0) {
 547                 ipoib_warn(priv, "ib_sa_path_rec_get failed: %d\n", path->query_id);
 548                 path->query = NULL;
 549                 return path->query_id;
 550         }
 551
 552         return 0;
 553 }
 554
 555 static void neigh_add_path(struct sk_buff *skb, struct net_device *dev)
 556 {
 557         struct ipoib_dev_priv *priv = netdev_priv(dev);
 558         struct ipoib_path *path;
 559         struct ipoib_neigh *neigh;
 560         unsigned long flags;
 561
 562         neigh = ipoib_neigh_alloc(skb->dst->neighbour, skb->dev);
 563         if (!neigh) {
 564                 ++dev->stats.tx_dropped;
 565                 dev_kfree_skb_any(skb);
 566                 return;
 567         }
 568
 569         spin_lock_irqsave(&priv->lock, flags);
 570
 571         path = __path_find(dev, skb->dst->neighbour->ha + 4);
 572         if (!path) {
 573                 path = path_rec_create(dev, skb->dst->neighbour->ha + 4);
 574                 if (!path)
 575                         goto err_path;
 576
 577                 __path_add(dev, path);
 578         }
 579
 580         list_add_tail(&neigh->list, &path->neigh_list);
 581
 582         if (path->ah) {
 583                 kref_get(&path->ah->ref);
 584                 neigh->ah = path->ah;
 585                 memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
 586                        sizeof(union ib_gid));
 587
 588                 if (ipoib_cm_enabled(dev, neigh->neighbour)) {
 589                         if (!ipoib_cm_get(neigh))
 590                                 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
 591                         if (!ipoib_cm_get(neigh)) {
 592                                 list_del(&neigh->list);
 593                                 if (neigh->ah)
 594                                         ipoib_put_ah(neigh->ah);
 595                                 ipoib_neigh_free(dev, neigh);
 596                                 goto err_drop;
 597                         }
 598                         if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE)
 599                                 __skb_queue_tail(&neigh->queue, skb);
 600                         else {
 601                                 ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
 602                                            skb_queue_len(&neigh->queue));
 603                                 goto err_drop;
 604                         }
 605                 } else
 606                         ipoib_send(dev, skb, path->ah, IPOIB_QPN(skb->dst->neighbour->ha));
 607         } else {
 608                 neigh->ah  = NULL;
 609
 610                 if (!path->query && path_rec_start(dev, path))
 611                         goto err_list;
 612
 613                 __skb_queue_tail(&neigh->queue, skb);
 614         }
 615
 616         spin_unlock_irqrestore(&priv->lock, flags);
 617         return;
 618
 619 err_list:
 620         list_del(&neigh->list);
 621
 622 err_path:
 623         ipoib_neigh_free(dev, neigh);
 624 err_drop:
 625         ++dev->stats.tx_dropped;
 626         dev_kfree_skb_any(skb);
 627
 628         spin_unlock_irqrestore(&priv->lock, flags);
 629 }
 630
 631 static void ipoib_path_lookup(struct sk_buff *skb, struct net_device *dev)
 632 {
 633         struct ipoib_dev_priv *priv = netdev_priv(skb->dev);
 634
 635         /* Look up path record for unicasts */
 636         if (skb->dst->neighbour->ha[4] != 0xff) {
 637                 neigh_add_path(skb, dev);
 638                 return;
 639         }
 640
 641         /* Add in the P_Key for multicasts */
 642         skb->dst->neighbour->ha[8] = (priv->pkey >> 8) & 0xff;
 643         skb->dst->neighbour->ha[9] = priv->pkey & 0xff;
 644         ipoib_mcast_send(dev, skb->dst->neighbour->ha + 4, skb);
 645 }
 646
 647 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
 648                              struct ipoib_pseudoheader *phdr)
 649 {
 650         struct ipoib_dev_priv *priv = netdev_priv(dev);
 651         struct ipoib_path *path;
 652         unsigned long flags;
 653
 654         spin_lock_irqsave(&priv->lock, flags);
 655
 656         path = __path_find(dev, phdr->hwaddr + 4);
 657         if (!path || !path->valid) {
