git.oblomov.eu Git - linux-2.6/blob - drivers/infiniband/ulp/ipoib/ipoib_multicast.c

   1 /*
   2  * Copyright (c) 2004, 2005 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  * $Id: ipoib_multicast.c 1362 2004-12-18 15:56:29Z roland $
  35  */
  36
  37 #include <linux/skbuff.h>
  38 #include <linux/rtnetlink.h>
  39 #include <linux/ip.h>
  40 #include <linux/in.h>
  41 #include <linux/igmp.h>
  42 #include <linux/inetdevice.h>
  43 #include <linux/delay.h>
  44 #include <linux/completion.h>
  45
  46 #include <net/dst.h>
  47
  48 #include "ipoib.h"
  49
  50 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
  51 static int mcast_debug_level;
  52
  53 module_param(mcast_debug_level, int, 0644);
  54 MODULE_PARM_DESC(mcast_debug_level,
  55                  "Enable multicast debug tracing if > 0");
  56 #endif
  57
  58 static DEFINE_MUTEX(mcast_mutex);
  59
  60 /* Used for all multicast joins (broadcast, IPv4 mcast and IPv6 mcast) */
  61 struct ipoib_mcast {
  62         struct ib_sa_mcmember_rec mcmember;
  63         struct ib_sa_multicast   *mc;
  64         struct ipoib_ah          *ah;
  65
  66         struct rb_node    rb_node;
  67         struct list_head  list;
  68
  69         unsigned long created;
  70         unsigned long backoff;
  71
  72         unsigned long flags;
  73         unsigned char logcount;
  74
  75         struct list_head  neigh_list;
  76
  77         struct sk_buff_head pkt_queue;
  78
  79         struct net_device *dev;
  80 };
  81
  82 struct ipoib_mcast_iter {
  83         struct net_device *dev;
  84         union ib_gid       mgid;
  85         unsigned long      created;
  86         unsigned int       queuelen;
  87         unsigned int       complete;
  88         unsigned int       send_only;
  89 };
  90
  91 static void ipoib_mcast_free(struct ipoib_mcast *mcast)
  92 {
  93         struct net_device *dev = mcast->dev;
  94         struct ipoib_dev_priv *priv = netdev_priv(dev);
  95         struct ipoib_neigh *neigh, *tmp;
  96         unsigned long flags;
  97         int tx_dropped = 0;
  98
  99         ipoib_dbg_mcast(netdev_priv(dev),
 100                         "deleting multicast group " IPOIB_GID_FMT "\n",
 101                         IPOIB_GID_ARG(mcast->mcmember.mgid));
 102
 103         spin_lock_irqsave(&priv->lock, flags);
 104
 105         list_for_each_entry_safe(neigh, tmp, &mcast->neigh_list, list) {
 106                 /*
 107                  * It's safe to call ipoib_put_ah() inside priv->lock
 108                  * here, because we know that mcast->ah will always
 109                  * hold one more reference, so ipoib_put_ah() will
 110                  * never do more than decrement the ref count.
 111                  */
 112                 if (neigh->ah)
 113                         ipoib_put_ah(neigh->ah);
 114                 ipoib_neigh_free(dev, neigh);
 115         }
 116
 117         spin_unlock_irqrestore(&priv->lock, flags);
 118
 119         if (mcast->ah)
 120                 ipoib_put_ah(mcast->ah);
 121
 122         while (!skb_queue_empty(&mcast->pkt_queue)) {
 123                 ++tx_dropped;
 124                 dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
 125         }
 126
 127         spin_lock_irqsave(&priv->tx_lock, flags);
 128         priv->stats.tx_dropped += tx_dropped;
 129         spin_unlock_irqrestore(&priv->tx_lock, flags);
 130
 131         kfree(mcast);
 132 }
 133
 134 static struct ipoib_mcast *ipoib_mcast_alloc(struct net_device *dev,
 135                                              int can_sleep)
 136 {
 137         struct ipoib_mcast *mcast;
 138
 139         mcast = kzalloc(sizeof *mcast, can_sleep ? GFP_KERNEL : GFP_ATOMIC);
 140         if (!mcast)
 141                 return NULL;
 142
 143         mcast->dev = dev;
 144         mcast->created = jiffies;
 145         mcast->backoff = 1;
 146
 147         INIT_LIST_HEAD(&mcast->list);
 148         INIT_LIST_HEAD(&mcast->neigh_list);
 149         skb_queue_head_init(&mcast->pkt_queue);
 150
 151         return mcast;
 152 }
 153
 154 static struct ipoib_mcast *__ipoib_mcast_find(struct net_device *dev, void *mgid)
 155 {
 156         struct ipoib_dev_priv *priv = netdev_priv(dev);
 157         struct rb_node *n = priv->multicast_tree.rb_node;
 158
 159         while (n) {
 160                 struct ipoib_mcast *mcast;
 161                 int ret;
 162
 163                 mcast = rb_entry(n, struct ipoib_mcast, rb_node);
 164
 165                 ret = memcmp(mgid, mcast->mcmember.mgid.raw,
 166                              sizeof (union ib_gid));
 167                 if (ret < 0)
 168                         n = n->rb_left;
 169                 else if (ret > 0)
 170                         n = n->rb_right;
 171                 else
