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         struct ipoib_dev_priv *priv = netdev_priv(dev);
  73         struct ipoib_neigh *neigh, *tmp;
  74         int tx_dropped = 0;
  75
  76         ipoib_dbg_mcast(netdev_priv(dev), "deleting multicast group %pI6\n",
  77                         mcast->mcmember.mgid.raw);
  78
  79         spin_lock_irq(&priv->lock);
  80
  81         list_for_each_entry_safe(neigh, tmp, &mcast->neigh_list, list) {
  82                 /*
  83                  * It's safe to call ipoib_put_ah() inside priv->lock
  84                  * here, because we know that mcast->ah will always
  85                  * hold one more reference, so ipoib_put_ah() will
  86                  * never do more than decrement the ref count.
  87                  */
  88                 if (neigh->ah)
  89                         ipoib_put_ah(neigh->ah);
  90                 ipoib_neigh_free(dev, neigh);
  91         }
  92
  93         spin_unlock_irq(&priv->lock);
  94
  95         if (mcast->ah)
  96                 ipoib_put_ah(mcast->ah);
  97
  98         while (!skb_queue_empty(&mcast->pkt_queue)) {
  99                 ++tx_dropped;
 100                 dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
 101         }
 102
 103         netif_tx_lock_bh(dev);
 104         dev->stats.tx_dropped += tx_dropped;
 105         netif_tx_unlock_bh(dev);
 106
 107         kfree(mcast);
 108 }
 109
 110 static struct ipoib_mcast *ipoib_mcast_alloc(struct net_device *dev,
 111                                              int can_sleep)
 112 {
 113         struct ipoib_mcast *mcast;
 114
 115         mcast = kzalloc(sizeof *mcast, can_sleep ? GFP_KERNEL : GFP_ATOMIC);
 116         if (!mcast)
 117                 return NULL;
 118
 119         mcast->dev = dev;
 120         mcast->created = jiffies;
 121         mcast->backoff = 1;
 122
 123         INIT_LIST_HEAD(&mcast->list);
 124         INIT_LIST_HEAD(&mcast->neigh_list);
 125         skb_queue_head_init(&mcast->pkt_queue);
 126
 127         return mcast;
 128 }
 129
 130 static struct ipoib_mcast *__ipoib_mcast_find(struct net_device *dev, void *mgid)
 131 {
 132         struct ipoib_dev_priv *priv = netdev_priv(dev);
 133         struct rb_node *n = priv->multicast_tree.rb_node;
 134
 135         while (n) {
 136                 struct ipoib_mcast *mcast;
 137                 int ret;
 138
 139                 mcast = rb_entry(n, struct ipoib_mcast, rb_node);
 140
 141                 ret = memcmp(mgid, mcast->mcmember.mgid.raw,
 142                              sizeof (union ib_gid));
 143                 if (ret < 0)
 144                         n = n->rb_left;
 145                 else if (ret > 0)
 146                         n = n->rb_right;
 147                 else
 148                         return mcast;
 149         }
 150
 151         return NULL;
 152 }
 153
 154 static int __ipoib_mcast_add(struct net_device *dev, struct ipoib_mcast *mcast)
 155 {
 156         struct ipoib_dev_priv *priv = netdev_priv(dev);
 157         struct rb_node **n = &priv->multicast_tree.rb_node, *pn = NULL;
 158
 159         while (*n) {
 160                 struct ipoib_mcast *tmcast;
 161                 int ret;
 162
 163                 pn = *n;
 164                 tmcast = rb_entry(pn, struct ipoib_mcast, rb_node);
 165
 166                 ret = memcmp(mcast->mcmember.mgid.raw, tmcast->mcmember.mgid.raw,
 167                              sizeof (union ib_gid));
 168                 if (ret < 0)
 169                         n = &pn->rb_left;
 170                 else if (ret > 0)
 171                         n = &pn->rb_right;
 172                 else
 173                         return -EEXIST;
 174         }
 175
 176         rb_link_node(&mcast->rb_node, pn, n);
 177         rb_insert_color(&mcast->rb_node, &priv->multicast_tree);
 178
 179         return 0;
 180 }
 181
 182 static int ipoib_mcast_join_finish(struct ipoib_mcast *mcast,
 183                                    struct ib_sa_mcmember_rec *mcmember)
 184 {
 185         struct net_device *dev = mcast->dev;
 186         struct ipoib_dev_priv *priv = netdev_priv(dev);
 187         struct ipoib_ah *ah;
 188         int ret;
 189         int set_qkey = 0;
 190
 191         mcast->mcmember = *mcmember;
 192
 193         /* Set the cached Q_Key before we attach if it's the broadcast group */
 194         if (!memcmp(mcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
 195                     sizeof (union ib_gid))) {
 196                 spin_lock_irq(&priv->lock);
 197                 if (!priv->broadcast) {
 198                         spin_unlock_irq(&priv->lock);
 199                         return -EAGAIN;
 200                 }
