SAFE public projects git trees. - safe/jmp/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 ipoib_ah          *ah;
  64
  65         struct rb_node    rb_node;
  66         struct list_head  list;
  67         struct completion done;
  68
  69         int                 query_id;
  70         struct ib_sa_query *query;
  71
  72         unsigned long created;
  73         unsigned long backoff;
  74
  75         unsigned long flags;
  76         unsigned char logcount;
  77
  78         struct list_head  neigh_list;
  79
  80         struct sk_buff_head pkt_queue;
  81
  82         struct net_device *dev;
  83 };
  84
  85 struct ipoib_mcast_iter {
  86         struct net_device *dev;
  87         union ib_gid       mgid;
  88         unsigned long      created;
  89         unsigned int       queuelen;
  90         unsigned int       complete;
  91         unsigned int       send_only;
  92 };
  93
  94 static void ipoib_mcast_free(struct ipoib_mcast *mcast)
  95 {
  96         struct net_device *dev = mcast->dev;
  97         struct ipoib_dev_priv *priv = netdev_priv(dev);
  98         struct ipoib_neigh *neigh, *tmp;
  99         unsigned long flags;
 100         int tx_dropped = 0;
 101
 102         ipoib_dbg_mcast(netdev_priv(dev),
 103                         "deleting multicast group " IPOIB_GID_FMT "\n",
 104                         IPOIB_GID_ARG(mcast->mcmember.mgid));
 105
 106         spin_lock_irqsave(&priv->lock, flags);
 107
 108         list_for_each_entry_safe(neigh, tmp, &mcast->neigh_list, list) {
 109                 /*
 110                  * It's safe to call ipoib_put_ah() inside priv->lock
 111                  * here, because we know that mcast->ah will always
 112                  * hold one more reference, so ipoib_put_ah() will
 113                  * never do more than decrement the ref count.
 114                  */
 115                 if (neigh->ah)
 116                         ipoib_put_ah(neigh->ah);
 117                 *to_ipoib_neigh(neigh->neighbour) = NULL;
 118                 neigh->neighbour->ops->destructor = NULL;
 119                 kfree(neigh);
 120         }
 121
 122         spin_unlock_irqrestore(&priv->lock, flags);
 123
 124         if (mcast->ah)
 125                 ipoib_put_ah(mcast->ah);
 126
 127         while (!skb_queue_empty(&mcast->pkt_queue)) {
 128                 ++tx_dropped;
 129                 dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
 130         }
 131
 132         spin_lock_irqsave(&priv->tx_lock, flags);
 133         priv->stats.tx_dropped += tx_dropped;
 134         spin_unlock_irqrestore(&priv->tx_lock, flags);
 135
 136         kfree(mcast);
 137 }
 138
 139 static struct ipoib_mcast *ipoib_mcast_alloc(struct net_device *dev,
 140                                              int can_sleep)
 141 {
 142         struct ipoib_mcast *mcast;
 143
 144         mcast = kzalloc(sizeof *mcast, can_sleep ? GFP_KERNEL : GFP_ATOMIC);
 145         if (!mcast)
 146                 return NULL;
 147
 148         mcast->dev = dev;
 149         mcast->created = jiffies;
 150         mcast->backoff = 1;
 151
 152         INIT_LIST_HEAD(&mcast->list);
 153         INIT_LIST_HEAD(&mcast->neigh_list);
 154         skb_queue_head_init(&mcast->pkt_queue);
 155
 156         return mcast;
 157 }
 158
 159 static struct ipoib_mcast *__ipoib_mcast_find(struct net_device *dev, union ib_gid *mgid)
 160 {
 161         struct ipoib_dev_priv *priv = netdev_priv(dev);
 162         struct rb_node *n = priv->multicast_tree.rb_node;
 163
 164         while (n) {
 165                 struct ipoib_mcast *mcast;
 166                 int ret;
 167
 168                 mcast = rb_entry(n, struct ipoib_mcast, rb_node);
 169
 170                 ret = memcmp(mgid->raw, mcast->mcmember.mgid.raw,
 171                              sizeof (union ib_gid));
 172                 if (ret < 0)
 173                         n = n->rb_left;
 174                 else if (ret > 0)
 175                         n = n->rb_right;
 176                 else
 177                         return mcast;
 178         }
 179
 180         return NULL;
 181 }
 182
 183 static int __ipoib_mcast_add(struct net_device *dev, struct ipoib_mcast *mcast)
 184 {
 185         struct ipoib_dev_priv *priv = netdev_priv(dev);
 186         struct rb_node **n = &priv->multicast_tree.rb_node, *pn = NULL;
 187
 188         while (*n) {
 189                 struct ipoib_mcast *tmcast;
 190                 int ret;
 191
 192                 pn = *n;
 193                 tmcast = rb_entry(pn, struct ipoib_mcast, rb_node);
 194
 195                 ret = memcmp(mcast->mcmember.mgid.raw, tmcast->mcmember.mgid.raw,
 196                              sizeof (union ib_gid));
 197                 if (ret < 0)
 198                         n = &pn->rb_left;
 199                 else if (ret > 0)
