kernel/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_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                 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE)
 644                         __skb_queue_tail(&neigh->queue, skb);
 645                 else
 646                         goto err_drop;
 647         }
 648
 649         spin_unlock_irqrestore(&priv->lock, flags);
 650         ipoib_neigh_put(neigh);
 651         return;
 652
 653 err_path:
 654         ipoib_neigh_free(neigh);
 655 err_drop:
 656         ++dev->stats.tx_dropped;
 657         dev_kfree_skb_any(skb);
 658
 659         spin_unlock_irqrestore(&priv->lock, flags);
 660         ipoib_neigh_put(neigh);
 661 }
 662
 663 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
 664                              struct ipoib_cb *cb)
 665 {
 666         struct ipoib_dev_priv *priv = netdev_priv(dev);
 667         struct ipoib_path *path;
 668         unsigned long flags;
 669
 670         spin_lock_irqsave(&priv->lock, flags);
 671
 672         path = __path_find(dev, cb->hwaddr + 4);
 673         if (!path || !path->valid) {
 674                 int new_path = 0;
 675
 676                 if (!path) {
 677                         path = path_rec_create(dev, cb->hwaddr + 4);
 678                         new_path = 1;
 679                 }
 680                 if (path) {
 681                         if (skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 682                                 __skb_queue_tail(&path->queue, skb);
 683                         } else {
 684                                 ++dev->stats.tx_dropped;
 685                                 dev_kfree_skb_any(skb);
 686                         }
 687
 688                         if (!path->query && path_rec_start(dev, path)) {
 689                                 spin_unlock_irqrestore(&priv->lock, flags);
 690                                 if (new_path)
 691                                         path_free(dev, path);
 692                                 return;
 693                         } else
 694                                 __path_add(dev, path);
 695                 } else {
 696                         ++dev->stats.tx_dropped;
 697                         dev_kfree_skb_any(skb);
 698                 }
 699
 700                 spin_unlock_irqrestore(&priv->lock, flags);
 701                 return;
 702         }
 703
 704         if (path->ah) {
 705                 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
 706                           be16_to_cpu(path->pathrec.dlid));
 707
 708                 spin_unlock_irqrestore(&priv->lock, flags);
 709                 ipoib_send(dev, skb, path->ah, IPOIB_QPN(cb->hwaddr));
 710                 return;
 711         } else if ((path->query || !path_rec_start(dev, path)) &&
 712                    skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 713                 __skb_queue_tail(&path->queue, skb);
 714         } else {
 715                 ++dev->stats.tx_dropped;
 716                 dev_kfree_skb_any(skb);
 717         }
 718
 719         spin_unlock_irqrestore(&priv->lock, flags);
 720 }
 721
 722 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
 723 {
 724         struct ipoib_dev_priv *priv = netdev_priv(dev);
 725         struct ipoib_neigh *neigh;
 726         struct ipoib_cb *cb = ipoib_skb_cb(skb);
 727         struct ipoib_header *header;
 728         unsigned long flags;
 729
 730         header = (struct ipoib_header *) skb->data;
 731
 732         if (unlikely(cb->hwaddr[4] == 0xff)) {
 733                 /* multicast, arrange "if" according to probability */
 734                 if ((header->proto != htons(ETH_P_IP)) &&
 735                     (header->proto != htons(ETH_P_IPV6)) &&
 736                     (header->proto != htons(ETH_P_ARP)) &&
 737                     (header->proto != htons(ETH_P_RARP)) &&
 738                     (header->proto != htons(ETH_P_TIPC))) {
 739                         /* ethertype not supported by IPoIB */
 740                         ++dev->stats.tx_dropped;
 741                         dev_kfree_skb_any(skb);
 742                         return NETDEV_TX_OK;
 743                 }
 744                 /* Add in the P_Key for multicast*/
 745                 cb->hwaddr[8] = (priv->pkey >> 8) & 0xff;
 746                 cb->hwaddr[9] = priv->pkey & 0xff;
 747
 748                 neigh = ipoib_neigh_get(dev, cb->hwaddr);
 749                 if (likely(neigh))
 750                         goto send_using_neigh;
 751                 ipoib_mcast_send(dev, cb->hwaddr, skb);
 752                 return NETDEV_TX_OK;
 753         }
 754
 755         /* unicast, arrange "switch" according to probability */
 756         switch (header->proto) {
 757         case htons(ETH_P_IP):
 758         case htons(ETH_P_IPV6):
 759         case htons(ETH_P_TIPC):
 760                 neigh = ipoib_neigh_get(dev, cb->hwaddr);
 761                 if (unlikely(!neigh)) {
 762                         neigh_add_path(skb, cb->hwaddr, dev);
 763                         return NETDEV_TX_OK;
 764                 }
 765                 break;
 766         case htons(ETH_P_ARP):
 767         case htons(ETH_P_RARP):
 768                 /* for unicast ARP and RARP should always perform path find */
 769                 unicast_arp_send(skb, dev, cb);
 770                 return NETDEV_TX_OK;
 771         default:
 772                 /* ethertype not supported by IPoIB */
 773                 ++dev->stats.tx_dropped;
 774                 dev_kfree_skb_any(skb);
 775                 return NETDEV_TX_OK;
 776         }
 777
 778 send_using_neigh:
 779         /* note we now hold a ref to neigh */
 780         if (ipoib_cm_get(neigh)) {
 781                 if (ipoib_cm_up(neigh)) {
 782                         ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
 783                         goto unref;
 784                 }
 785         } else if (neigh->ah) {
 786                 ipoib_send(dev, skb, neigh->ah, IPOIB_QPN(cb->hwaddr));
 787                 goto unref;
 788         }
 789
 790         if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
 791                 spin_lock_irqsave(&priv->lock, flags);
 792                 __skb_queue_tail(&neigh->queue, skb);
 793                 spin_unlock_irqrestore(&priv->lock, flags);
 794         } else {
 795                 ++dev->stats.tx_dropped;
 796                 dev_kfree_skb_any(skb);
 797         }
 798
 799 unref:
 800         ipoib_neigh_put(neigh);
 801
 802         return NETDEV_TX_OK;
 803 }
 804
 805 static void ipoib_timeout(struct net_device *dev)
 806 {
 807         struct ipoib_dev_priv *priv = netdev_priv(dev);
 808
 809         ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
 810                    jiffies_to_msecs(jiffies - dev->trans_start));
 811         ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
 812                    netif_queue_stopped(dev),
 813                    priv->tx_head, priv->tx_tail);
 814         /* XXX reset QP, etc. */
 815 }
 816
 817 static int ipoib_hard_header(struct sk_buff *skb,
 818                              struct net_device *dev,
 819                              unsigned short type,
 820                              const void *daddr, const void *saddr, unsigned len)
 821 {
 822         struct ipoib_header *header;
 823         struct ipoib_cb *cb = ipoib_skb_cb(skb);
 824
 825         header = (struct ipoib_header *) skb_push(skb, sizeof *header);
 826
 827         header->proto = htons(type);
 828         header->reserved = 0;
 829
 830         /*
 831          * we don't rely on dst_entry structure,  always stuff the
 832          * destination address into skb->cb so we can figure out where
 833          * to send the packet later.
 834          */
 835         memcpy(cb->hwaddr, daddr, INFINIBAND_ALEN);
 836
 837         return sizeof *header;
 838 }
 839
 840 static void ipoib_set_mcast_list(struct net_device *dev)
 841 {
 842         struct ipoib_dev_priv *priv = netdev_priv(dev);
 843
 844         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
 845                 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
 846                 return;
 847         }
 848
 849         queue_work(priv->wq, &priv->restart_task);
 850 }
 851
 852 static int ipoib_get_iflink(const struct net_device *dev)
 853 {
 854         struct ipoib_dev_priv *priv = netdev_priv(dev);
 855
 856         /* parent interface */
 857         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags))
 858                 return dev->ifindex;
 859
 860         /* child/vlan interface */
 861         return priv->parent->ifindex;
 862 }
 863
 864 static u32 ipoib_addr_hash(struct ipoib_neigh_hash *htbl, u8 *daddr)
 865 {
 866         /*
 867          * Use only the address parts that contributes to spreading
 868          * The subnet prefix is not used as one can not connect to
 869          * same remote port (GUID) using the same remote QPN via two
 870          * different subnets.
 871          */
 872          /* qpn octets[1:4) & port GUID octets[12:20) */
 873         u32 *d32 = (u32 *) daddr;
 874         u32 hv;
 875
 876         hv = jhash_3words(d32[3], d32[4], IPOIB_QPN_MASK & d32[0], 0);
 877         return hv & htbl->mask;
 878 }
 879
 880 struct ipoib_neigh *ipoib_neigh_get(struct net_device *dev, u8 *daddr)
 881 {
 882         struct ipoib_dev_priv *priv = netdev_priv(dev);
 883         struct ipoib_neigh_table *ntbl = &priv->ntbl;
 884         struct ipoib_neigh_hash *htbl;
 885         struct ipoib_neigh *neigh = NULL;
 886         u32 hash_val;
 887
 888         rcu_read_lock_bh();
 889
 890         htbl = rcu_dereference_bh(ntbl->htbl);
 891
 892         if (!htbl)
 893                 goto out_unlock;
 894
 895         hash_val = ipoib_addr_hash(htbl, daddr);
 896         for (neigh = rcu_dereference_bh(htbl->buckets[hash_val]);
 897              neigh != NULL;
 898              neigh = rcu_dereference_bh(neigh->hnext)) {
 899                 if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) {
 900                         /* found, take one ref on behalf of the caller */
 901                         if (!atomic_inc_not_zero(&neigh->refcnt)) {
 902                                 /* deleted */
 903                                 neigh = NULL;
 904                                 goto out_unlock;
 905                         }
 906                         neigh->alive = jiffies;
 907                         goto out_unlock;
 908                 }
 909         }
 910
 911 out_unlock:
 912         rcu_read_unlock_bh();
 913         return neigh;
 914 }
 915
 916 static void __ipoib_reap_neigh(struct ipoib_dev_priv *priv)
 917 {
 918         struct ipoib_neigh_table *ntbl = &priv->ntbl;
 919         struct ipoib_neigh_hash *htbl;
 920         unsigned long neigh_obsolete;
 921         unsigned long dt;
 922         unsigned long flags;
 923         int i;
 924
 925         if (test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
 926                 return;
 927
 928         spin_lock_irqsave(&priv->lock, flags);
 929
 930         htbl = rcu_dereference_protected(ntbl->htbl,
 931                                          lockdep_is_held(&priv->lock));
 932
 933         if (!htbl)
 934                 goto out_unlock;
 935
 936         /* neigh is obsolete if it was idle for two GC periods */
 937         dt = 2 * arp_tbl.gc_interval;
 938         neigh_obsolete = jiffies - dt;
 939         /* handle possible race condition */
 940         if (test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
 941                 goto out_unlock;
 942
 943         for (i = 0; i < htbl->size; i++) {
 944                 struct ipoib_neigh *neigh;
 945                 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
 946
 947                 while ((neigh = rcu_dereference_protected(*np,
 948                                                           lockdep_is_held(&priv->lock))) != NULL) {
 949                         /* was the neigh idle for two GC periods */
 950                         if (time_after(neigh_obsolete, neigh->alive)) {
 951                                 rcu_assign_pointer(*np,
 952                                                    rcu_dereference_protected(neigh->hnext,
 953                                                                              lockdep_is_held(&priv->lock)));
 954                                 /* remove from path/mc list */
 955                                 list_del(&neigh->list);
 956                                 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
 957                         } else {
 958                                 np = &neigh->hnext;
 959                         }
 960
 961                 }
 962         }
 963
 964 out_unlock:
 965         spin_unlock_irqrestore(&priv->lock, flags);
 966 }
 967
 968 static void ipoib_reap_neigh(struct work_struct *work)
 969 {
 970         struct ipoib_dev_priv *priv =
 971                 container_of(work, struct ipoib_dev_priv, neigh_reap_task.work);
 972
 973         __ipoib_reap_neigh(priv);
 974
 975         if (!test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
 976                 queue_delayed_work(priv->wq, &priv->neigh_reap_task,
 977                                    arp_tbl.gc_interval);
 978 }
 979
 980
 981 static struct ipoib_neigh *ipoib_neigh_ctor(u8 *daddr,
 982                                       struct net_device *dev)
 983 {
 984         struct ipoib_neigh *neigh;
 985
 986         neigh = kzalloc(sizeof *neigh, GFP_ATOMIC);
 987         if (!neigh)
 988                 return NULL;
 989
 990         neigh->dev = dev;
 991         memcpy(&neigh->daddr, daddr, sizeof(neigh->daddr));
 992         skb_queue_head_init(&neigh->queue);
 993         INIT_LIST_HEAD(&neigh->list);
 994         ipoib_cm_set(neigh, NULL);
 995         /* one ref on behalf of the caller */
 996         atomic_set(&neigh->refcnt, 1);
 997
 998         return neigh;
 999 }
1000
1001 struct ipoib_neigh *ipoib_neigh_alloc(u8 *daddr,
1002                                       struct net_device *dev)
1003 {
1004         struct ipoib_dev_priv *priv = netdev_priv(dev);
1005         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1006         struct ipoib_neigh_hash *htbl;
1007         struct ipoib_neigh *neigh;
1008         u32 hash_val;
1009
1010         htbl = rcu_dereference_protected(ntbl->htbl,
1011                                          lockdep_is_held(&priv->lock));
1012         if (!htbl) {
1013                 neigh = NULL;
1014                 goto out_unlock;
1015         }
1016
1017         /* need to add a new neigh, but maybe some other thread succeeded?
1018          * recalc hash, maybe hash resize took place so we do a search
1019          */
1020         hash_val = ipoib_addr_hash(htbl, daddr);
1021         for (neigh = rcu_dereference_protected(htbl->buckets[hash_val],
1022                                                lockdep_is_held(&priv->lock));
1023              neigh != NULL;
1024              neigh = rcu_dereference_protected(neigh->hnext,
1025                                                lockdep_is_held(&priv->lock))) {
1026                 if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) {
1027                         /* found, take one ref on behalf of the caller */
1028                         if (!atomic_inc_not_zero(&neigh->refcnt)) {
1029                                 /* deleted */
1030                                 neigh = NULL;
1031                                 break;
1032                         }
1033                         neigh->alive = jiffies;
1034                         goto out_unlock;
1035                 }
1036         }
1037
1038         neigh = ipoib_neigh_ctor(daddr, dev);
1039         if (!neigh)
1040                 goto out_unlock;
1041
1042         /* one ref on behalf of the hash table */
1043         atomic_inc(&neigh->refcnt);
1044         neigh->alive = jiffies;
1045         /* put in hash */
1046         rcu_assign_pointer(neigh->hnext,
1047                            rcu_dereference_protected(htbl->buckets[hash_val],
1048                                                      lockdep_is_held(&priv->lock)));
1049         rcu_assign_pointer(htbl->buckets[hash_val], neigh);
1050         atomic_inc(&ntbl->entries);
1051
1052 out_unlock:
1053
1054         return neigh;
1055 }
1056
1057 void ipoib_neigh_dtor(struct ipoib_neigh *neigh)
1058 {
1059         /* neigh reference count was dropprd to zero */
1060         struct net_device *dev = neigh->dev;
1061         struct ipoib_dev_priv *priv = netdev_priv(dev);
1062         struct sk_buff *skb;
1063         if (neigh->ah)
1064                 ipoib_put_ah(neigh->ah);
1065         while ((skb = __skb_dequeue(&neigh->queue))) {
1066                 ++dev->stats.tx_dropped;
1067                 dev_kfree_skb_any(skb);
1068         }
1069         if (ipoib_cm_get(neigh))
1070                 ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
1071         ipoib_dbg(netdev_priv(dev),
1072                   "neigh free for %06x %pI6\n",
1073                   IPOIB_QPN(neigh->daddr),
1074                   neigh->daddr + 4);
1075         kfree(neigh);
1076         if (atomic_dec_and_test(&priv->ntbl.entries)) {
1077                 if (test_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags))
1078                         complete(&priv->ntbl.flushed);
1079         }
1080 }
1081
1082 static void ipoib_neigh_reclaim(struct rcu_head *rp)
1083 {
1084         /* Called as a result of removal from hash table */
1085         struct ipoib_neigh *neigh = container_of(rp, struct ipoib_neigh, rcu);
1086         /* note TX context may hold another ref */
1087         ipoib_neigh_put(neigh);
1088 }
1089
1090 void ipoib_neigh_free(struct ipoib_neigh *neigh)
1091 {
1092         struct net_device *dev = neigh->dev;
1093         struct ipoib_dev_priv *priv = netdev_priv(dev);
1094         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1095         struct ipoib_neigh_hash *htbl;
1096         struct ipoib_neigh __rcu **np;
1097         struct ipoib_neigh *n;
1098         u32 hash_val;
1099
1100         htbl = rcu_dereference_protected(ntbl->htbl,
1101                                         lockdep_is_held(&priv->lock));
1102         if (!htbl)
1103                 return;
1104
1105         hash_val = ipoib_addr_hash(htbl, neigh->daddr);
1106         np = &htbl->buckets[hash_val];
1107         for (n = rcu_dereference_protected(*np,
1108                                             lockdep_is_held(&priv->lock));
1109              n != NULL;
1110              n = rcu_dereference_protected(*np,
1111                                         lockdep_is_held(&priv->lock))) {
1112                 if (n == neigh) {
1113                         /* found */
1114                         rcu_assign_pointer(*np,
1115                                            rcu_dereference_protected(neigh->hnext,
1116                                                                      lockdep_is_held(&priv->lock)));
1117                         /* remove from parent list */
1118                         list_del(&neigh->list);
1119                         call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1120                         return;
1121                 } else {
1122                         np = &n->hnext;
1123                 }
1124         }
1125 }
1126
1127 static int ipoib_neigh_hash_init(struct ipoib_dev_priv *priv)
1128 {
1129         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1130         struct ipoib_neigh_hash *htbl;
1131         struct ipoib_neigh **buckets;
1132         u32 size;
1133
1134         clear_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags);
1135         ntbl->htbl = NULL;
1136         htbl = kzalloc(sizeof(*htbl), GFP_KERNEL);
1137         if (!htbl)
1138                 return -ENOMEM;
1139         set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1140         size = roundup_pow_of_two(arp_tbl.gc_thresh3);
1141         buckets = kzalloc(size * sizeof(*buckets), GFP_KERNEL);
1142         if (!buckets) {
1143                 kfree(htbl);
1144                 return -ENOMEM;
1145         }
1146         htbl->size = size;
1147         htbl->mask = (size - 1);
1148         htbl->buckets = buckets;
1149         ntbl->htbl = htbl;
1150         htbl->ntbl = ntbl;
1151         atomic_set(&ntbl->entries, 0);
1152
1153         /* start garbage collection */
1154         clear_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1155         queue_delayed_work(priv->wq, &priv->neigh_reap_task,
1156                            arp_tbl.gc_interval);
1157
1158         return 0;
1159 }
1160
1161 static void neigh_hash_free_rcu(struct rcu_head *head)
1162 {
1163         struct ipoib_neigh_hash *htbl = container_of(head,
1164                                                     struct ipoib_neigh_hash,
1165                                                     rcu);
1166         struct ipoib_neigh __rcu **buckets = htbl->buckets;
1167         struct ipoib_neigh_table *ntbl = htbl->ntbl;
1168
1169         kfree(buckets);
1170         kfree(htbl);
1171         complete(&ntbl->deleted);
1172 }
1173
1174 void ipoib_del_neighs_by_gid(struct net_device *dev, u8 *gid)
1175 {
1176         struct ipoib_dev_priv *priv = netdev_priv(dev);
1177         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1178         struct ipoib_neigh_hash *htbl;
1179         unsigned long flags;
1180         int i;
1181
1182         /* remove all neigh connected to a given path or mcast */
1183         spin_lock_irqsave(&priv->lock, flags);
1184
1185         htbl = rcu_dereference_protected(ntbl->htbl,
1186                                          lockdep_is_held(&priv->lock));
1187
1188         if (!htbl)
1189                 goto out_unlock;
1190
1191         for (i = 0; i < htbl->size; i++) {
1192                 struct ipoib_neigh *neigh;
1193                 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1194
1195                 while ((neigh = rcu_dereference_protected(*np,
1196                                                           lockdep_is_held(&priv->lock))) != NULL) {
1197                         /* delete neighs belong to this parent */
1198                         if (!memcmp(gid, neigh->daddr + 4, sizeof (union ib_gid))) {
1199                                 rcu_assign_pointer(*np,
1200                                                    rcu_dereference_protected(neigh->hnext,
1201                                                                              lockdep_is_held(&priv->lock)));
1202                                 /* remove from parent list */
1203                                 list_del(&neigh->list);
1204                                 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1205                         } else {
1206                                 np = &neigh->hnext;
1207                         }
1208
1209                 }
1210         }
1211 out_unlock:
1212         spin_unlock_irqrestore(&priv->lock, flags);
1213 }
1214
1215 static void ipoib_flush_neighs(struct ipoib_dev_priv *priv)
1216 {
1217         struct ipoib_neigh_table *ntbl = &priv->ntbl;
1218         struct ipoib_neigh_hash *htbl;
1219         unsigned long flags;
1220         int i, wait_flushed = 0;
1221
1222         init_completion(&priv->ntbl.flushed);
1223
1224         spin_lock_irqsave(&priv->lock, flags);
1225
1226         htbl = rcu_dereference_protected(ntbl->htbl,
1227                                         lockdep_is_held(&priv->lock));
1228         if (!htbl)
1229                 goto out_unlock;
1230
1231         wait_flushed = atomic_read(&priv->ntbl.entries);
1232         if (!wait_flushed)
1233                 goto free_htbl;
1234
1235         for (i = 0; i < htbl->size; i++) {
1236                 struct ipoib_neigh *neigh;
1237                 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1238
1239                 while ((neigh = rcu_dereference_protected(*np,
1240                                        lockdep_is_held(&priv->lock))) != NULL) {
1241                         rcu_assign_pointer(*np,
1242                                            rcu_dereference_protected(neigh->hnext,
1243                                                                      lockdep_is_held(&priv->lock)));
1244                         /* remove from path/mc list */
1245                         list_del(&neigh->list);
1246                         call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1247                 }
1248         }
1249
1250 free_htbl:
1251         rcu_assign_pointer(ntbl->htbl, NULL);
1252         call_rcu(&htbl->rcu, neigh_hash_free_rcu);
1253
1254 out_unlock:
1255         spin_unlock_irqrestore(&priv->lock, flags);
1256         if (wait_flushed)
1257                 wait_for_completion(&priv->ntbl.flushed);
1258 }
1259
1260 static void ipoib_neigh_hash_uninit(struct net_device *dev)
1261 {
1262         struct ipoib_dev_priv *priv = netdev_priv(dev);
1263         int stopped;
1264
1265         ipoib_dbg(priv, "ipoib_neigh_hash_uninit\n");
1266         init_completion(&priv->ntbl.deleted);
1267         set_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags);
1268
1269         /* Stop GC if called at init fail need to cancel work */
1270         stopped = test_and_set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1271         if (!stopped)
1272                 cancel_delayed_work(&priv->neigh_reap_task);
1273
1274         ipoib_flush_neighs(priv);
1275
1276         wait_for_completion(&priv->ntbl.deleted);
1277 }
1278
1279
1280 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
1281 {
1282         struct ipoib_dev_priv *priv = netdev_priv(dev);
1283
1284         /* Allocate RX/TX "rings" to hold queued skbs */
1285         priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
1286                                 GFP_KERNEL);
1287         if (!priv->rx_ring) {
1288                 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
1289                        ca->name, ipoib_recvq_size);
1290                 goto out;
1291         }
1292
1293         priv->tx_ring = vzalloc(ipoib_sendq_size * sizeof *priv->tx_ring);
1294         if (!priv->tx_ring) {
1295                 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
1296                        ca->name, ipoib_sendq_size);
1297                 goto out_rx_ring_cleanup;
1298         }
1299
1300         /* priv->tx_head, tx_tail & tx_outstanding are already 0 */
1301
1302         if (ipoib_ib_dev_init(dev, ca, port))
1303                 goto out_tx_ring_cleanup;
1304
1305         /*
1306          * Must be after ipoib_ib_dev_init so we can allocate a per
1307          * device wq there and use it here
1308          */
1309         if (ipoib_neigh_hash_init(priv) < 0)
1310                 goto out_dev_uninit;
1311
1312         return 0;
1313
1314 out_dev_uninit:
1315         ipoib_ib_dev_cleanup(dev);
1316
1317 out_tx_ring_cleanup:
1318         vfree(priv->tx_ring);
1319
1320 out_rx_ring_cleanup:
1321         kfree(priv->rx_ring);
1322
1323 out:
1324         return -ENOMEM;
1325 }
1326
1327 void ipoib_dev_cleanup(struct net_device *dev)
1328 {
1329         struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
1330         LIST_HEAD(head);
1331
1332         ASSERT_RTNL();
1333
1334         ipoib_delete_debug_files(dev);
1335
1336         /* Delete any child interfaces first */
1337         list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
1338                 /* Stop GC on child */
1339                 set_bit(IPOIB_STOP_NEIGH_GC, &cpriv->flags);
1340                 cancel_delayed_work(&cpriv->neigh_reap_task);
1341                 unregister_netdevice_queue(cpriv->dev, &head);
1342         }
1343         unregister_netdevice_many(&head);
1344
1345         /*
1346          * Must be before ipoib_ib_dev_cleanup or we delete an in use
1347          * work queue
1348          */
1349         ipoib_neigh_hash_uninit(dev);
1350
1351         ipoib_ib_dev_cleanup(dev);
1352
1353         kfree(priv->rx_ring);
1354         vfree(priv->tx_ring);
1355
1356         priv->rx_ring = NULL;
1357         priv->tx_ring = NULL;
1358 }
1359
1360 static const struct header_ops ipoib_header_ops = {
1361         .create = ipoib_hard_header,
1362 };
1363
1364 static const struct net_device_ops ipoib_netdev_ops = {
1365         .ndo_uninit              = ipoib_uninit,
1366         .ndo_open                = ipoib_open,
1367         .ndo_stop                = ipoib_stop,
1368         .ndo_change_mtu          = ipoib_change_mtu,
1369         .ndo_fix_features        = ipoib_fix_features,
1370         .ndo_start_xmit          = ipoib_start_xmit,
1371         .ndo_tx_timeout          = ipoib_timeout,
1372         .ndo_set_rx_mode         = ipoib_set_mcast_list,
1373         .ndo_get_iflink          = ipoib_get_iflink,
1374 };
1375
1376 void ipoib_setup(struct net_device *dev)
1377 {
1378         struct ipoib_dev_priv *priv = netdev_priv(dev);
1379
1380         dev->netdev_ops          = &ipoib_netdev_ops;
1381         dev->header_ops          = &ipoib_header_ops;
1382
1383         ipoib_set_ethtool_ops(dev);
1384
1385         netif_napi_add(dev, &priv->napi, ipoib_poll, NAPI_POLL_WEIGHT);
1386
1387         dev->watchdog_timeo      = HZ;
1388
1389         dev->flags              |= IFF_BROADCAST | IFF_MULTICAST;
1390
1391         dev->hard_header_len     = IPOIB_ENCAP_LEN;
1392         dev->addr_len            = INFINIBAND_ALEN;
1393         dev->type                = ARPHRD_INFINIBAND;
1394         dev->tx_queue_len        = ipoib_sendq_size * 2;
1395         dev->features            = (NETIF_F_VLAN_CHALLENGED     |
1396                                     NETIF_F_HIGHDMA);
1397         netif_keep_dst(dev);
1398
1399         memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
1400
1401         priv->dev = dev;
1402
1403         spin_lock_init(&priv->lock);
1404
1405         init_rwsem(&priv->vlan_rwsem);
1406
1407         INIT_LIST_HEAD(&priv->path_list);
1408         INIT_LIST_HEAD(&priv->child_intfs);
1409         INIT_LIST_HEAD(&priv->dead_ahs);
1410         INIT_LIST_HEAD(&priv->multicast_list);
1411
1412         INIT_DELAYED_WORK(&priv->mcast_task,   ipoib_mcast_join_task);
1413         INIT_WORK(&priv->carrier_on_task, ipoib_mcast_carrier_on_task);
1414         INIT_WORK(&priv->flush_light,   ipoib_ib_dev_flush_light);
1415         INIT_WORK(&priv->flush_normal,   ipoib_ib_dev_flush_normal);
1416         INIT_WORK(&priv->flush_heavy,   ipoib_ib_dev_flush_heavy);
1417         INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
1418         INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
1419         INIT_DELAYED_WORK(&priv->neigh_reap_task, ipoib_reap_neigh);
1420 }
1421
1422 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
1423 {
1424         struct net_device *dev;
1425
1426         dev = alloc_netdev((int)sizeof(struct ipoib_dev_priv), name,
1427                            NET_NAME_UNKNOWN, ipoib_setup);
1428         if (!dev)
1429                 return NULL;
1430
1431         return netdev_priv(dev);
1432 }
1433
1434 static ssize_t show_pkey(struct device *dev,
1435                          struct device_attribute *attr, char *buf)
1436 {
1437         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1438
1439         return sprintf(buf, "0x%04x\n", priv->pkey);
1440 }
1441 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1442
1443 static ssize_t show_umcast(struct device *dev,
1444                            struct device_attribute *attr, char *buf)
1445 {
1446         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1447
1448         return sprintf(buf, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
1449 }
1450
1451 void ipoib_set_umcast(struct net_device *ndev, int umcast_val)
1452 {
1453         struct ipoib_dev_priv *priv = netdev_priv(ndev);
1454
1455         if (umcast_val > 0) {
1456                 set_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1457                 ipoib_warn(priv, "ignoring multicast groups joined directly "
1458                                 "by userspace\n");
1459         } else
1460                 clear_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1461 }
1462
1463 static ssize_t set_umcast(struct device *dev,
1464                           struct device_attribute *attr,
1465                           const char *buf, size_t count)
1466 {
1467         unsigned long umcast_val = simple_strtoul(buf, NULL, 0);
1468
1469         ipoib_set_umcast(to_net_dev(dev), umcast_val);
1470
1471         return count;
1472 }
1473 static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast);
1474
1475 int ipoib_add_umcast_attr(struct net_device *dev)
1476 {
1477         return device_create_file(&dev->dev, &dev_attr_umcast);
1478 }
1479
1480 static ssize_t create_child(struct device *dev,
1481                             struct device_attribute *attr,
1482                             const char *buf, size_t count)
1483 {
1484         int pkey;
1485         int ret;
1486
1487         if (sscanf(buf, "%i", &pkey) != 1)
1488                 return -EINVAL;
1489
1490         if (pkey <= 0 || pkey > 0xffff || pkey == 0x8000)
1491                 return -EINVAL;
1492
1493         /*
1494          * Set the full membership bit, so that we join the right
1495          * broadcast group, etc.
1496          */
1497         pkey |= 0x8000;
1498
1499         ret = ipoib_vlan_add(to_net_dev(dev), pkey);
1500
1501         return ret ? ret : count;
1502 }
1503 static DEVICE_ATTR(create_child, S_IWUSR, NULL, create_child);
1504
1505 static ssize_t delete_child(struct device *dev,
1506                             struct device_attribute *attr,
1507                             const char *buf, size_t count)
1508 {
1509         int pkey;
1510         int ret;
1511
1512         if (sscanf(buf, "%i", &pkey) != 1)
1513                 return -EINVAL;
1514
1515         if (pkey < 0 || pkey > 0xffff)
1516                 return -EINVAL;
1517
1518         ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
1519
1520         return ret ? ret : count;
1521
1522 }
1523 static DEVICE_ATTR(delete_child, S_IWUSR, NULL, delete_child);
1524
1525 int ipoib_add_pkey_attr(struct net_device *dev)
1526 {
1527         return device_create_file(&dev->dev, &dev_attr_pkey);
1528 }
1529
1530 int ipoib_set_dev_features(struct ipoib_dev_priv *priv, struct ib_device *hca)
1531 {
1532         struct ib_device_attr *device_attr;
1533         int result = -ENOMEM;
1534
1535         device_attr = kmalloc(sizeof *device_attr, GFP_KERNEL);
1536         if (!device_attr) {
1537                 printk(KERN_WARNING "%s: allocation of %zu bytes failed\n",
1538                        hca->name, sizeof *device_attr);
1539                 return result;
1540         }
1541
1542         result = ib_query_device(hca, device_attr);
1543         if (result) {
1544                 printk(KERN_WARNING "%s: ib_query_device failed (ret = %d)\n",
1545                        hca->name, result);
1546                 kfree(device_attr);
1547                 return result;
1548         }
1549         priv->hca_caps = device_attr->device_cap_flags;
1550
1551         kfree(device_attr);
1552
1553         if (priv->hca_caps & IB_DEVICE_UD_IP_CSUM) {
1554                 priv->dev->hw_features = NETIF_F_SG |
1555                         NETIF_F_IP_CSUM | NETIF_F_RXCSUM;
1556
1557                 if (priv->hca_caps & IB_DEVICE_UD_TSO)
1558                         priv->dev->hw_features |= NETIF_F_TSO;
1559
1560                 priv->dev->features |= priv->dev->hw_features;
1561         }
1562
1563         return 0;
1564 }
1565
1566 static struct net_device *ipoib_add_port(const char *format,
1567                                          struct ib_device *hca, u8 port)
1568 {
1569         struct ipoib_dev_priv *priv;
1570         struct ib_port_attr attr;
1571         int result = -ENOMEM;
1572
1573         priv = ipoib_intf_alloc(format);
1574         if (!priv)
1575                 goto alloc_mem_failed;
1576
1577         SET_NETDEV_DEV(priv->dev, hca->dma_device);
1578         priv->dev->dev_id = port - 1;
1579
1580         if (!ib_query_port(hca, port, &attr))
1581                 priv->max_ib_mtu = ib_mtu_enum_to_int(attr.max_mtu);
1582         else {
1583                 printk(KERN_WARNING "%s: ib_query_port %d failed\n",
1584                        hca->name, port);
1585                 goto device_init_failed;
1586         }
1587
1588         /* MTU will be reset when mcast join happens */
1589         priv->dev->mtu  = IPOIB_UD_MTU(priv->max_ib_mtu);
1590         priv->mcast_mtu  = priv->admin_mtu = priv->dev->mtu;
1591
1592         priv->dev->neigh_priv_len = sizeof(struct ipoib_neigh);
1593
1594         result = ib_query_pkey(hca, port, 0, &priv->pkey);
1595         if (result) {
1596                 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
1597                        hca->name, port, result);
1598                 goto device_init_failed;
1599         }
1600
1601         if (ipoib_set_dev_features(priv, hca))
1602                 goto device_init_failed;
1603
1604         /*
1605          * Set the full membership bit, so that we join the right
1606          * broadcast group, etc.
1607          */
1608         priv->pkey |= 0x8000;
1609
1610         priv->dev->broadcast[8] = priv->pkey >> 8;
1611         priv->dev->broadcast[9] = priv->pkey & 0xff;
1612
1613         result = ib_query_gid(hca, port, 0, &priv->local_gid);
1614         if (result) {
1615                 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1616                        hca->name, port, result);
1617                 goto device_init_failed;
1618         } else
1619                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1620
1621         result = ipoib_dev_init(priv->dev, hca, port);
1622         if (result < 0) {
1623                 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1624                        hca->name, port, result);
1625                 goto device_init_failed;
1626         }
1627
1628         INIT_IB_EVENT_HANDLER(&priv->event_handler,
1629                               priv->ca, ipoib_event);
1630         result = ib_register_event_handler(&priv->event_handler);
1631         if (result < 0) {
1632                 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1633                        "port %d (ret = %d)\n",
1634                        hca->name, port, result);
1635                 goto event_failed;
1636         }
1637
1638         result = register_netdev(priv->dev);
1639         if (result) {
1640                 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1641                        hca->name, port, result);
1642                 goto register_failed;
1643         }
1644
1645         ipoib_create_debug_files(priv->dev);
1646
1647         if (ipoib_cm_add_mode_attr(priv->dev))
1648                 goto sysfs_failed;
1649         if (ipoib_add_pkey_attr(priv->dev))
1650                 goto sysfs_failed;
1651         if (ipoib_add_umcast_attr(priv->dev))
1652                 goto sysfs_failed;
1653         if (device_create_file(&priv->dev->dev, &dev_attr_create_child))
1654                 goto sysfs_failed;
1655         if (device_create_file(&priv->dev->dev, &dev_attr_delete_child))
1656                 goto sysfs_failed;
1657
1658         return priv->dev;
1659
1660 sysfs_failed:
1661         ipoib_delete_debug_files(priv->dev);
1662         unregister_netdev(priv->dev);
1663
1664 register_failed:
1665         ib_unregister_event_handler(&priv->event_handler);
1666         flush_workqueue(ipoib_workqueue);
1667         /* Stop GC if started before flush */
1668         set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1669         cancel_delayed_work(&priv->neigh_reap_task);
1670         flush_workqueue(priv->wq);
1671
1672 event_failed:
1673         ipoib_dev_cleanup(priv->dev);
1674
1675 device_init_failed:
1676         free_netdev(priv->dev);
1677
1678 alloc_mem_failed:
1679         return ERR_PTR(result);
1680 }
1681
1682 static void ipoib_add_one(struct ib_device *device)
1683 {
1684         struct list_head *dev_list;
1685         struct net_device *dev;
1686         struct ipoib_dev_priv *priv;
1687         int s, e, p;
1688
1689         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1690                 return;
1691
1692         dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1693         if (!dev_list)
1694                 return;
1695
1696         INIT_LIST_HEAD(dev_list);
1697
1698         if (device->node_type == RDMA_NODE_IB_SWITCH) {
1699                 s = 0;
1700                 e = 0;
1701         } else {
1702                 s = 1;
1703                 e = device->phys_port_cnt;
1704         }
1705
1706         for (p = s; p <= e; ++p) {
1707                 if (rdma_port_get_link_layer(device, p) != IB_LINK_LAYER_INFINIBAND)
1708                         continue;
1709                 dev = ipoib_add_port("ib%d", device, p);
1710                 if (!IS_ERR(dev)) {
1711                         priv = netdev_priv(dev);
1712                         list_add_tail(&priv->list, dev_list);
1713                 }
1714         }
1715
1716         ib_set_client_data(device, &ipoib_client, dev_list);
1717 }
1718
1719 static void ipoib_remove_one(struct ib_device *device)
1720 {
1721         struct ipoib_dev_priv *priv, *tmp;
1722         struct list_head *dev_list;
1723
1724         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1725                 return;
1726
1727         dev_list = ib_get_client_data(device, &ipoib_client);
1728         if (!dev_list)
1729                 return;
1730
1731         list_for_each_entry_safe(priv, tmp, dev_list, list) {
1732                 ib_unregister_event_handler(&priv->event_handler);
1733                 flush_workqueue(ipoib_workqueue);
1734
1735                 rtnl_lock();
1736                 dev_change_flags(priv->dev, priv->dev->flags & ~IFF_UP);
1737                 rtnl_unlock();
1738
1739                 /* Stop GC */
1740                 set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1741                 cancel_delayed_work(&priv->neigh_reap_task);
1742                 flush_workqueue(priv->wq);
1743
1744                 unregister_netdev(priv->dev);
1745                 free_netdev(priv->dev);
1746         }
1747
1748         kfree(dev_list);
1749 }
1750
1751 static int __init ipoib_init_module(void)
1752 {
1753         int ret;
1754
1755         ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1756         ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1757         ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1758
1759         ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1760         ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1761         ipoib_sendq_size = max3(ipoib_sendq_size, 2 * MAX_SEND_CQE, IPOIB_MIN_QUEUE_SIZE);
1762 #ifdef CONFIG_INFINIBAND_IPOIB_CM
1763         ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
1764 #endif
1765
1766         /*
1767          * When copying small received packets, we only copy from the
1768          * linear data part of the SKB, so we rely on this condition.
1769          */
1770         BUILD_BUG_ON(IPOIB_CM_COPYBREAK > IPOIB_CM_HEAD_SIZE);
1771
1772         ret = ipoib_register_debugfs();
1773         if (ret)
1774                 return ret;
1775
1776         /*
1777          * We create a global workqueue here that is used for all flush
1778          * operations.  However, if you attempt to flush a workqueue
1779          * from a task on that same workqueue, it deadlocks the system.
1780          * We want to be able to flush the tasks associated with a
1781          * specific net device, so we also create a workqueue for each
1782          * netdevice.  We queue up the tasks for that device only on
1783          * its private workqueue, and we only queue up flush events
1784          * on our global flush workqueue.  This avoids the deadlocks.
1785          */
1786         ipoib_workqueue = create_singlethread_workqueue("ipoib_flush");
1787         if (!ipoib_workqueue) {
1788                 ret = -ENOMEM;
1789                 goto err_fs;
1790         }
1791
1792         ib_sa_register_client(&ipoib_sa_client);
1793
1794         ret = ib_register_client(&ipoib_client);
1795         if (ret)
1796                 goto err_sa;
1797
1798         ret = ipoib_netlink_init();
1799         if (ret)
1800                 goto err_client;
1801
1802         return 0;
1803
1804 err_client:
1805         ib_unregister_client(&ipoib_client);
1806
1807 err_sa:
1808         ib_sa_unregister_client(&ipoib_sa_client);
1809         destroy_workqueue(ipoib_workqueue);
1810
1811 err_fs:
1812         ipoib_unregister_debugfs();
1813
1814         return ret;
1815 }
1816
1817 static void __exit ipoib_cleanup_module(void)
1818 {
1819         ipoib_netlink_fini();
1820         ib_unregister_client(&ipoib_client);
1821         ib_sa_unregister_client(&ipoib_sa_client);
1822         ipoib_unregister_debugfs();
1823         destroy_workqueue(ipoib_workqueue);
1824 }
1825
1826 module_init(ipoib_init_module);
1827 module_exit(ipoib_cleanup_module);