 172                         return mcast;
 173         }
 174
 175         return NULL;
 176 }
 177
 178 static int __ipoib_mcast_add(struct net_device *dev, struct ipoib_mcast *mcast)
 179 {
 180         struct ipoib_dev_priv *priv = netdev_priv(dev);
 181         struct rb_node **n = &priv->multicast_tree.rb_node, *pn = NULL;
 182
 183         while (*n) {
 184                 struct ipoib_mcast *tmcast;
 185                 int ret;
 186
 187                 pn = *n;
 188                 tmcast = rb_entry(pn, struct ipoib_mcast, rb_node);
 189
 190                 ret = memcmp(mcast->mcmember.mgid.raw, tmcast->mcmember.mgid.raw,
 191                              sizeof (union ib_gid));
 192                 if (ret < 0)
 193                         n = &pn->rb_left;
 194                 else if (ret > 0)
 195                         n = &pn->rb_right;
 196                 else
 197                         return -EEXIST;
 198         }
 199
 200         rb_link_node(&mcast->rb_node, pn, n);
 201         rb_insert_color(&mcast->rb_node, &priv->multicast_tree);
 202
 203         return 0;
 204 }
 205
 206 static int ipoib_mcast_join_finish(struct ipoib_mcast *mcast,
 207                                    struct ib_sa_mcmember_rec *mcmember)
 208 {
 209         struct net_device *dev = mcast->dev;
 210         struct ipoib_dev_priv *priv = netdev_priv(dev);
 211         struct ipoib_ah *ah;
 212         int ret;
 213
 214         mcast->mcmember = *mcmember;
 215
 216         /* Set the cached Q_Key before we attach if it's the broadcast group */
 217         if (!memcmp(mcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
 218                     sizeof (union ib_gid))) {
 219                 priv->qkey = be32_to_cpu(priv->broadcast->mcmember.qkey);
 220                 priv->tx_wr.wr.ud.remote_qkey = priv->qkey;
 221         }
 222
 223         if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 224                 if (test_and_set_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 225                         ipoib_warn(priv, "multicast group " IPOIB_GID_FMT
 226                                    " already attached\n",
 227                                    IPOIB_GID_ARG(mcast->mcmember.mgid));
 228
 229                         return 0;
 230                 }
 231
 232                 ret = ipoib_mcast_attach(dev, be16_to_cpu(mcast->mcmember.mlid),
 233                                          &mcast->mcmember.mgid);
 234                 if (ret < 0) {
 235                         ipoib_warn(priv, "couldn't attach QP to multicast group "
 236                                    IPOIB_GID_FMT "\n",
 237                                    IPOIB_GID_ARG(mcast->mcmember.mgid));
 238
 239                         clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags);
 240                         return ret;
 241                 }
 242         }
 243
 244         {
 245                 struct ib_ah_attr av = {
 246                         .dlid          = be16_to_cpu(mcast->mcmember.mlid),
 247                         .port_num      = priv->port,
 248                         .sl            = mcast->mcmember.sl,
 249                         .ah_flags      = IB_AH_GRH,
 250                         .static_rate   = mcast->mcmember.rate,
 251                         .grh           = {
 252                                 .flow_label    = be32_to_cpu(mcast->mcmember.flow_label),
 253                                 .hop_limit     = mcast->mcmember.hop_limit,
 254                                 .sgid_index    = 0,
 255                                 .traffic_class = mcast->mcmember.traffic_class
 256                         }
 257                 };
 258                 av.grh.dgid = mcast->mcmember.mgid;
 259
 260                 ah = ipoib_create_ah(dev, priv->pd, &av);
 261                 if (!ah) {
 262                         ipoib_warn(priv, "ib_address_create failed\n");
 263                 } else {
 264                         spin_lock_irq(&priv->lock);
 265                         mcast->ah = ah;
 266                         spin_unlock_irq(&priv->lock);
 267
 268                         ipoib_dbg_mcast(priv, "MGID " IPOIB_GID_FMT
 269                                         " AV %p, LID 0x%04x, SL %d\n",
 270                                         IPOIB_GID_ARG(mcast->mcmember.mgid),
 271                                         mcast->ah->ah,
 272                                         be16_to_cpu(mcast->mcmember.mlid),
 273                                         mcast->mcmember.sl);
 274                 }
 275         }
 276
 277         /* actually send any queued packets */
 278         spin_lock_irq(&priv->tx_lock);
 279         while (!skb_queue_empty(&mcast->pkt_queue)) {