 200                         n = &pn->rb_right;
 201                 else
 202                         return -EEXIST;
 203         }
 204
 205         rb_link_node(&mcast->rb_node, pn, n);
 206         rb_insert_color(&mcast->rb_node, &priv->multicast_tree);
 207
 208         return 0;
 209 }
 210
 211 static int ipoib_mcast_join_finish(struct ipoib_mcast *mcast,
 212                                    struct ib_sa_mcmember_rec *mcmember)
 213 {
 214         struct net_device *dev = mcast->dev;
 215         struct ipoib_dev_priv *priv = netdev_priv(dev);
 216         struct ipoib_ah *ah;
 217         int ret;
 218
 219         mcast->mcmember = *mcmember;
 220
 221         /* Set the cached Q_Key before we attach if it's the broadcast group */
 222         if (!memcmp(mcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
 223                     sizeof (union ib_gid))) {
 224                 priv->qkey = be32_to_cpu(priv->broadcast->mcmember.qkey);
 225                 priv->tx_wr.wr.ud.remote_qkey = priv->qkey;
 226         }
 227
 228         if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 229                 if (test_and_set_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 230                         ipoib_warn(priv, "multicast group " IPOIB_GID_FMT
 231                                    " already attached\n",
 232                                    IPOIB_GID_ARG(mcast->mcmember.mgid));
 233
 234                         return 0;
 235                 }
 236
 237                 ret = ipoib_mcast_attach(dev, be16_to_cpu(mcast->mcmember.mlid),
 238                                          &mcast->mcmember.mgid);
 239                 if (ret < 0) {
 240                         ipoib_warn(priv, "couldn't attach QP to multicast group "
 241                                    IPOIB_GID_FMT "\n",
 242                                    IPOIB_GID_ARG(mcast->mcmember.mgid));
 243
 244                         clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags);
 245                         return ret;
 246                 }
 247         }
 248
 249         {
 250                 struct ib_ah_attr av = {
 251                         .dlid          = be16_to_cpu(mcast->mcmember.mlid),
 252                         .port_num      = priv->port,
 253                         .sl            = mcast->mcmember.sl,
 254                         .ah_flags      = IB_AH_GRH,
 255                         .grh           = {
 256                                 .flow_label    = be32_to_cpu(mcast->mcmember.flow_label),
 257                                 .hop_limit     = mcast->mcmember.hop_limit,
 258                                 .sgid_index    = 0,
 259                                 .traffic_class = mcast->mcmember.traffic_class
 260                         }
 261                 };
 262                 int path_rate = ib_sa_rate_enum_to_int(mcast->mcmember.rate);
 263
 264                 av.grh.dgid = mcast->mcmember.mgid;
 265
 266                 if (path_rate > 0 && priv->local_rate > path_rate)
 267                         av.static_rate = (priv->local_rate - 1) / path_rate;
 268
 269                 ipoib_dbg_mcast(priv, "static_rate %d for local port %dX, mcmember %dX\n",
 270                                 av.static_rate, priv->local_rate,
 271                                 ib_sa_rate_enum_to_int(mcast->mcmember.rate));
 272
 273                 ah = ipoib_create_ah(dev, priv->pd, &av);
 274                 if (!ah) {
 275                         ipoib_warn(priv, "ib_address_create failed\n");
 276                 } else {
 277                         ipoib_dbg_mcast(priv, "MGID " IPOIB_GID_FMT
 278                                         " AV %p, LID 0x%04x, SL %d\n",
 279                                         IPOIB_GID_ARG(mcast->mcmember.mgid),
 280                                         mcast->ah->ah,
 281                                         be16_to_cpu(mcast->mcmember.mlid),
 282                                         mcast->mcmember.sl);
 283                 }
 284
 285                 spin_lock_irq(&priv->lock);
 286                 mcast->ah = ah;
 287                 spin_unlock_irq(&priv->lock);
 288         }
 289
 290         /* actually send any queued packets */
 291         spin_lock_irq(&priv->tx_lock);
 292         while (!skb_queue_empty(&mcast->pkt_queue)) {
 293                 struct sk_buff *skb = skb_dequeue(&mcast->pkt_queue);
 294                 spin_unlock_irq(&priv->tx_lock);
 295
 296                 skb->dev = dev;
 297
 298                 if (!skb->dst || !skb->dst->neighbour) {
 299                         /* put pseudoheader back on for next time */
 300                         skb_push(skb, sizeof (struct ipoib_pseudoheader));
 301                 }
 302
 303                 if (dev_queue_xmit(skb))