 201                 priv->qkey = be32_to_cpu(priv->broadcast->mcmember.qkey);
 202                 spin_unlock_irq(&priv->lock);
 203                 priv->tx_wr.wr.ud.remote_qkey = priv->qkey;
 204                 set_qkey = 1;
 205         }
 206
 207         if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 208                 if (test_and_set_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 209                         ipoib_warn(priv, "multicast group %pI6 already attached\n",
 210                                    mcast->mcmember.mgid.raw);
 211
 212                         return 0;
 213                 }
 214
 215                 ret = ipoib_mcast_attach(dev, be16_to_cpu(mcast->mcmember.mlid),
 216                                          &mcast->mcmember.mgid, set_qkey);
 217                 if (ret < 0) {
 218                         ipoib_warn(priv, "couldn't attach QP to multicast group %pI6\n",
 219                                    mcast->mcmember.mgid.raw);
 220
 221                         clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags);
 222                         return ret;
 223                 }
 224         }
 225
 226         {
 227                 struct ib_ah_attr av = {
 228                         .dlid          = be16_to_cpu(mcast->mcmember.mlid),
 229                         .port_num      = priv->port,
 230                         .sl            = mcast->mcmember.sl,
 231                         .ah_flags      = IB_AH_GRH,
 232                         .static_rate   = mcast->mcmember.rate,
 233                         .grh           = {
 234                                 .flow_label    = be32_to_cpu(mcast->mcmember.flow_label),
 235                                 .hop_limit     = mcast->mcmember.hop_limit,
 236                                 .sgid_index    = 0,
 237                                 .traffic_class = mcast->mcmember.traffic_class
 238                         }
 239                 };
 240                 av.grh.dgid = mcast->mcmember.mgid;
 241
 242                 ah = ipoib_create_ah(dev, priv->pd, &av);
 243                 if (!ah) {
 244                         ipoib_warn(priv, "ib_address_create failed\n");
 245                 } else {
 246                         spin_lock_irq(&priv->lock);
 247                         mcast->ah = ah;
 248                         spin_unlock_irq(&priv->lock);
 249
 250                         ipoib_dbg_mcast(priv, "MGID %pI6 AV %p, LID 0x%04x, SL %d\n",
 251                                         mcast->mcmember.mgid.raw,
 252                                         mcast->ah->ah,
 253                                         be16_to_cpu(mcast->mcmember.mlid),
 254                                         mcast->mcmember.sl);
 255                 }
 256         }
 257
 258         /* actually send any queued packets */
 259         netif_tx_lock_bh(dev);
 260         while (!skb_queue_empty(&mcast->pkt_queue)) {
 261                 struct sk_buff *skb = skb_dequeue(&mcast->pkt_queue);
 262                 struct dst_entry *dst = skb_dst(skb);
 263                 struct neighbour *n = NULL;
 264
 265                 netif_tx_unlock_bh(dev);
 266
 267                 skb->dev = dev;
 268                 if (dst)
 269                         n = dst_get_neighbour_raw(dst);
 270                 if (!dst || !n) {
 271                         /* put pseudoheader back on for next time */
 272                         skb_push(skb, sizeof (struct ipoib_pseudoheader));
 273                 }
 274
 275                 if (dev_queue_xmit(skb))
 276                         ipoib_warn(priv, "dev_queue_xmit failed to requeue packet\n");
 277                 netif_tx_lock_bh(dev);
 278         }
 279         netif_tx_unlock_bh(dev);
 280
 281         return 0;
 282 }
 283
 284 static int
 285 ipoib_mcast_sendonly_join_complete(int status,
 286                                    struct ib_sa_multicast *multicast)
 287 {
 288         struct ipoib_mcast *mcast = multicast->context;
 289         struct net_device *dev = mcast->dev;
 290
 291         /* We trap for port events ourselves. */
 292         if (status == -ENETRESET)
 293                 return 0;
 294
 295         if (!status)
 296                 status = ipoib_mcast_join_finish(mcast, &multicast->rec);
 297
 298         if (status) {
 299                 if (mcast->logcount++ < 20)
 300                         ipoib_dbg_mcast(netdev_priv(dev), "multicast join failed for %pI6, status %d\n",
 301                                         mcast->mcmember.mgid.raw, status);
 302
 303                 /* Flush out any queued packets */
 304                 netif_tx_lock_bh(dev);
 305                 while (!skb_queue_empty(&mcast->pkt_queue)) {
