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 <linux/jhash.h>
  50 #include <net/arp.h>
  51
  52 MODULE_AUTHOR("Roland Dreier");
  53 MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
  54 MODULE_LICENSE("Dual BSD/GPL");
  55
  56 int ipoib_sendq_size __read_mostly = IPOIB_TX_RING_SIZE;
  57 int ipoib_recvq_size __read_mostly = IPOIB_RX_RING_SIZE;
  58
  59 module_param_named(send_queue_size, ipoib_sendq_size, int, 0444);
  60 MODULE_PARM_DESC(send_queue_size, "Number of descriptors in send queue");
  61 module_param_named(recv_queue_size, ipoib_recvq_size, int, 0444);
  62 MODULE_PARM_DESC(recv_queue_size, "Number of descriptors in receive queue");
  63
  64 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
  65 int ipoib_debug_level;
  66
  67 module_param_named(debug_level, ipoib_debug_level, int, 0644);
  68 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
  69 #endif
  70
  71 struct ipoib_path_iter {
  72         struct net_device *dev;
  73         struct ipoib_path  path;
  74 };
  75
  76 static const u8 ipv4_bcast_addr[] = {
  77         0x00, 0xff, 0xff, 0xff,
  78         0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
  79         0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff
  80 };
  81
  82 struct workqueue_struct *ipoib_workqueue;
  83
  84 struct ib_sa_client ipoib_sa_client;
  85
  86 static void ipoib_add_one(struct ib_device *device);
  87 static void ipoib_remove_one(struct ib_device *device);
  88 static void ipoib_neigh_reclaim(struct rcu_head *rp);
  89
  90 static struct ib_client ipoib_client = {
  91         .name   = "ipoib",
  92         .add    = ipoib_add_one,
  93         .remove = ipoib_remove_one
  94 };
  95
  96 int ipoib_open(struct net_device *dev)
  97 {
  98         struct ipoib_dev_priv *priv = netdev_priv(dev);
  99
 100         ipoib_dbg(priv, "bringing up interface\n");
 101
 102         set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 103
 104         if (ipoib_pkey_dev_delay_open(dev))
 105                 return 0;
 106
 107         if (ipoib_ib_dev_open(dev))
 108                 goto err_disable;
 109
 110         if (ipoib_ib_dev_up(dev))
 111                 goto err_stop;
 112
 113         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 114                 struct ipoib_dev_priv *cpriv;
 115
 116                 /* Bring up any child interfaces too */
 117                 mutex_lock(&priv->vlan_mutex);
 118                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 119                         int flags;
 120
 121                         flags = cpriv->dev->flags;
 122                         if (flags & IFF_UP)
 123                                 continue;
 124
 125                         dev_change_flags(cpriv->dev, flags | IFF_UP);
 126                 }
 127                 mutex_unlock(&priv->vlan_mutex);
 128         }
 129
 130         netif_start_queue(dev);
 131
 132         return 0;
 133
 134 err_stop:
 135         ipoib_ib_dev_stop(dev, 1);
 136
 137 err_disable:
 138         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 139
 140         return -EINVAL;
 141 }
 142
 143 static int ipoib_stop(struct net_device *dev)
 144 {
 145         struct ipoib_dev_priv *priv = netdev_priv(dev);
 146
 147         ipoib_dbg(priv, "stopping interface\n");
 148
 149         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 150
 151         netif_stop_queue(dev);
 152
 153         ipoib_ib_dev_down(dev, 1);
 154         ipoib_ib_dev_stop(dev, 0);
 155
 156         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 157                 struct ipoib_dev_priv *cpriv;
 158
 159                 /* Bring down any child interfaces too */
 160                 mutex_lock(&priv->vlan_mutex);
 161                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 162                         int flags;
 163
 164                         flags = cpriv->dev->flags;
 165                         if (!(flags & IFF_UP))
 166                                 continue;
 167
 168                         dev_change_flags(cpriv->dev, flags & ~IFF_UP);
 169                 }
 170                 mutex_unlock(&priv->vlan_mutex);
 171         }
 172
 173         return 0;
 174 }
 175
 176 static void ipoib_uninit(struct net_device *dev)
 177 {
 178         ipoib_dev_cleanup(dev);
 179 }
 180
 181 static netdev_features_t ipoib_fix_features(struct net_device *dev, netdev_features_t features)
 182 {
 183         struct ipoib_dev_priv *priv = netdev_priv(dev);
 184
 185         if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
 186                 features &= ~(NETIF_F_SG | NETIF_F_IP_CSUM | NETIF_F_TSO);
 187
 188         return features;
 189 }
 190
 191 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
 192 {
 193         struct ipoib_dev_priv *priv = netdev_priv(dev);
 194
 195         /* dev->mtu > 2K ==> connected mode */
 196         if (ipoib_cm_admin_enabled(dev)) {
 197                 if (new_mtu > ipoib_cm_max_mtu(dev))
 198                         return -EINVAL;
 199
 200                 if (new_mtu > priv->mcast_mtu)
 201                         ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
 202                                    priv->mcast_mtu);
 203
 204                 dev->mtu = new_mtu;
 205                 return 0;
 206         }
 207
 208         if (new_mtu > IPOIB_UD_MTU(priv->max_ib_mtu))
 209                 return -EINVAL;
 210
 211         priv->admin_mtu = new_mtu;
 212
 213         dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
 214
 215         return 0;
 216 }
 217
 218 int ipoib_set_mode(struct net_device *dev, const char *buf)
 219 {
 220         struct ipoib_dev_priv *priv = netdev_priv(dev);
 221
 222         /* flush paths if we switch modes so that connections are restarted */
 223         if (IPOIB_CM_SUPPORTED(dev->dev_addr) && !strcmp(buf, "connected\n")) {
 224                 set_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
 225                 ipoib_warn(priv, "enabling connected mode "
 226                            "will cause multicast packet drops\n");
 227                 netdev_update_features(dev);
 228                 rtnl_unlock();
 229                 priv->tx_wr.send_flags &= ~IB_SEND_IP_CSUM;
 230
 231                 ipoib_flush_paths(dev);
 232                 rtnl_lock();
 233                 return 0;
 234         }
 235
 236         if (!strcmp(buf, "datagram\n")) {
 237                 clear_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
 238                 netdev_update_features(dev);
 239                 dev_set_mtu(dev, min(priv->mcast_mtu, dev->mtu));
 240                 rtnl_unlock();
 241                 ipoib_flush_paths(dev);
 242                 rtnl_lock();
 243                 return 0;
 244         }
 245
 246         return -EINVAL;
 247 }
 248
 249 static struct ipoib_path *__path_find(struct net_device *dev, void *gid)
 250 {
 251         struct ipoib_dev_priv *priv = netdev_priv(dev);
 252         struct rb_node *n = priv->path_tree.rb_node;
 253         struct ipoib_path *path;
 254         int ret;
 255
 256         while (n) {
 257                 path = rb_entry(n, struct ipoib_path, rb_node);
 258
 259                 ret = memcmp(gid, path->pathrec.dgid.raw,
 260                              sizeof (union ib_gid));
 261
 262                 if (ret < 0)
 263                         n = n->rb_left;
 264                 else if (ret > 0)
 265                         n = n->rb_right;
 266                 else
 267                         return path;
 268         }
 269
 270         return NULL;
 271 }
 272
 273 static int __path_add(struct net_device *dev, struct ipoib_path *path)
 274 {
 275         struct ipoib_dev_priv *priv = netdev_priv(dev);
 276         struct rb_node **n = &priv->path_tree.rb_node;
 277         struct rb_node *pn = NULL;
 278         struct ipoib_path *tpath;
 279         int ret;
 280
 281         while (*n) {
 282                 pn = *n;
 283                 tpath = rb_entry(pn, struct ipoib_path, rb_node);
 284
 285                 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
 286                              sizeof (union ib_gid));
 287                 if (ret < 0)
 288                         n = &pn->rb_left;
 289                 else if (ret > 0)
 290                         n = &pn->rb_right;
 291                 else
 292                         return -EEXIST;
 293         }
 294
 295         rb_link_node(&path->rb_node, pn, n);
 296         rb_insert_color(&path->rb_node, &priv->path_tree);
 297
 298         list_add_tail(&path->list, &priv->path_list);
 299
 300         return 0;
 301 }
 302
 303 static void path_free(struct net_device *dev, struct ipoib_path *path)
 304 {
 305         struct sk_buff *skb;
 306
 307         while ((skb = __skb_dequeue(&path->queue)))
 308                 dev_kfree_skb_irq(skb);
 309
 310         ipoib_dbg(netdev_priv(dev), "path_free\n");
 311
 312         /* remove all neigh connected to this path */
 313         ipoib_del_neighs_by_gid(dev, path->pathrec.dgid.raw);
 314
 315         if (path->ah)
 316                 ipoib_put_ah(path->ah);
 317
 318         kfree(path);
 319 }
 320
 321 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 322
 323 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
 324 {
 325         struct ipoib_path_iter *iter;
 326
 327         iter = kmalloc(sizeof *iter, GFP_KERNEL);
 328         if (!iter)
 329                 return NULL;
 330
 331         iter->dev = dev;
 332         memset(iter->path.pathrec.dgid.raw, 0, 16);
 333
 334         if (ipoib_path_iter_next(iter)) {
 335                 kfree(iter);
 336                 return NULL;
 337         }
 338
 339         return iter;
 340 }
 341
 342 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
 343 {
 344         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
 345         struct rb_node *n;
 346         struct ipoib_path *path;
 347         int ret = 1;
 348
 349         spin_lock_irq(&priv->lock);
 350
 351         n = rb_first(&priv->path_tree);
 352
 353         while (n) {
 354                 path = rb_entry(n, struct ipoib_path, rb_node);
 355
 356                 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
 357                            sizeof (union ib_gid)) < 0) {
 358                         iter->path = *path;
 359                         ret = 0;
 360                         break;
 361                 }
 362
 363                 n = rb_next(n);
 364         }
 365
 366         spin_unlock_irq(&priv->lock);
 367
 368         return ret;
 369 }
 370
 371 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
 372                           struct ipoib_path *path)
 373 {
 374         *path = iter->path;
 375 }
 376
 377 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
 378
 379 void ipoib_mark_paths_invalid(struct net_device *dev)
 380 {
 381         struct ipoib_dev_priv *priv = netdev_priv(dev);
 382         struct ipoib_path *path, *tp;
 383
 384         spin_lock_irq(&priv->lock);
 385
 386         list_for_each_entry_safe(path, tp, &priv->path_list, list) {
 387                 ipoib_dbg(priv, "mark path LID 0x%04x GID %pI6 invalid\n",
 388                         be16_to_cpu(path->pathrec.dlid),
 389                         path->pathrec.dgid.raw);
 390                 path->valid =  0;
 391         }
 392
 393         spin_unlock_irq(&priv->lock);
 394 }
 395
 396 void ipoib_flush_paths(struct net_device *dev)
 397 {
 398         struct ipoib_dev_priv *priv = netdev_priv(dev);
 399         struct ipoib_path *path, *tp;
 400         LIST_HEAD(remove_list);
 401         unsigned long flags;
 402
 403         netif_tx_lock_bh(dev);
 404         spin_lock_irqsave(&priv->lock, flags);
 405
 406         list_splice_init(&priv->path_list, &remove_list);
 407
 408         list_for_each_entry(path, &remove_list, list)
 409                 rb_erase(&path->rb_node, &priv->path_tree);
 410
 411         list_for_each_entry_safe(path, tp, &remove_list, list) {
 412                 if (path->query)
 413                         ib_sa_cancel_query(path->query_id, path->query);
 414                 spin_unlock_irqrestore(&priv->lock, flags);
 415                 netif_tx_unlock_bh(dev);
 416                 wait_for_completion(&path->done);
 417                 path_free(dev, path);
 418                 netif_tx_lock_bh(dev);
 419                 spin_lock_irqsave(&priv->lock, flags);
 420         }
 421
 422         spin_unlock_irqrestore(&priv->lock, flags);
 423         netif_tx_unlock_bh(dev);
 424 }
 425
 426 static void path_rec_completion(int status,
 427                                 struct ib_sa_path_rec *pathrec,
 428                                 void *path_ptr)
 429 {
 430         struct ipoib_path *path = path_ptr;
 431         struct net_device *dev = path->dev;
 432         struct ipoib_dev_priv *priv = netdev_priv(dev);
 433         struct ipoib_ah *ah = NULL;
 434         struct ipoib_ah *old_ah = NULL;
 435         struct ipoib_neigh *neigh, *tn;
 436         struct sk_buff_head skqueue;
 437         struct sk_buff *skb;
 438         unsigned long flags;
 439
 440         if (!status)
 441                 ipoib_dbg(priv, "PathRec LID 0x%04x for GID %pI6\n",
 442                           be16_to_cpu(pathrec->dlid), pathrec->dgid.raw);
 443         else
 444                 ipoib_dbg(priv, "PathRec status %d for GID %pI6\n",
 445                           status, path->pathrec.dgid.raw);
 446
 447         skb_queue_head_init(&skqueue);
 448
 449         if (!status) {
 450                 struct ib_ah_attr av;
 451
 452                 if (!ib_init_ah_from_path(priv->ca, priv->port, pathrec, &av))
 453                         ah = ipoib_create_ah(dev, priv->pd, &av);
 454         }
 455
 456         spin_lock_irqsave(&priv->lock, flags);
 457
 458         if (!IS_ERR_OR_NULL(ah)) {
 459                 path->pathrec = *pathrec;
 460
 461                 old_ah   = path->ah;
 462                 path->ah = ah;
 463
 464                 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
 465                           ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
 466
 467                 while ((skb = __skb_dequeue(&path->queue)))
 468                         __skb_queue_tail(&skqueue, skb);
 469
 470                 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
 471                         if (neigh->ah) {
 472                                 WARN_ON(neigh->ah != old_ah);
 473                                 /*
 474                                  * Dropping the ah reference inside
 475                                  * priv->lock is safe here, because we
 476                                  * will hold one more reference from
 477                                  * the original value of path->ah (ie
 478                                  * old_ah).
 479                                  */
 480                                 ipoib_put_ah(neigh->ah);
 481                         }
 482                         kref_get(&path->ah->ref);
 483                         neigh->ah = path->ah;
 484
 485                         if (ipoib_cm_enabled(dev, neigh->daddr)) {
 486                                 if (!ipoib_cm_get(neigh))
 487                                         ipoib_cm_set(neigh, ipoib_cm_create_tx(dev,
 488                                                                                path,
 489                                                                                neigh));
 490                                 if (!ipoib_cm_get(neigh)) {
 491                                         list_del(&neigh->list);
 492                                         ipoib_neigh_free(neigh);
 493                                         continue;
 494                                 }
 495                         }
 496
 497                         while ((skb = __skb_dequeue(&neigh->queue)))
 498                                 __skb_queue_tail(&skqueue, skb);
 499                 }
 500                 path->valid = 1;
 501         }
 502
 503         path->query = NULL;
 504         complete(&path->done);
 505
 506         spin_unlock_irqrestore(&priv->lock, flags);
 507
 508         if (old_ah)
 509                 ipoib_put_ah(old_ah);
 510
 511         while ((skb = __skb_dequeue(&skqueue))) {
 512                 skb->dev = dev;
 513                 if (dev_queue_xmit(skb))
 514                         ipoib_warn(priv, "dev_queue_xmit failed "
 515                                    "to requeue packet\n");
 516         }
 517 }
 518
 519 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
 520 {
 521         struct ipoib_dev_priv *priv = netdev_priv(dev);
 522         struct ipoib_path *path;
 523
 524         if (!priv->broadcast)
 525                 return NULL;
 526
 527         path = kzalloc(sizeof *path, GFP_ATOMIC);
 528         if (!path)
 529                 return NULL;
 530
 531         path->dev = dev;
 532
 533         skb_queue_head_init(&path->queue);
 534
 535         INIT_LIST_HEAD(&path->neigh_list);
 536
 537         memcpy(path->pathrec.dgid.raw, gid, sizeof (union ib_gid));
 538         path->pathrec.sgid          = priv->local_gid;
 539         path->pathrec.pkey          = cpu_to_be16(priv->pkey);
 540         path->pathrec.numb_path     = 1;
 541         path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
 542
 543         return path;
 544 }
 545
 546 static int path_rec_start(struct net_device *dev,
 547                           struct ipoib_path *path)
 548 {
 549         struct ipoib_dev_priv *priv = netdev_priv(dev);
 550
 551         ipoib_dbg(priv, "Start path record lookup for %pI6\n",
 552                   path->pathrec.dgid.raw);
 553
 554         init_completion(&path->done);
 555
 556         path->query_id =
 557                 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
 558                                    &path->pathrec,
 559                                    IB_SA_PATH_REC_DGID          |
 560                                    IB_SA_PATH_REC_SGID          |
 561                                    IB_SA_PATH_REC_NUMB_PATH     |
 562                                    IB_SA_PATH_REC_TRAFFIC_CLASS |
 563                                    IB_SA_PATH_REC_PKEY,
 564                                    1000, GFP_ATOMIC,
 565                                    path_rec_completion,
 566                                    path, &path->query);
 567         if (path->query_id < 0) {
 568                 ipoib_warn(priv, "ib_sa_path_rec_get failed: %d\n", path->query_id);
 569                 path->query = NULL;
 570                 complete(&path->done);
 571                 return path->query_id;
 572         }
 573
 574         return 0;
 575 }
 576
 577 static void neigh_add_path(struct sk_buff *skb, u8 *daddr,
 578                            struct net_device *dev)
 579 {
 580         struct ipoib_dev_priv *priv = netdev_priv(dev);
 581         struct ipoib_path *path;
 582         struct ipoib_neigh *neigh;
 583         unsigned long flags;
 584
 585         spin_lock_irqsave(&priv->lock, flags);
 586         neigh = ipoib_neigh_alloc(daddr, dev);
 587         if (!neigh) {
 588                 spin_unlock_irqrestore(&priv->lock, flags);
 589                 ++dev->stats.tx_dropped;
 590                 dev_kfree_skb_any(skb);
 591                 return;
 592         }
 593
 594         path = __path_find(dev, daddr + 4);
 595         if (!path) {
 596                 path = path_rec_create(dev, daddr + 4);
 597                 if (!path)
 598                         goto err_path;
 599
 600                 __path_add(dev, path);
 601         }
 602
 603         list_add_tail(&neigh->list, &path->neigh_list);
 604
 605         if (path->ah) {
 606                 kref_get(&path->ah->ref);
 607                 neigh->ah = path->ah;
 608
 609                 if (ipoib_cm_enabled(dev, neigh->daddr)) {
 610                         if (!ipoib_cm_get(neigh))
 611                                 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
 612                         if (!ipoib_cm_get(neigh)) {
 613                                 list_del(&neigh->list);
 614                                 ipoib_neigh_free(neigh);
 615                                 goto err_drop;
 616                         }
 617                         if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE)
 618                                 __skb_queue_tail(&neigh->queue, skb);
 619                         else {
 620                                 ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
 621                                            skb_queue_len(&neigh->queue));
 622                                 goto err_drop;
 623                         }
 624                 } else {
 625                         spin_unlock_irqrestore(&priv->lock, flags);
 626                         ipoib_send(dev, skb, path->ah, IPOIB_QPN(daddr));
 627                         ipoib_neigh_put(neigh);
 628                         return;
 629                 }
 630         } else {
 631                 neigh->ah  = NULL;
 632
 633                 if (!path->query && path_rec_start(dev, path))
 634                         goto err_list;
 635
 636                 __skb_queue_tail(&neigh->queue, skb);
 637         }
 638
 639         spin_unlock_irqrestore(&priv->lock, flags);
 640         ipoib_neigh_put(neigh);
 641         return;
 642
 643 err_list:
 644         list_del(&neigh->list);
 645
 646 err_path:
 647         ipoib_neigh_free(neigh);
 648 err_drop:
 649         ++dev->stats.tx_dropped;
 650         dev_kfree_skb_any(skb);
 651
 652         spin_unlock_irqrestore(&priv->lock, flags);
 653         ipoib_neigh_put(neigh);
 654 }
 655
 656 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
 657                              struct ipoib_cb *cb)
 658 {
 659         struct ipoib_dev_priv *priv = netdev_priv(dev);
 660         struct ipoib_path *path;
 661         unsigned long flags;
 662
 663         spin_lock_irqsave(&priv->lock, flags);
 664
 665         path = __path_find(dev, cb->hwaddr + 4);
 666         if (!path || !path->valid) {
 667                 int new_path = 0;
 668
 669                 if (!path) {
 670                         path = path_rec_create(dev, cb->hwaddr + 4);
 671                         new_path = 1;
 672                 }
 673                 if (path) {
 674                         __skb_queue_tail(&path->queue, skb);
 675
 676                         if (!path->query && path_rec_start(dev, path)) {
 677                                 spin_unlock_irqrestore(&priv->lock, flags);
 678                                 if (new_path)
 679                                         path_free(dev, path);
 680                                 return;
 681                         } else
 682                                 __path_add(dev, path);
 683                 } else {
 684                         ++dev->stats.tx_dropped;
 685                         dev_kfree_skb_any(skb);
 686                 }
 687
 688                 spin_unlock_irqrestore(&priv->lock, flags);
 689                 return;
 690         }
 691
 692         if (path->ah) {
 693                 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
 694                           be16_to_cpu(path->pathrec.dlid));
 695
 696                 spin_unlock_irqrestore(&priv->lock, flags);
 697                 ipoib_send(dev, skb, path->ah, IPOIB_QPN(cb->hwaddr));
 698                 return;
 699         } else if ((path->query || !path_rec_start(dev, path)) &&
 700                    skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 701                 __skb_queue_tail(&path->queue, skb);
 702         } else {
 703                 ++dev->stats.tx_dropped;
 704                 dev_kfree_skb_any(skb);
 705         }
 706
 707         spin_unlock_irqrestore(&priv->lock, flags);
 708 }
 709
 710 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
 711 {
 712         struct ipoib_dev_priv *priv = netdev_priv(dev);
 713         struct ipoib_neigh *neigh;
 714         struct ipoib_cb *cb = (struct ipoib_cb *) skb->cb;
 715         struct ipoib_header *header;
 716         unsigned long flags;
 717
 718         header = (struct ipoib_header *) skb->data;
 719
 720         if (unlikely(cb->hwaddr[4] == 0xff)) {
 721                 /* multicast, arrange "if" according to probability */
 722                 if ((header->proto != htons(ETH_P_IP)) &&
 723                     (header->proto != htons(ETH_P_IPV6)) &&
 724                     (header->proto != htons(ETH_P_ARP)) &&
 725                     (header->proto != htons(ETH_P_RARP))) {
 726                         /* ethertype not supported by IPoIB */
 727                         ++dev->stats.tx_dropped;
 728                         dev_kfree_skb_any(skb);
 729                         return NETDEV_TX_OK;
 730                 }
 731                 /* Add in the P_Key for multicast*/
 732                 cb->hwaddr[8] = (priv->pkey >> 8) & 0xff;
 733                 cb->hwaddr[9] = priv->pkey & 0xff;
 734
 735                 neigh = ipoib_neigh_get(dev, cb->hwaddr);
 736                 if (likely(neigh))
 737                         goto send_using_neigh;
 738                 ipoib_mcast_send(dev, cb->hwaddr, skb);
 739                 return NETDEV_TX_OK;
 740         }
 741
 742         /* unicast, arrange "switch" according to probability */
 743         switch (header->proto) {
 744         case htons(ETH_P_IP):
 745         case htons(ETH_P_IPV6):
 746                 neigh = ipoib_neigh_get(dev, cb->hwaddr);
 747                 if (unlikely(!neigh)) {
 748                         neigh_add_path(skb, cb->hwaddr, dev);
 749                         return NETDEV_TX_OK;
 750                 }
 751                 break;
 752         case htons(ETH_P_ARP):
 753         case htons(ETH_P_RARP):
 754                 /* for unicast ARP and RARP should always perform path find */
 755                 unicast_arp_send(skb, dev, cb);
 756                 return NETDEV_TX_OK;
 757         default:
 758                 /* ethertype not supported by IPoIB */
 759                 ++dev->stats.tx_dropped;
 760                 dev_kfree_skb_any(skb);
 761                 return NETDEV_TX_OK;
 762         }
 763
 764 send_using_neigh:
 765         /* note we now hold a ref to neigh */
 766         if (ipoib_cm_get(neigh)) {
 767                 if (ipoib_cm_up(neigh)) {
 768                         ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
 769                         goto unref;
 770                 }
 771         } else if (neigh->ah) {
 772                 ipoib_send(dev, skb, neigh->ah, IPOIB_QPN(cb->hwaddr));
 773                 goto unref;
 774         }
 775
 776         if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 777                 spin_lock_irqsave(&priv->lock, flags);
 778                 __skb_queue_tail(&neigh->queue, skb);
 779                 spin_unlock_irqrestore(&priv->lock, flags);
 780         } else {
 781                 ++dev->stats.tx_dropped;
 782                 dev_kfree_skb_any(skb);
 783         }
 784
 785 unref:
 786         ipoib_neigh_put(neigh);
 787
 788         return NETDEV_TX_OK;
 789 }
 790
 791 static void ipoib_timeout(struct net_device *dev)
 792 {
 793         struct ipoib_dev_priv *priv = netdev_priv(dev);
 794
 795         ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
 796                    jiffies_to_msecs(jiffies - dev->trans_start));
 797         ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
 798                    netif_queue_stopped(dev),
 799                    priv->tx_head, priv->tx_tail);
 800         /* XXX reset QP, etc. */
 801 }
 802
 803 static int ipoib_hard_header(struct sk_buff *skb,
 804                              struct net_device *dev,
 805                              unsigned short type,
 806                              const void *daddr, const void *saddr, unsigned len)
 807 {
 808         struct ipoib_header *header;
 809         struct ipoib_cb *cb = (struct ipoib_cb *) skb->cb;
 810
 811         header = (struct ipoib_header *) skb_push(skb, sizeof *header);
 812
 813         header->proto = htons(type);
 814         header->reserved = 0;
 815
 816         /*
 817          * we don't rely on dst_entry structure,  always stuff the
 818          * destination address into skb->cb so we can figure out where
 819          * to send the packet later.
 820          */
 821         memcpy(cb->hwaddr, daddr, INFINIBAND_ALEN);
 822
 823         return 0;
 824 }
 825
 826 static void ipoib_set_mcast_list(struct net_device *dev)
 827 {
 828         struct ipoib_dev_priv *priv = netdev_priv(dev);
 829
 830         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
 831                 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
 832                 return;
 833         }
 834
 835         queue_work(ipoib_workqueue, &priv->restart_task);
 836 }
 837
 838 static u32 ipoib_addr_hash(struct ipoib_neigh_hash *htbl, u8 *daddr)
 839 {
 840         /*
 841          * Use only the address parts that contributes to spreading
 842          * The subnet prefix is not used as one can not connect to
 843          * same remote port (GUID) using the same remote QPN via two
 844          * different subnets.
 845          */
 846          /* qpn octets[1:4) & port GUID octets[12:20) */
 847         u32 *d32 = (u32 *) daddr;
 848         u32 hv;
 849
 850         hv = jhash_3words(d32[3], d32[4], IPOIB_QPN_MASK & d32[0], 0);
 851         return hv & htbl->mask;
 852 }
 853
 854 struct ipoib_neigh *ipoib_neigh_get(struct net_device *dev, u8 *daddr)
 855 {
 856         struct ipoib_dev_priv *priv = netdev_priv(dev);
 857         struct ipoib_neigh_table *ntbl = &priv->ntbl;
 858         struct ipoib_neigh_hash *htbl;
 859         struct ipoib_neigh *neigh = NULL;
 860         u32 hash_val;
 861
 862         rcu_read_lock_bh();
 863
 864         htbl = rcu_dereference_bh(ntbl->htbl);
 865
 866         if (!htbl)
 867                 goto out_unlock;
 868
 869         hash_val = ipoib_addr_hash(htbl, daddr);
 870         for (neigh = rcu_dereference_bh(htbl->buckets[hash_val]);
 871              neigh != NULL;
 872              neigh = rcu_dereference_bh(neigh->hnext)) {
 873                 if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) {
 874                         /* found, take one ref on behalf of the caller */
 875                         if (!atomic_inc_not_zero(&neigh->refcnt)) {
 876                                 /* deleted */
 877                                 neigh = NULL;
 878                                 goto out_unlock;
 879                         }
 880                         neigh->alive = jiffies;
 881                         goto out_unlock;
 882                 }
 883         }
 884
 885 out_unlock:
 886         rcu_read_unlock_bh();
 887         return neigh;
 888 }
 889
 890 static void __ipoib_reap_neigh(struct ipoib_dev_priv *priv)
 891 {
 892         struct ipoib_neigh_table *ntbl = &priv->ntbl;
 893         struct ipoib_neigh_hash *htbl;
 894         unsigned long neigh_obsolete;
 895         unsigned long dt;
 896         unsigned long flags;
 897         int i;
 898
 899         if (test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
 900                 return;
 901
 902         spin_lock_irqsave(&priv->lock, flags);
 903
 904         htbl = rcu_dereference_protected(ntbl->htbl,
 905                                          lockdep_is_held(&priv->lock));
 906
 907         if (!htbl)
 908                 goto out_unlock;
 909
 910         /* neigh is obsolete if it was idle for two GC periods */
 911         dt = 2 * arp_tbl.gc_interval;
 912         neigh_obsolete = jiffies - dt;
 913         /* handle possible race condition */
 914         if (test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
 915                 goto out_unlock;
 916
 917         for (i = 0; i < htbl->size; i++) {
 918                 struct ipoib_neigh *neigh;
 919                 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
 920
 921                 while ((neigh = rcu_dereference_protected(*np,
 922                                                           lockdep_is_held(&priv->lock))) != NULL) {
 923                         /* was the neigh idle for two GC periods */
 924                         if (time_after(neigh_obsolete, neigh->alive)) {
 925                                 rcu_assign_pointer(*np,
 926                                                    rcu_dereference_protected(neigh->hnext,
 927                                                                              lockdep_is_held(&priv->lock)));
 928                                 /* remove from path/mc list */
 929                                 list_del(&neigh->list);
 930                                 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
 931                         } else {
 932                                 np = &neigh->hnext;
 933                         }
 934
 935                 }
 936         }
 937
 938 out_unlock:
 939         spin_unlock_irqrestore(&priv->lock, flags);
 940 }
 941
 942 static void ipoib_reap_neigh(struct work_struct *work)
 943 {
 944         struct ipoib_dev_priv *priv =
 945                 container_of(work, struct ipoib_dev_priv, neigh_reap_task.work);
 946
 947         __ipoib_reap_neigh(priv);
 948
 949         if (!test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
 950                 queue_delayed_work(ipoib_workqueue, &priv->neigh_reap_task,
 951                                    arp_tbl.gc_interval);
 952 }
 953
 954
 955 static struct ipoib_neigh *ipoib_neigh_ctor(u8 *daddr,
 956                                       struct net_device *dev)
 957 {
 958         struct ipoib_neigh *neigh;
 959
 960         neigh = kzalloc(sizeof *neigh, GFP_ATOMIC);
 961         if (!neigh)
 962                 return NULL;
 963
 964         neigh->dev = dev;
 965         memcpy(&neigh->daddr, daddr, sizeof(neigh->daddr));
 966         skb_queue_head_init(&neigh->queue);
 967         INIT_LIST_HEAD(&neigh->list);
 968         ipoib_cm_set(neigh, NULL);
 969         /* one ref on behalf of the caller */
 970         atomic_set(&neigh->refcnt, 1);
 971
 972         return neigh;
 973 }
 974
 975 struct ipoib_neigh *ipoib_neigh_alloc(u8 *daddr,
 976                                       struct net_device *dev)
 977 {
 978         struct ipoib_dev_priv *priv = netdev_priv(dev);
 979         struct ipoib_neigh_table *ntbl = &priv->ntbl;
 980         struct ipoib_neigh_hash *htbl;
 981         struct ipoib_neigh *neigh;
 982         u32 hash_val;
 983
 984         htbl = rcu_dereference_protected(ntbl->htbl,
 985                                          lockdep_is_held(&priv->lock));
 986         if (!htbl) {
 987                 neigh = NULL;
 988                 goto out_unlock;
 989         }
 990
 991         /* need to add a new neigh, but maybe some other thread succeeded?
 992          * recalc hash, maybe hash resize took place so we do a search
 993          */
 994         hash_val = ipoib_addr_hash(htbl, daddr);
 995         for (neigh = rcu_dereference_protected(htbl->buckets[hash_val],
 996                                                lockdep_is_held(&priv->lock));
 997              neigh != NULL;
 998              neigh = rcu_dereference_protected(neigh->hnext,
 999                                                lockdep_is_held(&priv->lock))) {
1000                 if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) {
1001                         /* found, take one ref on behalf of the caller */
1002                         if (!atomic_inc_not_zero(&neigh->refcnt)) {
1003                                 /* deleted */
1004                                 neigh = NULL;
1005                                 break;
1006                         }
1007                         neigh->alive = jiffies;
1008                         goto out_unlock;
1009                 }
1010         }
1011
1012         neigh = ipoib_neigh_ctor(daddr, dev);
1013         if (!neigh)
1014                 goto out_unlock;
1015
1016         /* one ref on behalf of the hash table */
1017         atomic_inc(&neigh->refcnt);
1018         neigh->alive = jiffies;
1019         /* put in hash */
1020         rcu_assign_pointer(neigh->hnext,
1021                            rcu_dereference_protected(htbl->buckets[hash_val],
1022                                                      lockdep_is_held(&priv->lock)));
1023         rcu_assign_pointer(htbl->buckets[hash_val], neigh);
1024         atomic_inc(&ntbl->entries);
1025
1026 out_unlock:
1027
1028         return neigh;
1029 }
1030
1031 void ipoib_neigh_dtor(struct ipoib_neigh *neigh)
1032 {
1033         /* neigh reference count was dropprd to zero */
1034         struct net_device *dev = neigh->dev;
1035         struct ipoib_dev_priv *priv = netdev_priv(dev);
1036         struct sk_buff *skb;
1037         if (neigh->ah)
1038                 ipoib_put_ah(neigh->ah);
1039         while ((skb = __skb_dequeue(&neigh->queue))) {
1040                 ++dev->stats.tx_dropped;
1041                 dev_kfree_skb_any(skb);
1042         }
1043         if (ipoib_cm_get(neigh))
1044                 ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
1045         ipoib_dbg(netdev_priv(dev),
1046                   "neigh free for %06x %pI6\n",
1047                   IPOIB_QPN(neigh->daddr),
1048                   neigh->daddr + 4);
1049         kfree(neigh);
1050         if (atomic_dec_and_test(&priv->ntbl.entries)) {
1051                 if (test_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags))
1052                         complete(&priv->ntbl.flushed);
1053         }
1054 }
1055
1056 static void ipoib_neigh_reclaim(struct rcu_head *rp)
1057 {
1058         /* Called as a result of removal from hash table */
1059         struct ipoib_neigh *neigh = container_of(rp, struct ipoib_neigh, rcu);
1060         /* note TX context may hold another ref */
1061         ipoib_neigh_put(neigh);
1062 }
1063
1064 void ipoib_neigh_free(struct ipoib_neigh *neigh)
1065 {
1066         struct net_device *dev = neigh->dev;
1067         struct ipoib_dev_priv *priv = netdev_priv(dev);
1068         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1069         struct ipoib_neigh_hash *htbl;
1070         struct ipoib_neigh __rcu **np;
1071         struct ipoib_neigh *n;
1072         u32 hash_val;
1073
1074         htbl = rcu_dereference_protected(ntbl->htbl,
1075                                         lockdep_is_held(&priv->lock));
1076         if (!htbl)
1077                 return;
1078
1079         hash_val = ipoib_addr_hash(htbl, neigh->daddr);
1080         np = &htbl->buckets[hash_val];
1081         for (n = rcu_dereference_protected(*np,
1082                                             lockdep_is_held(&priv->lock));
1083              n != NULL;
1084              n = rcu_dereference_protected(*np,
1085                                         lockdep_is_held(&priv->lock))) {
1086                 if (n == neigh) {
1087                         /* found */
1088                         rcu_assign_pointer(*np,
1089                                            rcu_dereference_protected(neigh->hnext,
1090                                                                      lockdep_is_held(&priv->lock)));
1091                         call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1092                         return;
1093                 } else {
1094                         np = &n->hnext;
1095                 }
1096         }
1097 }
1098
1099 static int ipoib_neigh_hash_init(struct ipoib_dev_priv *priv)
1100 {
1101         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1102         struct ipoib_neigh_hash *htbl;
1103         struct ipoib_neigh **buckets;
1104         u32 size;
1105
1106         clear_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags);
1107         ntbl->htbl = NULL;
1108         htbl = kzalloc(sizeof(*htbl), GFP_KERNEL);
1109         if (!htbl)
1110                 return -ENOMEM;
1111         set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1112         size = roundup_pow_of_two(arp_tbl.gc_thresh3);
1113         buckets = kzalloc(size * sizeof(*buckets), GFP_KERNEL);
1114         if (!buckets) {
1115                 kfree(htbl);
1116                 return -ENOMEM;
1117         }
1118         htbl->size = size;
1119         htbl->mask = (size - 1);
1120         htbl->buckets = buckets;
1121         ntbl->htbl = htbl;
1122         htbl->ntbl = ntbl;
1123         atomic_set(&ntbl->entries, 0);
1124
1125         /* start garbage collection */
1126         clear_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1127         queue_delayed_work(ipoib_workqueue, &priv->neigh_reap_task,
1128                            arp_tbl.gc_interval);
1129
1130         return 0;
1131 }
1132
1133 static void neigh_hash_free_rcu(struct rcu_head *head)
1134 {
1135         struct ipoib_neigh_hash *htbl = container_of(head,
1136                                                     struct ipoib_neigh_hash,
1137                                                     rcu);
1138         struct ipoib_neigh __rcu **buckets = htbl->buckets;
1139         struct ipoib_neigh_table *ntbl = htbl->ntbl;
1140
1141         kfree(buckets);
1142         kfree(htbl);
1143         complete(&ntbl->deleted);
1144 }
1145
1146 void ipoib_del_neighs_by_gid(struct net_device *dev, u8 *gid)
1147 {
1148         struct ipoib_dev_priv *priv = netdev_priv(dev);
1149         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1150         struct ipoib_neigh_hash *htbl;
1151         unsigned long flags;
1152         int i;
1153
1154         /* remove all neigh connected to a given path or mcast */
1155         spin_lock_irqsave(&priv->lock, flags);
1156
1157         htbl = rcu_dereference_protected(ntbl->htbl,
1158                                          lockdep_is_held(&priv->lock));
1159
1160         if (!htbl)
1161                 goto out_unlock;
1162
1163         for (i = 0; i < htbl->size; i++) {
1164                 struct ipoib_neigh *neigh;
1165                 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1166
1167                 while ((neigh = rcu_dereference_protected(*np,
1168                                                           lockdep_is_held(&priv->lock))) != NULL) {
1169                         /* delete neighs belong to this parent */
1170                         if (!memcmp(gid, neigh->daddr + 4, sizeof (union ib_gid))) {
1171                                 rcu_assign_pointer(*np,
1172                                                    rcu_dereference_protected(neigh->hnext,
1173                                                                              lockdep_is_held(&priv->lock)));
1174                                 /* remove from parent list */
1175                                 list_del(&neigh->list);
1176                                 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1177                         } else {
1178                                 np = &neigh->hnext;
1179                         }
1180
1181                 }
1182         }
1183 out_unlock:
1184         spin_unlock_irqrestore(&priv->lock, flags);
1185 }
1186
1187 static void ipoib_flush_neighs(struct ipoib_dev_priv *priv)
1188 {
1189         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1190         struct ipoib_neigh_hash *htbl;
1191         unsigned long flags;
1192         int i, wait_flushed = 0;
1193
1194         init_completion(&priv->ntbl.flushed);
1195
1196         spin_lock_irqsave(&priv->lock, flags);
1197
1198         htbl = rcu_dereference_protected(ntbl->htbl,
1199                                         lockdep_is_held(&priv->lock));
1200         if (!htbl)
1201                 goto out_unlock;
1202
1203         wait_flushed = atomic_read(&priv->ntbl.entries);
1204         if (!wait_flushed)
1205                 goto free_htbl;
1206
1207         for (i = 0; i < htbl->size; i++) {
1208                 struct ipoib_neigh *neigh;
1209                 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1210
1211                 while ((neigh = rcu_dereference_protected(*np,
1212                                        lockdep_is_held(&priv->lock))) != NULL) {
1213                         rcu_assign_pointer(*np,
1214                                            rcu_dereference_protected(neigh->hnext,
1215                                                                      lockdep_is_held(&priv->lock)));
1216                         /* remove from path/mc list */
1217                         list_del(&neigh->list);
1218                         call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1219                 }
1220         }
1221
1222 free_htbl:
1223         rcu_assign_pointer(ntbl->htbl, NULL);
1224         call_rcu(&htbl->rcu, neigh_hash_free_rcu);
1225
1226 out_unlock:
1227         spin_unlock_irqrestore(&priv->lock, flags);
1228         if (wait_flushed)
1229                 wait_for_completion(&priv->ntbl.flushed);
1230 }
1231
1232 static void ipoib_neigh_hash_uninit(struct net_device *dev)
1233 {
1234         struct ipoib_dev_priv *priv = netdev_priv(dev);
1235         int stopped;
1236
1237         ipoib_dbg(priv, "ipoib_neigh_hash_uninit\n");
1238         init_completion(&priv->ntbl.deleted);
1239         set_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags);
1240
1241         /* Stop GC if called at init fail need to cancel work */
1242         stopped = test_and_set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1243         if (!stopped)
1244                 cancel_delayed_work(&priv->neigh_reap_task);
1245
1246         ipoib_flush_neighs(priv);
1247
1248         wait_for_completion(&priv->ntbl.deleted);
1249 }
1250
1251
1252 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
1253 {
1254         struct ipoib_dev_priv *priv = netdev_priv(dev);
1255
1256         if (ipoib_neigh_hash_init(priv) < 0)
1257                 goto out;
1258         /* Allocate RX/TX "rings" to hold queued skbs */
1259         priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
1260                                 GFP_KERNEL);
1261         if (!priv->rx_ring) {
1262                 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
1263                        ca->name, ipoib_recvq_size);
1264                 goto out_neigh_hash_cleanup;
1265         }
1266
1267         priv->tx_ring = vzalloc(ipoib_sendq_size * sizeof *priv->tx_ring);
1268         if (!priv->tx_ring) {
1269                 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
1270                        ca->name, ipoib_sendq_size);
1271                 goto out_rx_ring_cleanup;
1272         }
1273
1274         /* priv->tx_head, tx_tail & tx_outstanding are already 0 */
1275
1276         if (ipoib_ib_dev_init(dev, ca, port))
1277                 goto out_tx_ring_cleanup;
1278
1279         return 0;
1280
1281 out_tx_ring_cleanup:
1282         vfree(priv->tx_ring);
1283
1284 out_rx_ring_cleanup:
1285         kfree(priv->rx_ring);
1286
1287 out_neigh_hash_cleanup:
1288         ipoib_neigh_hash_uninit(dev);
1289 out:
1290         return -ENOMEM;
1291 }
1292
1293 void ipoib_dev_cleanup(struct net_device *dev)
1294 {
1295         struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
1296         LIST_HEAD(head);
1297
1298         ASSERT_RTNL();
1299
1300         ipoib_delete_debug_files(dev);
1301
1302         /* Delete any child interfaces first */
1303         list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
1304                 /* Stop GC on child */
1305                 set_bit(IPOIB_STOP_NEIGH_GC, &cpriv->flags);
1306                 cancel_delayed_work(&cpriv->neigh_reap_task);
1307                 unregister_netdevice_queue(cpriv->dev, &head);
1308         }
1309         unregister_netdevice_many(&head);
1310
1311         ipoib_ib_dev_cleanup(dev);
1312
1313         kfree(priv->rx_ring);
1314         vfree(priv->tx_ring);
1315
1316         priv->rx_ring = NULL;
1317         priv->tx_ring = NULL;
1318
1319         ipoib_neigh_hash_uninit(dev);
1320 }
1321
1322 static const struct header_ops ipoib_header_ops = {
1323         .create = ipoib_hard_header,
1324 };
1325
1326 static const struct net_device_ops ipoib_netdev_ops = {
1327         .ndo_uninit              = ipoib_uninit,
1328         .ndo_open                = ipoib_open,
1329         .ndo_stop                = ipoib_stop,
1330         .ndo_change_mtu          = ipoib_change_mtu,
1331         .ndo_fix_features        = ipoib_fix_features,
1332         .ndo_start_xmit          = ipoib_start_xmit,
1333         .ndo_tx_timeout          = ipoib_timeout,
1334         .ndo_set_rx_mode         = ipoib_set_mcast_list,
1335 };
1336
1337 void ipoib_setup(struct net_device *dev)
1338 {
1339         struct ipoib_dev_priv *priv = netdev_priv(dev);
1340
1341         dev->netdev_ops          = &ipoib_netdev_ops;
1342         dev->header_ops          = &ipoib_header_ops;
1343
1344         ipoib_set_ethtool_ops(dev);
1345
1346         netif_napi_add(dev, &priv->napi, ipoib_poll, 100);
1347
1348         dev->watchdog_timeo      = HZ;
1349
1350         dev->flags              |= IFF_BROADCAST | IFF_MULTICAST;
1351
1352         dev->hard_header_len     = IPOIB_ENCAP_LEN;
1353         dev->addr_len            = INFINIBAND_ALEN;
1354         dev->type                = ARPHRD_INFINIBAND;
1355         dev->tx_queue_len        = ipoib_sendq_size * 2;
1356         dev->features            = (NETIF_F_VLAN_CHALLENGED     |
1357                                     NETIF_F_HIGHDMA);
1358         dev->priv_flags         &= ~IFF_XMIT_DST_RELEASE;
1359
1360         memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
1361
1362         netif_carrier_off(dev);
1363
1364         priv->dev = dev;
1365
1366         spin_lock_init(&priv->lock);
1367
1368         mutex_init(&priv->vlan_mutex);
1369
1370         INIT_LIST_HEAD(&priv->path_list);
1371         INIT_LIST_HEAD(&priv->child_intfs);
1372         INIT_LIST_HEAD(&priv->dead_ahs);
1373         INIT_LIST_HEAD(&priv->multicast_list);
1374
1375         INIT_DELAYED_WORK(&priv->pkey_poll_task, ipoib_pkey_poll);
1376         INIT_DELAYED_WORK(&priv->mcast_task,   ipoib_mcast_join_task);
1377         INIT_WORK(&priv->carrier_on_task, ipoib_mcast_carrier_on_task);
1378         INIT_WORK(&priv->flush_light,   ipoib_ib_dev_flush_light);
1379         INIT_WORK(&priv->flush_normal,   ipoib_ib_dev_flush_normal);
1380         INIT_WORK(&priv->flush_heavy,   ipoib_ib_dev_flush_heavy);
1381         INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
1382         INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
1383         INIT_DELAYED_WORK(&priv->neigh_reap_task, ipoib_reap_neigh);
1384 }
1385
1386 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
1387 {
1388         struct net_device *dev;
1389
1390         dev = alloc_netdev((int) sizeof (struct ipoib_dev_priv), name,
1391                            ipoib_setup);
1392         if (!dev)
1393                 return NULL;
1394
1395         return netdev_priv(dev);
1396 }
1397
1398 static ssize_t show_pkey(struct device *dev,
1399                          struct device_attribute *attr, char *buf)
1400 {
1401         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1402
1403         return sprintf(buf, "0x%04x\n", priv->pkey);
1404 }
1405 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1406
1407 static ssize_t show_umcast(struct device *dev,
1408                            struct device_attribute *attr, char *buf)
1409 {
1410         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1411
1412         return sprintf(buf, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
1413 }
1414
1415 void ipoib_set_umcast(struct net_device *ndev, int umcast_val)
1416 {
1417         struct ipoib_dev_priv *priv = netdev_priv(ndev);
1418
1419         if (umcast_val > 0) {
1420                 set_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1421                 ipoib_warn(priv, "ignoring multicast groups joined directly "
1422                                 "by userspace\n");
1423         } else
1424                 clear_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1425 }
1426
1427 static ssize_t set_umcast(struct device *dev,
1428                           struct device_attribute *attr,
1429                           const char *buf, size_t count)
1430 {
1431         unsigned long umcast_val = simple_strtoul(buf, NULL, 0);
1432
1433         ipoib_set_umcast(to_net_dev(dev), umcast_val);
1434
1435         return count;
1436 }
1437 static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast);
1438
1439 int ipoib_add_umcast_attr(struct net_device *dev)
1440 {
1441         return device_create_file(&dev->dev, &dev_attr_umcast);
1442 }
1443
1444 static ssize_t create_child(struct device *dev,
1445                             struct device_attribute *attr,
1446                             const char *buf, size_t count)
1447 {
1448         int pkey;
1449         int ret;
1450
1451         if (sscanf(buf, "%i", &pkey) != 1)
1452                 return -EINVAL;
1453
1454         if (pkey < 0 || pkey > 0xffff)
1455                 return -EINVAL;
1456
1457         /*
1458          * Set the full membership bit, so that we join the right
1459          * broadcast group, etc.
1460          */
1461         pkey |= 0x8000;
1462
1463         ret = ipoib_vlan_add(to_net_dev(dev), pkey);
1464
1465         return ret ? ret : count;
1466 }
1467 static DEVICE_ATTR(create_child, S_IWUSR, NULL, create_child);
1468
1469 static ssize_t delete_child(struct device *dev,
1470                             struct device_attribute *attr,
1471                             const char *buf, size_t count)
1472 {
1473         int pkey;
1474         int ret;
1475
1476         if (sscanf(buf, "%i", &pkey) != 1)
1477                 return -EINVAL;
1478
1479         if (pkey < 0 || pkey > 0xffff)
1480                 return -EINVAL;
1481
1482         ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
1483
1484         return ret ? ret : count;
1485
1486 }
1487 static DEVICE_ATTR(delete_child, S_IWUSR, NULL, delete_child);
1488
1489 int ipoib_add_pkey_attr(struct net_device *dev)
1490 {
1491         return device_create_file(&dev->dev, &dev_attr_pkey);
1492 }
1493
1494 int ipoib_set_dev_features(struct ipoib_dev_priv *priv, struct ib_device *hca)
1495 {
1496         struct ib_device_attr *device_attr;
1497         int result = -ENOMEM;
1498
1499         device_attr = kmalloc(sizeof *device_attr, GFP_KERNEL);
1500         if (!device_attr) {
1501                 printk(KERN_WARNING "%s: allocation of %zu bytes failed\n",
1502                        hca->name, sizeof *device_attr);
1503                 return result;
1504         }
1505
1506         result = ib_query_device(hca, device_attr);
1507         if (result) {
1508                 printk(KERN_WARNING "%s: ib_query_device failed (ret = %d)\n",
1509                        hca->name, result);
1510                 kfree(device_attr);
1511                 return result;
1512         }
1513         priv->hca_caps = device_attr->device_cap_flags;
1514
1515         kfree(device_attr);
1516
1517         if (priv->hca_caps & IB_DEVICE_UD_IP_CSUM) {
1518                 priv->dev->hw_features = NETIF_F_SG |
1519                         NETIF_F_IP_CSUM | NETIF_F_RXCSUM;
1520
1521                 if (priv->hca_caps & IB_DEVICE_UD_TSO)
1522                         priv->dev->hw_features |= NETIF_F_TSO;
1523
1524                 priv->dev->features |= priv->dev->hw_features;
1525         }
1526
1527         return 0;
1528 }
1529
1530 static struct net_device *ipoib_add_port(const char *format,
1531                                          struct ib_device *hca, u8 port)
1532 {
1533         struct ipoib_dev_priv *priv;
1534         struct ib_port_attr attr;
1535         int result = -ENOMEM;
1536
1537         priv = ipoib_intf_alloc(format);
1538         if (!priv)
1539                 goto alloc_mem_failed;
1540
1541         SET_NETDEV_DEV(priv->dev, hca->dma_device);
1542         priv->dev->dev_id = port - 1;
1543
1544         if (!ib_query_port(hca, port, &attr))
1545                 priv->max_ib_mtu = ib_mtu_enum_to_int(attr.max_mtu);
1546         else {
1547                 printk(KERN_WARNING "%s: ib_query_port %d failed\n",
1548                        hca->name, port);
1549                 goto device_init_failed;
1550         }
1551
1552         /* MTU will be reset when mcast join happens */
1553         priv->dev->mtu  = IPOIB_UD_MTU(priv->max_ib_mtu);
1554         priv->mcast_mtu  = priv->admin_mtu = priv->dev->mtu;
1555
1556         priv->dev->neigh_priv_len = sizeof(struct ipoib_neigh);
1557
1558         result = ib_query_pkey(hca, port, 0, &priv->pkey);
1559         if (result) {
1560                 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
1561                        hca->name, port, result);
1562                 goto device_init_failed;
1563         }
1564
1565         if (ipoib_set_dev_features(priv, hca))
1566                 goto device_init_failed;
1567
1568         /*
1569          * Set the full membership bit, so that we join the right
1570          * broadcast group, etc.
1571          */
1572         priv->pkey |= 0x8000;
1573
1574         priv->dev->broadcast[8] = priv->pkey >> 8;
1575         priv->dev->broadcast[9] = priv->pkey & 0xff;
1576
1577         result = ib_query_gid(hca, port, 0, &priv->local_gid);
1578         if (result) {
1579                 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1580                        hca->name, port, result);
1581                 goto device_init_failed;
1582         } else
1583                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1584
1585         result = ipoib_dev_init(priv->dev, hca, port);
1586         if (result < 0) {
1587                 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1588                        hca->name, port, result);
1589                 goto device_init_failed;
1590         }
1591
1592         INIT_IB_EVENT_HANDLER(&priv->event_handler,
1593                               priv->ca, ipoib_event);
1594         result = ib_register_event_handler(&priv->event_handler);
1595         if (result < 0) {
1596                 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1597                        "port %d (ret = %d)\n",
1598                        hca->name, port, result);
1599                 goto event_failed;
1600         }
1601
1602         result = register_netdev(priv->dev);
1603         if (result) {
1604                 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1605                        hca->name, port, result);
1606                 goto register_failed;
1607         }
1608
1609         ipoib_create_debug_files(priv->dev);
1610
1611         if (ipoib_cm_add_mode_attr(priv->dev))
1612                 goto sysfs_failed;
1613         if (ipoib_add_pkey_attr(priv->dev))
1614                 goto sysfs_failed;
1615         if (ipoib_add_umcast_attr(priv->dev))
1616                 goto sysfs_failed;
1617         if (device_create_file(&priv->dev->dev, &dev_attr_create_child))
1618                 goto sysfs_failed;
1619         if (device_create_file(&priv->dev->dev, &dev_attr_delete_child))
1620                 goto sysfs_failed;
1621
1622         return priv->dev;
1623
1624 sysfs_failed:
1625         ipoib_delete_debug_files(priv->dev);
1626         unregister_netdev(priv->dev);
1627
1628 register_failed:
1629         ib_unregister_event_handler(&priv->event_handler);
1630         /* Stop GC if started before flush */
1631         set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1632         cancel_delayed_work(&priv->neigh_reap_task);
1633         flush_workqueue(ipoib_workqueue);
1634
1635 event_failed:
1636         ipoib_dev_cleanup(priv->dev);
1637
1638 device_init_failed:
1639         free_netdev(priv->dev);
1640
1641 alloc_mem_failed:
1642         return ERR_PTR(result);
1643 }
1644
1645 static void ipoib_add_one(struct ib_device *device)
1646 {
1647         struct list_head *dev_list;
1648         struct net_device *dev;
1649         struct ipoib_dev_priv *priv;
1650         int s, e, p;
1651
1652         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1653                 return;
1654
1655         dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1656         if (!dev_list)
1657                 return;
1658
1659         INIT_LIST_HEAD(dev_list);
1660
1661         if (device->node_type == RDMA_NODE_IB_SWITCH) {
1662                 s = 0;
1663                 e = 0;
1664         } else {
1665                 s = 1;
1666                 e = device->phys_port_cnt;
1667         }
1668
1669         for (p = s; p <= e; ++p) {
1670                 if (rdma_port_get_link_layer(device, p) != IB_LINK_LAYER_INFINIBAND)
1671                         continue;
1672                 dev = ipoib_add_port("ib%d", device, p);
1673                 if (!IS_ERR(dev)) {
1674                         priv = netdev_priv(dev);
1675                         list_add_tail(&priv->list, dev_list);
1676                 }
1677         }
1678
1679         ib_set_client_data(device, &ipoib_client, dev_list);
1680 }
1681
1682 static void ipoib_remove_one(struct ib_device *device)
1683 {
1684         struct ipoib_dev_priv *priv, *tmp;
1685         struct list_head *dev_list;
1686
1687         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1688                 return;
1689
1690         dev_list = ib_get_client_data(device, &ipoib_client);
1691
1692         list_for_each_entry_safe(priv, tmp, dev_list, list) {
1693                 ib_unregister_event_handler(&priv->event_handler);
1694
1695                 rtnl_lock();
1696                 dev_change_flags(priv->dev, priv->dev->flags & ~IFF_UP);
1697                 rtnl_unlock();
1698
1699                 /* Stop GC */
1700                 set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1701                 cancel_delayed_work(&priv->neigh_reap_task);
1702                 flush_workqueue(ipoib_workqueue);
1703
1704                 unregister_netdev(priv->dev);
1705                 free_netdev(priv->dev);
1706         }
1707
1708         kfree(dev_list);
1709 }
1710
1711 static int __init ipoib_init_module(void)
1712 {
1713         int ret;
1714
1715         ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1716         ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1717         ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1718
1719         ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1720         ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1721         ipoib_sendq_size = max3(ipoib_sendq_size, 2 * MAX_SEND_CQE, IPOIB_MIN_QUEUE_SIZE);
1722 #ifdef CONFIG_INFINIBAND_IPOIB_CM
1723         ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
1724 #endif
1725
1726         /*
1727          * When copying small received packets, we only copy from the
1728          * linear data part of the SKB, so we rely on this condition.
1729          */
1730         BUILD_BUG_ON(IPOIB_CM_COPYBREAK > IPOIB_CM_HEAD_SIZE);
1731
1732         ret = ipoib_register_debugfs();
1733         if (ret)
1734                 return ret;
1735
1736         /*
1737          * We create our own workqueue mainly because we want to be
1738          * able to flush it when devices are being removed.  We can't
1739          * use schedule_work()/flush_scheduled_work() because both
1740          * unregister_netdev() and linkwatch_event take the rtnl lock,
1741          * so flush_scheduled_work() can deadlock during device
1742          * removal.
1743          */
1744         ipoib_workqueue = create_singlethread_workqueue("ipoib");
1745         if (!ipoib_workqueue) {
1746                 ret = -ENOMEM;
1747                 goto err_fs;
1748         }
1749
1750         ib_sa_register_client(&ipoib_sa_client);
1751
1752         ret = ib_register_client(&ipoib_client);
1753         if (ret)
1754                 goto err_sa;
1755
1756         ret = ipoib_netlink_init();
1757         if (ret)
1758                 goto err_client;
1759
1760         return 0;
1761
1762 err_client:
1763         ib_unregister_client(&ipoib_client);
1764
1765 err_sa:
1766         ib_sa_unregister_client(&ipoib_sa_client);
1767         destroy_workqueue(ipoib_workqueue);
1768
1769 err_fs:
1770         ipoib_unregister_debugfs();
1771
1772         return ret;
1773 }
1774
1775 static void __exit ipoib_cleanup_module(void)
1776 {
1777         ipoib_netlink_fini();
1778         ib_unregister_client(&ipoib_client);
1779         ib_sa_unregister_client(&ipoib_sa_client);
1780         ipoib_unregister_debugfs();
1781         destroy_workqueue(ipoib_workqueue);
1782 }
1783
1784 module_init(ipoib_init_module);
1785 module_exit(ipoib_cleanup_module);