 304                         ipoib_warn(priv, "dev_queue_xmit failed to requeue packet\n");
 305                 spin_lock_irq(&priv->tx_lock);
 306         }
 307         spin_unlock_irq(&priv->tx_lock);
 308
 309         return 0;
 310 }
 311
 312 static void
 313 ipoib_mcast_sendonly_join_complete(int status,
 314                                    struct ib_sa_mcmember_rec *mcmember,
 315                                    void *mcast_ptr)
 316 {
 317         struct ipoib_mcast *mcast = mcast_ptr;
 318         struct net_device *dev = mcast->dev;
 319         struct ipoib_dev_priv *priv = netdev_priv(dev);
 320
 321         if (!status)
 322                 ipoib_mcast_join_finish(mcast, mcmember);
 323         else {
 324                 if (mcast->logcount++ < 20)
 325                         ipoib_dbg_mcast(netdev_priv(dev), "multicast join failed for "
 326                                         IPOIB_GID_FMT ", status %d\n",
 327                                         IPOIB_GID_ARG(mcast->mcmember.mgid), status);
 328
 329                 /* Flush out any queued packets */
 330                 spin_lock_irq(&priv->tx_lock);
 331                 while (!skb_queue_empty(&mcast->pkt_queue)) {
 332                         ++priv->stats.tx_dropped;
 333                         dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
 334                 }
 335                 spin_unlock_irq(&priv->tx_lock);
 336
 337                 /* Clear the busy flag so we try again */
 338                 clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 339         }
 340
 341         complete(&mcast->done);
 342 }
 343
 344 static int ipoib_mcast_sendonly_join(struct ipoib_mcast *mcast)
 345 {
 346         struct net_device *dev = mcast->dev;
 347         struct ipoib_dev_priv *priv = netdev_priv(dev);
 348         struct ib_sa_mcmember_rec rec = {
 349 #if 0                           /* Some SMs don't support send-only yet */
 350                 .join_state = 4
 351 #else
 352                 .join_state = 1
 353 #endif
 354         };
 355         int ret = 0;
 356
 357         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
 358                 ipoib_dbg_mcast(priv, "device shutting down, no multicast joins\n");
 359                 return -ENODEV;
 360         }
 361
 362         if (test_and_set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)) {
 363                 ipoib_dbg_mcast(priv, "multicast entry busy, skipping\n");
 364                 return -EBUSY;
 365         }
 366
 367         rec.mgid     = mcast->mcmember.mgid;
 368         rec.port_gid = priv->local_gid;
 369         rec.pkey     = cpu_to_be16(priv->pkey);
 370
 371         init_completion(&mcast->done);
 372
 373         ret = ib_sa_mcmember_rec_set(priv->ca, priv->port, &rec,
 374                                      IB_SA_MCMEMBER_REC_MGID            |
 375                                      IB_SA_MCMEMBER_REC_PORT_GID        |
 376                                      IB_SA_MCMEMBER_REC_PKEY            |
 377                                      IB_SA_MCMEMBER_REC_JOIN_STATE,
 378                                      1000, GFP_ATOMIC,
 379                                      ipoib_mcast_sendonly_join_complete,
 380                                      mcast, &mcast->query);
 381         if (ret < 0) {
 382                 ipoib_warn(priv, "ib_sa_mcmember_rec_set failed (ret = %d)\n",
 383                            ret);
 384         } else {
 385                 ipoib_dbg_mcast(priv, "no multicast record for " IPOIB_GID_FMT
 386                                 ", starting join\n",
 387                                 IPOIB_GID_ARG(mcast->mcmember.mgid));
 388
 389                 mcast->query_id = ret;
 390         }
 391
 392         return ret;
 393 }
 394
 395 static void ipoib_mcast_join_complete(int status,
 396                                       struct ib_sa_mcmember_rec *mcmember,
 397                                       void *mcast_ptr)
 398 {
 399         struct ipoib_mcast *mcast = mcast_ptr;
 400         struct net_device *dev = mcast->dev;
 401         struct ipoib_dev_priv *priv = netdev_priv(dev);
 402
 403         ipoib_dbg_mcast(priv, "join completion for " IPOIB_GID_FMT
 404                         " (status %d)\n",
 405                         IPOIB_GID_ARG(mcast->mcmember.mgid), status);
 406
 407         if (!status && !ipoib_mcast_join_finish(mcast, mcmember)) {
 408                 mcast->backoff = 1;
 409                 mutex_lock(&mcast_mutex);
 410                 if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 411                         queue_work(ipoib_workqueue, &priv->mcast_task);
 412                 mutex_unlock(&mcast_mutex);
 413                 complete(&mcast->done);
 414                 return;
 415         }
 416
 417         if (status == -EINTR) {
 418                 complete(&mcast->done);