 306                         ++dev->stats.tx_dropped;
 307                         dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
 308                 }
 309                 netif_tx_unlock_bh(dev);
 310
 311                 /* Clear the busy flag so we try again */
 312                 status = test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY,
 313                                             &mcast->flags);
 314         }
 315         return status;
 316 }
 317
 318 static int ipoib_mcast_sendonly_join(struct ipoib_mcast *mcast)
 319 {
 320         struct net_device *dev = mcast->dev;
 321         struct ipoib_dev_priv *priv = netdev_priv(dev);
 322         struct ib_sa_mcmember_rec rec = {
 323 #if 0                           /* Some SMs don't support send-only yet */
 324                 .join_state = 4
 325 #else
 326                 .join_state = 1
 327 #endif
 328         };
 329         int ret = 0;
 330
 331         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
 332                 ipoib_dbg_mcast(priv, "device shutting down, no multicast joins\n");
 333                 return -ENODEV;
 334         }
 335
 336         if (test_and_set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)) {
 337                 ipoib_dbg_mcast(priv, "multicast entry busy, skipping\n");
 338                 return -EBUSY;
 339         }
 340
 341         rec.mgid     = mcast->mcmember.mgid;
 342         rec.port_gid = priv->local_gid;
 343         rec.pkey     = cpu_to_be16(priv->pkey);
 344
 345         mcast->mc = ib_sa_join_multicast(&ipoib_sa_client, priv->ca,
 346                                          priv->port, &rec,
 347                                          IB_SA_MCMEMBER_REC_MGID        |
 348                                          IB_SA_MCMEMBER_REC_PORT_GID    |
 349                                          IB_SA_MCMEMBER_REC_PKEY        |
 350                                          IB_SA_MCMEMBER_REC_JOIN_STATE,
 351                                          GFP_ATOMIC,
 352                                          ipoib_mcast_sendonly_join_complete,
 353                                          mcast);
 354         if (IS_ERR(mcast->mc)) {
 355                 ret = PTR_ERR(mcast->mc);
 356                 clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 357                 ipoib_warn(priv, "ib_sa_join_multicast failed (ret = %d)\n",
 358                            ret);
 359         } else {
 360                 ipoib_dbg_mcast(priv, "no multicast record for %pI6, starting join\n",
 361                                 mcast->mcmember.mgid.raw);
 362         }
 363
 364         return ret;
 365 }
 366
 367 void ipoib_mcast_carrier_on_task(struct work_struct *work)
 368 {
 369         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
 370                                                    carrier_on_task);
 371         struct ib_port_attr attr;
 372
 373         /*
 374          * Take rtnl_lock to avoid racing with ipoib_stop() and
 375          * turning the carrier back on while a device is being
 376          * removed.
 377          */
 378         if (ib_query_port(priv->ca, priv->port, &attr) ||
 379             attr.state != IB_PORT_ACTIVE) {
 380                 ipoib_dbg(priv, "Keeping carrier off until IB port is active\n");
 381                 return;
 382         }
 383
 384         rtnl_lock();
 385         netif_carrier_on(priv->dev);
 386         rtnl_unlock();
 387 }
 388
 389 static int ipoib_mcast_join_complete(int status,
 390                                      struct ib_sa_multicast *multicast)
 391 {
 392         struct ipoib_mcast *mcast = multicast->context;
 393         struct net_device *dev = mcast->dev;
 394         struct ipoib_dev_priv *priv = netdev_priv(dev);
 395
 396         ipoib_dbg_mcast(priv, "join completion for %pI6 (status %d)\n",
 397                         mcast->mcmember.mgid.raw, status);
 398
 399         /* We trap for port events ourselves. */
 400         if (status == -ENETRESET)
 401                 return 0;
 402
 403         if (!status)
 404                 status = ipoib_mcast_join_finish(mcast, &multicast->rec);
 405
 406         if (!status) {
 407                 mcast->backoff = 1;
 408                 mutex_lock(&mcast_mutex);
 409                 if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 410                         queue_delayed_work(ipoib_workqueue,
 411                                            &priv->mcast_task, 0);
 412                 mutex_unlock(&mcast_mutex);
 413
 414                 /*
 415                  * Defer carrier on work to ipoib_workqueue to avoid a
 416                  * deadlock on rtnl_lock here.
