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
  35 #include <linux/skbuff.h>
  36 #include <linux/rtnetlink.h>
  37 #include <linux/moduleparam.h>
  38 #include <linux/ip.h>
  39 #include <linux/in.h>
  40 #include <linux/igmp.h>
  41 #include <linux/inetdevice.h>
  42 #include <linux/delay.h>
  43 #include <linux/completion.h>
  44 #include <linux/slab.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 struct ipoib_mcast_iter {
  61         struct net_device *dev;
  62         union ib_gid       mgid;
  63         unsigned long      created;
  64         unsigned int       queuelen;
  65         unsigned int       complete;
  66         unsigned int       send_only;
  67 };
  68
  69 static void ipoib_mcast_free(struct ipoib_mcast *mcast)
  70 {
  71         struct net_device *dev = mcast->dev;
  72         int tx_dropped = 0;
  73
  74         ipoib_dbg_mcast(netdev_priv(dev), "deleting multicast group %pI6\n",
  75                         mcast->mcmember.mgid.raw);
  76
  77         /* remove all neigh connected to this mcast */
  78         ipoib_del_neighs_by_gid(dev, mcast->mcmember.mgid.raw);
  79
  80         if (mcast->ah)
  81                 ipoib_put_ah(mcast->ah);
  82
  83         while (!skb_queue_empty(&mcast->pkt_queue)) {
  84                 ++tx_dropped;
  85                 dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
  86         }
  87
  88         netif_tx_lock_bh(dev);
  89         dev->stats.tx_dropped += tx_dropped;
  90         netif_tx_unlock_bh(dev);
  91
  92         kfree(mcast);
  93 }
  94
  95 static struct ipoib_mcast *ipoib_mcast_alloc(struct net_device *dev,
  96                                              int can_sleep)
  97 {
  98         struct ipoib_mcast *mcast;
  99
 100         mcast = kzalloc(sizeof *mcast, can_sleep ? GFP_KERNEL : GFP_ATOMIC);
 101         if (!mcast)
 102                 return NULL;
 103
 104         mcast->dev = dev;
 105         mcast->created = jiffies;
 106         mcast->backoff = 1;
 107
 108         INIT_LIST_HEAD(&mcast->list);
 109         INIT_LIST_HEAD(&mcast->neigh_list);
 110         skb_queue_head_init(&mcast->pkt_queue);
 111
 112         return mcast;
 113 }
 114
 115 static struct ipoib_mcast *__ipoib_mcast_find(struct net_device *dev, void *mgid)
 116 {
 117         struct ipoib_dev_priv *priv = netdev_priv(dev);
 118         struct rb_node *n = priv->multicast_tree.rb_node;
 119
 120         while (n) {
 121                 struct ipoib_mcast *mcast;
 122                 int ret;
 123
 124                 mcast = rb_entry(n, struct ipoib_mcast, rb_node);
 125
 126                 ret = memcmp(mgid, mcast->mcmember.mgid.raw,
 127                              sizeof (union ib_gid));
 128                 if (ret < 0)
 129                         n = n->rb_left;
 130                 else if (ret > 0)
 131                         n = n->rb_right;
 132                 else
 133                         return mcast;
 134         }
 135
 136         return NULL;
 137 }
 138
 139 static int __ipoib_mcast_add(struct net_device *dev, struct ipoib_mcast *mcast)
 140 {
 141         struct ipoib_dev_priv *priv = netdev_priv(dev);
 142         struct rb_node **n = &priv->multicast_tree.rb_node, *pn = NULL;
 143
 144         while (*n) {
 145                 struct ipoib_mcast *tmcast;
 146                 int ret;
 147
 148                 pn = *n;
 149                 tmcast = rb_entry(pn, struct ipoib_mcast, rb_node);
 150
 151                 ret = memcmp(mcast->mcmember.mgid.raw, tmcast->mcmember.mgid.raw,
 152                              sizeof (union ib_gid));
 153                 if (ret < 0)
 154                         n = &pn->rb_left;
 155                 else if (ret > 0)
 156                         n = &pn->rb_right;
 157                 else
 158                         return -EEXIST;
 159         }
 160
 161         rb_link_node(&mcast->rb_node, pn, n);
 162         rb_insert_color(&mcast->rb_node, &priv->multicast_tree);
 163
 164         return 0;
 165 }
 166
 167 static int ipoib_mcast_join_finish(struct ipoib_mcast *mcast,
 168                                    struct ib_sa_mcmember_rec *mcmember)
 169 {
 170         struct net_device *dev = mcast->dev;
 171         struct ipoib_dev_priv *priv = netdev_priv(dev);
 172         struct ipoib_ah *ah;
 173         int ret;
 174         int set_qkey = 0;
 175
 176         mcast->mcmember = *mcmember;
 177
 178         /* Set the multicast MTU and cached Q_Key before we attach if it's
 179          * the broadcast group.