 419                 return;
 420         }
 421
 422         if (status && mcast->logcount++ < 20) {
 423                 if (status == -ETIMEDOUT || status == -EINTR) {
 424                         ipoib_dbg_mcast(priv, "multicast join failed for " IPOIB_GID_FMT
 425                                         ", status %d\n",
 426                                         IPOIB_GID_ARG(mcast->mcmember.mgid),
 427                                         status);
 428                 } else {
 429                         ipoib_warn(priv, "multicast join failed for "
 430                                    IPOIB_GID_FMT ", status %d\n",
 431                                    IPOIB_GID_ARG(mcast->mcmember.mgid),
 432                                    status);
 433                 }
 434         }
 435
 436         mcast->backoff *= 2;
 437         if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
 438                 mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
 439
 440         mcast->query = NULL;
 441
 442         mutex_lock(&mcast_mutex);
 443         if (test_bit(IPOIB_MCAST_RUN, &priv->flags)) {
 444                 if (status == -ETIMEDOUT)
 445                         queue_work(ipoib_workqueue, &priv->mcast_task);
 446                 else
 447                         queue_delayed_work(ipoib_workqueue, &priv->mcast_task,
 448                                            mcast->backoff * HZ);
 449         } else
 450                 complete(&mcast->done);
 451         mutex_unlock(&mcast_mutex);
 452
 453         return;
 454 }
 455
 456 static void ipoib_mcast_join(struct net_device *dev, struct ipoib_mcast *mcast,
 457                              int create)
 458 {
 459         struct ipoib_dev_priv *priv = netdev_priv(dev);
 460         struct ib_sa_mcmember_rec rec = {
 461                 .join_state = 1
 462         };
 463         ib_sa_comp_mask comp_mask;
 464         int ret = 0;
 465
 466         ipoib_dbg_mcast(priv, "joining MGID " IPOIB_GID_FMT "\n",
 467                         IPOIB_GID_ARG(mcast->mcmember.mgid));
 468
 469         rec.mgid     = mcast->mcmember.mgid;
 470         rec.port_gid = priv->local_gid;
 471         rec.pkey     = cpu_to_be16(priv->pkey);
 472
 473         comp_mask =
 474                 IB_SA_MCMEMBER_REC_MGID         |
 475                 IB_SA_MCMEMBER_REC_PORT_GID     |
 476                 IB_SA_MCMEMBER_REC_PKEY         |
 477                 IB_SA_MCMEMBER_REC_JOIN_STATE;
 478
 479         if (create) {
 480                 comp_mask |=
 481                         IB_SA_MCMEMBER_REC_QKEY         |
 482                         IB_SA_MCMEMBER_REC_SL           |
 483                         IB_SA_MCMEMBER_REC_FLOW_LABEL   |
 484                         IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
 485
 486                 rec.qkey          = priv->broadcast->mcmember.qkey;
 487                 rec.sl            = priv->broadcast->mcmember.sl;
 488                 rec.flow_label    = priv->broadcast->mcmember.flow_label;
 489                 rec.traffic_class = priv->broadcast->mcmember.traffic_class;
 490         }
 491
 492         init_completion(&mcast->done);
 493
 494         ret = ib_sa_mcmember_rec_set(priv->ca, priv->port, &rec, comp_mask,
 495                                      mcast->backoff * 1000, GFP_ATOMIC,
 496                                      ipoib_mcast_join_complete,
 497                                      mcast, &mcast->query);
 498
 499         if (ret < 0) {
 500                 ipoib_warn(priv, "ib_sa_mcmember_rec_set failed, status %d\n", ret);
 501
 502                 mcast->backoff *= 2;
 503                 if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
 504                         mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
 505
 506                 mutex_lock(&mcast_mutex);
 507                 if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 508                         queue_delayed_work(ipoib_workqueue,
 509                                            &priv->mcast_task,
 510                                            mcast->backoff * HZ);
 511                 mutex_unlock(&mcast_mutex);
 512         } else
 513                 mcast->query_id = ret;
 514 }
 515
 516 void ipoib_mcast_join_task(void *dev_ptr)
 517 {
 518         struct net_device *dev = dev_ptr;
 519         struct ipoib_dev_priv *priv = netdev_priv(dev);
 520
 521         if (!test_bit(IPOIB_MCAST_RUN, &priv->flags))
 522                 return;
 523
 524         if (ib_query_gid(priv->ca, priv->port, 0, &priv->local_gid))
 525                 ipoib_warn(priv, "ib_gid_entry_get() failed\n");
 526         else
 527                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
 528
 529         {
 530                 struct ib_port_attr attr;
 531