 658                 if (!path)
 659                         path = path_rec_create(dev, phdr->hwaddr + 4);
 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_rec_start(dev, path)) {
 666                                 spin_unlock_irqrestore(&priv->lock, flags);
 667                                 path_free(dev, path);
 668                                 return;
 669                         } else
 670                                 __path_add(dev, path);
 671                 } else {
 672                         ++dev->stats.tx_dropped;
 673                         dev_kfree_skb_any(skb);
 674                 }
 675
 676                 spin_unlock_irqrestore(&priv->lock, flags);
 677                 return;
 678         }
 679
 680         if (path->ah) {
 681                 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
 682                           be16_to_cpu(path->pathrec.dlid));
 683
 684                 ipoib_send(dev, skb, path->ah, IPOIB_QPN(phdr->hwaddr));
 685         } else if ((path->query || !path_rec_start(dev, path)) &&
 686                    skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 687                 /* put pseudoheader back on for next time */
 688                 skb_push(skb, sizeof *phdr);
 689                 __skb_queue_tail(&path->queue, skb);
 690         } else {
 691                 ++dev->stats.tx_dropped;
 692                 dev_kfree_skb_any(skb);
 693         }
 694
 695         spin_unlock_irqrestore(&priv->lock, flags);
 696 }
 697
 698 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
 699 {
 700         struct ipoib_dev_priv *priv = netdev_priv(dev);
 701         struct ipoib_neigh *neigh;
 702         unsigned long flags;
 703
 704         if (likely(skb->dst && skb->dst->neighbour)) {
 705                 if (unlikely(!*to_ipoib_neigh(skb->dst->neighbour))) {
 706                         ipoib_path_lookup(skb, dev);
 707                         return NETDEV_TX_OK;
 708                 }
 709
 710                 neigh = *to_ipoib_neigh(skb->dst->neighbour);
 711
 712                 if (neigh->ah)
 713                         if (unlikely((memcmp(&neigh->dgid.raw,
 714                                             skb->dst->neighbour->ha + 4,
 715                                             sizeof(union ib_gid))) ||
 716                                          (neigh->dev != dev))) {
 717                                 spin_lock_irqsave(&priv->lock, flags);
 718                                 /*
 719                                  * It's safe to call ipoib_put_ah() inside
 720                                  * priv->lock here, because we know that
 721                                  * path->ah will always hold one more reference,
 722                                  * so ipoib_put_ah() will never do more than
 723                                  * decrement the ref count.
 724                                  */
 725                                 ipoib_put_ah(neigh->ah);
 726                                 list_del(&neigh->list);
 727                                 ipoib_neigh_free(dev, neigh);
 728                                 spin_unlock_irqrestore(&priv->lock, flags);
 729                                 ipoib_path_lookup(skb, dev);
 730                                 return NETDEV_TX_OK;
 731                         }
 732
 733                 if (ipoib_cm_get(neigh)) {
 734                         if (ipoib_cm_up(neigh)) {
 735                                 ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
 736                                 return NETDEV_TX_OK;
 737                         }
 738                 } else if (neigh->ah) {
 739                         ipoib_send(dev, skb, neigh->ah, IPOIB_QPN(skb->dst->neighbour->ha));
 740                         return NETDEV_TX_OK;
 741                 }
 742
 743                 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 744                         spin_lock_irqsave(&priv->lock, flags);
 745                         __skb_queue_tail(&neigh->queue, skb);
 746                         spin_unlock_irqrestore(&priv->lock, flags);
 747                 } else {
 748                         ++dev->stats.tx_dropped;
 749                         dev_kfree_skb_any(skb);
 750                 }
 751         } else {
 752                 struct ipoib_pseudoheader *phdr =
 753                         (struct ipoib_pseudoheader *) skb->data;
 754                 skb_pull(skb, sizeof *phdr);
 755
 756                 if (phdr->hwaddr[4] == 0xff) {
 757                         /* Add in the P_Key for multicast*/
 758                         phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
 759                         phdr->hwaddr[9] = priv->pkey & 0xff;
 760
 761                         ipoib_mcast_send(dev, phdr->hwaddr + 4, skb);
 762                 } else {
 763                         /* unicast GID -- should be ARP or RARP reply */
 764
 765                         if ((be16_to_cpup((__be16 *) skb->data) != ETH_P_ARP) &&
 766                             (be16_to_cpup((__be16 *) skb->data) != ETH_P_RARP)) {
 767                                 ipoib_warn(priv, "Unicast, no %s: type %04x, QPN %06x "
 768                                            IPOIB_GID_FMT "\n",
 769                                            skb->dst ? "neigh" : "dst",
 770                                            be16_to_cpup((__be16 *) skb->data),
 771                                            IPOIB_QPN(phdr->hwaddr),
 772                                            IPOIB_GID_RAW_ARG(phdr->hwaddr + 4));
 773                                 dev_kfree_skb_any(skb);
 774                                 ++dev->stats.tx_dropped;
 775                                 return NETDEV_TX_OK;
 776                         }
 777
 778                         unicast_arp_send(skb, dev, phdr);
 779                 }
 780         }
 781
 782         return NETDEV_TX_OK;
 783 }
 784
 785 static void ipoib_timeout(struct net_device *dev)
 786 {
 787         struct ipoib_dev_priv *priv = netdev_priv(dev);
 788
 789         ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
 790                    jiffies_to_msecs(jiffies - dev->trans_start));
 791         ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
 792                    netif_queue_stopped(dev),
 793                    priv->tx_head, priv->tx_tail);
 794         /* XXX reset QP, etc. */
 795 }
 796
 797 static int ipoib_hard_header(struct sk_buff *skb,
 798                              struct net_device *dev,
 799                              unsigned short type,
 800                              const void *daddr, const void *saddr, unsigned len)
 801 {
 802         struct ipoib_header *header;
 803
 804         header = (struct ipoib_header *) skb_push(skb, sizeof *header);
 805
 806         header->proto = htons(type);
 807         header->reserved = 0;
 808
 809         /*
 810          * If we don't have a neighbour structure, stuff the
 811          * destination address onto the front of the skb so we can
 812          * figure out where to send the packet later.
 813          */
 814         if ((!skb->dst || !skb->dst->neighbour) && daddr) {
 815                 struct ipoib_pseudoheader *phdr =
 816                         (struct ipoib_pseudoheader *) skb_push(skb, sizeof *phdr);
 817                 memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
 818         }
 819
 820         return 0;
 821 }
 822
 823 static void ipoib_set_mcast_list(struct net_device *dev)
 824 {
 825         struct ipoib_dev_priv *priv = netdev_priv(dev);
 826
 827         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
 828                 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
 829                 return;
 830         }
 831
 832         queue_work(ipoib_workqueue, &priv->restart_task);
 833 }
 834
 835 static void ipoib_neigh_cleanup(struct neighbour *n)
 836 {
 837         struct ipoib_neigh *neigh;
 838         struct ipoib_dev_priv *priv = netdev_priv(n->dev);
 839         unsigned long flags;
 840         struct ipoib_ah *ah = NULL;
 841
 842         neigh = *to_ipoib_neigh(n);
 843         if (neigh)
 844                 priv = netdev_priv(neigh->dev);
 845         else
 846                 return;
 847         ipoib_dbg(priv,
 848                   "neigh_cleanup for %06x " IPOIB_GID_FMT "\n",
 849                   IPOIB_QPN(n->ha),
 850                   IPOIB_GID_RAW_ARG(n->ha + 4));
 851
 852         spin_lock_irqsave(&priv->lock, flags);
 853
 854         if (neigh->ah)
 855                 ah = neigh->ah;
 856         list_del(&neigh->list);
 857         ipoib_neigh_free(n->dev, neigh);
 858
 859         spin_unlock_irqrestore(&priv->lock, flags);
 860
 861         if (ah)
 862                 ipoib_put_ah(ah);
 863 }
 864
 865 struct ipoib_neigh *ipoib_neigh_alloc(struct neighbour *neighbour,
 866                                       struct net_device *dev)
 867 {
 868         struct ipoib_neigh *neigh;
 869
 870         neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
 871         if (!neigh)
 872                 return NULL;
 873
 874         neigh->neighbour = neighbour;
 875         neigh->dev = dev;
 876         *to_ipoib_neigh(neighbour) = neigh;
 877         skb_queue_head_init(&neigh->queue);
 878         ipoib_cm_set(neigh, NULL);
 879
 880         return neigh;
 881 }
 882
 883 void ipoib_neigh_free(struct net_device *dev, struct ipoib_neigh *neigh)
 884 {
 885         struct sk_buff *skb;
 886         *to_ipoib_neigh(neigh->neighbour) = NULL;
 887         while ((skb = __skb_dequeue(&neigh->queue))) {
 888                 ++dev->stats.tx_dropped;
 889                 dev_kfree_skb_any(skb);
 890         }
 891         if (ipoib_cm_get(neigh))
 892                 ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
 893         kfree(neigh);
 894 }
 895
 896 static int ipoib_neigh_setup_dev(struct net_device *dev, struct neigh_parms *parms)
 897 {
 898         parms->neigh_cleanup = ipoib_neigh_cleanup;
 899
 900         return 0;
 901 }
 902
 903 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
 904 {
 905         struct ipoib_dev_priv *priv = netdev_priv(dev);
 906
 907         /* Allocate RX/TX "rings" to hold queued skbs */
 908         priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
 909                                 GFP_KERNEL);
 910         if (!priv->rx_ring) {
 911                 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
 912                        ca->name, ipoib_recvq_size);
 913                 goto out;
 914         }
 915
 916         priv->tx_ring = vmalloc(ipoib_sendq_size * sizeof *priv->tx_ring);
 917         if (!priv->tx_ring) {
 918                 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
 919                        ca->name, ipoib_sendq_size);
 920                 goto out_rx_ring_cleanup;
 921         }
 922         memset(priv->tx_ring, 0, ipoib_sendq_size * sizeof *priv->tx_ring);
 923
 924         /* priv->tx_head, tx_tail & tx_outstanding are already 0 */
 925
 926         if (ipoib_ib_dev_init(dev, ca, port))
 927                 goto out_tx_ring_cleanup;
 928
 929         return 0;
 930
 931 out_tx_ring_cleanup:
 932         vfree(priv->tx_ring);
 933
 934 out_rx_ring_cleanup:
 935         kfree(priv->rx_ring);
 936
 937 out:
 938         return -ENOMEM;
 939 }
 940
 941 void ipoib_dev_cleanup(struct net_device *dev)
 942 {
 943         struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
 944
 945         ipoib_delete_debug_files(dev);
 946
 947         /* Delete any child interfaces first */
 948         list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
 949                 unregister_netdev(cpriv->dev);
 950                 ipoib_dev_cleanup(cpriv->dev);
 951                 free_netdev(cpriv->dev);
 952         }
 953
 954         ipoib_ib_dev_cleanup(dev);
 955
 956         kfree(priv->rx_ring);
 957         vfree(priv->tx_ring);
 958
 959         priv->rx_ring = NULL;
 960         priv->tx_ring = NULL;
 961 }
 962
 963 static const struct header_ops ipoib_header_ops = {
 964         .create = ipoib_hard_header,
 965 };
 966
 967 static int get_skb_hdr(struct sk_buff *skb, void **iphdr,
 968                        void **tcph, u64 *hdr_flags, void *priv)
 969 {
 970         unsigned int ip_len;
 971         struct iphdr *iph;
 972
 973         if (unlikely(skb->protocol != htons(ETH_P_IP)))
 974                 return -1;
 975
 976         /*
 977          * In the future we may add an else clause that verifies the
 978          * checksum and allows devices which do not calculate checksum
 979          * to use LRO.
 980          */
 981         if (unlikely(skb->ip_summed != CHECKSUM_UNNECESSARY))
 982                 return -1;
 983
 984         /* Check for non-TCP packet */
 985         skb_reset_network_header(skb);
 986         iph = ip_hdr(skb);
 987         if (iph->protocol != IPPROTO_TCP)
 988                 return -1;
 989
 990         ip_len = ip_hdrlen(skb);
 991         skb_set_transport_header(skb, ip_len);
 992         *tcph = tcp_hdr(skb);
 993
 994         /* check if IP header and TCP header are complete */
 995         if (ntohs(iph->tot_len) < ip_len + tcp_hdrlen(skb))
 996                 return -1;
 997
 998         *hdr_flags = LRO_IPV4 | LRO_TCP;
 999         *iphdr = iph;
1000
1001         return 0;
1002 }
1003
1004 static void ipoib_lro_setup(struct ipoib_dev_priv *priv)
1005 {
1006         priv->lro.lro_mgr.max_aggr       = lro_max_aggr;
1007         priv->lro.lro_mgr.max_desc       = IPOIB_MAX_LRO_DESCRIPTORS;
1008         priv->lro.lro_mgr.lro_arr        = priv->lro.lro_desc;
1009         priv->lro.lro_mgr.get_skb_header = get_skb_hdr;
1010         priv->lro.lro_mgr.features       = LRO_F_NAPI;
1011         priv->lro.lro_mgr.dev            = priv->dev;
1012         priv->lro.lro_mgr.ip_summed_aggr = CHECKSUM_UNNECESSARY;
1013 }
1014
1015 static void ipoib_setup(struct net_device *dev)
1016 {
1017         struct ipoib_dev_priv *priv = netdev_priv(dev);
1018
1019         dev->open                = ipoib_open;
1020         dev->stop                = ipoib_stop;
1021         dev->change_mtu          = ipoib_change_mtu;
1022         dev->hard_start_xmit     = ipoib_start_xmit;
1023         dev->tx_timeout          = ipoib_timeout;
1024         dev->header_ops          = &ipoib_header_ops;
1025         dev->set_multicast_list  = ipoib_set_mcast_list;
1026         dev->neigh_setup         = ipoib_neigh_setup_dev;
1027
1028         ipoib_set_ethtool_ops(dev);
1029
1030         netif_napi_add(dev, &priv->napi, ipoib_poll, 100);
1031
1032         dev->watchdog_timeo      = HZ;
1033
1034         dev->flags              |= IFF_BROADCAST | IFF_MULTICAST;
1035
1036         /*
1037          * We add in INFINIBAND_ALEN to allow for the destination
1038          * address "pseudoheader" for skbs without neighbour struct.
1039          */
1040         dev->hard_header_len     = IPOIB_ENCAP_LEN + INFINIBAND_ALEN;
1041         dev->addr_len            = INFINIBAND_ALEN;
1042         dev->type                = ARPHRD_INFINIBAND;
1043         dev->tx_queue_len        = ipoib_sendq_size * 2;
1044         dev->features            = (NETIF_F_VLAN_CHALLENGED     |
1045                                     NETIF_F_HIGHDMA);
1046
1047         memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
1048
1049         netif_carrier_off(dev);
1050
1051         priv->dev = dev;
1052
1053         ipoib_lro_setup(priv);
1054
1055         spin_lock_init(&priv->lock);
1056
1057         mutex_init(&priv->vlan_mutex);
1058
1059         INIT_LIST_HEAD(&priv->path_list);
1060         INIT_LIST_HEAD(&priv->child_intfs);
1061         INIT_LIST_HEAD(&priv->dead_ahs);
1062         INIT_LIST_HEAD(&priv->multicast_list);
1063
1064         INIT_DELAYED_WORK(&priv->pkey_poll_task, ipoib_pkey_poll);
1065         INIT_DELAYED_WORK(&priv->mcast_task,   ipoib_mcast_join_task);
1066         INIT_WORK(&priv->carrier_on_task, ipoib_mcast_carrier_on_task);
1067         INIT_WORK(&priv->flush_light,   ipoib_ib_dev_flush_light);
1068         INIT_WORK(&priv->flush_normal,   ipoib_ib_dev_flush_normal);
1069         INIT_WORK(&priv->flush_heavy,   ipoib_ib_dev_flush_heavy);
1070         INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
1071         INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
1072 }
1073
1074 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
1075 {
1076         struct net_device *dev;
1077
1078         dev = alloc_netdev((int) sizeof (struct ipoib_dev_priv), name,
1079                            ipoib_setup);
1080         if (!dev)
1081                 return NULL;
1082
1083         return netdev_priv(dev);
1084 }
1085
1086 static ssize_t show_pkey(struct device *dev,
1087                          struct device_attribute *attr, char *buf)
1088 {
1089         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1090
1091         return sprintf(buf, "0x%04x\n", priv->pkey);
1092 }
1093 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1094
1095 static ssize_t show_umcast(struct device *dev,
1096                            struct device_attribute *attr, char *buf)
1097 {
1098         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1099
1100         return sprintf(buf, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
1101 }
1102
1103 static ssize_t set_umcast(struct device *dev,
1104                           struct device_attribute *attr,
1105                           const char *buf, size_t count)
1106 {
1107         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1108         unsigned long umcast_val = simple_strtoul(buf, NULL, 0);
1109
1110         if (umcast_val > 0) {
1111                 set_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1112                 ipoib_warn(priv, "ignoring multicast groups joined directly "
1113                                 "by userspace\n");
1114         } else
1115                 clear_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1116
1117         return count;
1118 }
1119 static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast);
1120
1121 int ipoib_add_umcast_attr(struct net_device *dev)
1122 {
1123         return device_create_file(&dev->dev, &dev_attr_umcast);
1124 }
1125
1126 static ssize_t create_child(struct device *dev,
1127                             struct device_attribute *attr,
1128                             const char *buf, size_t count)
1129 {
1130         int pkey;
1131         int ret;
1132
1133         if (sscanf(buf, "%i", &pkey) != 1)
1134                 return -EINVAL;
1135
1136         if (pkey < 0 || pkey > 0xffff)
1137                 return -EINVAL;
1138
1139         /*
1140          * Set the full membership bit, so that we join the right
1141          * broadcast group, etc.
1142          */
1143         pkey |= 0x8000;
1144
1145         ret = ipoib_vlan_add(to_net_dev(dev), pkey);
1146
1147         return ret ? ret : count;
1148 }
1149 static DEVICE_ATTR(create_child, S_IWUGO, NULL, create_child);
1150
1151 static ssize_t delete_child(struct device *dev,
1152                             struct device_attribute *attr,
1153                             const char *buf, size_t count)
1154 {
1155         int pkey;
1156         int ret;
1157
1158         if (sscanf(buf, "%i", &pkey) != 1)
1159                 return -EINVAL;
1160
1161         if (pkey < 0 || pkey > 0xffff)
1162                 return -EINVAL;
1163
1164         ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
1165
1166         return ret ? ret : count;
1167
1168 }
1169 static DEVICE_ATTR(delete_child, S_IWUGO, NULL, delete_child);
1170
1171 int ipoib_add_pkey_attr(struct net_device *dev)
1172 {
1173         return device_create_file(&dev->dev, &dev_attr_pkey);
1174 }
1175
1176 int ipoib_set_dev_features(struct ipoib_dev_priv *priv, struct ib_device *hca)
1177 {
1178         struct ib_device_attr *device_attr;
1179         int result = -ENOMEM;
1180
1181         device_attr = kmalloc(sizeof *device_attr, GFP_KERNEL);
1182         if (!device_attr) {
1183                 printk(KERN_WARNING "%s: allocation of %zu bytes failed\n",
1184                        hca->name, sizeof *device_attr);
1185                 return result;
1186         }
1187
1188         result = ib_query_device(hca, device_attr);
1189         if (result) {
1190                 printk(KERN_WARNING "%s: ib_query_device failed (ret = %d)\n",
1191                        hca->name, result);
1192                 kfree(device_attr);
1193                 return result;
1194         }
1195         priv->hca_caps = device_attr->device_cap_flags;
1196
1197         kfree(device_attr);
1198
1199         if (priv->hca_caps & IB_DEVICE_UD_IP_CSUM) {
1200                 set_bit(IPOIB_FLAG_CSUM, &priv->flags);
1201                 priv->dev->features |= NETIF_F_SG | NETIF_F_IP_CSUM;
1202         }
1203
1204         if (lro)
1205                 priv->dev->features |= NETIF_F_LRO;
1206
1207         if (priv->dev->features & NETIF_F_SG && priv->hca_caps & IB_DEVICE_UD_TSO)
1208                 priv->dev->features |= NETIF_F_TSO;
1209
1210         return 0;
1211 }
1212
1213
1214 static struct net_device *ipoib_add_port(const char *format,
1215                                          struct ib_device *hca, u8 port)
1216 {
1217         struct ipoib_dev_priv *priv;
1218         struct ib_port_attr attr;
1219         int result = -ENOMEM;
1220
1221         priv = ipoib_intf_alloc(format);
1222         if (!priv)
1223                 goto alloc_mem_failed;
1224
1225         SET_NETDEV_DEV(priv->dev, hca->dma_device);
1226
1227         if (!ib_query_port(hca, port, &attr))
1228                 priv->max_ib_mtu = ib_mtu_enum_to_int(attr.max_mtu);
1229         else {
1230                 printk(KERN_WARNING "%s: ib_query_port %d failed\n",
1231                        hca->name, port);
1232                 goto device_init_failed;
1233         }
1234
1235         /* MTU will be reset when mcast join happens */
1236         priv->dev->mtu  = IPOIB_UD_MTU(priv->max_ib_mtu);
1237         priv->mcast_mtu  = priv->admin_mtu = priv->dev->mtu;
1238
1239         result = ib_query_pkey(hca, port, 0, &priv->pkey);
1240         if (result) {
1241                 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
1242                        hca->name, port, result);
1243                 goto device_init_failed;
1244         }
1245
1246         if (ipoib_set_dev_features(priv, hca))
1247                 goto device_init_failed;
1248
1249         /*
1250          * Set the full membership bit, so that we join the right
1251          * broadcast group, etc.
1252          */
1253         priv->pkey |= 0x8000;
1254
1255         priv->dev->broadcast[8] = priv->pkey >> 8;
1256         priv->dev->broadcast[9] = priv->pkey & 0xff;
1257
1258         result = ib_query_gid(hca, port, 0, &priv->local_gid);
1259         if (result) {
1260                 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1261                        hca->name, port, result);
1262                 goto device_init_failed;
1263         } else
1264                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1265
1266         result = ipoib_dev_init(priv->dev, hca, port);
1267         if (result < 0) {
1268                 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1269                        hca->name, port, result);
1270                 goto device_init_failed;
1271         }
1272
1273         INIT_IB_EVENT_HANDLER(&priv->event_handler,
1274                               priv->ca, ipoib_event);
1275         result = ib_register_event_handler(&priv->event_handler);
1276         if (result < 0) {
1277                 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1278                        "port %d (ret = %d)\n",
1279                        hca->name, port, result);
1280                 goto event_failed;
1281         }
1282
1283         result = register_netdev(priv->dev);
1284         if (result) {
1285                 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1286                        hca->name, port, result);
1287                 goto register_failed;
1288         }
1289
1290         ipoib_create_debug_files(priv->dev);
1291
1292         if (ipoib_cm_add_mode_attr(priv->dev))
1293                 goto sysfs_failed;
1294         if (ipoib_add_pkey_attr(priv->dev))
1295                 goto sysfs_failed;
1296         if (ipoib_add_umcast_attr(priv->dev))
1297                 goto sysfs_failed;
1298         if (device_create_file(&priv->dev->dev, &dev_attr_create_child))
1299                 goto sysfs_failed;
1300         if (device_create_file(&priv->dev->dev, &dev_attr_delete_child))
1301                 goto sysfs_failed;
1302
1303         return priv->dev;
1304
1305 sysfs_failed:
1306         ipoib_delete_debug_files(priv->dev);
1307         unregister_netdev(priv->dev);
1308
1309 register_failed:
1310         ib_unregister_event_handler(&priv->event_handler);
1311         flush_workqueue(ipoib_workqueue);
1312
1313 event_failed:
1314         ipoib_dev_cleanup(priv->dev);
1315
1316 device_init_failed:
1317         free_netdev(priv->dev);
1318
1319 alloc_mem_failed:
1320         return ERR_PTR(result);
1321 }
1322
1323 static void ipoib_add_one(struct ib_device *device)
1324 {
1325         struct list_head *dev_list;
1326         struct net_device *dev;
1327         struct ipoib_dev_priv *priv;
1328         int s, e, p;
1329
1330         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1331                 return;
1332
1333         dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1334         if (!dev_list)
1335                 return;
1336
1337         INIT_LIST_HEAD(dev_list);
1338
1339         if (device->node_type == RDMA_NODE_IB_SWITCH) {
1340                 s = 0;
1341                 e = 0;
1342         } else {
1343                 s = 1;
1344                 e = device->phys_port_cnt;
1345         }
1346
1347         for (p = s; p <= e; ++p) {
1348                 dev = ipoib_add_port("ib%d", device, p);
1349                 if (!IS_ERR(dev)) {
1350                         priv = netdev_priv(dev);
1351                         list_add_tail(&priv->list, dev_list);
1352                 }
1353         }
1354
1355         ib_set_client_data(device, &ipoib_client, dev_list);
1356 }
1357
1358 static void ipoib_remove_one(struct ib_device *device)
1359 {
1360         struct ipoib_dev_priv *priv, *tmp;
1361         struct list_head *dev_list;
1362
1363         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1364                 return;
1365
1366         dev_list = ib_get_client_data(device, &ipoib_client);
1367
1368         list_for_each_entry_safe(priv, tmp, dev_list, list) {
1369                 ib_unregister_event_handler(&priv->event_handler);
1370
1371                 rtnl_lock();
1372                 dev_change_flags(priv->dev, priv->dev->flags & ~IFF_UP);
1373                 rtnl_unlock();
1374
1375                 flush_workqueue(ipoib_workqueue);
1376
1377                 unregister_netdev(priv->dev);
1378                 ipoib_dev_cleanup(priv->dev);
1379                 free_netdev(priv->dev);
1380         }
1381
1382         kfree(dev_list);
1383 }
1384
1385 static int __init ipoib_init_module(void)
1386 {
1387         int ret;
1388
1389         ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1390         ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1391         ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1392
1393         ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1394         ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1395         ipoib_sendq_size = max(ipoib_sendq_size, max(2 * MAX_SEND_CQE,
1396                                                      IPOIB_MIN_QUEUE_SIZE));
1397 #ifdef CONFIG_INFINIBAND_IPOIB_CM
1398         ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
1399 #endif
1400
1401         /*
1402          * When copying small received packets, we only copy from the
1403          * linear data part of the SKB, so we rely on this condition.
1404          */
1405         BUILD_BUG_ON(IPOIB_CM_COPYBREAK > IPOIB_CM_HEAD_SIZE);
1406
1407         ret = ipoib_register_debugfs();
1408         if (ret)
1409                 return ret;
1410
1411         /*
1412          * We create our own workqueue mainly because we want to be
1413          * able to flush it when devices are being removed.  We can't
1414          * use schedule_work()/flush_scheduled_work() because both
1415          * unregister_netdev() and linkwatch_event take the rtnl lock,
1416          * so flush_scheduled_work() can deadlock during device
1417          * removal.
1418          */
1419         ipoib_workqueue = create_singlethread_workqueue("ipoib");
1420         if (!ipoib_workqueue) {
1421                 ret = -ENOMEM;
1422                 goto err_fs;
1423         }
1424
1425         ib_sa_register_client(&ipoib_sa_client);
1426
1427         ret = ib_register_client(&ipoib_client);
1428         if (ret)
1429                 goto err_sa;
1430
1431         return 0;
1432
1433 err_sa:
1434         ib_sa_unregister_client(&ipoib_sa_client);
1435         destroy_workqueue(ipoib_workqueue);
1436
1437 err_fs:
1438         ipoib_unregister_debugfs();
1439
1440         return ret;
1441 }
1442
1443 static void __exit ipoib_cleanup_module(void)
1444 {
1445         ib_unregister_client(&ipoib_client);
1446         ib_sa_unregister_client(&ipoib_sa_client);
1447         ipoib_unregister_debugfs();
1448         destroy_workqueue(ipoib_workqueue);
1449 }
1450
1451 module_init(ipoib_init_module);
1452 module_exit(ipoib_cleanup_module);