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