 532                 if (!ib_query_port(priv->ca, priv->port, &attr)) {
 533                         priv->local_lid  = attr.lid;
 534                         priv->local_rate = attr.active_speed *
 535                                 ib_width_enum_to_int(attr.active_width);
 536                 } else
 537                         ipoib_warn(priv, "ib_query_port failed\n");
 538         }
 539
 540         if (!priv->broadcast) {
 541                 struct ipoib_mcast *broadcast;
 542
 543                 broadcast = ipoib_mcast_alloc(dev, 1);
 544                 if (!broadcast) {
 545                         ipoib_warn(priv, "failed to allocate broadcast group\n");
 546                         mutex_lock(&mcast_mutex);
 547                         if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 548                                 queue_delayed_work(ipoib_workqueue,
 549                                                    &priv->mcast_task, HZ);
 550                         mutex_unlock(&mcast_mutex);
 551                         return;
 552                 }
 553
 554                 spin_lock_irq(&priv->lock);
 555                 memcpy(broadcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
 556                        sizeof (union ib_gid));
 557                 priv->broadcast = broadcast;
 558
 559                 __ipoib_mcast_add(dev, priv->broadcast);
 560                 spin_unlock_irq(&priv->lock);
 561         }
 562
 563         if (!test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
 564                 ipoib_mcast_join(dev, priv->broadcast, 0);
 565                 return;
 566         }
 567
 568         while (1) {
 569                 struct ipoib_mcast *mcast = NULL;
 570
 571                 spin_lock_irq(&priv->lock);
 572                 list_for_each_entry(mcast, &priv->multicast_list, list) {
 573                         if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)
 574                             && !test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)
 575                             && !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 576                                 /* Found the next unjoined group */
 577                                 break;
 578                         }
 579                 }
 580                 spin_unlock_irq(&priv->lock);
 581
 582                 if (&mcast->list == &priv->multicast_list) {
 583                         /* All done */
 584                         break;
 585                 }
 586
 587                 ipoib_mcast_join(dev, mcast, 1);
 588                 return;
 589         }
 590
 591         priv->mcast_mtu = ib_mtu_enum_to_int(priv->broadcast->mcmember.mtu) -
 592                 IPOIB_ENCAP_LEN;
 593         dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
 594
 595         ipoib_dbg_mcast(priv, "successfully joined all multicast groups\n");
 596
 597         clear_bit(IPOIB_MCAST_RUN, &priv->flags);
 598         netif_carrier_on(dev);
 599 }
 600
 601 int ipoib_mcast_start_thread(struct net_device *dev)
 602 {
 603         struct ipoib_dev_priv *priv = netdev_priv(dev);
 604
 605         ipoib_dbg_mcast(priv, "starting multicast thread\n");
 606
 607         mutex_lock(&mcast_mutex);
 608         if (!test_and_set_bit(IPOIB_MCAST_RUN, &priv->flags))
 609                 queue_work(ipoib_workqueue, &priv->mcast_task);
 610         mutex_unlock(&mcast_mutex);
 611
 612         spin_lock_irq(&priv->lock);
 613         set_bit(IPOIB_MCAST_STARTED, &priv->flags);
 614         spin_unlock_irq(&priv->lock);
 615
 616         return 0;
 617 }
 618
 619 int ipoib_mcast_stop_thread(struct net_device *dev, int flush)
 620 {
 621         struct ipoib_dev_priv *priv = netdev_priv(dev);
 622         struct ipoib_mcast *mcast;
 623
 624         ipoib_dbg_mcast(priv, "stopping multicast thread\n");
 625
 626         spin_lock_irq(&priv->lock);
 627         clear_bit(IPOIB_MCAST_STARTED, &priv->flags);
 628         spin_unlock_irq(&priv->lock);
 629
 630         mutex_lock(&mcast_mutex);
 631         clear_bit(IPOIB_MCAST_RUN, &priv->flags);
 632         cancel_delayed_work(&priv->mcast_task);
 633         mutex_unlock(&mcast_mutex);
 634
 635         if (flush)
 636                 flush_workqueue(ipoib_workqueue);
 637
 638         if (priv->broadcast && priv->broadcast->query) {
 639                 ib_sa_cancel_query(priv->broadcast->query_id, priv->broadcast->query);
 640                 priv->broadcast->query = NULL;
 641                 ipoib_dbg_mcast(priv, "waiting for bcast\n");
 642                 wait_for_completion(&priv->broadcast->done);
 643         }
 644
 645         list_for_each_entry(mcast, &priv->multicast_list, list) {
 646                 if (mcast->query) {
 647                         ib_sa_cancel_query(mcast->query_id, mcast->query);
 648                         mcast->query = NULL;
