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