 280                 struct sk_buff *skb = skb_dequeue(&mcast->pkt_queue);
 281                 spin_unlock_irq(&priv->tx_lock);
 282
 283                 skb->dev = dev;
 284
 285                 if (!skb->dst || !skb->dst->neighbour) {
 286                         /* put pseudoheader back on for next time */
 287                         skb_push(skb, sizeof (struct ipoib_pseudoheader));
 288                 }
 289
 290                 if (dev_queue_xmit(skb))
 291                         ipoib_warn(priv, "dev_queue_xmit failed to requeue packet\n");
 292                 spin_lock_irq(&priv->tx_lock);
 293         }
 294         spin_unlock_irq(&priv->tx_lock);
 295
 296         return 0;
 297 }
 298
 299 static int
 300 ipoib_mcast_sendonly_join_complete(int status,
 301                                    struct ib_sa_multicast *multicast)
 302 {
 303         struct ipoib_mcast *mcast = multicast->context;
 304         struct net_device *dev = mcast->dev;
 305         struct ipoib_dev_priv *priv = netdev_priv(dev);
 306
 307         /* We trap for port events ourselves. */
 308         if (status == -ENETRESET)
 309                 return 0;
 310
 311         if (!status)
 312                 status = ipoib_mcast_join_finish(mcast, &multicast->rec);
 313
 314         if (status) {
 315                 if (mcast->logcount++ < 20)
 316                         ipoib_dbg_mcast(netdev_priv(dev), "multicast join failed for "
 317                                         IPOIB_GID_FMT ", status %d\n",
 318                                         IPOIB_GID_ARG(mcast->mcmember.mgid), status);
 319
 320                 /* Flush out any queued packets */
 321                 spin_lock_irq(&priv->tx_lock);
 322                 while (!skb_queue_empty(&mcast->pkt_queue)) {
 323                         ++priv->stats.tx_dropped;
 324                         dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
 325                 }
 326                 spin_unlock_irq(&priv->tx_lock);
 327
 328                 /* Clear the busy flag so we try again */
 329                 status = test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY,
 330                                             &mcast->flags);
 331         }
 332         return status;
 333 }
 334
 335 static int ipoib_mcast_sendonly_join(struct ipoib_mcast *mcast)
 336 {
 337         struct net_device *dev = mcast->dev;
 338         struct ipoib_dev_priv *priv = netdev_priv(dev);
 339         struct ib_sa_mcmember_rec rec = {
 340 #if 0                           /* Some SMs don't support send-only yet */
 341                 .join_state = 4
 342 #else
 343                 .join_state = 1
 344 #endif
 345         };
 346         int ret = 0;
 347
 348         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
 349                 ipoib_dbg_mcast(priv, "device shutting down, no multicast joins\n");
 350                 return -ENODEV;
 351         }
 352
 353         if (test_and_set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)) {
 354                 ipoib_dbg_mcast(priv, "multicast entry busy, skipping\n");
 355                 return -EBUSY;
 356         }
 357
 358         rec.mgid     = mcast->mcmember.mgid;
 359         rec.port_gid = priv->local_gid;
 360         rec.pkey     = cpu_to_be16(priv->pkey);
 361
 362         mcast->mc = ib_sa_join_multicast(&ipoib_sa_client, priv->ca,
 363                                          priv->port, &rec,
 364                                          IB_SA_MCMEMBER_REC_MGID        |
 365                                          IB_SA_MCMEMBER_REC_PORT_GID    |
 366                                          IB_SA_MCMEMBER_REC_PKEY        |
 367                                          IB_SA_MCMEMBER_REC_JOIN_STATE,
 368                                          GFP_ATOMIC,
 369                                          ipoib_mcast_sendonly_join_complete,
 370                                          mcast);
 371         if (IS_ERR(mcast->mc)) {
 372                 ret = PTR_ERR(mcast->mc);
 373                 clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 374                 ipoib_warn(priv, "ib_sa_join_multicast failed (ret = %d)\n",
 375                            ret);
 376         } else {
 377                 ipoib_dbg_mcast(priv, "no multicast record for " IPOIB_GID_FMT
 378                                 ", starting join\n",
 379                                 IPOIB_GID_ARG(mcast->mcmember.mgid));
 380         }
 381
 382         return ret;
 383 }
 384
 385 static int ipoib_mcast_join_complete(int status,
 386                                      struct ib_sa_multicast *multicast)
 387 {
 388         struct ipoib_mcast *mcast = multicast->context;
 389         struct net_device *dev = mcast->dev;