 180          */
 181         if (!memcmp(mcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
 182                     sizeof (union ib_gid))) {
 183                 spin_lock_irq(&priv->lock);
 184                 if (!priv->broadcast) {
 185                         spin_unlock_irq(&priv->lock);
 186                         return -EAGAIN;
 187                 }
 188                 priv->mcast_mtu = IPOIB_UD_MTU(ib_mtu_enum_to_int(priv->broadcast->mcmember.mtu));
 189                 priv->qkey = be32_to_cpu(priv->broadcast->mcmember.qkey);
 190                 spin_unlock_irq(&priv->lock);
 191                 priv->tx_wr.wr.ud.remote_qkey = priv->qkey;
 192                 set_qkey = 1;
 193
 194                 if (!ipoib_cm_admin_enabled(dev)) {
 195                         rtnl_lock();
 196                         dev_set_mtu(dev, min(priv->mcast_mtu, priv->admin_mtu));
 197                         rtnl_unlock();
 198                 }
 199         }
 200
 201         if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 202                 if (test_and_set_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 203                         ipoib_warn(priv, "multicast group %pI6 already attached\n",
 204                                    mcast->mcmember.mgid.raw);
 205
 206                         return 0;
 207                 }
 208
 209                 ret = ipoib_mcast_attach(dev, be16_to_cpu(mcast->mcmember.mlid),
 210                                          &mcast->mcmember.mgid, set_qkey);
 211                 if (ret < 0) {
 212                         ipoib_warn(priv, "couldn't attach QP to multicast group %pI6\n",
 213                                    mcast->mcmember.mgid.raw);
 214
 215                         clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags);
 216                         return ret;
 217                 }
 218         }
 219
 220         {
 221                 struct ib_ah_attr av = {
 222                         .dlid          = be16_to_cpu(mcast->mcmember.mlid),
 223                         .port_num      = priv->port,
 224                         .sl            = mcast->mcmember.sl,
 225                         .ah_flags      = IB_AH_GRH,
 226                         .static_rate   = mcast->mcmember.rate,
 227                         .grh           = {
 228                                 .flow_label    = be32_to_cpu(mcast->mcmember.flow_label),
 229                                 .hop_limit     = mcast->mcmember.hop_limit,
 230                                 .sgid_index    = 0,
 231                                 .traffic_class = mcast->mcmember.traffic_class
 232                         }
 233                 };
 234                 av.grh.dgid = mcast->mcmember.mgid;
 235
 236                 ah = ipoib_create_ah(dev, priv->pd, &av);
 237                 if (IS_ERR(ah)) {
 238                         ipoib_warn(priv, "ib_address_create failed %ld\n",
 239                                 -PTR_ERR(ah));
 240                         /* use original error */
 241                         return PTR_ERR(ah);
 242                 } else {
 243                         spin_lock_irq(&priv->lock);
 244                         mcast->ah = ah;
 245                         spin_unlock_irq(&priv->lock);
 246
 247                         ipoib_dbg_mcast(priv, "MGID %pI6 AV %p, LID 0x%04x, SL %d\n",
 248                                         mcast->mcmember.mgid.raw,
 249                                         mcast->ah->ah,
 250                                         be16_to_cpu(mcast->mcmember.mlid),
 251                                         mcast->mcmember.sl);
 252                 }
 253         }
 254
 255         /* actually send any queued packets */
 256         netif_tx_lock_bh(dev);
 257         while (!skb_queue_empty(&mcast->pkt_queue)) {
 258                 struct sk_buff *skb = skb_dequeue(&mcast->pkt_queue);
 259
 260                 netif_tx_unlock_bh(dev);
 261
 262                 skb->dev = dev;
 263                 if (dev_queue_xmit(skb))
 264                         ipoib_warn(priv, "dev_queue_xmit failed to requeue packet\n");
 265
 266                 netif_tx_lock_bh(dev);
 267         }
 268         netif_tx_unlock_bh(dev);
 269
 270         return 0;
 271 }
 272
 273 static int
 274 ipoib_mcast_sendonly_join_complete(int status,
 275                                    struct ib_sa_multicast *multicast)
 276 {
 277         struct ipoib_mcast *mcast = multicast->context;
 278         struct net_device *dev = mcast->dev;
 279
 280         /* We trap for port events ourselves. */
 281         if (status == -ENETRESET)
 282                 return 0;
 283
 284         if (!status)
 285                 status = ipoib_mcast_join_finish(mcast, &multicast->rec);
 286
 287         if (status) {
 288                 if (mcast->logcount++ < 20)
 289                         ipoib_dbg_mcast(netdev_priv(dev), "multicast join failed for %pI6, status %d\n",
 290                                         mcast->mcmember.mgid.raw, status);
 291
 292                 /* Flush out any queued packets */
 293                 netif_tx_lock_bh(dev);
 294                 while (!skb_queue_empty(&mcast->pkt_queue)) {
 295                         ++dev->stats.tx_dropped;
 296                         dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
 297                 }
 298                 netif_tx_unlock_bh(dev);
 299
 300                 /* Clear the busy flag so we try again */
 301                 status = test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY,
 302                                             &mcast->flags);
 303         }
 304         return status;
 305 }
 306
 307 static int ipoib_mcast_sendonly_join(struct ipoib_mcast *mcast)
 308 {
 309         struct net_device *dev = mcast->dev;
 310         struct ipoib_dev_priv *priv = netdev_priv(dev);
 311         struct ib_sa_mcmember_rec rec = {
 312 #if 0                           /* Some SMs don't support send-only yet */
 313                 .join_state = 4
 314 #else
 315                 .join_state = 1
 316 #endif
 317         };
 318         int ret = 0;
 319
 320         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
 321                 ipoib_dbg_mcast(priv, "device shutting down, no multicast joins\n");
 322                 return -ENODEV;
 323         }
 324
 325         if (test_and_set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)) {
 326                 ipoib_dbg_mcast(priv, "multicast entry busy, skipping\n");
 327                 return -EBUSY;
 328         }
 329
 330         rec.mgid     = mcast->mcmember.mgid;
 331         rec.port_gid = priv->local_gid;
 332         rec.pkey     = cpu_to_be16(priv->pkey);
 333
 334         mcast->mc = ib_sa_join_multicast(&ipoib_sa_client, priv->ca,
 335                                          priv->port, &rec,
 336                                          IB_SA_MCMEMBER_REC_MGID        |
 337                                          IB_SA_MCMEMBER_REC_PORT_GID    |
 338                                          IB_SA_MCMEMBER_REC_PKEY        |
 339                                          IB_SA_MCMEMBER_REC_JOIN_STATE,
 340                                          GFP_ATOMIC,
 341                                          ipoib_mcast_sendonly_join_complete,
 342                                          mcast);
 343         if (IS_ERR(mcast->mc)) {
 344                 ret = PTR_ERR(mcast->mc);
 345                 clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 346                 ipoib_warn(priv, "ib_sa_join_multicast failed (ret = %d)\n",
 347                            ret);
 348         } else {
 349                 ipoib_dbg_mcast(priv, "no multicast record for %pI6, starting join\n",
 350                                 mcast->mcmember.mgid.raw);
 351         }
 352
 353         return ret;
 354 }
 355
 356 void ipoib_mcast_carrier_on_task(struct work_struct *work)
 357 {
 358         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
 359                                                    carrier_on_task);
 360         struct ib_port_attr attr;
 361
 362         /*
 363          * Take rtnl_lock to avoid racing with ipoib_stop() and
 364          * turning the carrier back on while a device is being
 365          * removed.
 366          */
 367         if (ib_query_port(priv->ca, priv->port, &attr) ||
 368             attr.state != IB_PORT_ACTIVE) {
 369                 ipoib_dbg(priv, "Keeping carrier off until IB port is active\n");
 370                 return;
 371         }
 372
 373         rtnl_lock();
 374         netif_carrier_on(priv->dev);
 375         rtnl_unlock();
 376 }
 377
 378 static int ipoib_mcast_join_complete(int status,
 379                                      struct ib_sa_multicast *multicast)
 380 {
 381         struct ipoib_mcast *mcast = multicast->context;
 382         struct net_device *dev = mcast->dev;
 383         struct ipoib_dev_priv *priv = netdev_priv(dev);
 384
 385         ipoib_dbg_mcast(priv, "join completion for %pI6 (status %d)\n",
 386                         mcast->mcmember.mgid.raw, status);
 387
 388         /* We trap for port events ourselves. */
 389         if (status == -ENETRESET) {
 390                 status = 0;
 391                 goto out;
 392         }
 393
 394         if (!status)
 395                 status = ipoib_mcast_join_finish(mcast, &multicast->rec);
 396
 397         if (!status) {
 398                 mcast->backoff = 1;
 399                 mutex_lock(&mcast_mutex);
 400                 if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 401                         queue_delayed_work(ipoib_workqueue,
 402                                            &priv->mcast_task, 0);
 403                 mutex_unlock(&mcast_mutex);
 404
 405                 /*
 406                  * Defer carrier on work to ipoib_workqueue to avoid a
 407                  * deadlock on rtnl_lock here.
 408                  */
 409                 if (mcast == priv->broadcast)
 410                         queue_work(ipoib_workqueue, &priv->carrier_on_task);
 411
 412                 status = 0;
 413                 goto out;
 414         }
 415
 416         if (mcast->logcount++ < 20) {
 417                 if (status == -ETIMEDOUT || status == -EAGAIN) {
 418                         ipoib_dbg_mcast(priv, "multicast join failed for %pI6, status %d\n",
 419                                         mcast->mcmember.mgid.raw, status);
 420                 } else {
 421                         ipoib_warn(priv, "multicast join failed for %pI6, status %d\n",
 422                                    mcast->mcmember.mgid.raw, status);
 423                 }
 424         }
 425
 426         mcast->backoff *= 2;
 427         if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
 428                 mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
 429
 430         /* Clear the busy flag so we try again */
 431         status = test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 432
 433         mutex_lock(&mcast_mutex);
 434         spin_lock_irq(&priv->lock);
 435         if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 436                 queue_delayed_work(ipoib_workqueue, &priv->mcast_task,
 437                                    mcast->backoff * HZ);
 438         spin_unlock_irq(&priv->lock);
 439         mutex_unlock(&mcast_mutex);
 440 out:
 441         complete(&mcast->done);
 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 %pI6\n", mcast->mcmember.mgid.raw);
 456
 457         rec.mgid     = mcast->mcmember.mgid;
 458         rec.port_gid = priv->local_gid;
 459         rec.pkey     = cpu_to_be16(priv->pkey);
 460
 461         comp_mask =
 462                 IB_SA_MCMEMBER_REC_MGID         |
 463                 IB_SA_MCMEMBER_REC_PORT_GID     |
 464                 IB_SA_MCMEMBER_REC_PKEY         |
 465                 IB_SA_MCMEMBER_REC_JOIN_STATE;
 466
 467         if (create) {
 468                 comp_mask |=
 469                         IB_SA_MCMEMBER_REC_QKEY                 |
 470                         IB_SA_MCMEMBER_REC_MTU_SELECTOR         |
 471                         IB_SA_MCMEMBER_REC_MTU                  |
 472                         IB_SA_MCMEMBER_REC_TRAFFIC_CLASS        |
 473                         IB_SA_MCMEMBER_REC_RATE_SELECTOR        |
 474                         IB_SA_MCMEMBER_REC_RATE                 |
 475                         IB_SA_MCMEMBER_REC_SL                   |
 476                         IB_SA_MCMEMBER_REC_FLOW_LABEL           |
 477                         IB_SA_MCMEMBER_REC_HOP_LIMIT;
 478
 479                 rec.qkey          = priv->broadcast->mcmember.qkey;
 480                 rec.mtu_selector  = IB_SA_EQ;
 481                 rec.mtu           = priv->broadcast->mcmember.mtu;
 482                 rec.traffic_class = priv->broadcast->mcmember.traffic_class;
 483                 rec.rate_selector = IB_SA_EQ;
 484                 rec.rate          = priv->broadcast->mcmember.rate;
 485                 rec.sl            = priv->broadcast->mcmember.sl;
 486                 rec.flow_label    = priv->broadcast->mcmember.flow_label;
 487                 rec.hop_limit     = priv->broadcast->mcmember.hop_limit;
 488         }
 489
 490         set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 491         init_completion(&mcast->done);
 492         set_bit(IPOIB_MCAST_JOIN_STARTED, &mcast->flags);
 493
 494         mcast->mc = ib_sa_join_multicast(&ipoib_sa_client, priv->ca, priv->port,
 495                                          &rec, comp_mask, GFP_KERNEL,
 496                                          ipoib_mcast_join_complete, mcast);
 497         if (IS_ERR(mcast->mc)) {
 498                 clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 499                 complete(&mcast->done);
 500                 ret = PTR_ERR(mcast->mc);
 501                 ipoib_warn(priv, "ib_sa_join_multicast failed, status %d\n", ret);
 502
 503                 mcast->backoff *= 2;
 504                 if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
 505                         mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
 506
 507                 mutex_lock(&mcast_mutex);
 508                 if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 509                         queue_delayed_work(ipoib_workqueue,
 510                                            &priv->mcast_task,
 511                                            mcast->backoff * HZ);
 512                 mutex_unlock(&mcast_mutex);
 513         }
 514 }
 515
 516 void ipoib_mcast_join_task(struct work_struct *work)
 517 {
 518         struct ipoib_dev_priv *priv =
 519                 container_of(work, struct ipoib_dev_priv, mcast_task.work);
 520         struct net_device *dev = priv->dev;
 521
 522         if (!test_bit(IPOIB_MCAST_RUN, &priv->flags))
 523                 return;
 524
 525         if (ib_query_gid(priv->ca, priv->port, 0, &priv->local_gid))
 526                 ipoib_warn(priv, "ib_query_gid() failed\n");
 527         else
 528                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
 529
 530         {
 531                 struct ib_port_attr attr;
 532
 533                 if (!ib_query_port(priv->ca, priv->port, &attr))
 534                         priv->local_lid = attr.lid;
 535                 else
 536                         ipoib_warn(priv, "ib_query_port failed\n");
 537         }
 538
 539         if (!priv->broadcast) {
 540                 struct ipoib_mcast *broadcast;
 541
 542                 if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 543                         return;
 544
 545                 broadcast = ipoib_mcast_alloc(dev, 1);
 546                 if (!broadcast) {
 547                         ipoib_warn(priv, "failed to allocate broadcast group\n");
 548                         mutex_lock(&mcast_mutex);
 549                         if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 550                                 queue_delayed_work(ipoib_workqueue,
 551                                                    &priv->mcast_task, HZ);
 552                         mutex_unlock(&mcast_mutex);
 553                         return;
 554                 }
 555
 556                 spin_lock_irq(&priv->lock);
 557                 memcpy(broadcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
 558                        sizeof (union ib_gid));
 559                 priv->broadcast = broadcast;
 560
 561                 __ipoib_mcast_add(dev, priv->broadcast);
 562                 spin_unlock_irq(&priv->lock);
 563         }
 564
 565         if (!test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
 566                 if (!test_bit(IPOIB_MCAST_FLAG_BUSY, &priv->broadcast->flags))
 567                         ipoib_mcast_join(dev, priv->broadcast, 0);
 568                 return;
 569         }
 570
 571         while (1) {
 572                 struct ipoib_mcast *mcast = NULL;
 573
 574                 spin_lock_irq(&priv->lock);
 575                 list_for_each_entry(mcast, &priv->multicast_list, list) {
 576                         if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)
 577                             && !test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)
 578                             && !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 579                                 /* Found the next unjoined group */
 580                                 break;
 581                         }
 582                 }
 583                 spin_unlock_irq(&priv->lock);
 584
 585                 if (&mcast->list == &priv->multicast_list) {
 586                         /* All done */
 587                         break;
 588                 }
 589
 590                 ipoib_mcast_join(dev, mcast, 1);
 591                 return;
 592         }
 593
 594         ipoib_dbg_mcast(priv, "successfully joined all multicast groups\n");
 595
 596         clear_bit(IPOIB_MCAST_RUN, &priv->flags);
 597 }
 598
 599 int ipoib_mcast_start_thread(struct net_device *dev)
 600 {
 601         struct ipoib_dev_priv *priv = netdev_priv(dev);
 602
 603         ipoib_dbg_mcast(priv, "starting multicast thread\n");
 604
 605         mutex_lock(&mcast_mutex);
 606         if (!test_and_set_bit(IPOIB_MCAST_RUN, &priv->flags))
 607                 queue_delayed_work(ipoib_workqueue, &priv->mcast_task, 0);
 608         mutex_unlock(&mcast_mutex);
 609
 610         return 0;
 611 }
 612
 613 int ipoib_mcast_stop_thread(struct net_device *dev, int flush)
 614 {
 615         struct ipoib_dev_priv *priv = netdev_priv(dev);
 616
 617         ipoib_dbg_mcast(priv, "stopping multicast thread\n");
 618
 619         mutex_lock(&mcast_mutex);
 620         clear_bit(IPOIB_MCAST_RUN, &priv->flags);
 621         cancel_delayed_work(&priv->mcast_task);
 622         mutex_unlock(&mcast_mutex);
 623
 624         if (flush)
 625                 flush_workqueue(ipoib_workqueue);
 626
 627         return 0;
 628 }
 629
 630 static int ipoib_mcast_leave(struct net_device *dev, struct ipoib_mcast *mcast)
 631 {
 632         struct ipoib_dev_priv *priv = netdev_priv(dev);
 633         int ret = 0;
 634
 635         if (test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
 636                 ib_sa_free_multicast(mcast->mc);
 637
 638         if (test_and_clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 639                 ipoib_dbg_mcast(priv, "leaving MGID %pI6\n",
 640                                 mcast->mcmember.mgid.raw);
 641
 642                 /* Remove ourselves from the multicast group */
 643                 ret = ib_detach_mcast(priv->qp, &mcast->mcmember.mgid,
 644                                       be16_to_cpu(mcast->mcmember.mlid));
 645                 if (ret)
 646                         ipoib_warn(priv, "ib_detach_mcast failed (result = %d)\n", ret);
 647         }
 648
 649         return 0;
 650 }
 651
 652 void ipoib_mcast_send(struct net_device *dev, u8 *daddr, struct sk_buff *skb)
 653 {
 654         struct ipoib_dev_priv *priv = netdev_priv(dev);
 655         struct ipoib_mcast *mcast;
 656         unsigned long flags;
 657         void *mgid = daddr + 4;
 658
 659         spin_lock_irqsave(&priv->lock, flags);
 660
 661         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)         ||
 662             !priv->broadcast                                    ||
 663             !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
 664                 ++dev->stats.tx_dropped;
 665                 dev_kfree_skb_any(skb);
 666                 goto unlock;
 667         }
 668
 669         mcast = __ipoib_mcast_find(dev, mgid);
 670         if (!mcast) {
 671                 /* Let's create a new send only group now */
 672                 ipoib_dbg_mcast(priv, "setting up send only multicast group for %pI6\n",
 673                                 mgid);
 674
 675                 mcast = ipoib_mcast_alloc(dev, 0);
 676                 if (!mcast) {
 677                         ipoib_warn(priv, "unable to allocate memory for "
 678                                    "multicast structure\n");
 679                         ++dev->stats.tx_dropped;
 680                         dev_kfree_skb_any(skb);
 681                         goto out;
 682                 }
 683
 684                 set_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags);
 685                 memcpy(mcast->mcmember.mgid.raw, mgid, sizeof (union ib_gid));
 686                 __ipoib_mcast_add(dev, mcast);
 687                 list_add_tail(&mcast->list, &priv->multicast_list);
 688         }
 689
 690         if (!mcast->ah) {
 691                 if (skb_queue_len(&mcast->pkt_queue) < IPOIB_MAX_MCAST_QUEUE)
 692                         skb_queue_tail(&mcast->pkt_queue, skb);
 693                 else {
 694                         ++dev->stats.tx_dropped;
 695                         dev_kfree_skb_any(skb);
 696                 }
 697
 698                 if (test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
 699                         ipoib_dbg_mcast(priv, "no address vector, "
 700                                         "but multicast join already started\n");
 701                 else if (test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags))
 702                         ipoib_mcast_sendonly_join(mcast);
 703
 704                 /*
 705                  * If lookup completes between here and out:, don't
 706                  * want to send packet twice.
 707                  */
 708                 mcast = NULL;
 709         }
 710
 711 out:
 712         if (mcast && mcast->ah) {
 713                 struct ipoib_neigh *neigh;
 714
 715                 spin_unlock_irqrestore(&priv->lock, flags);
 716                 neigh = ipoib_neigh_get(dev, daddr);
 717                 spin_lock_irqsave(&priv->lock, flags);
 718                 if (!neigh) {
 719                         neigh = ipoib_neigh_alloc(daddr, dev);
 720                         if (neigh) {
 721                                 kref_get(&mcast->ah->ref);
 722                                 neigh->ah       = mcast->ah;
 723                                 list_add_tail(&neigh->list, &mcast->neigh_list);
 724                         }
 725                 }
 726                 spin_unlock_irqrestore(&priv->lock, flags);
 727                 ipoib_send(dev, skb, mcast->ah, IB_MULTICAST_QPN);
 728                 if (neigh)
 729                         ipoib_neigh_put(neigh);
 730                 return;
 731         }
 732
 733 unlock:
 734         spin_unlock_irqrestore(&priv->lock, flags);
 735 }
 736
 737 void ipoib_mcast_dev_flush(struct net_device *dev)
 738 {
 739         struct ipoib_dev_priv *priv = netdev_priv(dev);
 740         LIST_HEAD(remove_list);
 741         struct ipoib_mcast *mcast, *tmcast;
 742         unsigned long flags;
 743
 744         ipoib_dbg_mcast(priv, "flushing multicast list\n");
 745
 746         spin_lock_irqsave(&priv->lock, flags);
 747
 748         list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
 749                 list_del(&mcast->list);
 750                 rb_erase(&mcast->rb_node, &priv->multicast_tree);
 751                 list_add_tail(&mcast->list, &remove_list);
 752         }
 753
 754         if (priv->broadcast) {
 755                 rb_erase(&priv->broadcast->rb_node, &priv->multicast_tree);
 756                 list_add_tail(&priv->broadcast->list, &remove_list);
 757                 priv->broadcast = NULL;
 758         }
 759
 760         spin_unlock_irqrestore(&priv->lock, flags);
 761
 762         /* seperate between the wait to the leave*/
 763         list_for_each_entry_safe(mcast, tmcast, &remove_list, list)
 764                 if (test_bit(IPOIB_MCAST_JOIN_STARTED, &mcast->flags))
 765                         wait_for_completion(&mcast->done);
 766
 767         list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
 768                 ipoib_mcast_leave(dev, mcast);
 769                 ipoib_mcast_free(mcast);
 770         }
 771 }
 772
 773 static int ipoib_mcast_addr_is_valid(const u8 *addr, const u8 *broadcast)
 774 {
 775         /* reserved QPN, prefix, scope */
 776         if (memcmp(addr, broadcast, 6))
 777                 return 0;
 778         /* signature lower, pkey */
 779         if (memcmp(addr + 7, broadcast + 7, 3))
 780                 return 0;
 781         return 1;
 782 }
 783
 784 void ipoib_mcast_restart_task(struct work_struct *work)
 785 {
 786         struct ipoib_dev_priv *priv =
 787                 container_of(work, struct ipoib_dev_priv, restart_task);
 788         struct net_device *dev = priv->dev;
 789         struct netdev_hw_addr *ha;
 790         struct ipoib_mcast *mcast, *tmcast;
 791         LIST_HEAD(remove_list);
 792         unsigned long flags;
 793         struct ib_sa_mcmember_rec rec;
 794
 795         ipoib_dbg_mcast(priv, "restarting multicast task\n");
 796
 797         ipoib_mcast_stop_thread(dev, 0);
 798
 799         local_irq_save(flags);
 800         netif_addr_lock(dev);
 801         spin_lock(&priv->lock);
 802
 803         /*
 804          * Unfortunately, the networking core only gives us a list of all of
 805          * the multicast hardware addresses. We need to figure out which ones
 806          * are new and which ones have been removed
 807          */
 808
 809         /* Clear out the found flag */
 810         list_for_each_entry(mcast, &priv->multicast_list, list)
 811                 clear_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
 812
 813         /* Mark all of the entries that are found or don't exist */
 814         netdev_for_each_mc_addr(ha, dev) {
 815                 union ib_gid mgid;
 816
 817                 if (!ipoib_mcast_addr_is_valid(ha->addr, dev->broadcast))
 818                         continue;
 819
 820                 memcpy(mgid.raw, ha->addr + 4, sizeof mgid);
 821
 822                 mcast = __ipoib_mcast_find(dev, &mgid);
 823                 if (!mcast || test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 824                         struct ipoib_mcast *nmcast;
 825
 826                         /* ignore group which is directly joined by userspace */
 827                         if (test_bit(IPOIB_FLAG_UMCAST, &priv->flags) &&
 828                             !ib_sa_get_mcmember_rec(priv->ca, priv->port, &mgid, &rec)) {
 829                                 ipoib_dbg_mcast(priv, "ignoring multicast entry for mgid %pI6\n",
 830                                                 mgid.raw);
 831                                 continue;
 832                         }
 833
 834                         /* Not found or send-only group, let's add a new entry */
 835                         ipoib_dbg_mcast(priv, "adding multicast entry for mgid %pI6\n",
 836                                         mgid.raw);
 837
 838                         nmcast = ipoib_mcast_alloc(dev, 0);
 839                         if (!nmcast) {
 840                                 ipoib_warn(priv, "unable to allocate memory for multicast structure\n");
 841                                 continue;
 842                         }
 843
 844                         set_bit(IPOIB_MCAST_FLAG_FOUND, &nmcast->flags);
 845
 846                         nmcast->mcmember.mgid = mgid;
 847
 848                         if (mcast) {
 849                                 /* Destroy the send only entry */
 850                                 list_move_tail(&mcast->list, &remove_list);
 851
 852                                 rb_replace_node(&mcast->rb_node,
 853                                                 &nmcast->rb_node,
 854                                                 &priv->multicast_tree);
 855                         } else
 856                                 __ipoib_mcast_add(dev, nmcast);
 857
 858                         list_add_tail(&nmcast->list, &priv->multicast_list);
 859                 }
 860
 861                 if (mcast)
 862                         set_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
 863         }
 864
 865         /* Remove all of the entries don't exist anymore */
 866         list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
 867                 if (!test_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags) &&
 868                     !test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 869                         ipoib_dbg_mcast(priv, "deleting multicast group %pI6\n",
 870                                         mcast->mcmember.mgid.raw);
 871
 872                         rb_erase(&mcast->rb_node, &priv->multicast_tree);
 873
 874                         /* Move to the remove list */
 875                         list_move_tail(&mcast->list, &remove_list);
 876                 }
 877         }
 878
 879         spin_unlock(&priv->lock);
 880         netif_addr_unlock(dev);
 881         local_irq_restore(flags);
 882
 883         /* We have to cancel outside of the spinlock */
 884         list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
 885                 ipoib_mcast_leave(mcast->dev, mcast);
 886                 ipoib_mcast_free(mcast);
 887         }
 888
 889         if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 890                 ipoib_mcast_start_thread(dev);
 891 }
 892
 893 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 894
 895 struct ipoib_mcast_iter *ipoib_mcast_iter_init(struct net_device *dev)
 896 {
 897         struct ipoib_mcast_iter *iter;
 898
 899         iter = kmalloc(sizeof *iter, GFP_KERNEL);
 900         if (!iter)
 901                 return NULL;
 902
 903         iter->dev = dev;
 904         memset(iter->mgid.raw, 0, 16);
 905
 906         if (ipoib_mcast_iter_next(iter)) {
 907                 kfree(iter);
 908                 return NULL;
 909         }
 910
 911         return iter;
 912 }
 913
 914 int ipoib_mcast_iter_next(struct ipoib_mcast_iter *iter)
 915 {
 916         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
 917         struct rb_node *n;
 918         struct ipoib_mcast *mcast;
 919         int ret = 1;
 920
 921         spin_lock_irq(&priv->lock);
 922
 923         n = rb_first(&priv->multicast_tree);
 924
 925         while (n) {
 926                 mcast = rb_entry(n, struct ipoib_mcast, rb_node);
 927
 928                 if (memcmp(iter->mgid.raw, mcast->mcmember.mgid.raw,
 929                            sizeof (union ib_gid)) < 0) {
 930                         iter->mgid      = mcast->mcmember.mgid;
 931                         iter->created   = mcast->created;
 932                         iter->queuelen  = skb_queue_len(&mcast->pkt_queue);
 933                         iter->complete  = !!mcast->ah;
 934                         iter->send_only = !!(mcast->flags & (1 << IPOIB_MCAST_FLAG_SENDONLY));
 935
 936                         ret = 0;
 937
 938                         break;
 939                 }
 940
 941                 n = rb_next(n);
 942         }
 943
 944         spin_unlock_irq(&priv->lock);
 945
 946         return ret;
 947 }
 948
 949 void ipoib_mcast_iter_read(struct ipoib_mcast_iter *iter,
 950                            union ib_gid *mgid,
 951                            unsigned long *created,
 952                            unsigned int *queuelen,
 953                            unsigned int *complete,
 954                            unsigned int *send_only)
 955 {
 956         *mgid      = iter->mgid;
 957         *created   = iter->created;
 958         *queuelen  = iter->queuelen;
 959         *complete  = iter->complete;
 960         *send_only = iter->send_only;
 961 }
 962
 963 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */