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 #define DRV_VERSION "1.0.0"
  53
  54 const char ipoib_driver_version[] = DRV_VERSION;
  55
  56 MODULE_AUTHOR("Roland Dreier");
  57 MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
  58 MODULE_LICENSE("Dual BSD/GPL");
  59 MODULE_VERSION(DRV_VERSION);
  60
  61 int ipoib_sendq_size __read_mostly = IPOIB_TX_RING_SIZE;
  62 int ipoib_recvq_size __read_mostly = IPOIB_RX_RING_SIZE;
  63
  64 module_param_named(send_queue_size, ipoib_sendq_size, int, 0444);
  65 MODULE_PARM_DESC(send_queue_size, "Number of descriptors in send queue");
  66 module_param_named(recv_queue_size, ipoib_recvq_size, int, 0444);
  67 MODULE_PARM_DESC(recv_queue_size, "Number of descriptors in receive queue");
  68
  69 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
  70 int ipoib_debug_level;
  71
  72 module_param_named(debug_level, ipoib_debug_level, int, 0644);
  73 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
  74 #endif
  75
  76 struct ipoib_path_iter {
  77         struct net_device *dev;
  78         struct ipoib_path  path;
  79 };
  80
  81 static const u8 ipv4_bcast_addr[] = {
  82         0x00, 0xff, 0xff, 0xff,
  83         0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
  84         0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff
  85 };
  86
  87 struct workqueue_struct *ipoib_workqueue;
  88
  89 struct ib_sa_client ipoib_sa_client;
  90
  91 static void ipoib_add_one(struct ib_device *device);
  92 static void ipoib_remove_one(struct ib_device *device);
  93 static void ipoib_neigh_reclaim(struct rcu_head *rp);
  94
  95 static struct ib_client ipoib_client = {
  96         .name   = "ipoib",
  97         .add    = ipoib_add_one,
  98         .remove = ipoib_remove_one
  99 };
 100
 101 int ipoib_open(struct net_device *dev)
 102 {
 103         struct ipoib_dev_priv *priv = netdev_priv(dev);
 104
 105         ipoib_dbg(priv, "bringing up interface\n");
 106
 107         netif_carrier_off(dev);
 108
 109         set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 110
 111         if (ipoib_ib_dev_open(dev)) {
 112                 if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags))
 113                         return 0;
 114                 goto err_disable;
 115         }
 116
 117         if (ipoib_ib_dev_up(dev))
 118                 goto err_stop;
 119
 120         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 121                 struct ipoib_dev_priv *cpriv;
 122
 123                 /* Bring up any child interfaces too */
 124                 down_read(&priv->vlan_rwsem);
 125                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 126                         int flags;
 127
 128                         flags = cpriv->dev->flags;
 129                         if (flags & IFF_UP)
 130                                 continue;
 131
 132                         dev_change_flags(cpriv->dev, flags | IFF_UP);
 133                 }
 134                 up_read(&priv->vlan_rwsem);
 135         }
 136
 137         netif_start_queue(dev);
 138
 139         return 0;
 140
 141 err_stop:
 142         ipoib_ib_dev_stop(dev);
 143
 144 err_disable:
 145         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 146
 147         return -EINVAL;
 148 }
 149
 150 static int ipoib_stop(struct net_device *dev)
 151 {
 152         struct ipoib_dev_priv *priv = netdev_priv(dev);
 153
 154         ipoib_dbg(priv, "stopping interface\n");
 155
 156         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 157
 158         netif_stop_queue(dev);
 159
 160         ipoib_ib_dev_down(dev);
 161         ipoib_ib_dev_stop(dev);
 162
 163         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 164                 struct ipoib_dev_priv *cpriv;
 165
 166                 /* Bring down any child interfaces too */
 167                 down_read(&priv->vlan_rwsem);
 168                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
 169                         int flags;
 170
 171                         flags = cpriv->dev->flags;
 172                         if (!(flags & IFF_UP))
 173                                 continue;
 174
 175                         dev_change_flags(cpriv->dev, flags & ~IFF_UP);
 176                 }
 177                 up_read(&priv->vlan_rwsem);
 178         }
 179
 180         return 0;
 181 }
 182
 183 static void ipoib_uninit(struct net_device *dev)
 184 {
 185         ipoib_dev_cleanup(dev);
 186 }
 187
 188 static netdev_features_t ipoib_fix_features(struct net_device *dev, netdev_features_t features)
 189 {
 190         struct ipoib_dev_priv *priv = netdev_priv(dev);
 191
 192         if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
 193                 features &= ~(NETIF_F_IP_CSUM | NETIF_F_TSO);
 194
 195         return features;
 196 }
 197
 198 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
 199 {
 200         struct ipoib_dev_priv *priv = netdev_priv(dev);
 201
 202         /* dev->mtu > 2K ==> connected mode */
 203         if (ipoib_cm_admin_enabled(dev)) {
 204                 if (new_mtu > ipoib_cm_max_mtu(dev))
 205                         return -EINVAL;
 206
 207                 if (new_mtu > priv->mcast_mtu)
 208                         ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
 209                                    priv->mcast_mtu);
 210
 211                 dev->mtu = new_mtu;
 212                 return 0;
 213         }
 214
 215         if (new_mtu > IPOIB_UD_MTU(priv->max_ib_mtu))
 216                 return -EINVAL;
 217
 218         priv->admin_mtu = new_mtu;
 219
 220         dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
 221
 222         return 0;
 223 }
 224
 225 int ipoib_set_mode(struct net_device *dev, const char *buf)
 226 {
 227         struct ipoib_dev_priv *priv = netdev_priv(dev);
 228
 229         /* flush paths if we switch modes so that connections are restarted */
 230         if (IPOIB_CM_SUPPORTED(dev->dev_addr) && !strcmp(buf, "connected\n")) {
 231                 set_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
 232                 ipoib_warn(priv, "enabling connected mode "
 233                            "will cause multicast packet drops\n");
 234                 netdev_update_features(dev);
 235                 dev_set_mtu(dev, ipoib_cm_max_mtu(dev));
 236                 rtnl_unlock();
 237                 priv->tx_wr.send_flags &= ~IB_SEND_IP_CSUM;
 238
 239                 ipoib_flush_paths(dev);
 240                 rtnl_lock();
 241                 return 0;
 242         }
 243
 244         if (!strcmp(buf, "datagram\n")) {
 245                 clear_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
 246                 netdev_update_features(dev);
 247                 dev_set_mtu(dev, min(priv->mcast_mtu, dev->mtu));
 248                 rtnl_unlock();
 249                 ipoib_flush_paths(dev);
 250                 rtnl_lock();
 251                 return 0;
 252         }
 253
 254         return -EINVAL;
 255 }
 256
 257 static struct ipoib_path *__path_find(struct net_device *dev, void *gid)
 258 {
 259         struct ipoib_dev_priv *priv = netdev_priv(dev);
 260         struct rb_node *n = priv->path_tree.rb_node;
 261         struct ipoib_path *path;
 262         int ret;
 263
 264         while (n) {
 265                 path = rb_entry(n, struct ipoib_path, rb_node);
 266
 267                 ret = memcmp(gid, path->pathrec.dgid.raw,
 268                              sizeof (union ib_gid));
 269