 649                         ipoib_dbg_mcast(priv, "waiting for MGID " IPOIB_GID_FMT "\n",
 650                                         IPOIB_GID_ARG(mcast->mcmember.mgid));
 651                         wait_for_completion(&mcast->done);
 652                 }
 653         }
 654
 655         return 0;
 656 }
 657
 658 static int ipoib_mcast_leave(struct net_device *dev, struct ipoib_mcast *mcast)
 659 {
 660         struct ipoib_dev_priv *priv = netdev_priv(dev);
 661         struct ib_sa_mcmember_rec rec = {
 662                 .join_state = 1
 663         };
 664         int ret = 0;
 665
 666         if (!test_and_clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags))
 667                 return 0;
 668
 669         ipoib_dbg_mcast(priv, "leaving MGID " IPOIB_GID_FMT "\n",
 670                         IPOIB_GID_ARG(mcast->mcmember.mgid));
 671
 672         rec.mgid     = mcast->mcmember.mgid;
 673         rec.port_gid = priv->local_gid;
 674         rec.pkey     = cpu_to_be16(priv->pkey);
 675
 676         /* Remove ourselves from the multicast group */
 677         ret = ipoib_mcast_detach(dev, be16_to_cpu(mcast->mcmember.mlid),
 678                                  &mcast->mcmember.mgid);
 679         if (ret)
 680                 ipoib_warn(priv, "ipoib_mcast_detach failed (result = %d)\n", ret);
 681
 682         /*
 683          * Just make one shot at leaving and don't wait for a reply;
 684          * if we fail, too bad.
 685          */
 686         ret = ib_sa_mcmember_rec_delete(priv->ca, priv->port, &rec,
 687                                         IB_SA_MCMEMBER_REC_MGID         |
 688                                         IB_SA_MCMEMBER_REC_PORT_GID     |
 689                                         IB_SA_MCMEMBER_REC_PKEY         |
 690                                         IB_SA_MCMEMBER_REC_JOIN_STATE,
 691                                         0, GFP_ATOMIC, NULL,
 692                                         mcast, &mcast->query);
 693         if (ret < 0)
 694                 ipoib_warn(priv, "ib_sa_mcmember_rec_delete failed "
 695                            "for leave (result = %d)\n", ret);
 696
 697         return 0;
 698 }
 699
 700 void ipoib_mcast_send(struct net_device *dev, union ib_gid *mgid,
 701                       struct sk_buff *skb)
 702 {
 703         struct ipoib_dev_priv *priv = netdev_priv(dev);
 704         struct ipoib_mcast *mcast;
 705
 706         /*
 707          * We can only be called from ipoib_start_xmit, so we're
 708          * inside tx_lock -- no need to save/restore flags.
 709          */
 710         spin_lock(&priv->lock);
 711
 712         if (!test_bit(IPOIB_MCAST_STARTED, &priv->flags)        ||
 713             !priv->broadcast                                    ||
 714             !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
 715                 ++priv->stats.tx_dropped;
 716                 dev_kfree_skb_any(skb);
 717                 goto unlock;
 718         }
 719
 720         mcast = __ipoib_mcast_find(dev, mgid);
 721         if (!mcast) {
 722                 /* Let's create a new send only group now */
 723                 ipoib_dbg_mcast(priv, "setting up send only multicast group for "
 724                                 IPOIB_GID_FMT "\n", IPOIB_GID_ARG(*mgid));
 725
 726                 mcast = ipoib_mcast_alloc(dev, 0);
 727                 if (!mcast) {
 728                         ipoib_warn(priv, "unable to allocate memory for "
 729                                    "multicast structure\n");
 730                         ++priv->stats.tx_dropped;
 731                         dev_kfree_skb_any(skb);
 732                         goto out;
 733                 }
 734
 735                 set_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags);
 736                 mcast->mcmember.mgid = *mgid;
 737                 __ipoib_mcast_add(dev, mcast);
 738                 list_add_tail(&mcast->list, &priv->multicast_list);
 739         }
 740
 741         if (!mcast->ah) {
 742                 if (skb_queue_len(&mcast->pkt_queue) < IPOIB_MAX_MCAST_QUEUE)
 743                         skb_queue_tail(&mcast->pkt_queue, skb);
 744                 else {
 745                         ++priv->stats.tx_dropped;
 746                         dev_kfree_skb_any(skb);
 747                 }
 748
 749                 if (mcast->query)
 750                         ipoib_dbg_mcast(priv, "no address vector, "
 751                                         "but multicast join already started\n");
 752                 else if (test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags))
 753                         ipoib_mcast_sendonly_join(mcast);
 754
 755                 /*
 756                  * If lookup completes between here and out:, don't
 757                  * want to send packet twice.
 758                  */
 759                 mcast = NULL;
 760         }
 761
 762 out:
 763         if (mcast && mcast->ah) {
 764                 if (skb->dst            &&
 765                     skb->dst->neighbour &&
 766                     !*to_ipoib_neigh(skb->dst->neighbour)) {
 767                         struct ipoib_neigh *neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
 768
 769                         if (neigh) {
 770                                 kref_get(&mcast->ah->ref);
 771                                 neigh->ah       = mcast->ah;
 772                                 neigh->neighbour = skb->dst->neighbour;
 773                                 *to_ipoib_neigh(skb->dst->neighbour) = neigh;
 774                                 list_add_tail(&neigh->list, &mcast->neigh_list);
 775                         }
 776                 }
 777
 778                 ipoib_send(dev, skb, mcast->ah, IB_MULTICAST_QPN);
 779         }
 780
 781 unlock:
 782         spin_unlock(&priv->lock);
 783 }
 784
 785 void ipoib_mcast_dev_flush(struct net_device *dev)
 786 {
 787         struct ipoib_dev_priv *priv = netdev_priv(dev);
 788         LIST_HEAD(remove_list);
 789         struct ipoib_mcast *mcast, *tmcast;
 790         unsigned long flags;
 791
 792         ipoib_dbg_mcast(priv, "flushing multicast list\n");
 793
 794         spin_lock_irqsave(&priv->lock, flags);
 795
 796         list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
 797                 list_del(&mcast->list);
 798                 rb_erase(&mcast->rb_node, &priv->multicast_tree);
 799                 list_add_tail(&mcast->list, &remove_list);
 800         }
 801
 802         if (priv->broadcast) {
 803                 rb_erase(&priv->broadcast->rb_node, &priv->multicast_tree);
 804                 list_add_tail(&priv->broadcast->list, &remove_list);
 805                 priv->broadcast = NULL;
 806         }
 807
 808         spin_unlock_irqrestore(&priv->lock, flags);
 809
 810         list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
 811                 ipoib_mcast_leave(dev, mcast);
 812                 ipoib_mcast_free(mcast);
 813         }
 814 }
 815
 816 void ipoib_mcast_restart_task(void *dev_ptr)
 817 {
 818         struct net_device *dev = dev_ptr;
 819         struct ipoib_dev_priv *priv = netdev_priv(dev);
 820         struct dev_mc_list *mclist;
 821         struct ipoib_mcast *mcast, *tmcast;
 822         LIST_HEAD(remove_list);
 823         unsigned long flags;
 824
 825         ipoib_dbg_mcast(priv, "restarting multicast task\n");
 826
 827         ipoib_mcast_stop_thread(dev, 0);
 828
 829         spin_lock_irqsave(&dev->xmit_lock, flags);
 830         spin_lock(&priv->lock);
 831
 832         /*
 833          * Unfortunately, the networking core only gives us a list of all of
 834          * the multicast hardware addresses. We need to figure out which ones
 835          * are new and which ones have been removed
 836          */
 837
 838         /* Clear out the found flag */
 839         list_for_each_entry(mcast, &priv->multicast_list, list)
 840                 clear_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
 841
 842         /* Mark all of the entries that are found or don't exist */
 843         for (mclist = dev->mc_list; mclist; mclist = mclist->next) {
 844                 union ib_gid mgid;
 845
 846                 memcpy(mgid.raw, mclist->dmi_addr + 4, sizeof mgid);
 847
 848                 /* Add in the P_Key */
 849                 mgid.raw[4] = (priv->pkey >> 8) & 0xff;
 850                 mgid.raw[5] = priv->pkey & 0xff;
 851
 852                 mcast = __ipoib_mcast_find(dev, &mgid);
 853                 if (!mcast || test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 854                         struct ipoib_mcast *nmcast;
 855
 856                         /* Not found or send-only group, let's add a new entry */
 857                         ipoib_dbg_mcast(priv, "adding multicast entry for mgid "
 858                                         IPOIB_GID_FMT "\n", IPOIB_GID_ARG(mgid));
 859
 860                         nmcast = ipoib_mcast_alloc(dev, 0);
 861                         if (!nmcast) {
 862                                 ipoib_warn(priv, "unable to allocate memory for multicast structure\n");
 863                                 continue;
 864                         }
 865
 866                         set_bit(IPOIB_MCAST_FLAG_FOUND, &nmcast->flags);
 867
 868                         nmcast->mcmember.mgid = mgid;
 869
 870                         if (mcast) {
 871                                 /* Destroy the send only entry */
 872                                 list_del(&mcast->list);
 873                                 list_add_tail(&mcast->list, &remove_list);
 874
 875                                 rb_replace_node(&mcast->rb_node,
 876                                                 &nmcast->rb_node,
 877                                                 &priv->multicast_tree);
 878                         } else
 879                                 __ipoib_mcast_add(dev, nmcast);
 880
 881                         list_add_tail(&nmcast->list, &priv->multicast_list);
 882                 }
 883
 884                 if (mcast)
 885                         set_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
 886         }
 887
 888         /* Remove all of the entries don't exist anymore */
 889         list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
 890                 if (!test_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags) &&
 891                     !test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 892                         ipoib_dbg_mcast(priv, "deleting multicast group " IPOIB_GID_FMT "\n",
 893                                         IPOIB_GID_ARG(mcast->mcmember.mgid));
 894
 895                         rb_erase(&mcast->rb_node, &priv->multicast_tree);
 896
 897                         /* Move to the remove list */
 898                         list_del(&mcast->list);
 899                         list_add_tail(&mcast->list, &remove_list);
 900                 }
 901         }
 902
 903         spin_unlock(&priv->lock);
 904         spin_unlock_irqrestore(&dev->xmit_lock, flags);
 905
 906         /* We have to cancel outside of the spinlock */
 907         list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
 908                 ipoib_mcast_leave(mcast->dev, mcast);
 909                 ipoib_mcast_free(mcast);
 910         }
 911
 912         if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 913                 ipoib_mcast_start_thread(dev);
 914 }
 915
 916 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 917
 918 struct ipoib_mcast_iter *ipoib_mcast_iter_init(struct net_device *dev)
 919 {
 920         struct ipoib_mcast_iter *iter;
 921
 922         iter = kmalloc(sizeof *iter, GFP_KERNEL);
 923         if (!iter)
 924                 return NULL;
 925
 926         iter->dev = dev;
 927         memset(iter->mgid.raw, 0, 16);
 928
 929         if (ipoib_mcast_iter_next(iter)) {
 930                 kfree(iter);
 931                 return NULL;
 932         }
 933
 934         return iter;
 935 }
 936
 937 int ipoib_mcast_iter_next(struct ipoib_mcast_iter *iter)
 938 {
 939         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
 940         struct rb_node *n;
 941         struct ipoib_mcast *mcast;
 942         int ret = 1;
 943
 944         spin_lock_irq(&priv->lock);
 945
 946         n = rb_first(&priv->multicast_tree);
 947
 948         while (n) {
 949                 mcast = rb_entry(n, struct ipoib_mcast, rb_node);
 950
 951                 if (memcmp(iter->mgid.raw, mcast->mcmember.mgid.raw,
 952                            sizeof (union ib_gid)) < 0) {
 953                         iter->mgid      = mcast->mcmember.mgid;
 954                         iter->created   = mcast->created;
 955                         iter->queuelen  = skb_queue_len(&mcast->pkt_queue);
 956                         iter->complete  = !!mcast->ah;
 957                         iter->send_only = !!(mcast->flags & (1 << IPOIB_MCAST_FLAG_SENDONLY));
 958
 959                         ret = 0;
 960
 961                         break;
 962                 }
 963
 964                 n = rb_next(n);
 965         }
 966
 967         spin_unlock_irq(&priv->lock);
 968
 969         return ret;
 970 }
 971
 972 void ipoib_mcast_iter_read(struct ipoib_mcast_iter *iter,
 973                            union ib_gid *mgid,
 974                            unsigned long *created,
 975                            unsigned int *queuelen,
 976                            unsigned int *complete,
 977                            unsigned int *send_only)
 978 {
 979         *mgid      = iter->mgid;
 980         *created   = iter->created;
 981         *queuelen  = iter->queuelen;
 982         *complete  = iter->complete;
 983         *send_only = iter->send_only;
 984 }
 985
 986 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */