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