 270                 if (ret < 0)
 271                         n = n->rb_left;
 272                 else if (ret > 0)
 273                         n = n->rb_right;
 274                 else
 275                         return path;
 276         }
 277
 278         return NULL;
 279 }
 280
 281 static int __path_add(struct net_device *dev, struct ipoib_path *path)
 282 {
 283         struct ipoib_dev_priv *priv = netdev_priv(dev);
 284         struct rb_node **n = &priv->path_tree.rb_node;
 285         struct rb_node *pn = NULL;
 286         struct ipoib_path *tpath;
 287         int ret;
 288
 289         while (*n) {
 290                 pn = *n;
 291                 tpath = rb_entry(pn, struct ipoib_path, rb_node);
 292
 293                 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
 294                              sizeof (union ib_gid));
 295                 if (ret < 0)
 296                         n = &pn->rb_left;
 297                 else if (ret > 0)
 298                         n = &pn->rb_right;
 299                 else
 300                         return -EEXIST;
 301         }
 302
 303         rb_link_node(&path->rb_node, pn, n);
 304         rb_insert_color(&path->rb_node, &priv->path_tree);
 305
 306         list_add_tail(&path->list, &priv->path_list);
 307
 308         return 0;
 309 }
 310
 311 static void path_free(struct net_device *dev, struct ipoib_path *path)
 312 {
 313         struct sk_buff *skb;
 314
 315         while ((skb = __skb_dequeue(&path->queue)))
 316                 dev_kfree_skb_irq(skb);
 317
 318         ipoib_dbg(netdev_priv(dev), "path_free\n");
 319
 320         /* remove all neigh connected to this path */
 321         ipoib_del_neighs_by_gid(dev, path->pathrec.dgid.raw);
 322
 323         if (path->ah)
 324                 ipoib_put_ah(path->ah);
 325
 326         kfree(path);
 327 }
 328
 329 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 330
 331 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
 332 {
 333         struct ipoib_path_iter *iter;
 334
 335         iter = kmalloc(sizeof *iter, GFP_KERNEL);
 336         if (!iter)
 337                 return NULL;
 338
 339         iter->dev = dev;
 340         memset(iter->path.pathrec.dgid.raw, 0, 16);
 341
 342         if (ipoib_path_iter_next(iter)) {
 343                 kfree(iter);
 344                 return NULL;
 345         }
 346
 347         return iter;
 348 }
 349
 350 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
 351 {
 352         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
 353         struct rb_node *n;
 354         struct ipoib_path *path;
 355         int ret = 1;
 356
 357         spin_lock_irq(&priv->lock);
 358
 359         n = rb_first(&priv->path_tree);
 360
 361         while (n) {
 362                 path = rb_entry(n, struct ipoib_path, rb_node);
 363
 364                 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
 365                            sizeof (union ib_gid)) < 0) {
 366                         iter->path = *path;
 367                         ret = 0;
 368                         break;
 369                 }
 370
 371                 n = rb_next(n);
 372         }
 373
 374         spin_unlock_irq(&priv->lock);
 375
 376         return ret;
 377 }
 378
 379 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
 380                           struct ipoib_path *path)
 381 {
 382         *path = iter->path;
 383 }
 384
 385 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
 386
 387 void ipoib_mark_paths_invalid(struct net_device *dev)
 388 {
 389         struct ipoib_dev_priv *priv = netdev_priv(dev);
 390         struct ipoib_path *path, *tp;
 391
 392         spin_lock_irq(&priv->lock);
 393
 394         list_for_each_entry_safe(path, tp, &priv->path_list, list) {
 395                 ipoib_dbg(priv, "mark path LID 0x%04x GID %pI6 invalid\n",
 396                         be16_to_cpu(path->pathrec.dlid),
 397                         path->pathrec.dgid.raw);
 398                 path->valid =  0;
 399         }
 400
 401         spin_unlock_irq(&priv->lock);
 402 }
 403
 404 void ipoib_flush_paths(struct net_device *dev)
 405 {
 406         struct ipoib_dev_priv *priv = netdev_priv(dev);
 407         struct ipoib_path *path, *tp;
 408         LIST_HEAD(remove_list);
 409         unsigned long flags;
 410
 411         netif_tx_lock_bh(dev);
 412         spin_lock_irqsave(&priv->lock, flags);
 413
 414         list_splice_init(&priv->path_list, &remove_list);
 415
 416         list_for_each_entry(path, &remove_list, list)
 417                 rb_erase(&path->rb_node, &priv->path_tree);
 418
 419         list_for_each_entry_safe(path, tp, &remove_list, list) {
 420                 if (path->query)
 421                         ib_sa_cancel_query(path->query_id, path->query);
 422                 spin_unlock_irqrestore(&priv->lock, flags);
 423                 netif_tx_unlock_bh(dev);
 424                 wait_for_completion(&path->done);
 425                 path_free(dev, path);
 426                 netif_tx_lock_bh(dev);
 427                 spin_lock_irqsave(&priv->lock, flags);
 428         }
 429
 430         spin_unlock_irqrestore(&priv->lock, flags);
 431         netif_tx_unlock_bh(dev);
 432 }
 433
 434 static void path_rec_completion(int status,
 435                                 struct ib_sa_path_rec *pathrec,
 436                                 void *path_ptr)
 437 {
 438         struct ipoib_path *path = path_ptr;
 439         struct net_device *dev = path->dev;
 440         struct ipoib_dev_priv *priv = netdev_priv(dev);
 441         struct ipoib_ah *ah = NULL;
 442         struct ipoib_ah *old_ah = NULL;
 443         struct ipoib_neigh *neigh, *tn;
 444         struct sk_buff_head skqueue;
 445         struct sk_buff *skb;
 446         unsigned long flags;
 447
 448         if (!status)
 449                 ipoib_dbg(priv, "PathRec LID 0x%04x for GID %pI6\n",
 450                           be16_to_cpu(pathrec->dlid), pathrec->dgid.raw);
 451         else
 452                 ipoib_dbg(priv, "PathRec status %d for GID %pI6\n",
 453                           status, path->pathrec.dgid.raw);
 454
 455         skb_queue_head_init(&skqueue);
 456
 457         if (!status) {
 458                 struct ib_ah_attr av;
 459
 460                 if (!ib_init_ah_from_path(priv->ca, priv->port, pathrec, &av))
 461                         ah = ipoib_create_ah(dev, priv->pd, &av);
 462         }
 463
 464         spin_lock_irqsave(&priv->lock, flags);
 465
 466         if (!IS_ERR_OR_NULL(ah)) {
 467                 path->pathrec = *pathrec;
 468
 469                 old_ah   = path->ah;
 470                 path->ah = ah;
 471
 472                 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
 473                           ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
 474
 475                 while ((skb = __skb_dequeue(&path->queue)))
 476                         __skb_queue_tail(&skqueue, skb);
 477
 478                 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
 479                         if (neigh->ah) {
 480                                 WARN_ON(neigh->ah != old_ah);
 481                                 /*
 482                                  * Dropping the ah reference inside
 483                                  * priv->lock is safe here, because we
 484                                  * will hold one more reference from
 485                                  * the original value of path->ah (ie
 486                                  * old_ah).
 487                                  */
 488                                 ipoib_put_ah(neigh->ah);
 489                         }
 490                         kref_get(&path->ah->ref);
 491                         neigh->ah = path->ah;
 492
 493                         if (ipoib_cm_enabled(dev, neigh->daddr)) {
 494                                 if (!ipoib_cm_get(neigh))
 495                                         ipoib_cm_set(neigh, ipoib_cm_create_tx(dev,
 496                                                                                path,
 497                                                                                neigh));
 498                                 if (!ipoib_cm_get(neigh)) {
 499                                         ipoib_neigh_free(neigh);
 500                                         continue;
 501                                 }
 502                         }
 503
 504                         while ((skb = __skb_dequeue(&neigh->queue)))
 505                                 __skb_queue_tail(&skqueue, skb);
 506                 }
 507                 path->valid = 1;
 508         }
 509
 510         path->query = NULL;
 511         complete(&path->done);
 512
 513         spin_unlock_irqrestore(&priv->lock, flags);
 514
 515         if (IS_ERR_OR_NULL(ah))
 516                 ipoib_del_neighs_by_gid(dev, path->pathrec.dgid.raw);
 517
 518         if (old_ah)
 519                 ipoib_put_ah(old_ah);
 520
 521         while ((skb = __skb_dequeue(&skqueue))) {
 522                 skb->dev = dev;
 523                 if (dev_queue_xmit(skb))
 524                         ipoib_warn(priv, "dev_queue_xmit failed "
 525                                    "to requeue packet\n");
 526         }
 527 }
 528
 529 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
 530 {
 531         struct ipoib_dev_priv *priv = netdev_priv(dev);
 532         struct ipoib_path *path;
 533
 534         if (!priv->broadcast)
 535                 return NULL;
 536
 537         path = kzalloc(sizeof *path, GFP_ATOMIC);
 538         if (!path)
 539                 return NULL;
 540
 541         path->dev = dev;
 542
 543         skb_queue_head_init(&path->queue);
 544
 545         INIT_LIST_HEAD(&path->neigh_list);
 546
 547         memcpy(path->pathrec.dgid.raw, gid, sizeof (union ib_gid));
 548         path->pathrec.sgid          = priv->local_gid;
 549         path->pathrec.pkey          = cpu_to_be16(priv->pkey);
 550         path->pathrec.numb_path     = 1;
 551         path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
 552
 553         return path;
 554 }
 555
 556 static int path_rec_start(struct net_device *dev,
 557                           struct ipoib_path *path)
 558 {
 559         struct ipoib_dev_priv *priv = netdev_priv(dev);
 560
 561         ipoib_dbg(priv, "Start path record lookup for %pI6\n",
 562                   path->pathrec.dgid.raw);
 563
 564         init_completion(&path->done);
 565
 566         path->query_id =
 567                 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
 568                                    &path->pathrec,
 569                                    IB_SA_PATH_REC_DGID          |
 570                                    IB_SA_PATH_REC_SGID          |
 571                                    IB_SA_PATH_REC_NUMB_PATH     |
 572                                    IB_SA_PATH_REC_TRAFFIC_CLASS |
 573                                    IB_SA_PATH_REC_PKEY,
 574                                    1000, GFP_ATOMIC,
 575                                    path_rec_completion,
 576                                    path, &path->query);
 577         if (path->query_id < 0) {
 578                 ipoib_warn(priv, "ib_sa_path_rec_get failed: %d\n", path->query_id);
 579                 path->query = NULL;
 580                 complete(&path->done);
 581                 return path->query_id;
 582         }
 583
 584         return 0;
 585 }
 586
 587 static void neigh_add_path(struct sk_buff *skb, u8 *daddr,
 588                            struct net_device *dev)
 589 {
 590         struct ipoib_dev_priv *priv = netdev_priv(dev);
 591         struct ipoib_path *path;
 592         struct ipoib_neigh *neigh;
 593         unsigned long flags;
 594
 595         spin_lock_irqsave(&priv->lock, flags);
 596         neigh = ipoib_neigh_alloc(daddr, dev);
 597         if (!neigh) {
 598                 spin_unlock_irqrestore(&priv->lock, flags);
 599                 ++dev->stats.tx_dropped;
 600                 dev_kfree_skb_any(skb);
 601                 return;
 602         }
 603
 604         path = __path_find(dev, daddr + 4);
 605         if (!path) {
 606                 path = path_rec_create(dev, daddr + 4);
 607                 if (!path)
 608                         goto err_path;
 609
 610                 __path_add(dev, path);
 611         }
 612
 613         list_add_tail(&neigh->list, &path->neigh_list);
 614
 615         if (path->ah) {
 616                 kref_get(&path->ah->ref);
 617                 neigh->ah = path->ah;
 618
 619                 if (ipoib_cm_enabled(dev, neigh->daddr)) {
 620                         if (!ipoib_cm_get(neigh))
 621                                 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
 622                         if (!ipoib_cm_get(neigh)) {
 623                                 ipoib_neigh_free(neigh);
 624                                 goto err_drop;
 625                         }
 626                         if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE)
 627                                 __skb_queue_tail(&neigh->queue, skb);
 628                         else {
 629                                 ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
 630                                            skb_queue_len(&neigh->queue));
 631                                 goto err_drop;
 632                         }
 633                 } else {
 634                         spin_unlock_irqrestore(&priv->lock, flags);
 635                         ipoib_send(dev, skb, path->ah, IPOIB_QPN(daddr));
 636                         ipoib_neigh_put(neigh);
 637                         return;
 638                 }
 639         } else {
 640                 neigh->ah  = NULL;
 641
 642                 if (!path->query && path_rec_start(dev, path))
 643                         goto err_path;
 644                 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE)
 645                         __skb_queue_tail(&neigh->queue, skb);
 646                 else
 647                         goto err_drop;
 648         }
 649
 650         spin_unlock_irqrestore(&priv->lock, flags);
 651         ipoib_neigh_put(neigh);
 652         return;
 653
 654 err_path:
 655         ipoib_neigh_free(neigh);
 656 err_drop:
 657         ++dev->stats.tx_dropped;
 658         dev_kfree_skb_any(skb);
 659
 660         spin_unlock_irqrestore(&priv->lock, flags);
 661         ipoib_neigh_put(neigh);
 662 }
 663
 664 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
 665                              struct ipoib_cb *cb)
 666 {
 667         struct ipoib_dev_priv *priv = netdev_priv(dev);
 668         struct ipoib_path *path;
 669         unsigned long flags;
 670
 671         spin_lock_irqsave(&priv->lock, flags);
 672
 673         path = __path_find(dev, cb->hwaddr + 4);
 674         if (!path || !path->valid) {
 675                 int new_path = 0;
 676
 677                 if (!path) {
 678                         path = path_rec_create(dev, cb->hwaddr + 4);
 679                         new_path = 1;
 680                 }
 681                 if (path) {
 682                         if (skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 683                                 __skb_queue_tail(&path->queue, skb);
 684                         } else {
 685                                 ++dev->stats.tx_dropped;
 686                                 dev_kfree_skb_any(skb);
 687                         }
 688
 689                         if (!path->query && path_rec_start(dev, path)) {
 690                                 spin_unlock_irqrestore(&priv->lock, flags);
 691                                 if (new_path)
 692                                         path_free(dev, path);
 693                                 return;
 694                         } else
 695                                 __path_add(dev, path);
 696                 } else {
 697                         ++dev->stats.tx_dropped;
 698                         dev_kfree_skb_any(skb);
 699                 }
 700
 701                 spin_unlock_irqrestore(&priv->lock, flags);
 702                 return;
 703         }
 704
 705         if (path->ah) {
 706                 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
 707                           be16_to_cpu(path->pathrec.dlid));
 708
 709                 spin_unlock_irqrestore(&priv->lock, flags);
 710                 ipoib_send(dev, skb, path->ah, IPOIB_QPN(cb->hwaddr));
 711                 return;
 712         } else if ((path->query || !path_rec_start(dev, path)) &&
 713                    skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 714                 __skb_queue_tail(&path->queue, skb);
 715         } else {
 716                 ++dev->stats.tx_dropped;
 717                 dev_kfree_skb_any(skb);
 718         }
 719
 720         spin_unlock_irqrestore(&priv->lock, flags);
 721 }
 722
 723 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
 724 {
 725         struct ipoib_dev_priv *priv = netdev_priv(dev);
 726         struct ipoib_neigh *neigh;
 727         struct ipoib_cb *cb = ipoib_skb_cb(skb);
 728         struct ipoib_header *header;
 729         unsigned long flags;
 730
 731         header = (struct ipoib_header *) skb->data;
 732
 733         if (unlikely(cb->hwaddr[4] == 0xff)) {
 734                 /* multicast, arrange "if" according to probability */
 735                 if ((header->proto != htons(ETH_P_IP)) &&
 736                     (header->proto != htons(ETH_P_IPV6)) &&
 737                     (header->proto != htons(ETH_P_ARP)) &&
 738                     (header->proto != htons(ETH_P_RARP)) &&
 739                     (header->proto != htons(ETH_P_TIPC))) {
 740                         /* ethertype not supported by IPoIB */
 741                         ++dev->stats.tx_dropped;
 742                         dev_kfree_skb_any(skb);
 743                         return NETDEV_TX_OK;
 744                 }
 745                 /* Add in the P_Key for multicast*/
 746                 cb->hwaddr[8] = (priv->pkey >> 8) & 0xff;
 747                 cb->hwaddr[9] = priv->pkey & 0xff;
 748
 749                 neigh = ipoib_neigh_get(dev, cb->hwaddr);
 750                 if (likely(neigh))
 751                         goto send_using_neigh;
 752                 ipoib_mcast_send(dev, cb->hwaddr, skb);
 753                 return NETDEV_TX_OK;
 754         }
 755
 756         /* unicast, arrange "switch" according to probability */
 757         switch (header->proto) {
 758         case htons(ETH_P_IP):
 759         case htons(ETH_P_IPV6):
 760         case htons(ETH_P_TIPC):
 761                 neigh = ipoib_neigh_get(dev, cb->hwaddr);
 762                 if (unlikely(!neigh)) {
 763                         neigh_add_path(skb, cb->hwaddr, dev);
 764                         return NETDEV_TX_OK;
 765                 }
 766                 break;
 767         case htons(ETH_P_ARP):
 768         case htons(ETH_P_RARP):
 769                 /* for unicast ARP and RARP should always perform path find */
 770                 unicast_arp_send(skb, dev, cb);
 771                 return NETDEV_TX_OK;
 772         default:
 773                 /* ethertype not supported by IPoIB */
 774                 ++dev->stats.tx_dropped;
 775                 dev_kfree_skb_any(skb);
 776                 return NETDEV_TX_OK;
 777         }
 778
 779 send_using_neigh:
 780         /* note we now hold a ref to neigh */
 781         if (ipoib_cm_get(neigh)) {
 782                 if (ipoib_cm_up(neigh)) {
 783                         ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
 784                         goto unref;
 785                 }
 786         } else if (neigh->ah) {
 787                 ipoib_send(dev, skb, neigh->ah, IPOIB_QPN(cb->hwaddr));
 788                 goto unref;
 789         }
 790
 791         if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 792                 spin_lock_irqsave(&priv->lock, flags);
 793                 __skb_queue_tail(&neigh->queue, skb);
 794                 spin_unlock_irqrestore(&priv->lock, flags);
 795         } else {
 796                 ++dev->stats.tx_dropped;
 797                 dev_kfree_skb_any(skb);
 798         }
 799
 800 unref:
 801         ipoib_neigh_put(neigh);
 802
 803         return NETDEV_TX_OK;
 804 }
 805
 806 static void ipoib_timeout(struct net_device *dev)
 807 {
 808         struct ipoib_dev_priv *priv = netdev_priv(dev);
 809
 810         ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
 811                    jiffies_to_msecs(jiffies - dev->trans_start));
 812         ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
 813                    netif_queue_stopped(dev),
 814                    priv->tx_head, priv->tx_tail);
 815         /* XXX reset QP, etc. */
 816 }
 817
 818 static int ipoib_hard_header(struct sk_buff *skb,
 819                              struct net_device *dev,
 820                              unsigned short type,
 821                              const void *daddr, const void *saddr, unsigned len)
 822 {
 823         struct ipoib_header *header;
 824         struct ipoib_cb *cb = ipoib_skb_cb(skb);
 825
 826         header = (struct ipoib_header *) skb_push(skb, sizeof *header);
 827
 828         header->proto = htons(type);
 829         header->reserved = 0;
 830
 831         /*
 832          * we don't rely on dst_entry structure,  always stuff the
 833          * destination address into skb->cb so we can figure out where
 834          * to send the packet later.
 835          */
 836         memcpy(cb->hwaddr, daddr, INFINIBAND_ALEN);
 837
 838         return sizeof *header;
 839 }
 840
 841 static void ipoib_set_mcast_list(struct net_device *dev)
 842 {
 843         struct ipoib_dev_priv *priv = netdev_priv(dev);
 844
 845         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
 846                 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
 847                 return;
 848         }
 849
 850         queue_work(priv->wq, &priv->restart_task);
 851 }
 852
 853 static int ipoib_get_iflink(const struct net_device *dev)
 854 {
 855         struct ipoib_dev_priv *priv = netdev_priv(dev);
 856
 857         /* parent interface */
 858         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags))
 859                 return dev->ifindex;
 860
 861         /* child/vlan interface */
 862         return priv->parent->ifindex;
 863 }
 864
 865 static u32 ipoib_addr_hash(struct ipoib_neigh_hash *htbl, u8 *daddr)
 866 {
 867         /*
 868          * Use only the address parts that contributes to spreading
 869          * The subnet prefix is not used as one can not connect to
 870          * same remote port (GUID) using the same remote QPN via two
 871          * different subnets.
 872          */
 873          /* qpn octets[1:4) & port GUID octets[12:20) */
 874         u32 *d32 = (u32 *) daddr;
 875         u32 hv;
 876
 877         hv = jhash_3words(d32[3], d32[4], IPOIB_QPN_MASK & d32[0], 0);
 878         return hv & htbl->mask;
 879 }
 880
 881 struct ipoib_neigh *ipoib_neigh_get(struct net_device *dev, u8 *daddr)
 882 {
 883         struct ipoib_dev_priv *priv = netdev_priv(dev);
 884         struct ipoib_neigh_table *ntbl = &priv->ntbl;
 885         struct ipoib_neigh_hash *htbl;
 886         struct ipoib_neigh *neigh = NULL;
 887         u32 hash_val;
 888
 889         rcu_read_lock_bh();
 890
 891         htbl = rcu_dereference_bh(ntbl->htbl);
 892
 893         if (!htbl)
 894                 goto out_unlock;
 895
 896         hash_val = ipoib_addr_hash(htbl, daddr);
 897         for (neigh = rcu_dereference_bh(htbl->buckets[hash_val]);
 898              neigh != NULL;
 899              neigh = rcu_dereference_bh(neigh->hnext)) {
 900                 if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) {
 901                         /* found, take one ref on behalf of the caller */
 902                         if (!atomic_inc_not_zero(&neigh->refcnt)) {
 903                                 /* deleted */
 904                                 neigh = NULL;
 905                                 goto out_unlock;
 906                         }
 907                         neigh->alive = jiffies;
 908                         goto out_unlock;
 909                 }
 910         }
 911
 912 out_unlock:
 913         rcu_read_unlock_bh();
 914         return neigh;
 915 }
 916
 917 static void __ipoib_reap_neigh(struct ipoib_dev_priv *priv)
 918 {
 919         struct ipoib_neigh_table *ntbl = &priv->ntbl;
 920         struct ipoib_neigh_hash *htbl;
 921         unsigned long neigh_obsolete;
 922         unsigned long dt;
 923         unsigned long flags;
 924         int i;
 925
 926         if (test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
 927                 return;
 928
 929         spin_lock_irqsave(&priv->lock, flags);
 930
 931         htbl = rcu_dereference_protected(ntbl->htbl,
 932                                          lockdep_is_held(&priv->lock));
 933
 934         if (!htbl)
 935                 goto out_unlock;
 936
 937         /* neigh is obsolete if it was idle for two GC periods */
 938         dt = 2 * arp_tbl.gc_interval;
 939         neigh_obsolete = jiffies - dt;
 940         /* handle possible race condition */
 941         if (test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
 942                 goto out_unlock;
 943
 944         for (i = 0; i < htbl->size; i++) {
 945                 struct ipoib_neigh *neigh;
 946                 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
 947
 948                 while ((neigh = rcu_dereference_protected(*np,
 949                                                           lockdep_is_held(&priv->lock))) != NULL) {
 950                         /* was the neigh idle for two GC periods */
 951                         if (time_after(neigh_obsolete, neigh->alive)) {
 952                                 rcu_assign_pointer(*np,
 953                                                    rcu_dereference_protected(neigh->hnext,
 954                                                                              lockdep_is_held(&priv->lock)));
 955                                 /* remove from path/mc list */
 956                                 list_del(&neigh->list);
 957                                 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
 958                         } else {
 959                                 np = &neigh->hnext;
 960                         }
 961
 962                 }
 963         }
 964
 965 out_unlock:
 966         spin_unlock_irqrestore(&priv->lock, flags);
 967 }
 968
 969 static void ipoib_reap_neigh(struct work_struct *work)
 970 {
 971         struct ipoib_dev_priv *priv =
 972                 container_of(work, struct ipoib_dev_priv, neigh_reap_task.work);
 973
 974         __ipoib_reap_neigh(priv);
 975
 976         if (!test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
 977                 queue_delayed_work(priv->wq, &priv->neigh_reap_task,
 978                                    arp_tbl.gc_interval);
 979 }
 980
 981
 982 static struct ipoib_neigh *ipoib_neigh_ctor(u8 *daddr,
 983                                       struct net_device *dev)
 984 {
 985         struct ipoib_neigh *neigh;
 986
 987         neigh = kzalloc(sizeof *neigh, GFP_ATOMIC);
 988         if (!neigh)
 989                 return NULL;
 990
 991         neigh->dev = dev;
 992         memcpy(&neigh->daddr, daddr, sizeof(neigh->daddr));
 993         skb_queue_head_init(&neigh->queue);
 994         INIT_LIST_HEAD(&neigh->list);
 995         ipoib_cm_set(neigh, NULL);
 996         /* one ref on behalf of the caller */
 997         atomic_set(&neigh->refcnt, 1);
 998
 999         return neigh;
1000 }
1001
1002 struct ipoib_neigh *ipoib_neigh_alloc(u8 *daddr,
1003                                       struct net_device *dev)
1004 {
1005         struct ipoib_dev_priv *priv = netdev_priv(dev);
1006         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1007         struct ipoib_neigh_hash *htbl;
1008         struct ipoib_neigh *neigh;
1009         u32 hash_val;
1010
1011         htbl = rcu_dereference_protected(ntbl->htbl,
1012                                          lockdep_is_held(&priv->lock));
1013         if (!htbl) {
1014                 neigh = NULL;
1015                 goto out_unlock;
1016         }
1017
1018         /* need to add a new neigh, but maybe some other thread succeeded?
1019          * recalc hash, maybe hash resize took place so we do a search
1020          */
1021         hash_val = ipoib_addr_hash(htbl, daddr);
1022         for (neigh = rcu_dereference_protected(htbl->buckets[hash_val],
1023                                                lockdep_is_held(&priv->lock));
1024              neigh != NULL;
1025              neigh = rcu_dereference_protected(neigh->hnext,
1026                                                lockdep_is_held(&priv->lock))) {
1027                 if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) {
1028                         /* found, take one ref on behalf of the caller */
1029                         if (!atomic_inc_not_zero(&neigh->refcnt)) {
1030                                 /* deleted */
1031                                 neigh = NULL;
1032                                 break;
1033                         }
1034                         neigh->alive = jiffies;
1035                         goto out_unlock;
1036                 }
1037         }
1038
1039         neigh = ipoib_neigh_ctor(daddr, dev);
1040         if (!neigh)
1041                 goto out_unlock;
1042
1043         /* one ref on behalf of the hash table */
1044         atomic_inc(&neigh->refcnt);
1045         neigh->alive = jiffies;
1046         /* put in hash */
1047         rcu_assign_pointer(neigh->hnext,
1048                            rcu_dereference_protected(htbl->buckets[hash_val],
1049                                                      lockdep_is_held(&priv->lock)));
1050         rcu_assign_pointer(htbl->buckets[hash_val], neigh);
1051         atomic_inc(&ntbl->entries);
1052
1053 out_unlock:
1054
1055         return neigh;
1056 }
1057
1058 void ipoib_neigh_dtor(struct ipoib_neigh *neigh)
1059 {
1060         /* neigh reference count was dropprd to zero */
1061         struct net_device *dev = neigh->dev;
1062         struct ipoib_dev_priv *priv = netdev_priv(dev);
1063         struct sk_buff *skb;
1064         if (neigh->ah)
1065                 ipoib_put_ah(neigh->ah);
1066         while ((skb = __skb_dequeue(&neigh->queue))) {
1067                 ++dev->stats.tx_dropped;
1068                 dev_kfree_skb_any(skb);
1069         }
1070         if (ipoib_cm_get(neigh))
1071                 ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
1072         ipoib_dbg(netdev_priv(dev),
1073                   "neigh free for %06x %pI6\n",
1074                   IPOIB_QPN(neigh->daddr),
1075                   neigh->daddr + 4);
1076         kfree(neigh);
1077         if (atomic_dec_and_test(&priv->ntbl.entries)) {
1078                 if (test_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags))
1079                         complete(&priv->ntbl.flushed);
1080         }
1081 }
1082
1083 static void ipoib_neigh_reclaim(struct rcu_head *rp)
1084 {
1085         /* Called as a result of removal from hash table */
1086         struct ipoib_neigh *neigh = container_of(rp, struct ipoib_neigh, rcu);
1087         /* note TX context may hold another ref */
1088         ipoib_neigh_put(neigh);
1089 }
1090
1091 void ipoib_neigh_free(struct ipoib_neigh *neigh)
1092 {
1093         struct net_device *dev = neigh->dev;
1094         struct ipoib_dev_priv *priv = netdev_priv(dev);
1095         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1096         struct ipoib_neigh_hash *htbl;
1097         struct ipoib_neigh __rcu **np;
1098         struct ipoib_neigh *n;
1099         u32 hash_val;
1100
1101         htbl = rcu_dereference_protected(ntbl->htbl,
1102                                         lockdep_is_held(&priv->lock));
1103         if (!htbl)
1104                 return;
1105
1106         hash_val = ipoib_addr_hash(htbl, neigh->daddr);
1107         np = &htbl->buckets[hash_val];
1108         for (n = rcu_dereference_protected(*np,
1109                                             lockdep_is_held(&priv->lock));
1110              n != NULL;
1111              n = rcu_dereference_protected(*np,
1112                                         lockdep_is_held(&priv->lock))) {
1113                 if (n == neigh) {
1114                         /* found */
1115                         rcu_assign_pointer(*np,
1116                                            rcu_dereference_protected(neigh->hnext,
1117                                                                      lockdep_is_held(&priv->lock)));
1118                         /* remove from parent list */
1119                         list_del(&neigh->list);
1120                         call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1121                         return;
1122                 } else {
1123                         np = &n->hnext;
1124                 }
1125         }
1126 }
1127
1128 static int ipoib_neigh_hash_init(struct ipoib_dev_priv *priv)
1129 {
1130         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1131         struct ipoib_neigh_hash *htbl;
1132         struct ipoib_neigh __rcu **buckets;
1133         u32 size;
1134
1135         clear_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags);
1136         ntbl->htbl = NULL;
1137         htbl = kzalloc(sizeof(*htbl), GFP_KERNEL);
1138         if (!htbl)
1139                 return -ENOMEM;
1140         set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1141         size = roundup_pow_of_two(arp_tbl.gc_thresh3);
1142         buckets = kzalloc(size * sizeof(*buckets), GFP_KERNEL);
1143         if (!buckets) {
1144                 kfree(htbl);
1145                 return -ENOMEM;
1146         }
1147         htbl->size = size;
1148         htbl->mask = (size - 1);
1149         htbl->buckets = buckets;
1150         RCU_INIT_POINTER(ntbl->htbl, htbl);
1151         htbl->ntbl = ntbl;
1152         atomic_set(&ntbl->entries, 0);
1153
1154         /* start garbage collection */
1155         clear_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1156         queue_delayed_work(priv->wq, &priv->neigh_reap_task,
1157                            arp_tbl.gc_interval);
1158
1159         return 0;
1160 }
1161
1162 static void neigh_hash_free_rcu(struct rcu_head *head)
1163 {
1164         struct ipoib_neigh_hash *htbl = container_of(head,
1165                                                     struct ipoib_neigh_hash,
1166                                                     rcu);
1167         struct ipoib_neigh __rcu **buckets = htbl->buckets;
1168         struct ipoib_neigh_table *ntbl = htbl->ntbl;
1169
1170         kfree(buckets);
1171         kfree(htbl);
1172         complete(&ntbl->deleted);
1173 }
1174
1175 void ipoib_del_neighs_by_gid(struct net_device *dev, u8 *gid)
1176 {
1177         struct ipoib_dev_priv *priv = netdev_priv(dev);
1178         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1179         struct ipoib_neigh_hash *htbl;
1180         unsigned long flags;
1181         int i;
1182
1183         /* remove all neigh connected to a given path or mcast */
1184         spin_lock_irqsave(&priv->lock, flags);
1185
1186         htbl = rcu_dereference_protected(ntbl->htbl,
1187                                          lockdep_is_held(&priv->lock));
1188
1189         if (!htbl)
1190                 goto out_unlock;
1191
1192         for (i = 0; i < htbl->size; i++) {
1193                 struct ipoib_neigh *neigh;
1194                 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1195
1196                 while ((neigh = rcu_dereference_protected(*np,
1197                                                           lockdep_is_held(&priv->lock))) != NULL) {
1198                         /* delete neighs belong to this parent */
1199                         if (!memcmp(gid, neigh->daddr + 4, sizeof (union ib_gid))) {
1200                                 rcu_assign_pointer(*np,
1201                                                    rcu_dereference_protected(neigh->hnext,
1202                                                                              lockdep_is_held(&priv->lock)));
1203                                 /* remove from parent list */
1204                                 list_del(&neigh->list);
1205                                 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1206                         } else {
1207                                 np = &neigh->hnext;
1208                         }
1209
1210                 }
1211         }
1212 out_unlock:
1213         spin_unlock_irqrestore(&priv->lock, flags);
1214 }
1215
1216 static void ipoib_flush_neighs(struct ipoib_dev_priv *priv)
1217 {
1218         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1219         struct ipoib_neigh_hash *htbl;
1220         unsigned long flags;
1221         int i, wait_flushed = 0;
1222
1223         init_completion(&priv->ntbl.flushed);
1224
1225         spin_lock_irqsave(&priv->lock, flags);
1226
1227         htbl = rcu_dereference_protected(ntbl->htbl,
1228                                         lockdep_is_held(&priv->lock));
1229         if (!htbl)
1230                 goto out_unlock;
1231
1232         wait_flushed = atomic_read(&priv->ntbl.entries);
1233         if (!wait_flushed)
1234                 goto free_htbl;
1235
1236         for (i = 0; i < htbl->size; i++) {
1237                 struct ipoib_neigh *neigh;
1238                 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1239
1240                 while ((neigh = rcu_dereference_protected(*np,
1241                                        lockdep_is_held(&priv->lock))) != NULL) {
1242                         rcu_assign_pointer(*np,
1243                                            rcu_dereference_protected(neigh->hnext,
1244                                                                      lockdep_is_held(&priv->lock)));
1245                         /* remove from path/mc list */
1246                         list_del(&neigh->list);
1247                         call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1248                 }
1249         }
1250
1251 free_htbl:
1252         rcu_assign_pointer(ntbl->htbl, NULL);
1253         call_rcu(&htbl->rcu, neigh_hash_free_rcu);
1254
1255 out_unlock:
1256         spin_unlock_irqrestore(&priv->lock, flags);
1257         if (wait_flushed)
1258                 wait_for_completion(&priv->ntbl.flushed);
1259 }
1260
1261 static void ipoib_neigh_hash_uninit(struct net_device *dev)
1262 {
1263         struct ipoib_dev_priv *priv = netdev_priv(dev);
1264         int stopped;
1265
1266         ipoib_dbg(priv, "ipoib_neigh_hash_uninit\n");
1267         init_completion(&priv->ntbl.deleted);
1268         set_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags);
1269
1270         /* Stop GC if called at init fail need to cancel work */
1271         stopped = test_and_set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1272         if (!stopped)
1273                 cancel_delayed_work(&priv->neigh_reap_task);
1274
1275         ipoib_flush_neighs(priv);
1276
1277         wait_for_completion(&priv->ntbl.deleted);
1278 }
1279
1280
1281 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
1282 {
1283         struct ipoib_dev_priv *priv = netdev_priv(dev);
1284
1285         /* Allocate RX/TX "rings" to hold queued skbs */
1286         priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
1287                                 GFP_KERNEL);
1288         if (!priv->rx_ring) {
1289                 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
1290                        ca->name, ipoib_recvq_size);
1291                 goto out;
1292         }
1293
1294         priv->tx_ring = vzalloc(ipoib_sendq_size * sizeof *priv->tx_ring);
1295         if (!priv->tx_ring) {
1296                 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
1297                        ca->name, ipoib_sendq_size);
1298                 goto out_rx_ring_cleanup;
1299         }
1300
1301         /* priv->tx_head, tx_tail & tx_outstanding are already 0 */
1302
1303         if (ipoib_ib_dev_init(dev, ca, port))
1304                 goto out_tx_ring_cleanup;
1305
1306         /*
1307          * Must be after ipoib_ib_dev_init so we can allocate a per
1308          * device wq there and use it here
1309          */
1310         if (ipoib_neigh_hash_init(priv) < 0)
1311                 goto out_dev_uninit;
1312
1313         return 0;
1314
1315 out_dev_uninit:
1316         ipoib_ib_dev_cleanup(dev);
1317
1318 out_tx_ring_cleanup:
1319         vfree(priv->tx_ring);
1320
1321 out_rx_ring_cleanup:
1322         kfree(priv->rx_ring);
1323
1324 out:
1325         return -ENOMEM;
1326 }
1327
1328 void ipoib_dev_cleanup(struct net_device *dev)
1329 {
1330         struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
1331         LIST_HEAD(head);
1332
1333         ASSERT_RTNL();
1334
1335         ipoib_delete_debug_files(dev);
1336
1337         /* Delete any child interfaces first */
1338         list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
1339                 /* Stop GC on child */
1340                 set_bit(IPOIB_STOP_NEIGH_GC, &cpriv->flags);
1341                 cancel_delayed_work(&cpriv->neigh_reap_task);
1342                 unregister_netdevice_queue(cpriv->dev, &head);
1343         }
1344         unregister_netdevice_many(&head);
1345
1346         /*
1347          * Must be before ipoib_ib_dev_cleanup or we delete an in use
1348          * work queue
1349          */
1350         ipoib_neigh_hash_uninit(dev);
1351
1352         ipoib_ib_dev_cleanup(dev);
1353
1354         kfree(priv->rx_ring);
1355         vfree(priv->tx_ring);
1356
1357         priv->rx_ring = NULL;
1358         priv->tx_ring = NULL;
1359 }
1360
1361 static const struct header_ops ipoib_header_ops = {
1362         .create = ipoib_hard_header,
1363 };
1364
1365 static const struct net_device_ops ipoib_netdev_ops = {
1366         .ndo_uninit              = ipoib_uninit,
1367         .ndo_open                = ipoib_open,
1368         .ndo_stop                = ipoib_stop,
1369         .ndo_change_mtu          = ipoib_change_mtu,
1370         .ndo_fix_features        = ipoib_fix_features,
1371         .ndo_start_xmit          = ipoib_start_xmit,
1372         .ndo_tx_timeout          = ipoib_timeout,
1373         .ndo_set_rx_mode         = ipoib_set_mcast_list,
1374         .ndo_get_iflink          = ipoib_get_iflink,
1375 };
1376
1377 void ipoib_setup(struct net_device *dev)
1378 {
1379         struct ipoib_dev_priv *priv = netdev_priv(dev);
1380
1381         dev->netdev_ops          = &ipoib_netdev_ops;
1382         dev->header_ops          = &ipoib_header_ops;
1383
1384         ipoib_set_ethtool_ops(dev);
1385
1386         netif_napi_add(dev, &priv->napi, ipoib_poll, NAPI_POLL_WEIGHT);
1387
1388         dev->watchdog_timeo      = HZ;
1389
1390         dev->flags              |= IFF_BROADCAST | IFF_MULTICAST;
1391
1392         dev->hard_header_len     = IPOIB_ENCAP_LEN;
1393         dev->addr_len            = INFINIBAND_ALEN;
1394         dev->type                = ARPHRD_INFINIBAND;
1395         dev->tx_queue_len        = ipoib_sendq_size * 2;
1396         dev->features            = (NETIF_F_VLAN_CHALLENGED     |
1397                                     NETIF_F_HIGHDMA);
1398         netif_keep_dst(dev);
1399
1400         memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
1401
1402         priv->dev = dev;
1403
1404         spin_lock_init(&priv->lock);
1405
1406         init_rwsem(&priv->vlan_rwsem);
1407
1408         INIT_LIST_HEAD(&priv->path_list);
1409         INIT_LIST_HEAD(&priv->child_intfs);
1410         INIT_LIST_HEAD(&priv->dead_ahs);
1411         INIT_LIST_HEAD(&priv->multicast_list);
1412
1413         INIT_DELAYED_WORK(&priv->mcast_task,   ipoib_mcast_join_task);
1414         INIT_WORK(&priv->carrier_on_task, ipoib_mcast_carrier_on_task);
1415         INIT_WORK(&priv->flush_light,   ipoib_ib_dev_flush_light);
1416         INIT_WORK(&priv->flush_normal,   ipoib_ib_dev_flush_normal);
1417         INIT_WORK(&priv->flush_heavy,   ipoib_ib_dev_flush_heavy);
1418         INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
1419         INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
1420         INIT_DELAYED_WORK(&priv->neigh_reap_task, ipoib_reap_neigh);
1421 }
1422
1423 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
1424 {
1425         struct net_device *dev;
1426
1427         dev = alloc_netdev((int)sizeof(struct ipoib_dev_priv), name,
1428                            NET_NAME_UNKNOWN, ipoib_setup);
1429         if (!dev)
1430                 return NULL;
1431
1432         return netdev_priv(dev);
1433 }
1434
1435 static ssize_t show_pkey(struct device *dev,
1436                          struct device_attribute *attr, char *buf)
1437 {
1438         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1439
1440         return sprintf(buf, "0x%04x\n", priv->pkey);
1441 }
1442 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1443
1444 static ssize_t show_umcast(struct device *dev,
1445                            struct device_attribute *attr, char *buf)
1446 {
1447         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1448
1449         return sprintf(buf, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
1450 }
1451
1452 void ipoib_set_umcast(struct net_device *ndev, int umcast_val)
1453 {
1454         struct ipoib_dev_priv *priv = netdev_priv(ndev);
1455
1456         if (umcast_val > 0) {
1457                 set_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1458                 ipoib_warn(priv, "ignoring multicast groups joined directly "
1459                                 "by userspace\n");
1460         } else
1461                 clear_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1462 }
1463
1464 static ssize_t set_umcast(struct device *dev,
1465                           struct device_attribute *attr,
1466                           const char *buf, size_t count)
1467 {
1468         unsigned long umcast_val = simple_strtoul(buf, NULL, 0);
1469
1470         ipoib_set_umcast(to_net_dev(dev), umcast_val);
1471
1472         return count;
1473 }
1474 static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast);
1475
1476 int ipoib_add_umcast_attr(struct net_device *dev)
1477 {
1478         return device_create_file(&dev->dev, &dev_attr_umcast);
1479 }
1480
1481 static ssize_t create_child(struct device *dev,
1482                             struct device_attribute *attr,
1483                             const char *buf, size_t count)
1484 {
1485         int pkey;
1486         int ret;
1487
1488         if (sscanf(buf, "%i", &pkey) != 1)
1489                 return -EINVAL;
1490
1491         if (pkey <= 0 || pkey > 0xffff || pkey == 0x8000)
1492                 return -EINVAL;
1493
1494         /*
1495          * Set the full membership bit, so that we join the right
1496          * broadcast group, etc.
1497          */
1498         pkey |= 0x8000;
1499
1500         ret = ipoib_vlan_add(to_net_dev(dev), pkey);
1501
1502         return ret ? ret : count;
1503 }
1504 static DEVICE_ATTR(create_child, S_IWUSR, NULL, create_child);
1505
1506 static ssize_t delete_child(struct device *dev,
1507                             struct device_attribute *attr,
1508                             const char *buf, size_t count)
1509 {
1510         int pkey;
1511         int ret;
1512
1513         if (sscanf(buf, "%i", &pkey) != 1)
1514                 return -EINVAL;
1515
1516         if (pkey < 0 || pkey > 0xffff)
1517                 return -EINVAL;
1518
1519         ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
1520
1521         return ret ? ret : count;
1522
1523 }
1524 static DEVICE_ATTR(delete_child, S_IWUSR, NULL, delete_child);
1525
1526 int ipoib_add_pkey_attr(struct net_device *dev)
1527 {
1528         return device_create_file(&dev->dev, &dev_attr_pkey);
1529 }
1530
1531 int ipoib_set_dev_features(struct ipoib_dev_priv *priv, struct ib_device *hca)
1532 {
1533         struct ib_device_attr *device_attr;
1534         int result = -ENOMEM;
1535
1536         device_attr = kmalloc(sizeof *device_attr, GFP_KERNEL);
1537         if (!device_attr) {
1538                 printk(KERN_WARNING "%s: allocation of %zu bytes failed\n",
1539                        hca->name, sizeof *device_attr);
1540                 return result;
1541         }
1542
1543         result = ib_query_device(hca, device_attr);
1544         if (result) {
1545                 printk(KERN_WARNING "%s: ib_query_device failed (ret = %d)\n",
1546                        hca->name, result);
1547                 kfree(device_attr);
1548                 return result;
1549         }
1550         priv->hca_caps = device_attr->device_cap_flags;
1551
1552         kfree(device_attr);
1553
1554         if (priv->hca_caps & IB_DEVICE_UD_IP_CSUM) {
1555                 priv->dev->hw_features = NETIF_F_SG |
1556                         NETIF_F_IP_CSUM | NETIF_F_RXCSUM;
1557
1558                 if (priv->hca_caps & IB_DEVICE_UD_TSO)
1559                         priv->dev->hw_features |= NETIF_F_TSO;
1560
1561                 priv->dev->features |= priv->dev->hw_features;
1562         }
1563
1564         return 0;
1565 }
1566
1567 static struct net_device *ipoib_add_port(const char *format,
1568                                          struct ib_device *hca, u8 port)
1569 {
1570         struct ipoib_dev_priv *priv;
1571         struct ib_port_attr attr;
1572         int result = -ENOMEM;
1573
1574         priv = ipoib_intf_alloc(format);
1575         if (!priv)
1576                 goto alloc_mem_failed;
1577
1578         SET_NETDEV_DEV(priv->dev, hca->dma_device);
1579         priv->dev->dev_id = port - 1;
1580
1581         result = ib_query_port(hca, port, &attr);
1582         if (!result)
1583                 priv->max_ib_mtu = ib_mtu_enum_to_int(attr.max_mtu);
1584         else {
1585                 printk(KERN_WARNING "%s: ib_query_port %d failed\n",
1586                        hca->name, port);
1587                 goto device_init_failed;
1588         }
1589
1590         /* MTU will be reset when mcast join happens */
1591         priv->dev->mtu  = IPOIB_UD_MTU(priv->max_ib_mtu);
1592         priv->mcast_mtu  = priv->admin_mtu = priv->dev->mtu;
1593
1594         priv->dev->neigh_priv_len = sizeof(struct ipoib_neigh);
1595
1596         result = ib_query_pkey(hca, port, 0, &priv->pkey);
1597         if (result) {
1598                 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
1599                        hca->name, port, result);
1600                 goto device_init_failed;
1601         }
1602
1603         result = ipoib_set_dev_features(priv, hca);
1604         if (result)
1605                 goto device_init_failed;
1606
1607         /*
1608          * Set the full membership bit, so that we join the right
1609          * broadcast group, etc.
1610          */
1611         priv->pkey |= 0x8000;
1612
1613         priv->dev->broadcast[8] = priv->pkey >> 8;
1614         priv->dev->broadcast[9] = priv->pkey & 0xff;
1615
1616         result = ib_query_gid(hca, port, 0, &priv->local_gid);
1617         if (result) {
1618                 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1619                        hca->name, port, result);
1620                 goto device_init_failed;
1621         } else
1622                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1623
1624         result = ipoib_dev_init(priv->dev, hca, port);
1625         if (result < 0) {
1626                 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1627                        hca->name, port, result);
1628                 goto device_init_failed;
1629         }
1630
1631         INIT_IB_EVENT_HANDLER(&priv->event_handler,
1632                               priv->ca, ipoib_event);
1633         result = ib_register_event_handler(&priv->event_handler);
1634         if (result < 0) {
1635                 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1636                        "port %d (ret = %d)\n",
1637                        hca->name, port, result);
1638                 goto event_failed;
1639         }
1640
1641         result = register_netdev(priv->dev);
1642         if (result) {
1643                 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1644                        hca->name, port, result);
1645                 goto register_failed;
1646         }
1647
1648         ipoib_create_debug_files(priv->dev);
1649
1650         if (ipoib_cm_add_mode_attr(priv->dev))
1651                 goto sysfs_failed;
1652         if (ipoib_add_pkey_attr(priv->dev))
1653                 goto sysfs_failed;
1654         if (ipoib_add_umcast_attr(priv->dev))
1655                 goto sysfs_failed;
1656         if (device_create_file(&priv->dev->dev, &dev_attr_create_child))
1657                 goto sysfs_failed;
1658         if (device_create_file(&priv->dev->dev, &dev_attr_delete_child))
1659                 goto sysfs_failed;
1660
1661         return priv->dev;
1662
1663 sysfs_failed:
1664         ipoib_delete_debug_files(priv->dev);
1665         unregister_netdev(priv->dev);
1666
1667 register_failed:
1668         ib_unregister_event_handler(&priv->event_handler);
1669         flush_workqueue(ipoib_workqueue);
1670         /* Stop GC if started before flush */
1671         set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1672         cancel_delayed_work(&priv->neigh_reap_task);
1673         flush_workqueue(priv->wq);
1674
1675 event_failed:
1676         ipoib_dev_cleanup(priv->dev);
1677
1678 device_init_failed:
1679         free_netdev(priv->dev);
1680
1681 alloc_mem_failed:
1682         return ERR_PTR(result);
1683 }
1684
1685 static void ipoib_add_one(struct ib_device *device)
1686 {
1687         struct list_head *dev_list;
1688         struct net_device *dev;
1689         struct ipoib_dev_priv *priv;
1690         int p;
1691         int count = 0;
1692
1693         dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1694         if (!dev_list)
1695                 return;
1696
1697         INIT_LIST_HEAD(dev_list);
1698
1699         for (p = rdma_start_port(device); p <= rdma_end_port(device); ++p) {
1700                 if (!rdma_protocol_ib(device, p))
1701                         continue;
1702                 dev = ipoib_add_port("ib%d", device, p);
1703                 if (!IS_ERR(dev)) {
1704                         priv = netdev_priv(dev);
1705                         list_add_tail(&priv->list, dev_list);
1706                         count++;
1707                 }
1708         }
1709
1710         if (!count) {
1711                 kfree(dev_list);
1712                 return;
1713         }
1714
1715         ib_set_client_data(device, &ipoib_client, dev_list);
1716 }
1717
1718 static void ipoib_remove_one(struct ib_device *device)
1719 {
1720         struct ipoib_dev_priv *priv, *tmp;
1721         struct list_head *dev_list;
1722
1723         dev_list = ib_get_client_data(device, &ipoib_client);
1724         if (!dev_list)
1725                 return;
1726
1727         list_for_each_entry_safe(priv, tmp, dev_list, list) {
1728                 ib_unregister_event_handler(&priv->event_handler);
1729                 flush_workqueue(ipoib_workqueue);
1730
1731                 rtnl_lock();
1732                 dev_change_flags(priv->dev, priv->dev->flags & ~IFF_UP);
1733                 rtnl_unlock();
1734
1735                 /* Stop GC */
1736                 set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1737                 cancel_delayed_work(&priv->neigh_reap_task);
1738                 flush_workqueue(priv->wq);
1739
1740                 unregister_netdev(priv->dev);
1741                 free_netdev(priv->dev);
1742         }
1743
1744         kfree(dev_list);
1745 }
1746
1747 static int __init ipoib_init_module(void)
1748 {
1749         int ret;
1750
1751         ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1752         ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1753         ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1754
1755         ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1756         ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1757         ipoib_sendq_size = max3(ipoib_sendq_size, 2 * MAX_SEND_CQE, IPOIB_MIN_QUEUE_SIZE);
1758 #ifdef CONFIG_INFINIBAND_IPOIB_CM
1759         ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
1760 #endif
1761
1762         /*
1763          * When copying small received packets, we only copy from the
1764          * linear data part of the SKB, so we rely on this condition.
1765          */
1766         BUILD_BUG_ON(IPOIB_CM_COPYBREAK > IPOIB_CM_HEAD_SIZE);
1767
1768         ret = ipoib_register_debugfs();
1769         if (ret)
1770                 return ret;
1771
1772         /*
1773          * We create a global workqueue here that is used for all flush
1774          * operations.  However, if you attempt to flush a workqueue
1775          * from a task on that same workqueue, it deadlocks the system.
1776          * We want to be able to flush the tasks associated with a
1777          * specific net device, so we also create a workqueue for each
1778          * netdevice.  We queue up the tasks for that device only on
1779          * its private workqueue, and we only queue up flush events
1780          * on our global flush workqueue.  This avoids the deadlocks.
1781          */
1782         ipoib_workqueue = create_singlethread_workqueue("ipoib_flush");
1783         if (!ipoib_workqueue) {
1784                 ret = -ENOMEM;
1785                 goto err_fs;
1786         }
1787
1788         ib_sa_register_client(&ipoib_sa_client);
1789
1790         ret = ib_register_client(&ipoib_client);
1791         if (ret)
1792                 goto err_sa;
1793
1794         ret = ipoib_netlink_init();
1795         if (ret)
1796                 goto err_client;
1797
1798         return 0;
1799
1800 err_client:
1801         ib_unregister_client(&ipoib_client);
1802
1803 err_sa:
1804         ib_sa_unregister_client(&ipoib_sa_client);
1805         destroy_workqueue(ipoib_workqueue);
1806
1807 err_fs:
1808         ipoib_unregister_debugfs();
1809
1810         return ret;
1811 }
1812
1813 static void __exit ipoib_cleanup_module(void)
1814 {
1815         ipoib_netlink_fini();
1816         ib_unregister_client(&ipoib_client);
1817         ib_sa_unregister_client(&ipoib_sa_client);
1818         ipoib_unregister_debugfs();
1819         destroy_workqueue(ipoib_workqueue);
1820 }
1821
1822 module_init(ipoib_init_module);
1823 module_exit(ipoib_cleanup_module);