 390         struct ipoib_dev_priv *priv = netdev_priv(dev);
 391
 392         ipoib_dbg_mcast(priv, "join completion for " IPOIB_GID_FMT
 393                         " (status %d)\n",
 394                         IPOIB_GID_ARG(mcast->mcmember.mgid), status);
 395
 396         /* We trap for port events ourselves. */
 397         if (status == -ENETRESET)
 398                 return 0;
 399
 400         if (!status)
 401                 status = ipoib_mcast_join_finish(mcast, &multicast->rec);
 402
 403         if (!status) {
 404                 mcast->backoff = 1;
 405                 mutex_lock(&mcast_mutex);
 406                 if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 407                         queue_delayed_work(ipoib_workqueue,
 408                                            &priv->mcast_task, 0);
 409                 mutex_unlock(&mcast_mutex);
 410                 return 0;
 411         }
 412
 413         if (mcast->logcount++ < 20) {
 414                 if (status == -ETIMEDOUT) {
 415                         ipoib_dbg_mcast(priv, "multicast join failed for " IPOIB_GID_FMT
 416                                         ", status %d\n",
 417                                         IPOIB_GID_ARG(mcast->mcmember.mgid),
 418                                         status);
 419                 } else {
 420                         ipoib_warn(priv, "multicast join failed for "
 421                                    IPOIB_GID_FMT ", status %d\n",
 422                                    IPOIB_GID_ARG(mcast->mcmember.mgid),
 423                                    status);
 424                 }
 425         }
 426
 427         mcast->backoff *= 2;
 428         if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
 429                 mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
 430
 431         /* Clear the busy flag so we try again */
 432         status = test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 433
 434         mutex_lock(&mcast_mutex);
 435         spin_lock_irq(&priv->lock);
 436         if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 437                 queue_delayed_work(ipoib_workqueue, &priv->mcast_task,
 438                                    mcast->backoff * HZ);
 439         spin_unlock_irq(&priv->lock);
 440         mutex_unlock(&mcast_mutex);
 441
 442         return status;
 443 }
 444
 445 static void ipoib_mcast_join(struct net_device *dev, struct ipoib_mcast *mcast,
 446                              int create)
 447 {
 448         struct ipoib_dev_priv *priv = netdev_priv(dev);
 449         struct ib_sa_mcmember_rec rec = {
 450                 .join_state = 1
 451         };
 452         ib_sa_comp_mask comp_mask;
 453         int ret = 0;
 454
 455         ipoib_dbg_mcast(priv, "joining MGID " IPOIB_GID_FMT "\n",
 456                         IPOIB_GID_ARG(mcast->mcmember.mgid));
 457
 458         rec.mgid     = mcast->mcmember.mgid;
 459         rec.port_gid = priv->local_gid;
 460         rec.pkey     = cpu_to_be16(priv->pkey);
 461
 462         comp_mask =
 463                 IB_SA_MCMEMBER_REC_MGID         |
 464                 IB_SA_MCMEMBER_REC_PORT_GID     |
 465                 IB_SA_MCMEMBER_REC_PKEY         |
 466                 IB_SA_MCMEMBER_REC_JOIN_STATE;
 467
 468         if (create) {
 469                 comp_mask |=
 470                         IB_SA_MCMEMBER_REC_QKEY                 |
 471                         IB_SA_MCMEMBER_REC_MTU_SELECTOR         |
 472                         IB_SA_MCMEMBER_REC_MTU                  |
 473                         IB_SA_MCMEMBER_REC_TRAFFIC_CLASS        |
 474                         IB_SA_MCMEMBER_REC_RATE_SELECTOR        |
 475                         IB_SA_MCMEMBER_REC_RATE                 |
 476                         IB_SA_MCMEMBER_REC_SL                   |
 477                         IB_SA_MCMEMBER_REC_FLOW_LABEL           |
 478                         IB_SA_MCMEMBER_REC_HOP_LIMIT;
 479
 480                 rec.qkey          = priv->broadcast->mcmember.qkey;
 481                 rec.mtu_selector  = IB_SA_EQ;
 482                 rec.mtu           = priv->broadcast->mcmember.mtu;
 483                 rec.traffic_class = priv->broadcast->mcmember.traffic_class;
 484                 rec.rate_selector = IB_SA_EQ;
 485                 rec.rate          = priv->broadcast->mcmember.rate;
 486                 rec.sl            = priv->broadcast->mcmember.sl;
 487                 rec.flow_label    = priv->broadcast->mcmember.flow_label;
 488                 rec.hop_limit     = priv->broadcast->mcmember.hop_limit;
 489         }
 490
 491         set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 492         mcast->mc = ib_sa_join_multicast(&ipoib_sa_client, priv->ca, priv->port,
 493                                          &rec, comp_mask, GFP_KERNEL,
 494                                          ipoib_mcast_join_complete, mcast);
 495         if (IS_ERR(mcast->mc)) {
 496                 clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 497                 ret = PTR_ERR(mcast->mc);
 498                 ipoib_warn(priv, "ib_sa_join_multicast failed, status %d\n", ret);
 499
 500                 mcast->backoff *= 2;
 501                 if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
 502                         mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
 503
 504                 mutex_lock(&mcast_mutex);
 505                 if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 506                         queue_delayed_work(ipoib_workqueue,
 507                                            &priv->mcast_task,
 508                                            mcast->backoff * HZ);
 509                 mutex_unlock(&mcast_mutex);
 510         }
 511 }
 512
 513 void ipoib_mcast_join_task(struct work_struct *work)
 514 {
 515         struct ipoib_dev_priv *priv =
 516                 container_of(work, struct ipoib_dev_priv, mcast_task.work);
 517         struct net_device *dev = priv->dev;
 518
 519         if (!test_bit(IPOIB_MCAST_RUN, &priv->flags))
 520                 return;
 521
 522         if (ib_query_gid(priv->ca, priv->port, 0, &priv->local_gid))
 523                 ipoib_warn(priv, "ib_gid_entry_get() failed\n");
 524         else
 525                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
 526
 527         {
 528                 struct ib_port_attr attr;
 529
 530                 if (!ib_query_port(priv->ca, priv->port, &attr))
 531                         priv->local_lid = attr.lid;
 532                 else
 533                         ipoib_warn(priv, "ib_query_port failed\n");
 534         }
 535
 536         if (!priv->broadcast) {
 537                 struct ipoib_mcast *broadcast;
 538
 539                 broadcast = ipoib_mcast_alloc(dev, 1);
 540                 if (!broadcast) {
 541                         ipoib_warn(priv, "failed to allocate broadcast group\n");
 542                         mutex_lock(&mcast_mutex);
 543                         if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 544                                 queue_delayed_work(ipoib_workqueue,
 545                                                    &priv->mcast_task, HZ);
 546                         mutex_unlock(&mcast_mutex);
 547                         return;
 548                 }
 549
 550                 spin_lock_irq(&priv->lock);
 551                 memcpy(broadcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
 552                        sizeof (union ib_gid));
 553                 priv->broadcast = broadcast;
 554
 555                 __ipoib_mcast_add(dev, priv->broadcast);
 556                 spin_unlock_irq(&priv->lock);
 557         }
 558
 559         if (!test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
 560                 if (!test_bit(IPOIB_MCAST_FLAG_BUSY, &priv->broadcast->flags))
 561                         ipoib_mcast_join(dev, priv->broadcast, 0);
 562                 return;
 563         }
 564
 565         while (1) {
 566                 struct ipoib_mcast *mcast = NULL;
 567
 568                 spin_lock_irq(&priv->lock);
 569                 list_for_each_entry(mcast, &priv->multicast_list, list) {
 570                         if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)
 571                             && !test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)
 572                             && !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 573                                 /* Found the next unjoined group */
 574                                 break;
 575                         }
 576                 }
 577                 spin_unlock_irq(&priv->lock);
 578
 579                 if (&mcast->list == &priv->multicast_list) {
 580                         /* All done */
 581                         break;
 582                 }
 583
 584                 ipoib_mcast_join(dev, mcast, 1);
 585                 return;
 586         }
 587
 588         priv->mcast_mtu = ib_mtu_enum_to_int(priv->broadcast->mcmember.mtu) -
 589                 IPOIB_ENCAP_LEN;
 590
 591         if (!ipoib_cm_admin_enabled(dev))
 592                 dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
 593
 594         ipoib_dbg_mcast(priv, "successfully joined all multicast groups\n");
 595
 596         clear_bit(IPOIB_MCAST_RUN, &priv->flags);
 597         netif_carrier_on(dev);
 598 }
 599
 600 int ipoib_mcast_start_thread(struct net_device *dev)
 601 {
 602         struct ipoib_dev_priv *priv = netdev_priv(dev);
 603
 604         ipoib_dbg_mcast(priv, "starting multicast thread\n");
 605
 606         mutex_lock(&mcast_mutex);
 607         if (!test_and_set_bit(IPOIB_MCAST_RUN, &priv->flags))
 608                 queue_delayed_work(ipoib_workqueue, &priv->mcast_task, 0);
 609         mutex_unlock(&mcast_mutex);
 610
 611         spin_lock_irq(&priv->lock);
 612         set_bit(IPOIB_MCAST_STARTED, &priv->flags);
 613         spin_unlock_irq(&priv->lock);
 614
 615         return 0;
 616 }
 617
 618 int ipoib_mcast_stop_thread(struct net_device *dev, int flush)
 619 {
 620         struct ipoib_dev_priv *priv = netdev_priv(dev);
 621
 622         ipoib_dbg_mcast(priv, "stopping multicast thread\n");
 623
 624         spin_lock_irq(&priv->lock);
 625         clear_bit(IPOIB_MCAST_STARTED, &priv->flags);
 626         spin_unlock_irq(&priv->lock);
 627
 628         mutex_lock(&mcast_mutex);
 629         clear_bit(IPOIB_MCAST_RUN, &priv->flags);
 630         cancel_delayed_work(&priv->mcast_task);
 631         mutex_unlock(&mcast_mutex);
 632
 633         if (flush)
 634                 flush_workqueue(ipoib_workqueue);
 635
 636         return 0;
 637 }
 638
 639 static int ipoib_mcast_leave(struct net_device *dev, struct ipoib_mcast *mcast)
 640 {
 641         struct ipoib_dev_priv *priv = netdev_priv(dev);
 642         int ret = 0;
 643
 644         if (test_and_clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 645                 ipoib_dbg_mcast(priv, "leaving MGID " IPOIB_GID_FMT "\n",
 646                                 IPOIB_GID_ARG(mcast->mcmember.mgid));
 647
 648                 /* Remove ourselves from the multicast group */
 649                 ret = ipoib_mcast_detach(dev, be16_to_cpu(mcast->mcmember.mlid),
 650                                          &mcast->mcmember.mgid);
 651                 if (ret)
 652                         ipoib_warn(priv, "ipoib_mcast_detach failed (result = %d)\n", ret);
 653         }
 654
 655         if (test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
 656                 ib_sa_free_multicast(mcast->mc);
 657
 658         return 0;
 659 }
 660
 661 void ipoib_mcast_send(struct net_device *dev, void *mgid, struct sk_buff *skb)
 662 {
 663         struct ipoib_dev_priv *priv = netdev_priv(dev);
 664         struct ipoib_mcast *mcast;
 665
 666         /*
 667          * We can only be called from ipoib_start_xmit, so we're
 668          * inside tx_lock -- no need to save/restore flags.
 669          */
 670         spin_lock(&priv->lock);
 671
 672         if (!test_bit(IPOIB_MCAST_STARTED, &priv->flags)        ||
 673             !priv->broadcast                                    ||
 674             !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
 675                 ++priv->stats.tx_dropped;
 676                 dev_kfree_skb_any(skb);
 677                 goto unlock;
 678         }
 679
 680         mcast = __ipoib_mcast_find(dev, mgid);
 681         if (!mcast) {
 682                 /* Let's create a new send only group now */
 683                 ipoib_dbg_mcast(priv, "setting up send only multicast group for "
 684                                 IPOIB_GID_FMT "\n", IPOIB_GID_RAW_ARG(mgid));
 685
 686                 mcast = ipoib_mcast_alloc(dev, 0);
 687                 if (!mcast) {
 688                         ipoib_warn(priv, "unable to allocate memory for "
 689                                    "multicast structure\n");
 690                         ++priv->stats.tx_dropped;
 691                         dev_kfree_skb_any(skb);
 692                         goto out;
 693                 }
 694
 695                 set_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags);
 696                 memcpy(mcast->mcmember.mgid.raw, mgid, sizeof (union ib_gid));
 697                 __ipoib_mcast_add(dev, mcast);
 698                 list_add_tail(&mcast->list, &priv->multicast_list);
 699         }
 700
 701         if (!mcast->ah) {
 702                 if (skb_queue_len(&mcast->pkt_queue) < IPOIB_MAX_MCAST_QUEUE)
 703                         skb_queue_tail(&mcast->pkt_queue, skb);
 704                 else {
 705                         ++priv->stats.tx_dropped;
 706                         dev_kfree_skb_any(skb);
 707                 }
 708
 709                 if (test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
 710                         ipoib_dbg_mcast(priv, "no address vector, "
 711                                         "but multicast join already started\n");
 712                 else if (test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags))
 713                         ipoib_mcast_sendonly_join(mcast);
 714
 715                 /*
 716                  * If lookup completes between here and out:, don't
 717                  * want to send packet twice.
 718                  */
 719                 mcast = NULL;
 720         }
 721
 722 out:
 723         if (mcast && mcast->ah) {
 724                 if (skb->dst            &&
 725                     skb->dst->neighbour &&
 726                     !*to_ipoib_neigh(skb->dst->neighbour)) {
 727                         struct ipoib_neigh *neigh = ipoib_neigh_alloc(skb->dst->neighbour);
 728
 729                         if (neigh) {
 730                                 kref_get(&mcast->ah->ref);
 731                                 neigh->ah       = mcast->ah;
 732                                 list_add_tail(&neigh->list, &mcast->neigh_list);
 733                         }
 734                 }
 735
 736                 ipoib_send(dev, skb, mcast->ah, IB_MULTICAST_QPN);
 737         }
 738
 739 unlock:
 740         spin_unlock(&priv->lock);
 741 }
 742
 743 void ipoib_mcast_dev_flush(struct net_device *dev)
 744 {
 745         struct ipoib_dev_priv *priv = netdev_priv(dev);
 746         LIST_HEAD(remove_list);
 747         struct ipoib_mcast *mcast, *tmcast;
 748         unsigned long flags;
 749
 750         ipoib_dbg_mcast(priv, "flushing multicast list\n");
 751
 752         spin_lock_irqsave(&priv->lock, flags);
 753
 754         list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
 755                 list_del(&mcast->list);
 756                 rb_erase(&mcast->rb_node, &priv->multicast_tree);
 757                 list_add_tail(&mcast->list, &remove_list);
 758         }
 759
 760         if (priv->broadcast) {
 761                 rb_erase(&priv->broadcast->rb_node, &priv->multicast_tree);
 762                 list_add_tail(&priv->broadcast->list, &remove_list);
 763                 priv->broadcast = NULL;
 764         }
 765
 766         spin_unlock_irqrestore(&priv->lock, flags);
 767
 768         list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
 769                 ipoib_mcast_leave(dev, mcast);
 770                 ipoib_mcast_free(mcast);
 771         }
 772 }
 773
 774 void ipoib_mcast_restart_task(struct work_struct *work)
 775 {
 776         struct ipoib_dev_priv *priv =
 777                 container_of(work, struct ipoib_dev_priv, restart_task);
 778         struct net_device *dev = priv->dev;
 779         struct dev_mc_list *mclist;
 780         struct ipoib_mcast *mcast, *tmcast;
 781         LIST_HEAD(remove_list);
 782         unsigned long flags;
 783
 784         ipoib_dbg_mcast(priv, "restarting multicast task\n");
 785
 786         ipoib_mcast_stop_thread(dev, 0);
 787
 788         local_irq_save(flags);
 789         netif_tx_lock(dev);
 790         spin_lock(&priv->lock);
 791
 792         /*
 793          * Unfortunately, the networking core only gives us a list of all of
 794          * the multicast hardware addresses. We need to figure out which ones
 795          * are new and which ones have been removed
 796          */
 797
 798         /* Clear out the found flag */
 799         list_for_each_entry(mcast, &priv->multicast_list, list)
 800                 clear_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
 801
 802         /* Mark all of the entries that are found or don't exist */
 803         for (mclist = dev->mc_list; mclist; mclist = mclist->next) {
 804                 union ib_gid mgid;
 805
 806                 memcpy(mgid.raw, mclist->dmi_addr + 4, sizeof mgid);
 807
 808                 /* Add in the P_Key */
 809                 mgid.raw[4] = (priv->pkey >> 8) & 0xff;
 810                 mgid.raw[5] = priv->pkey & 0xff;
 811
 812                 mcast = __ipoib_mcast_find(dev, &mgid);
 813                 if (!mcast || test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 814                         struct ipoib_mcast *nmcast;
 815
 816                         /* Not found or send-only group, let's add a new entry */
 817                         ipoib_dbg_mcast(priv, "adding multicast entry for mgid "
 818                                         IPOIB_GID_FMT "\n", IPOIB_GID_ARG(mgid));
 819
 820                         nmcast = ipoib_mcast_alloc(dev, 0);
 821                         if (!nmcast) {
 822                                 ipoib_warn(priv, "unable to allocate memory for multicast structure\n");
 823                                 continue;
 824                         }
 825
 826                         set_bit(IPOIB_MCAST_FLAG_FOUND, &nmcast->flags);
 827
 828                         nmcast->mcmember.mgid = mgid;
 829
 830                         if (mcast) {
 831                                 /* Destroy the send only entry */
 832                                 list_move_tail(&mcast->list, &remove_list);
 833
 834                                 rb_replace_node(&mcast->rb_node,
 835                                                 &nmcast->rb_node,
 836                                                 &priv->multicast_tree);
 837                         } else
 838                                 __ipoib_mcast_add(dev, nmcast);
 839
 840                         list_add_tail(&nmcast->list, &priv->multicast_list);
 841                 }
 842
 843                 if (mcast)
 844                         set_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
 845         }
 846
 847         /* Remove all of the entries don't exist anymore */
 848         list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
 849                 if (!test_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags) &&
 850                     !test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 851                         ipoib_dbg_mcast(priv, "deleting multicast group " IPOIB_GID_FMT "\n",
 852                                         IPOIB_GID_ARG(mcast->mcmember.mgid));
 853
 854                         rb_erase(&mcast->rb_node, &priv->multicast_tree);
 855
 856                         /* Move to the remove list */
 857                         list_move_tail(&mcast->list, &remove_list);
 858                 }
 859         }
 860
 861         spin_unlock(&priv->lock);
 862         netif_tx_unlock(dev);
 863         local_irq_restore(flags);
 864
 865         /* We have to cancel outside of the spinlock */
 866         list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
 867                 ipoib_mcast_leave(mcast->dev, mcast);
 868                 ipoib_mcast_free(mcast);
 869         }
 870
 871         if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 872                 ipoib_mcast_start_thread(dev);
 873 }
 874
 875 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 876
 877 struct ipoib_mcast_iter *ipoib_mcast_iter_init(struct net_device *dev)
 878 {
 879         struct ipoib_mcast_iter *iter;
 880
 881         iter = kmalloc(sizeof *iter, GFP_KERNEL);
 882         if (!iter)
 883                 return NULL;
 884
 885         iter->dev = dev;
 886         memset(iter->mgid.raw, 0, 16);
 887
 888         if (ipoib_mcast_iter_next(iter)) {
 889                 kfree(iter);
 890                 return NULL;
 891         }
 892
 893         return iter;
 894 }
 895
 896 int ipoib_mcast_iter_next(struct ipoib_mcast_iter *iter)
 897 {
 898         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
 899         struct rb_node *n;
 900         struct ipoib_mcast *mcast;
 901         int ret = 1;
 902
 903         spin_lock_irq(&priv->lock);
 904
 905         n = rb_first(&priv->multicast_tree);
 906
 907         while (n) {
 908                 mcast = rb_entry(n, struct ipoib_mcast, rb_node);
 909
 910                 if (memcmp(iter->mgid.raw, mcast->mcmember.mgid.raw,
 911                            sizeof (union ib_gid)) < 0) {
 912                         iter->mgid      = mcast->mcmember.mgid;
 913                         iter->created   = mcast->created;
 914                         iter->queuelen  = skb_queue_len(&mcast->pkt_queue);
 915                         iter->complete  = !!mcast->ah;
 916                         iter->send_only = !!(mcast->flags & (1 << IPOIB_MCAST_FLAG_SENDONLY));
 917
 918                         ret = 0;
 919
 920                         break;
 921                 }
 922
 923                 n = rb_next(n);
 924         }
 925
 926         spin_unlock_irq(&priv->lock);
 927
 928         return ret;
 929 }
 930
 931 void ipoib_mcast_iter_read(struct ipoib_mcast_iter *iter,
 932                            union ib_gid *mgid,
 933                            unsigned long *created,
 934                            unsigned int *queuelen,
 935                            unsigned int *complete,
 936                            unsigned int *send_only)
 937 {
 938         *mgid      = iter->mgid;
 939         *created   = iter->created;
 940         *queuelen  = iter->queuelen;
 941         *complete  = iter->complete;
 942         *send_only = iter->send_only;
 943 }
 944
 945 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */