drivers/infiniband/ulp/ipoib/ipoib_main.c

   1 /*
   2  * Copyright (c) 2004 Topspin Communications.  All rights reserved.
   3  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
   4  * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
   5  *
   6  * This software is available to you under a choice of one of two
   7  * licenses.  You may choose to be licensed under the terms of the GNU
   8  * General Public License (GPL) Version 2, available from the file
   9  * COPYING in the main directory of this source tree, or the
  10  * OpenIB.org BSD license below:
  11  *
  12  *     Redistribution and use in source and binary forms, with or
  13  *     without modification, are permitted provided that the following
  14  *     conditions are met:
  15  *
  16  *      - Redistributions of source code must retain the above
  17  *        copyright notice, this list of conditions and the following
  18  *        disclaimer.
  19  *
  20  *      - Redistributions in binary form must reproduce the above
  21  *        copyright notice, this list of conditions and the following
  22  *        disclaimer in the documentation and/or other materials
  23  *        provided with the distribution.
  24  *
  25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  32  * SOFTWARE.
  33  */
  34
  35 #include "ipoib.h"
  36
  37 #include <linux/module.h>
  38
  39 #include <linux/init.h>
  40 #include <linux/slab.h>
  41 #include <linux/kernel.h>
  42 #include <linux/vmalloc.h>
  43
  44 #include <linux/if_arp.h>       /* For ARPHRD_xxx */
  45
  46 #include <linux/ip.h>
  47 #include <linux/in.h>
  48
  49 #include <net/dst.h>
  50
  51 MODULE_AUTHOR("Roland Dreier");
  52 MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
  53 MODULE_LICENSE("Dual BSD/GPL");
  54
  55 int ipoib_sendq_size __read_mostly = IPOIB_TX_RING_SIZE;
  56 int ipoib_recvq_size __read_mostly = IPOIB_RX_RING_SIZE;
  57
  58 module_param_named(send_queue_size, ipoib_sendq_size, int, 0444);
  59 MODULE_PARM_DESC(send_queue_size, "Number of descriptors in send queue");
  60 module_param_named(recv_queue_size, ipoib_recvq_size, int, 0444);
  61 MODULE_PARM_DESC(recv_queue_size, "Number of descriptors in receive queue");
  62
  63 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
  64 int ipoib_debug_level;
  65
  66 module_param_named(debug_level, ipoib_debug_level, int, 0644);
  67 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
  68 #endif
  69
  70 struct ipoib_path_iter {
  71         struct net_device *dev;
  72         struct ipoib_path  path;
  73 };
  74
  75 static const u8 ipv4_bcast_addr[] = {
  76         0x00, 0xff, 0xff, 0xff,
  77         0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
  78         0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff
  79 };
  80
  81 struct workqueue_struct *ipoib_workqueue;
  82
  83 struct ib_sa_client ipoib_sa_client;
  84
  85 static void ipoib_add_one(struct ib_device *device);
  86 static void ipoib_remove_one(struct ib_device *device);
  87
  88 static struct ib_client ipoib_client = {
  89         .name   = "ipoib",
  90         .add    = ipoib_add_one,
  91         .remove = ipoib_remove_one
  92 };
  93
  94 int ipoib_open(struct net_device *dev)
  95 {
  96         struct ipoib_dev_priv *priv = netdev_priv(dev);
  97
  98         ipoib_dbg(priv, "bringing up interface\n");
  99
 100         set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 101
 102         if (ipoib_pkey_dev_delay_open(dev))
 103                 return 0;
 104
 105         if (ipoib_ib_dev_open(dev))
 106                 goto err_disable;
 107
 108         if (ipoib_ib_dev_up(dev))
 109                 goto err_stop;
 110
 111         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 112                 struct ipoib_dev_priv *cpriv;
 113
 114                 /* Bring up any child interfaces too */
 115                 mutex_lock(&priv->vlan_mutex);
 116                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 117                         int flags;
 118
 119                         flags = cpriv->dev->flags;
 120                         if (flags & IFF_UP)
 121                                 continue;
 122
 123                         dev_change_flags(cpriv->dev, flags | IFF_UP);
 124                 }
 125                 mutex_unlock(&priv->vlan_mutex);
 126         }
 127
 128         netif_start_queue(dev);
 129
 130         return 0;
 131
 132 err_stop:
 133         ipoib_ib_dev_stop(dev, 1);
 134
 135 err_disable:
 136         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 137
 138         return -EINVAL;
 139 }
 140
 141 static int ipoib_stop(struct net_device *dev)
 142 {
 143         struct ipoib_dev_priv *priv = netdev_priv(dev);
 144
 145         ipoib_dbg(priv, "stopping interface\n");
 146
 147         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 148
 149         netif_stop_queue(dev);
 150
 151         ipoib_ib_dev_down(dev, 0);
 152         ipoib_ib_dev_stop(dev, 0);
 153
 154         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 155                 struct ipoib_dev_priv *cpriv;
 156
 157                 /* Bring down any child interfaces too */
 158                 mutex_lock(&priv->vlan_mutex);
 159                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 160                         int flags;
 161
 162                         flags = cpriv->dev->flags;
 163                         if (!(flags & IFF_UP))
 164                                 continue;
 165
 166                         dev_change_flags(cpriv->dev, flags & ~IFF_UP);
 167                 }
 168                 mutex_unlock(&priv->vlan_mutex);
 169         }
 170
 171         return 0;
 172 }
 173
 174 static u32 ipoib_fix_features(struct net_device *dev, u32 features)
 175 {
 176         struct ipoib_dev_priv *priv = netdev_priv(dev);
 177
 178         if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
 179                 features &= ~(NETIF_F_SG | NETIF_F_IP_CSUM | NETIF_F_TSO);
 180
 181         return features;
 182 }
 183
 184 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
 185 {
 186         struct ipoib_dev_priv *priv = netdev_priv(dev);
 187
 188         /* dev->mtu > 2K ==> connected mode */
 189         if (ipoib_cm_admin_enabled(dev)) {
 190                 if (new_mtu > ipoib_cm_max_mtu(dev))
 191                         return -EINVAL;
 192
 193                 if (new_mtu > priv->mcast_mtu)
 194                         ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
 195                                    priv->mcast_mtu);
 196
 197                 dev->mtu = new_mtu;
 198                 return 0;
 199         }
 200
 201         if (new_mtu > IPOIB_UD_MTU(priv->max_ib_mtu))
 202                 return -EINVAL;
 203
 204         priv->admin_mtu = new_mtu;
 205
 206         dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
 207
 208         return 0;
 209 }
 210
 211 static struct ipoib_path *__path_find(struct net_device *dev, void *gid)
 212 {
 213         struct ipoib_dev_priv *priv = netdev_priv(dev);
 214         struct rb_node *n = priv->path_tree.rb_node;
 215         struct ipoib_path *path;
 216         int ret;
 217
 218         while (n) {
 219                 path = rb_entry(n, struct ipoib_path, rb_node);
 220
 221                 ret = memcmp(gid, path->pathrec.dgid.raw,
 222                              sizeof (union ib_gid));
 223
 224                 if (ret < 0)
 225                         n = n->rb_left;
 226                 else if (ret > 0)
 227                         n = n->rb_right;
 228                 else
 229                         return path;
 230         }
 231
 232         return NULL;
 233 }
 234
 235 static int __path_add(struct net_device *dev, struct ipoib_path *path)
 236 {
 237         struct ipoib_dev_priv *priv = netdev_priv(dev);
 238         struct rb_node **n = &priv->path_tree.rb_node;
 239         struct rb_node *pn = NULL;
 240         struct ipoib_path *tpath;
 241         int ret;
 242
 243         while (*n) {
 244                 pn = *n;
 245                 tpath = rb_entry(pn, struct ipoib_path, rb_node);
 246
 247                 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
 248                              sizeof (union ib_gid));
 249                 if (ret < 0)
 250                         n = &pn->rb_left;
 251                 else if (ret > 0)
 252                         n = &pn->rb_right;
 253                 else
 254                         return -EEXIST;
 255         }
 256
 257         rb_link_node(&path->rb_node, pn, n);
 258         rb_insert_color(&path->rb_node, &priv->path_tree);
 259
 260         list_add_tail(&path->list, &priv->path_list);
 261
 262         return 0;
 263 }
 264
 265 static void path_free(struct net_device *dev, struct ipoib_path *path)
 266 {
 267         struct ipoib_dev_priv *priv = netdev_priv(dev);
 268         struct ipoib_neigh *neigh, *tn;
 269         struct sk_buff *skb;
 270         unsigned long flags;
 271
 272         while ((skb = __skb_dequeue(&path->queue)))
 273                 dev_kfree_skb_irq(skb);
 274
 275         spin_lock_irqsave(&priv->lock, flags);
 276
 277         list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
 278                 /*
 279                  * It's safe to call ipoib_put_ah() inside priv->lock
 280                  * here, because we know that path->ah will always
 281                  * hold one more reference, so ipoib_put_ah() will
 282                  * never do more than decrement the ref count.
 283                  */
 284                 if (neigh->ah)
 285                         ipoib_put_ah(neigh->ah);
 286
 287                 ipoib_neigh_free(dev, neigh);
 288         }
 289
 290         spin_unlock_irqrestore(&priv->lock, flags);
 291
 292         if (path->ah)
 293                 ipoib_put_ah(path->ah);
 294
 295         kfree(path);
 296 }
 297
 298 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 299
 300 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
 301 {
 302         struct ipoib_path_iter *iter;
 303
 304         iter = kmalloc(sizeof *iter, GFP_KERNEL);
 305         if (!iter)
 306                 return NULL;
 307
 308         iter->dev = dev;
 309         memset(iter->path.pathrec.dgid.raw, 0, 16);
 310
 311         if (ipoib_path_iter_next(iter)) {
 312                 kfree(iter);
 313                 return NULL;
 314         }
 315
 316         return iter;
 317 }
 318
 319 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
 320 {
 321         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
 322         struct rb_node *n;
 323         struct ipoib_path *path;
 324         int ret = 1;
 325
 326         spin_lock_irq(&priv->lock);
 327
 328         n = rb_first(&priv->path_tree);
 329
 330         while (n) {
 331                 path = rb_entry(n, struct ipoib_path, rb_node);
 332
 333                 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
 334                            sizeof (union ib_gid)) < 0) {
 335                         iter->path = *path;
 336                         ret = 0;
 337                         break;
 338                 }
 339
 340                 n = rb_next(n);
 341         }
 342
 343         spin_unlock_irq(&priv->lock);
 344
 345         return ret;
 346 }
 347
 348 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
 349                           struct ipoib_path *path)
 350 {
 351         *path = iter->path;
 352 }
 353
 354 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
 355
 356 void ipoib_mark_paths_invalid(struct net_device *dev)
 357 {
 358         struct ipoib_dev_priv *priv = netdev_priv(dev);
 359         struct ipoib_path *path, *tp;
 360
 361         spin_lock_irq(&priv->lock);
 362
 363         list_for_each_entry_safe(path, tp, &priv->path_list, list) {
 364                 ipoib_dbg(priv, "mark path LID 0x%04x GID %pI6 invalid\n",
 365                         be16_to_cpu(path->pathrec.dlid),
 366                         path->pathrec.dgid.raw);
 367                 path->valid =  0;
 368         }
 369
 370         spin_unlock_irq(&priv->lock);
 371 }
 372
 373 void ipoib_flush_paths(struct net_device *dev)
 374 {
 375         struct ipoib_dev_priv *priv = netdev_priv(dev);
 376         struct ipoib_path *path, *tp;
 377         LIST_HEAD(remove_list);
 378         unsigned long flags;
 379
 380         netif_tx_lock_bh(dev);
 381         spin_lock_irqsave(&priv->lock, flags);
 382
 383         list_splice_init(&priv->path_list, &remove_list);
 384
 385         list_for_each_entry(path, &remove_list, list)
 386                 rb_erase(&path->rb_node, &priv->path_tree);
 387
 388         list_for_each_entry_safe(path, tp, &remove_list, list) {
 389                 if (path->query)
 390                         ib_sa_cancel_query(path->query_id, path->query);
 391                 spin_unlock_irqrestore(&priv->lock, flags);
 392                 netif_tx_unlock_bh(dev);
 393                 wait_for_completion(&path->done);
 394                 path_free(dev, path);
 395                 netif_tx_lock_bh(dev);
 396                 spin_lock_irqsave(&priv->lock, flags);
 397         }
 398
 399         spin_unlock_irqrestore(&priv->lock, flags);
 400         netif_tx_unlock_bh(dev);
 401 }
 402
 403 static void path_rec_completion(int status,
 404                                 struct ib_sa_path_rec *pathrec,
 405                                 void *path_ptr)
 406 {
 407         struct ipoib_path *path = path_ptr;
 408         struct net_device *dev = path->dev;
 409         struct ipoib_dev_priv *priv = netdev_priv(dev);
 410         struct ipoib_ah *ah = NULL;
 411         struct ipoib_ah *old_ah = NULL;
 412         struct ipoib_neigh *neigh, *tn;
 413         struct sk_buff_head skqueue;
 414         struct sk_buff *skb;
 415         unsigned long flags;
 416
 417         if (!status)
 418                 ipoib_dbg(priv, "PathRec LID 0x%04x for GID %pI6\n",
 419                           be16_to_cpu(pathrec->dlid), pathrec->dgid.raw);
 420         else
 421                 ipoib_dbg(priv, "PathRec status %d for GID %pI6\n",
 422                           status, path->pathrec.dgid.raw);
 423
 424         skb_queue_head_init(&skqueue);
 425
 426         if (!status) {
 427                 struct ib_ah_attr av;
 428
 429                 if (!ib_init_ah_from_path(priv->ca, priv->port, pathrec, &av))
 430                         ah = ipoib_create_ah(dev, priv->pd, &av);
 431         }
 432
 433         spin_lock_irqsave(&priv->lock, flags);
 434
 435         if (ah) {
 436                 path->pathrec = *pathrec;
 437
 438                 old_ah   = path->ah;
 439                 path->ah = ah;
 440
 441                 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
 442                           ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
 443
 444                 while ((skb = __skb_dequeue(&path->queue)))
 445                         __skb_queue_tail(&skqueue, skb);
 446
 447                 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
 448                         if (neigh->ah) {
 449                                 WARN_ON(neigh->ah != old_ah);
 450                                 /*
 451                                  * Dropping the ah reference inside
 452                                  * priv->lock is safe here, because we
 453                                  * will hold one more reference from
 454                                  * the original value of path->ah (ie
 455                                  * old_ah).
 456                                  */
 457                                 ipoib_put_ah(neigh->ah);
 458                         }
 459                         kref_get(&path->ah->ref);
 460                         neigh->ah = path->ah;
 461                         memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
 462                                sizeof(union ib_gid));
 463
 464                         if (ipoib_cm_enabled(dev, neigh->neighbour)) {
 465                                 if (!ipoib_cm_get(neigh))
 466                                         ipoib_cm_set(neigh, ipoib_cm_create_tx(dev,
 467                                                                                path,
 468                                                                                neigh));
 469                                 if (!ipoib_cm_get(neigh)) {
 470                                         list_del(&neigh->list);
 471                                         if (neigh->ah)
 472                                                 ipoib_put_ah(neigh->ah);
 473                                         ipoib_neigh_free(dev, neigh);
 474                                         continue;
 475                                 }
 476                         }
 477
 478                         while ((skb = __skb_dequeue(&neigh->queue)))
 479                                 __skb_queue_tail(&skqueue, skb);
 480                 }
 481                 path->valid = 1;
 482         }
 483
 484         path->query = NULL;
 485         complete(&path->done);
 486
 487         spin_unlock_irqrestore(&priv->lock, flags);
 488
 489         if (old_ah)
 490                 ipoib_put_ah(old_ah);
 491
 492         while ((skb = __skb_dequeue(&skqueue))) {
 493                 skb->dev = dev;
 494                 if (dev_queue_xmit(skb))
 495                         ipoib_warn(priv, "dev_queue_xmit failed "
 496                                    "to requeue packet\n");
 497         }
 498 }
 499
 500 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
 501 {
 502         struct ipoib_dev_priv *priv = netdev_priv(dev);
 503         struct ipoib_path *path;
 504
 505         if (!priv->broadcast)
 506                 return NULL;
 507
 508         path = kzalloc(sizeof *path, GFP_ATOMIC);
 509         if (!path)
 510                 return NULL;
 511
 512         path->dev = dev;
 513
 514         skb_queue_head_init(&path->queue);
 515
 516         INIT_LIST_HEAD(&path->neigh_list);
 517
 518         memcpy(path->pathrec.dgid.raw, gid, sizeof (union ib_gid));
 519         path->pathrec.sgid          = priv->local_gid;
 520         path->pathrec.pkey          = cpu_to_be16(priv->pkey);
 521         path->pathrec.numb_path     = 1;
 522         path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
 523
 524         return path;
 525 }
 526
 527 static int path_rec_start(struct net_device *dev,
 528                           struct ipoib_path *path)
 529 {
 530         struct ipoib_dev_priv *priv = netdev_priv(dev);
 531
 532         ipoib_dbg(priv, "Start path record lookup for %pI6\n",
 533                   path->pathrec.dgid.raw);
 534
 535         init_completion(&path->done);
 536
 537         path->query_id =
 538                 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
 539                                    &path->pathrec,
 540                                    IB_SA_PATH_REC_DGID          |
 541                                    IB_SA_PATH_REC_SGID          |
 542                                    IB_SA_PATH_REC_NUMB_PATH     |
 543                                    IB_SA_PATH_REC_TRAFFIC_CLASS |
 544                                    IB_SA_PATH_REC_PKEY,
 545                                    1000, GFP_ATOMIC,
 546                                    path_rec_completion,
 547                                    path, &path->query);
 548         if (path->query_id < 0) {
 549                 ipoib_warn(priv, "ib_sa_path_rec_get failed: %d\n", path->query_id);
 550                 path->query = NULL;
 551                 complete(&path->done);
 552                 return path->query_id;
 553         }
 554
 555         return 0;
 556 }
 557
 558 static void neigh_add_path(struct sk_buff *skb, struct net_device *dev)
 559 {
 560         struct ipoib_dev_priv *priv = netdev_priv(dev);
 561         struct ipoib_path *path;
 562         struct ipoib_neigh *neigh;
 563         struct neighbour *n;
 564         unsigned long flags;
 565
 566         n = dst_get_neighbour(skb_dst(skb));
 567         neigh = ipoib_neigh_alloc(n, skb->dev);
 568         if (!neigh) {
 569                 ++dev->stats.tx_dropped;
 570                 dev_kfree_skb_any(skb);
 571                 return;
 572         }
 573
 574         spin_lock_irqsave(&priv->lock, flags);
 575
 576         path = __path_find(dev, n->ha + 4);
 577         if (!path) {
 578                 path = path_rec_create(dev, n->ha + 4);
 579                 if (!path)
 580                         goto err_path;
 581
 582                 __path_add(dev, path);
 583         }
 584
 585         list_add_tail(&neigh->list, &path->neigh_list);
 586
 587         if (path->ah) {
 588                 kref_get(&path->ah->ref);
 589                 neigh->ah = path->ah;
 590                 memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
 591                        sizeof(union ib_gid));
 592
 593                 if (ipoib_cm_enabled(dev, neigh->neighbour)) {
 594                         if (!ipoib_cm_get(neigh))
 595                                 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
 596                         if (!ipoib_cm_get(neigh)) {
 597                                 list_del(&neigh->list);
 598                                 if (neigh->ah)
 599                                         ipoib_put_ah(neigh->ah);
 600                                 ipoib_neigh_free(dev, neigh);
 601                                 goto err_drop;
 602                         }
 603                         if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE)
 604                                 __skb_queue_tail(&neigh->queue, skb);
 605                         else {
 606                                 ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
 607                                            skb_queue_len(&neigh->queue));
 608                                 goto err_drop;
 609                         }
 610                 } else {
 611                         spin_unlock_irqrestore(&priv->lock, flags);
 612                         ipoib_send(dev, skb, path->ah, IPOIB_QPN(n->ha));
 613                         return;
 614                 }
 615         } else {
 616                 neigh->ah  = NULL;
 617
 618                 if (!path->query && path_rec_start(dev, path))
 619                         goto err_list;
 620
 621                 __skb_queue_tail(&neigh->queue, skb);
 622         }
 623
 624         spin_unlock_irqrestore(&priv->lock, flags);
 625         return;
 626
 627 err_list:
 628         list_del(&neigh->list);
 629
 630 err_path:
 631         ipoib_neigh_free(dev, neigh);
 632 err_drop:
 633         ++dev->stats.tx_dropped;
 634         dev_kfree_skb_any(skb);
 635
 636         spin_unlock_irqrestore(&priv->lock, flags);
 637 }
 638
 639 static void ipoib_path_lookup(struct sk_buff *skb, struct net_device *dev)
 640 {
 641         struct ipoib_dev_priv *priv = netdev_priv(skb->dev);
 642         struct dst_entry *dst = skb_dst(skb);
 643         struct neighbour *n;
 644
 645         /* Look up path record for unicasts */
 646         n = dst_get_neighbour(dst);
 647         if (n->ha[4] != 0xff) {
 648                 neigh_add_path(skb, dev);
 649                 return;
 650         }
 651
 652         /* Add in the P_Key for multicasts */
 653         n->ha[8] = (priv->pkey >> 8) & 0xff;
 654         n->ha[9] = priv->pkey & 0xff;
 655         ipoib_mcast_send(dev, n->ha + 4, skb);
 656 }
 657
 658 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
 659                              struct ipoib_pseudoheader *phdr)
 660 {
 661         struct ipoib_dev_priv *priv = netdev_priv(dev);
 662         struct ipoib_path *path;
 663         unsigned long flags;
 664
 665         spin_lock_irqsave(&priv->lock, flags);
 666
 667         path = __path_find(dev, phdr->hwaddr + 4);
 668         if (!path || !path->valid) {
 669                 int new_path = 0;
 670
 671                 if (!path) {
 672                         path = path_rec_create(dev, phdr->hwaddr + 4);
 673                         new_path = 1;
 674                 }
 675                 if (path) {
 676                         /* put pseudoheader back on for next time */
 677                         skb_push(skb, sizeof *phdr);
 678                         __skb_queue_tail(&path->queue, skb);
 679
 680                         if (!path->query && path_rec_start(dev, path)) {
 681                                 spin_unlock_irqrestore(&priv->lock, flags);
 682                                 if (new_path)
 683                                         path_free(dev, path);
 684                                 return;
 685                         } else
 686                                 __path_add(dev, path);
 687                 } else {
 688                         ++dev->stats.tx_dropped;
 689                         dev_kfree_skb_any(skb);
 690                 }
 691
 692                 spin_unlock_irqrestore(&priv->lock, flags);
 693                 return;
 694         }
 695
 696         if (path->ah) {
 697                 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
 698                           be16_to_cpu(path->pathrec.dlid));
 699
 700                 spin_unlock_irqrestore(&priv->lock, flags);
 701                 ipoib_send(dev, skb, path->ah, IPOIB_QPN(phdr->hwaddr));
 702                 return;
 703         } else if ((path->query || !path_rec_start(dev, path)) &&
 704                    skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 705                 /* put pseudoheader back on for next time */
 706                 skb_push(skb, sizeof *phdr);
 707                 __skb_queue_tail(&path->queue, skb);
 708         } else {
 709                 ++dev->stats.tx_dropped;
 710                 dev_kfree_skb_any(skb);
 711         }
 712
 713         spin_unlock_irqrestore(&priv->lock, flags);
 714 }
 715
 716 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
 717 {
 718         struct ipoib_dev_priv *priv = netdev_priv(dev);
 719         struct ipoib_neigh *neigh;
 720         struct neighbour *n;
 721         unsigned long flags;
 722
 723         n = dst_get_neighbour(skb_dst(skb));
 724         if (likely(skb_dst(skb) && n)) {
 725                 if (unlikely(!*to_ipoib_neigh(n))) {
 726                         ipoib_path_lookup(skb, dev);
 727                         return NETDEV_TX_OK;
 728                 }
 729
 730                 neigh = *to_ipoib_neigh(n);
 731
 732                 if (unlikely((memcmp(&neigh->dgid.raw,
 733                                      n->ha + 4,
 734                                      sizeof(union ib_gid))) ||
 735                              (neigh->dev != dev))) {
 736                         spin_lock_irqsave(&priv->lock, flags);
 737                         /*
 738                          * It's safe to call ipoib_put_ah() inside
 739                          * priv->lock here, because we know that
 740                          * path->ah will always hold one more reference,
 741                          * so ipoib_put_ah() will never do more than
 742                          * decrement the ref count.
 743                          */
 744                         if (neigh->ah)
 745                                 ipoib_put_ah(neigh->ah);
 746                         list_del(&neigh->list);
 747                         ipoib_neigh_free(dev, neigh);
 748                         spin_unlock_irqrestore(&priv->lock, flags);
 749                         ipoib_path_lookup(skb, dev);
 750                         return NETDEV_TX_OK;
 751                 }
 752
 753                 if (ipoib_cm_get(neigh)) {
 754                         if (ipoib_cm_up(neigh)) {
 755                                 ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
 756                                 return NETDEV_TX_OK;
 757                         }
 758                 } else if (neigh->ah) {
 759                         ipoib_send(dev, skb, neigh->ah, IPOIB_QPN(n->ha));
 760                         return NETDEV_TX_OK;
 761                 }
 762
 763                 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 764                         spin_lock_irqsave(&priv->lock, flags);
 765                         __skb_queue_tail(&neigh->queue, skb);
 766                         spin_unlock_irqrestore(&priv->lock, flags);
 767                 } else {
 768                         ++dev->stats.tx_dropped;
 769                         dev_kfree_skb_any(skb);
 770                 }
 771         } else {
 772                 struct ipoib_pseudoheader *phdr =
 773                         (struct ipoib_pseudoheader *) skb->data;
 774                 skb_pull(skb, sizeof *phdr);
 775
 776                 if (phdr->hwaddr[4] == 0xff) {
 777                         /* Add in the P_Key for multicast*/
 778                         phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
 779                         phdr->hwaddr[9] = priv->pkey & 0xff;
 780
 781                         ipoib_mcast_send(dev, phdr->hwaddr + 4, skb);
 782                 } else {
 783                         /* unicast GID -- should be ARP or RARP reply */
 784
 785                         if ((be16_to_cpup((__be16 *) skb->data) != ETH_P_ARP) &&
 786                             (be16_to_cpup((__be16 *) skb->data) != ETH_P_RARP)) {
 787                                 ipoib_warn(priv, "Unicast, no %s: type %04x, QPN %06x %pI6\n",
 788                                            skb_dst(skb) ? "neigh" : "dst",
 789                                            be16_to_cpup((__be16 *) skb->data),
 790                                            IPOIB_QPN(phdr->hwaddr),
 791                                            phdr->hwaddr + 4);
 792                                 dev_kfree_skb_any(skb);
 793                                 ++dev->stats.tx_dropped;
 794                                 return NETDEV_TX_OK;
 795                         }
 796
 797                         unicast_arp_send(skb, dev, phdr);
 798                 }
 799         }
 800
 801         return NETDEV_TX_OK;
 802 }
 803
 804 static void ipoib_timeout(struct net_device *dev)
 805 {
 806         struct ipoib_dev_priv *priv = netdev_priv(dev);
 807
 808         ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
 809                    jiffies_to_msecs(jiffies - dev->trans_start));
 810         ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
 811                    netif_queue_stopped(dev),
 812                    priv->tx_head, priv->tx_tail);
 813         /* XXX reset QP, etc. */
 814 }
 815
 816 static int ipoib_hard_header(struct sk_buff *skb,
 817                              struct net_device *dev,
 818                              unsigned short type,
 819                              const void *daddr, const void *saddr, unsigned len)
 820 {
 821         struct ipoib_header *header;
 822         struct dst_entry *dst;
 823         struct neighbour *n;
 824
 825         header = (struct ipoib_header *) skb_push(skb, sizeof *header);
 826
 827         header->proto = htons(type);
 828         header->reserved = 0;
 829
 830         /*
 831          * If we don't have a neighbour structure, stuff the
 832          * destination address onto the front of the skb so we can
 833          * figure out where to send the packet later.
 834          */
 835         dst = skb_dst(skb);
 836         n = NULL;
 837         if (dst)
 838                 n = dst_get_neighbour(dst);
 839         if ((!dst || !n) && daddr) {
 840                 struct ipoib_pseudoheader *phdr =
 841                         (struct ipoib_pseudoheader *) skb_push(skb, sizeof *phdr);
 842                 memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
 843         }
 844
 845         return 0;
 846 }
 847
 848 static void ipoib_set_mcast_list(struct net_device *dev)
 849 {
 850         struct ipoib_dev_priv *priv = netdev_priv(dev);
 851
 852         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
 853                 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
 854                 return;
 855         }
 856
 857         queue_work(ipoib_workqueue, &priv->restart_task);
 858 }
 859
 860 static void ipoib_neigh_cleanup(struct neighbour *n)
 861 {
 862         struct ipoib_neigh *neigh;
 863         struct ipoib_dev_priv *priv = netdev_priv(n->dev);
 864         unsigned long flags;
 865         struct ipoib_ah *ah = NULL;
 866
 867         neigh = *to_ipoib_neigh(n);
 868         if (neigh)
 869                 priv = netdev_priv(neigh->dev);
 870         else
 871                 return;
 872         ipoib_dbg(priv,
 873                   "neigh_cleanup for %06x %pI6\n",
 874                   IPOIB_QPN(n->ha),
 875                   n->ha + 4);
 876
 877         spin_lock_irqsave(&priv->lock, flags);
 878
 879         if (neigh->ah)
 880                 ah = neigh->ah;
 881         list_del(&neigh->list);
 882         ipoib_neigh_free(n->dev, neigh);
 883
 884         spin_unlock_irqrestore(&priv->lock, flags);
 885
 886         if (ah)
 887                 ipoib_put_ah(ah);
 888 }
 889
 890 struct ipoib_neigh *ipoib_neigh_alloc(struct neighbour *neighbour,
 891                                       struct net_device *dev)
 892 {
 893         struct ipoib_neigh *neigh;
 894
 895         neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
 896         if (!neigh)
 897                 return NULL;
 898
 899         neigh->neighbour = neighbour;
 900         neigh->dev = dev;
 901         memset(&neigh->dgid.raw, 0, sizeof (union ib_gid));
 902         *to_ipoib_neigh(neighbour) = neigh;
 903         skb_queue_head_init(&neigh->queue);
 904         ipoib_cm_set(neigh, NULL);
 905
 906         return neigh;
 907 }
 908
 909 void ipoib_neigh_free(struct net_device *dev, struct ipoib_neigh *neigh)
 910 {
 911         struct sk_buff *skb;
 912         *to_ipoib_neigh(neigh->neighbour) = NULL;
 913         while ((skb = __skb_dequeue(&neigh->queue))) {
 914                 ++dev->stats.tx_dropped;
 915                 dev_kfree_skb_any(skb);
 916         }
 917         if (ipoib_cm_get(neigh))
 918                 ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
 919         kfree(neigh);
 920 }
 921
 922 static int ipoib_neigh_setup_dev(struct net_device *dev, struct neigh_parms *parms)
 923 {
 924         parms->neigh_cleanup = ipoib_neigh_cleanup;
 925
 926         return 0;
 927 }
 928
 929 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
 930 {
 931         struct ipoib_dev_priv *priv = netdev_priv(dev);
 932
 933         /* Allocate RX/TX "rings" to hold queued skbs */
 934         priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
 935                                 GFP_KERNEL);
 936         if (!priv->rx_ring) {
 937                 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
 938                        ca->name, ipoib_recvq_size);
 939                 goto out;
 940         }
 941
 942         priv->tx_ring = vzalloc(ipoib_sendq_size * sizeof *priv->tx_ring);
 943         if (!priv->tx_ring) {
 944                 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
 945                        ca->name, ipoib_sendq_size);
 946                 goto out_rx_ring_cleanup;
 947         }
 948
 949         /* priv->tx_head, tx_tail & tx_outstanding are already 0 */
 950
 951         if (ipoib_ib_dev_init(dev, ca, port))
 952                 goto out_tx_ring_cleanup;
 953
 954         return 0;
 955
 956 out_tx_ring_cleanup:
 957         vfree(priv->tx_ring);
 958
 959 out_rx_ring_cleanup:
 960         kfree(priv->rx_ring);
 961
 962 out:
 963         return -ENOMEM;
 964 }
 965
 966 void ipoib_dev_cleanup(struct net_device *dev)
 967 {
 968         struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
 969
 970         ipoib_delete_debug_files(dev);
 971
 972         /* Delete any child interfaces first */
 973         list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
 974                 unregister_netdev(cpriv->dev);
 975                 ipoib_dev_cleanup(cpriv->dev);
 976                 free_netdev(cpriv->dev);
 977         }
 978
 979         ipoib_ib_dev_cleanup(dev);
 980
 981         kfree(priv->rx_ring);
 982         vfree(priv->tx_ring);
 983
 984         priv->rx_ring = NULL;
 985         priv->tx_ring = NULL;
 986 }
 987
 988 static const struct header_ops ipoib_header_ops = {
 989         .create = ipoib_hard_header,
 990 };
 991
 992 static const struct net_device_ops ipoib_netdev_ops = {
 993         .ndo_open                = ipoib_open,
 994         .ndo_stop                = ipoib_stop,
 995         .ndo_change_mtu          = ipoib_change_mtu,
 996         .ndo_fix_features        = ipoib_fix_features,
 997         .ndo_start_xmit          = ipoib_start_xmit,
 998         .ndo_tx_timeout          = ipoib_timeout,
 999         .ndo_set_multicast_list  = ipoib_set_mcast_list,
1000         .ndo_neigh_setup         = ipoib_neigh_setup_dev,
1001 };
1002
1003 static void ipoib_setup(struct net_device *dev)
1004 {
1005         struct ipoib_dev_priv *priv = netdev_priv(dev);
1006
1007         dev->netdev_ops          = &ipoib_netdev_ops;
1008         dev->header_ops          = &ipoib_header_ops;
1009
1010         ipoib_set_ethtool_ops(dev);
1011
1012         netif_napi_add(dev, &priv->napi, ipoib_poll, 100);
1013
1014         dev->watchdog_timeo      = HZ;
1015
1016         dev->flags              |= IFF_BROADCAST | IFF_MULTICAST;
1017
1018         /*
1019          * We add in INFINIBAND_ALEN to allow for the destination
1020          * address "pseudoheader" for skbs without neighbour struct.
1021          */
1022         dev->hard_header_len     = IPOIB_ENCAP_LEN + INFINIBAND_ALEN;
1023         dev->addr_len            = INFINIBAND_ALEN;
1024         dev->type                = ARPHRD_INFINIBAND;
1025         dev->tx_queue_len        = ipoib_sendq_size * 2;
1026         dev->features            = (NETIF_F_VLAN_CHALLENGED     |
1027                                     NETIF_F_HIGHDMA);
1028         dev->priv_flags         &= ~IFF_XMIT_DST_RELEASE;
1029
1030         memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
1031
1032         netif_carrier_off(dev);
1033
1034         priv->dev = dev;
1035
1036         spin_lock_init(&priv->lock);
1037
1038         mutex_init(&priv->vlan_mutex);
1039
1040         INIT_LIST_HEAD(&priv->path_list);
1041         INIT_LIST_HEAD(&priv->child_intfs);
1042         INIT_LIST_HEAD(&priv->dead_ahs);
1043         INIT_LIST_HEAD(&priv->multicast_list);
1044
1045         INIT_DELAYED_WORK(&priv->pkey_poll_task, ipoib_pkey_poll);
1046         INIT_DELAYED_WORK(&priv->mcast_task,   ipoib_mcast_join_task);
1047         INIT_WORK(&priv->carrier_on_task, ipoib_mcast_carrier_on_task);
1048         INIT_WORK(&priv->flush_light,   ipoib_ib_dev_flush_light);
1049         INIT_WORK(&priv->flush_normal,   ipoib_ib_dev_flush_normal);
1050         INIT_WORK(&priv->flush_heavy,   ipoib_ib_dev_flush_heavy);
1051         INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
1052         INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
1053 }
1054
1055 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
1056 {
1057         struct net_device *dev;
1058
1059         dev = alloc_netdev((int) sizeof (struct ipoib_dev_priv), name,
1060                            ipoib_setup);
1061         if (!dev)
1062                 return NULL;
1063
1064         return netdev_priv(dev);
1065 }
1066
1067 static ssize_t show_pkey(struct device *dev,
1068                          struct device_attribute *attr, char *buf)
1069 {
1070         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1071
1072         return sprintf(buf, "0x%04x\n", priv->pkey);
1073 }
1074 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1075
1076 static ssize_t show_umcast(struct device *dev,
1077                            struct device_attribute *attr, char *buf)
1078 {
1079         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1080
1081         return sprintf(buf, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
1082 }
1083
1084 static ssize_t set_umcast(struct device *dev,
1085                           struct device_attribute *attr,
1086                           const char *buf, size_t count)
1087 {
1088         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1089         unsigned long umcast_val = simple_strtoul(buf, NULL, 0);
1090
1091         if (umcast_val > 0) {
1092                 set_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1093                 ipoib_warn(priv, "ignoring multicast groups joined directly "
1094                                 "by userspace\n");
1095         } else
1096                 clear_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1097
1098         return count;
1099 }
1100 static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast);
1101
1102 int ipoib_add_umcast_attr(struct net_device *dev)
1103 {
1104         return device_create_file(&dev->dev, &dev_attr_umcast);
1105 }
1106
1107 static ssize_t create_child(struct device *dev,
1108                             struct device_attribute *attr,
1109                             const char *buf, size_t count)
1110 {
1111         int pkey;
1112         int ret;
1113
1114         if (sscanf(buf, "%i", &pkey) != 1)
1115                 return -EINVAL;
1116
1117         if (pkey < 0 || pkey > 0xffff)
1118                 return -EINVAL;
1119
1120         /*
1121          * Set the full membership bit, so that we join the right
1122          * broadcast group, etc.
1123          */
1124         pkey |= 0x8000;
1125
1126         ret = ipoib_vlan_add(to_net_dev(dev), pkey);
1127
1128         return ret ? ret : count;
1129 }
1130 static DEVICE_ATTR(create_child, S_IWUSR, NULL, create_child);
1131
1132 static ssize_t delete_child(struct device *dev,
1133                             struct device_attribute *attr,
1134                             const char *buf, size_t count)
1135 {
1136         int pkey;
1137         int ret;
1138
1139         if (sscanf(buf, "%i", &pkey) != 1)
1140                 return -EINVAL;
1141
1142         if (pkey < 0 || pkey > 0xffff)
1143                 return -EINVAL;
1144
1145         ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
1146
1147         return ret ? ret : count;
1148
1149 }
1150 static DEVICE_ATTR(delete_child, S_IWUSR, NULL, delete_child);
1151
1152 int ipoib_add_pkey_attr(struct net_device *dev)
1153 {
1154         return device_create_file(&dev->dev, &dev_attr_pkey);
1155 }
1156
1157 int ipoib_set_dev_features(struct ipoib_dev_priv *priv, struct ib_device *hca)
1158 {
1159         struct ib_device_attr *device_attr;
1160         int result = -ENOMEM;
1161
1162         device_attr = kmalloc(sizeof *device_attr, GFP_KERNEL);
1163         if (!device_attr) {
1164                 printk(KERN_WARNING "%s: allocation of %zu bytes failed\n",
1165                        hca->name, sizeof *device_attr);
1166                 return result;
1167         }
1168
1169         result = ib_query_device(hca, device_attr);
1170         if (result) {
1171                 printk(KERN_WARNING "%s: ib_query_device failed (ret = %d)\n",
1172                        hca->name, result);
1173                 kfree(device_attr);
1174                 return result;
1175         }
1176         priv->hca_caps = device_attr->device_cap_flags;
1177
1178         kfree(device_attr);
1179
1180         if (priv->hca_caps & IB_DEVICE_UD_IP_CSUM) {
1181                 priv->dev->hw_features = NETIF_F_SG |
1182                         NETIF_F_IP_CSUM | NETIF_F_RXCSUM;
1183
1184                 if (priv->hca_caps & IB_DEVICE_UD_TSO)
1185                         priv->dev->hw_features |= NETIF_F_TSO;
1186
1187                 priv->dev->features |= priv->dev->hw_features;
1188         }
1189
1190         return 0;
1191 }
1192
1193 static struct net_device *ipoib_add_port(const char *format,
1194                                          struct ib_device *hca, u8 port)
1195 {
1196         struct ipoib_dev_priv *priv;
1197         struct ib_port_attr attr;
1198         int result = -ENOMEM;
1199
1200         priv = ipoib_intf_alloc(format);
1201         if (!priv)
1202                 goto alloc_mem_failed;
1203
1204         SET_NETDEV_DEV(priv->dev, hca->dma_device);
1205         priv->dev->dev_id = port - 1;
1206
1207         if (!ib_query_port(hca, port, &attr))
1208                 priv->max_ib_mtu = ib_mtu_enum_to_int(attr.max_mtu);
1209         else {
1210                 printk(KERN_WARNING "%s: ib_query_port %d failed\n",
1211                        hca->name, port);
1212                 goto device_init_failed;
1213         }
1214
1215         /* MTU will be reset when mcast join happens */
1216         priv->dev->mtu  = IPOIB_UD_MTU(priv->max_ib_mtu);
1217         priv->mcast_mtu  = priv->admin_mtu = priv->dev->mtu;
1218
1219         result = ib_query_pkey(hca, port, 0, &priv->pkey);
1220         if (result) {
1221                 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
1222                        hca->name, port, result);
1223                 goto device_init_failed;
1224         }
1225
1226         if (ipoib_set_dev_features(priv, hca))
1227                 goto device_init_failed;
1228
1229         /*
1230          * Set the full membership bit, so that we join the right
1231          * broadcast group, etc.
1232          */
1233         priv->pkey |= 0x8000;
1234
1235         priv->dev->broadcast[8] = priv->pkey >> 8;
1236         priv->dev->broadcast[9] = priv->pkey & 0xff;
1237
1238         result = ib_query_gid(hca, port, 0, &priv->local_gid);
1239         if (result) {
1240                 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1241                        hca->name, port, result);
1242                 goto device_init_failed;
1243         } else
1244                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1245
1246         result = ipoib_dev_init(priv->dev, hca, port);
1247         if (result < 0) {
1248                 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1249                        hca->name, port, result);
1250                 goto device_init_failed;
1251         }
1252
1253         INIT_IB_EVENT_HANDLER(&priv->event_handler,
1254                               priv->ca, ipoib_event);
1255         result = ib_register_event_handler(&priv->event_handler);
1256         if (result < 0) {
1257                 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1258                        "port %d (ret = %d)\n",
1259                        hca->name, port, result);
1260                 goto event_failed;
1261         }
1262
1263         result = register_netdev(priv->dev);
1264         if (result) {
1265                 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1266                        hca->name, port, result);
1267                 goto register_failed;
1268         }
1269
1270         ipoib_create_debug_files(priv->dev);
1271
1272         if (ipoib_cm_add_mode_attr(priv->dev))
1273                 goto sysfs_failed;
1274         if (ipoib_add_pkey_attr(priv->dev))
1275                 goto sysfs_failed;
1276         if (ipoib_add_umcast_attr(priv->dev))
1277                 goto sysfs_failed;
1278         if (device_create_file(&priv->dev->dev, &dev_attr_create_child))
1279                 goto sysfs_failed;
1280         if (device_create_file(&priv->dev->dev, &dev_attr_delete_child))
1281                 goto sysfs_failed;
1282
1283         return priv->dev;
1284
1285 sysfs_failed:
1286         ipoib_delete_debug_files(priv->dev);
1287         unregister_netdev(priv->dev);
1288
1289 register_failed:
1290         ib_unregister_event_handler(&priv->event_handler);
1291         flush_workqueue(ipoib_workqueue);
1292
1293 event_failed:
1294         ipoib_dev_cleanup(priv->dev);
1295
1296 device_init_failed:
1297         free_netdev(priv->dev);
1298
1299 alloc_mem_failed:
1300         return ERR_PTR(result);
1301 }
1302
1303 static void ipoib_add_one(struct ib_device *device)
1304 {
1305         struct list_head *dev_list;
1306         struct net_device *dev;
1307         struct ipoib_dev_priv *priv;
1308         int s, e, p;
1309
1310         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1311                 return;
1312
1313         dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1314         if (!dev_list)
1315                 return;
1316
1317         INIT_LIST_HEAD(dev_list);
1318
1319         if (device->node_type == RDMA_NODE_IB_SWITCH) {
1320                 s = 0;
1321                 e = 0;
1322         } else {
1323                 s = 1;
1324                 e = device->phys_port_cnt;
1325         }
1326
1327         for (p = s; p <= e; ++p) {
1328                 if (rdma_port_get_link_layer(device, p) != IB_LINK_LAYER_INFINIBAND)
1329                         continue;
1330                 dev = ipoib_add_port("ib%d", device, p);
1331                 if (!IS_ERR(dev)) {
1332                         priv = netdev_priv(dev);
1333                         list_add_tail(&priv->list, dev_list);
1334                 }
1335         }
1336
1337         ib_set_client_data(device, &ipoib_client, dev_list);
1338 }
1339
1340 static void ipoib_remove_one(struct ib_device *device)
1341 {
1342         struct ipoib_dev_priv *priv, *tmp;
1343         struct list_head *dev_list;
1344
1345         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1346                 return;
1347
1348         dev_list = ib_get_client_data(device, &ipoib_client);
1349
1350         list_for_each_entry_safe(priv, tmp, dev_list, list) {
1351                 ib_unregister_event_handler(&priv->event_handler);
1352
1353                 rtnl_lock();
1354                 dev_change_flags(priv->dev, priv->dev->flags & ~IFF_UP);
1355                 rtnl_unlock();
1356
1357                 flush_workqueue(ipoib_workqueue);
1358
1359                 unregister_netdev(priv->dev);
1360                 ipoib_dev_cleanup(priv->dev);
1361                 free_netdev(priv->dev);
1362         }
1363
1364         kfree(dev_list);
1365 }
1366
1367 static int __init ipoib_init_module(void)
1368 {
1369         int ret;
1370
1371         ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1372         ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1373         ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1374
1375         ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1376         ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1377         ipoib_sendq_size = max3(ipoib_sendq_size, 2 * MAX_SEND_CQE, IPOIB_MIN_QUEUE_SIZE);
1378 #ifdef CONFIG_INFINIBAND_IPOIB_CM
1379         ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
1380 #endif
1381
1382         /*
1383          * When copying small received packets, we only copy from the
1384          * linear data part of the SKB, so we rely on this condition.
1385          */
1386         BUILD_BUG_ON(IPOIB_CM_COPYBREAK > IPOIB_CM_HEAD_SIZE);
1387
1388         ret = ipoib_register_debugfs();
1389         if (ret)
1390                 return ret;
1391
1392         /*
1393          * We create our own workqueue mainly because we want to be
1394          * able to flush it when devices are being removed.  We can't
1395          * use schedule_work()/flush_scheduled_work() because both
1396          * unregister_netdev() and linkwatch_event take the rtnl lock,
1397          * so flush_scheduled_work() can deadlock during device
1398          * removal.
1399          */
1400         ipoib_workqueue = create_singlethread_workqueue("ipoib");
1401         if (!ipoib_workqueue) {
1402                 ret = -ENOMEM;
1403                 goto err_fs;
1404         }
1405
1406         ib_sa_register_client(&ipoib_sa_client);
1407
1408         ret = ib_register_client(&ipoib_client);
1409         if (ret)
1410                 goto err_sa;
1411
1412         return 0;
1413
1414 err_sa:
1415         ib_sa_unregister_client(&ipoib_sa_client);
1416         destroy_workqueue(ipoib_workqueue);
1417
1418 err_fs:
1419         ipoib_unregister_debugfs();
1420
1421         return ret;
1422 }
1423
1424 static void __exit ipoib_cleanup_module(void)
1425 {
1426         ib_unregister_client(&ipoib_client);
1427         ib_sa_unregister_client(&ipoib_sa_client);
1428         ipoib_unregister_debugfs();
1429         destroy_workqueue(ipoib_workqueue);
1430 }
1431
1432 module_init(ipoib_init_module);
1433 module_exit(ipoib_cleanup_module);