 417                  */
 418                 if (mcast == priv->broadcast)
 419                         queue_work(ipoib_workqueue, &priv->carrier_on_task);
 420
 421                 return 0;
 422         }
 423
 424         if (mcast->logcount++ < 20) {
 425                 if (status == -ETIMEDOUT || status == -EAGAIN) {
 426                         ipoib_dbg_mcast(priv, "multicast join failed for %pI6, status %d\n",
 427                                         mcast->mcmember.mgid.raw, status);
 428                 } else {
 429                         ipoib_warn(priv, "multicast join failed for %pI6, status %d\n",
 430                                    mcast->mcmember.mgid.raw, status);
 431                 }
 432         }
 433
 434         mcast->backoff *= 2;
 435         if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
 436                 mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
 437
 438         /* Clear the busy flag so we try again */
 439         status = test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 440
 441         mutex_lock(&mcast_mutex);
 442         spin_lock_irq(&priv->lock);
 443         if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 444                 queue_delayed_work(ipoib_workqueue, &priv->mcast_task,
 445                                    mcast->backoff * HZ);
 446         spin_unlock_irq(&priv->lock);
 447         mutex_unlock(&mcast_mutex);
 448
 449         return status;
 450 }
 451
 452 static void ipoib_mcast_join(struct net_device *dev, struct ipoib_mcast *mcast,
 453                              int create)
 454 {
 455         struct ipoib_dev_priv *priv = netdev_priv(dev);
 456         struct ib_sa_mcmember_rec rec = {
 457                 .join_state = 1
 458         };
 459         ib_sa_comp_mask comp_mask;
 460         int ret = 0;
 461
 462         ipoib_dbg_mcast(priv, "joining MGID %pI6\n", mcast->mcmember.mgid.raw);
 463
 464         rec.mgid     = mcast->mcmember.mgid;
 465         rec.port_gid = priv->local_gid;
 466         rec.pkey     = cpu_to_be16(priv->pkey);
 467
 468         comp_mask =
 469                 IB_SA_MCMEMBER_REC_MGID         |
 470                 IB_SA_MCMEMBER_REC_PORT_GID     |
 471                 IB_SA_MCMEMBER_REC_PKEY         |
 472                 IB_SA_MCMEMBER_REC_JOIN_STATE;
 473
 474         if (create) {
 475                 comp_mask |=
 476                         IB_SA_MCMEMBER_REC_QKEY                 |
 477                         IB_SA_MCMEMBER_REC_MTU_SELECTOR         |
 478                         IB_SA_MCMEMBER_REC_MTU                  |
 479                         IB_SA_MCMEMBER_REC_TRAFFIC_CLASS        |
 480                         IB_SA_MCMEMBER_REC_RATE_SELECTOR        |
 481                         IB_SA_MCMEMBER_REC_RATE                 |
 482                         IB_SA_MCMEMBER_REC_SL                   |
 483                         IB_SA_MCMEMBER_REC_FLOW_LABEL           |
 484                         IB_SA_MCMEMBER_REC_HOP_LIMIT;
 485
 486                 rec.qkey          = priv->broadcast->mcmember.qkey;
 487                 rec.mtu_selector  = IB_SA_EQ;
 488                 rec.mtu           = priv->broadcast->mcmember.mtu;
 489                 rec.traffic_class = priv->broadcast->mcmember.traffic_class;
 490                 rec.rate_selector = IB_SA_EQ;
 491                 rec.rate          = priv->broadcast->mcmember.rate;
 492                 rec.sl            = priv->broadcast->mcmember.sl;
 493                 rec.flow_label    = priv->broadcast->mcmember.flow_label;
 494                 rec.hop_limit     = priv->broadcast->mcmember.hop_limit;
 495         }
 496
 497         set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 498         mcast->mc = ib_sa_join_multicast(&ipoib_sa_client, priv->ca, priv->port,
 499                                          &rec, comp_mask, GFP_KERNEL,
 500                                          ipoib_mcast_join_complete, mcast);
 501         if (IS_ERR(mcast->mc)) {
 502                 clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 503                 ret = PTR_ERR(mcast->mc);
 504                 ipoib_warn(priv, "ib_sa_join_multicast failed, status %d\n", ret);
 505
 506                 mcast->backoff *= 2;
 507                 if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
 508                         mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
 509
 510                 mutex_lock(&mcast_mutex);
 511                 if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 512                         queue_delayed_work(ipoib_workqueue,
 513                                            &priv->mcast_task,
 514                                            mcast->backoff * HZ);
 515                 mutex_unlock(&mcast_mutex);
 516         }
 517 }
 518
 519 void ipoib_mcast_join_task(struct work_struct *work)
 520 {
 521         struct ipoib_dev_priv *priv =
 522                 container_of(work, struct ipoib_dev_priv, mcast_task.work);
 523         struct net_device *dev = priv->dev;
 524
 525         if (!test_bit(IPOIB_MCAST_RUN, &priv->flags))
 526                 return;
 527
 528         if (ib_query_gid(priv->ca, priv->port, 0, &priv->local_gid))
 529                 ipoib_warn(priv, "ib_query_gid() failed\n");
 530         else
 531                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
 532
 533         {
 534                 struct ib_port_attr attr;
 535
 536                 if (!ib_query_port(priv->ca, priv->port, &attr))
 537                         priv->local_lid = attr.lid;
 538                 else
 539                         ipoib_warn(priv, "ib_query_port failed\n");
 540         }
 541
 542         if (!priv->broadcast) {
 543                 struct ipoib_mcast *broadcast;
 544
 545                 if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 546                         return;
 547
 548                 broadcast = ipoib_mcast_alloc(dev, 1);
 549                 if (!broadcast) {
 550                         ipoib_warn(priv, "failed to allocate broadcast group\n");
 551                         mutex_lock(&mcast_mutex);
 552                         if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 553                                 queue_delayed_work(ipoib_workqueue,
 554                                                    &priv->mcast_task, HZ);
 555                         mutex_unlock(&mcast_mutex);
 556                         return;
 557                 }
 558
 559                 spin_lock_irq(&priv->lock);
 560                 memcpy(broadcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
 561                        sizeof (union ib_gid));
 562                 priv->broadcast = broadcast;
 563
 564                 __ipoib_mcast_add(dev, priv->broadcast);
 565                 spin_unlock_irq(&priv->lock);
 566         }
 567
 568         if (!test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
 569                 if (!test_bit(IPOIB_MCAST_FLAG_BUSY, &priv->broadcast->flags))
 570                         ipoib_mcast_join(dev, priv->broadcast, 0);
 571                 return;
 572         }
 573
 574         while (1) {
 575                 struct ipoib_mcast *mcast = NULL;
 576
 577                 spin_lock_irq(&priv->lock);
 578                 list_for_each_entry(mcast, &priv->multicast_list, list) {
 579                         if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)
 580                             && !test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)
 581                             && !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 582                                 /* Found the next unjoined group */
 583                                 break;
 584                         }
 585                 }
 586                 spin_unlock_irq(&priv->lock);
 587
 588                 if (&mcast->list == &priv->multicast_list) {
 589                         /* All done */
 590                         break;
 591                 }
 592
 593                 ipoib_mcast_join(dev, mcast, 1);
 594                 return;
 595         }
 596
 597         priv->mcast_mtu = IPOIB_UD_MTU(ib_mtu_enum_to_int(priv->broadcast->mcmember.mtu));
 598
 599         if (!ipoib_cm_admin_enabled(dev)) {
 600                 rtnl_lock();
 601                 dev_set_mtu(dev, min(priv->mcast_mtu, priv->admin_mtu));
 602                 rtnl_unlock();
 603         }
 604
 605         ipoib_dbg_mcast(priv, "successfully joined all multicast groups\n");
 606
 607         clear_bit(IPOIB_MCAST_RUN, &priv->flags);
 608 }
 609
 610 int ipoib_mcast_start_thread(struct net_device *dev)
 611 {
 612         struct ipoib_dev_priv *priv = netdev_priv(dev);
 613
 614         ipoib_dbg_mcast(priv, "starting multicast thread\n");
 615
 616         mutex_lock(&mcast_mutex);
 617         if (!test_and_set_bit(IPOIB_MCAST_RUN, &priv->flags))
 618                 queue_delayed_work(ipoib_workqueue, &priv->mcast_task, 0);
 619         mutex_unlock(&mcast_mutex);
 620
 621         return 0;
 622 }
 623
 624 int ipoib_mcast_stop_thread(struct net_device *dev, int flush)
 625 {
 626         struct ipoib_dev_priv *priv = netdev_priv(dev);
 627
 628         ipoib_dbg_mcast(priv, "stopping multicast thread\n");
 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         return 0;
 639 }
 640
 641 static int ipoib_mcast_leave(struct net_device *dev, struct ipoib_mcast *mcast)
 642 {
 643         struct ipoib_dev_priv *priv = netdev_priv(dev);
 644         int ret = 0;
 645
 646         if (test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
 647                 ib_sa_free_multicast(mcast->mc);
 648
 649         if (test_and_clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 650                 ipoib_dbg_mcast(priv, "leaving MGID %pI6\n",
 651                                 mcast->mcmember.mgid.raw);
 652
 653                 /* Remove ourselves from the multicast group */
 654                 ret = ib_detach_mcast(priv->qp, &mcast->mcmember.mgid,
 655                                       be16_to_cpu(mcast->mcmember.mlid));
 656                 if (ret)
 657                         ipoib_warn(priv, "ib_detach_mcast failed (result = %d)\n", ret);
 658         }
 659
 660         return 0;
 661 }
 662
 663 void ipoib_mcast_send(struct net_device *dev, void *mgid, struct sk_buff *skb)
 664 {
 665         struct ipoib_dev_priv *priv = netdev_priv(dev);
 666         struct ipoib_mcast *mcast;
 667         unsigned long flags;
 668
 669         spin_lock_irqsave(&priv->lock, flags);
 670
 671         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)         ||
 672             !priv->broadcast                                    ||
 673             !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
 674                 ++dev->stats.tx_dropped;
 675                 dev_kfree_skb_any(skb);
 676                 goto unlock;
 677         }
 678
 679         mcast = __ipoib_mcast_find(dev, mgid);
 680         if (!mcast) {
 681                 /* Let's create a new send only group now */
 682                 ipoib_dbg_mcast(priv, "setting up send only multicast group for %pI6\n",
 683                                 mgid);
 684
 685                 mcast = ipoib_mcast_alloc(dev, 0);
 686                 if (!mcast) {
 687                         ipoib_warn(priv, "unable to allocate memory for "
 688                                    "multicast structure\n");
 689                         ++dev->stats.tx_dropped;
 690                         dev_kfree_skb_any(skb);
 691                         goto out;
 692                 }
 693
 694                 set_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags);
 695                 memcpy(mcast->mcmember.mgid.raw, mgid, sizeof (union ib_gid));
 696                 __ipoib_mcast_add(dev, mcast);
 697                 list_add_tail(&mcast->list, &priv->multicast_list);
 698         }
 699
 700         if (!mcast->ah) {
 701                 if (skb_queue_len(&mcast->pkt_queue) < IPOIB_MAX_MCAST_QUEUE)
 702                         skb_queue_tail(&mcast->pkt_queue, skb);
 703                 else {
 704                         ++dev->stats.tx_dropped;
 705                         dev_kfree_skb_any(skb);
 706                 }
 707
 708                 if (test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
 709                         ipoib_dbg_mcast(priv, "no address vector, "
 710                                         "but multicast join already started\n");
 711                 else if (test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags))
 712                         ipoib_mcast_sendonly_join(mcast);
 713
 714                 /*
 715                  * If lookup completes between here and out:, don't
 716                  * want to send packet twice.
 717                  */
 718                 mcast = NULL;
 719         }
 720
 721 out:
 722         if (mcast && mcast->ah) {
 723                 struct dst_entry *dst = skb_dst(skb);
 724                 struct neighbour *n = NULL;
 725
 726                 rcu_read_lock();
 727                 if (dst)
 728                         n = dst_get_neighbour(dst);
 729                 if (n && !*to_ipoib_neigh(n)) {
 730                         struct ipoib_neigh *neigh = ipoib_neigh_alloc(n,
 731                                                                       skb->dev);
 732
 733                         if (neigh) {
 734                                 kref_get(&mcast->ah->ref);
 735                                 neigh->ah       = mcast->ah;
 736                                 list_add_tail(&neigh->list, &mcast->neigh_list);
 737                         }
 738                 }
 739                 rcu_read_unlock();
 740                 spin_unlock_irqrestore(&priv->lock, flags);
 741                 ipoib_send(dev, skb, mcast->ah, IB_MULTICAST_QPN);
 742                 return;
 743         }
 744
 745 unlock:
 746         spin_unlock_irqrestore(&priv->lock, flags);
 747 }
 748
 749 void ipoib_mcast_dev_flush(struct net_device *dev)
 750 {
 751         struct ipoib_dev_priv *priv = netdev_priv(dev);
 752         LIST_HEAD(remove_list);
 753         struct ipoib_mcast *mcast, *tmcast;
 754         unsigned long flags;
 755
 756         ipoib_dbg_mcast(priv, "flushing multicast list\n");
 757
 758         spin_lock_irqsave(&priv->lock, flags);
 759
 760         list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
 761                 list_del(&mcast->list);
 762                 rb_erase(&mcast->rb_node, &priv->multicast_tree);
 763                 list_add_tail(&mcast->list, &remove_list);
 764         }
 765
 766         if (priv->broadcast) {
 767                 rb_erase(&priv->broadcast->rb_node, &priv->multicast_tree);
 768                 list_add_tail(&priv->broadcast->list, &remove_list);
 769                 priv->broadcast = NULL;
 770         }
 771
 772         spin_unlock_irqrestore(&priv->lock, flags);
 773
 774         list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
 775                 ipoib_mcast_leave(dev, mcast);
 776                 ipoib_mcast_free(mcast);
 777         }
 778 }
 779
 780 static int ipoib_mcast_addr_is_valid(const u8 *addr, const u8 *broadcast)
 781 {
 782         /* reserved QPN, prefix, scope */
 783         if (memcmp(addr, broadcast, 6))
 784                 return 0;
 785         /* signature lower, pkey */
 786         if (memcmp(addr + 7, broadcast + 7, 3))
 787                 return 0;
 788         return 1;
 789 }
 790
 791 void ipoib_mcast_restart_task(struct work_struct *work)
 792 {
 793         struct ipoib_dev_priv *priv =
 794                 container_of(work, struct ipoib_dev_priv, restart_task);
 795         struct net_device *dev = priv->dev;
 796         struct netdev_hw_addr *ha;
 797         struct ipoib_mcast *mcast, *tmcast;
 798         LIST_HEAD(remove_list);
 799         unsigned long flags;
 800         struct ib_sa_mcmember_rec rec;
 801
 802         ipoib_dbg_mcast(priv, "restarting multicast task\n");
 803
 804         ipoib_mcast_stop_thread(dev, 0);
 805
 806         local_irq_save(flags);
 807         netif_addr_lock(dev);
 808         spin_lock(&priv->lock);
 809
 810         /*
 811          * Unfortunately, the networking core only gives us a list of all of
 812          * the multicast hardware addresses. We need to figure out which ones
 813          * are new and which ones have been removed
 814          */
 815
 816         /* Clear out the found flag */
 817         list_for_each_entry(mcast, &priv->multicast_list, list)
 818                 clear_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
 819
 820         /* Mark all of the entries that are found or don't exist */
 821         netdev_for_each_mc_addr(ha, dev) {
 822                 union ib_gid mgid;
 823
 824                 if (!ipoib_mcast_addr_is_valid(ha->addr, dev->broadcast))
 825                         continue;
 826
 827                 memcpy(mgid.raw, ha->addr + 4, sizeof mgid);
 828
 829                 mcast = __ipoib_mcast_find(dev, &mgid);
 830                 if (!mcast || test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 831                         struct ipoib_mcast *nmcast;
 832
 833                         /* ignore group which is directly joined by userspace */
 834                         if (test_bit(IPOIB_FLAG_UMCAST, &priv->flags) &&
 835                             !ib_sa_get_mcmember_rec(priv->ca, priv->port, &mgid, &rec)) {
 836                                 ipoib_dbg_mcast(priv, "ignoring multicast entry for mgid %pI6\n",
 837                                                 mgid.raw);
 838                                 continue;
 839                         }
 840
 841                         /* Not found or send-only group, let's add a new entry */
 842                         ipoib_dbg_mcast(priv, "adding multicast entry for mgid %pI6\n",
 843                                         mgid.raw);
 844
 845                         nmcast = ipoib_mcast_alloc(dev, 0);
 846                         if (!nmcast) {
 847                                 ipoib_warn(priv, "unable to allocate memory for multicast structure\n");
 848                                 continue;
 849                         }
 850
 851                         set_bit(IPOIB_MCAST_FLAG_FOUND, &nmcast->flags);
 852
 853                         nmcast->mcmember.mgid = mgid;
 854
 855                         if (mcast) {
 856                                 /* Destroy the send only entry */
 857                                 list_move_tail(&mcast->list, &remove_list);
 858
 859                                 rb_replace_node(&mcast->rb_node,
 860                                                 &nmcast->rb_node,
 861                                                 &priv->multicast_tree);
 862                         } else
 863                                 __ipoib_mcast_add(dev, nmcast);
 864
 865                         list_add_tail(&nmcast->list, &priv->multicast_list);
 866                 }
 867
 868                 if (mcast)
 869                         set_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
 870         }
 871
 872         /* Remove all of the entries don't exist anymore */
 873         list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
 874                 if (!test_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags) &&
 875                     !test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 876                         ipoib_dbg_mcast(priv, "deleting multicast group %pI6\n",
 877                                         mcast->mcmember.mgid.raw);
 878
 879                         rb_erase(&mcast->rb_node, &priv->multicast_tree);
 880
 881                         /* Move to the remove list */
 882                         list_move_tail(&mcast->list, &remove_list);
 883                 }
 884         }
 885
 886         spin_unlock(&priv->lock);
 887         netif_addr_unlock(dev);
 888         local_irq_restore(flags);
 889
 890         /* We have to cancel outside of the spinlock */
 891         list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
 892                 ipoib_mcast_leave(mcast->dev, mcast);
 893                 ipoib_mcast_free(mcast);
 894         }
 895
 896         if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 897                 ipoib_mcast_start_thread(dev);
 898 }
 899
 900 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 901
 902 struct ipoib_mcast_iter *ipoib_mcast_iter_init(struct net_device *dev)
 903 {
 904         struct ipoib_mcast_iter *iter;
 905
 906         iter = kmalloc(sizeof *iter, GFP_KERNEL);
 907         if (!iter)
 908                 return NULL;
 909
 910         iter->dev = dev;
 911         memset(iter->mgid.raw, 0, 16);
 912
 913         if (ipoib_mcast_iter_next(iter)) {
 914                 kfree(iter);
 915                 return NULL;
 916         }
 917
 918         return iter;
 919 }
 920
 921 int ipoib_mcast_iter_next(struct ipoib_mcast_iter *iter)
 922 {
 923         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
 924         struct rb_node *n;
 925         struct ipoib_mcast *mcast;
 926         int ret = 1;
 927
 928         spin_lock_irq(&priv->lock);
 929
 930         n = rb_first(&priv->multicast_tree);
 931
 932         while (n) {
 933                 mcast = rb_entry(n, struct ipoib_mcast, rb_node);
 934
 935                 if (memcmp(iter->mgid.raw, mcast->mcmember.mgid.raw,
 936                            sizeof (union ib_gid)) < 0) {
 937                         iter->mgid      = mcast->mcmember.mgid;
 938                         iter->created   = mcast->created;
 939                         iter->queuelen  = skb_queue_len(&mcast->pkt_queue);
 940                         iter->complete  = !!mcast->ah;
 941                         iter->send_only = !!(mcast->flags & (1 << IPOIB_MCAST_FLAG_SENDONLY));
 942
 943                         ret = 0;
 944
 945                         break;
 946                 }
 947
 948                 n = rb_next(n);
 949         }
 950
 951         spin_unlock_irq(&priv->lock);
 952
 953         return ret;
 954 }
 955
 956 void ipoib_mcast_iter_read(struct ipoib_mcast_iter *iter,
 957                            union ib_gid *mgid,
 958                            unsigned long *created,
 959                            unsigned int *queuelen,
 960                            unsigned int *complete,
 961                            unsigned int *send_only)
 962 {
 963         *mgid      = iter->mgid;
 964         *created   = iter->created;
 965         *queuelen  = iter->queuelen;
 966         *complete  = iter->complete;
 967         *send_only = iter->send_only;
 968 }
 969
 970 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */