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 multicast MTU and cached Q_Key before we attach if it's
 194          * the broadcast group.
 195          */
 196         if (!memcmp(mcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
 197                     sizeof (union ib_gid))) {
 198                 spin_lock_irq(&priv->lock);
 199                 if (!priv->broadcast) {
 200                         spin_unlock_irq(&priv->lock);
 201                         return -EAGAIN;
 202                 }
 203                 priv->mcast_mtu = IPOIB_UD_MTU(ib_mtu_enum_to_int(priv->broadcast->mcmember.mtu));
 204                 priv->qkey = be32_to_cpu(priv->broadcast->mcmember.qkey);
 205                 spin_unlock_irq(&priv->lock);
 206                 priv->tx_wr.wr.ud.remote_qkey = priv->qkey;
 207                 set_qkey = 1;
 208
 209                 if (!ipoib_cm_admin_enabled(dev)) {
 210                         rtnl_lock();
 211                         dev_set_mtu(dev, min(priv->mcast_mtu, priv->admin_mtu));
 212                         rtnl_unlock();
 213                 }
 214         }
 215
 216         if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 217                 if (test_and_set_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 218                         ipoib_warn(priv, "multicast group %pI6 already attached\n",
 219                                    mcast->mcmember.mgid.raw);
 220
 221                         return 0;
 222                 }
 223
 224                 ret = ipoib_mcast_attach(dev, be16_to_cpu(mcast->mcmember.mlid),
 225                                          &mcast->mcmember.mgid, set_qkey);
 226                 if (ret < 0) {
 227                         ipoib_warn(priv, "couldn't attach QP to multicast group %pI6\n",
 228                                    mcast->mcmember.mgid.raw);
 229
 230                         clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags);
 231                         return ret;
 232                 }
 233         }
 234
 235         {
 236                 struct ib_ah_attr av = {
 237                         .dlid          = be16_to_cpu(mcast->mcmember.mlid),
 238                         .port_num      = priv->port,
 239                         .sl            = mcast->mcmember.sl,
 240                         .ah_flags      = IB_AH_GRH,
 241                         .static_rate   = mcast->mcmember.rate,
 242                         .grh           = {
 243                                 .flow_label    = be32_to_cpu(mcast->mcmember.flow_label),
 244                                 .hop_limit     = mcast->mcmember.hop_limit,
 245                                 .sgid_index    = 0,
 246                                 .traffic_class = mcast->mcmember.traffic_class
 247                         }
 248                 };
 249                 av.grh.dgid = mcast->mcmember.mgid;
 250
 251                 ah = ipoib_create_ah(dev, priv->pd, &av);
 252                 if (IS_ERR(ah)) {
 253                         ipoib_warn(priv, "ib_address_create failed %ld\n",
 254                                 -PTR_ERR(ah));
 255                         /* use original error */
 256                         return PTR_ERR(ah);
 257                 } else {
 258                         spin_lock_irq(&priv->lock);
 259                         mcast->ah = ah;
 260                         spin_unlock_irq(&priv->lock);
 261
 262                         ipoib_dbg_mcast(priv, "MGID %pI6 AV %p, LID 0x%04x, SL %d\n",
 263                                         mcast->mcmember.mgid.raw,
 264                                         mcast->ah->ah,
 265                                         be16_to_cpu(mcast->mcmember.mlid),
 266                                         mcast->mcmember.sl);
 267                 }
 268         }
 269
 270         /* actually send any queued packets */
 271         netif_tx_lock_bh(dev);
 272         while (!skb_queue_empty(&mcast->pkt_queue)) {
 273                 struct sk_buff *skb = skb_dequeue(&mcast->pkt_queue);
 274
 275                 netif_tx_unlock_bh(dev);
 276
 277                 skb->dev = dev;
 278
 279                 if (dev_queue_xmit(skb))
 280                         ipoib_warn(priv, "dev_queue_xmit failed to requeue packet\n");
 281
 282                 netif_tx_lock_bh(dev);
 283         }
 284         netif_tx_unlock_bh(dev);
 285
 286         return 0;
 287 }
 288
 289 static int
 290 ipoib_mcast_sendonly_join_complete(int status,
 291                                    struct ib_sa_multicast *multicast)
 292 {
 293         struct ipoib_mcast *mcast = multicast->context;
 294         struct net_device *dev = mcast->dev;
 295
 296         /* We trap for port events ourselves. */
 297         if (status == -ENETRESET)
 298                 return 0;
 299
 300         if (!status)
 301                 status = ipoib_mcast_join_finish(mcast, &multicast->rec);
 302
 303         if (status) {
 304                 if (mcast->logcount++ < 20)
 305                         ipoib_dbg_mcast(netdev_priv(dev), "multicast join failed for %pI6, status %d\n",
 306                                         mcast->mcmember.mgid.raw, status);
 307
 308                 /* Flush out any queued packets */
 309                 netif_tx_lock_bh(dev);
 310                 while (!skb_queue_empty(&mcast->pkt_queue)) {
 311                         ++dev->stats.tx_dropped;
 312                         dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
 313                 }
 314                 netif_tx_unlock_bh(dev);
 315
 316                 /* Clear the busy flag so we try again */
 317                 status = test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY,
 318                                             &mcast->flags);
 319         }
 320         return status;
 321 }
 322
 323 static int ipoib_mcast_sendonly_join(struct ipoib_mcast *mcast)
 324 {
 325         struct net_device *dev = mcast->dev;
 326         struct ipoib_dev_priv *priv = netdev_priv(dev);
 327         struct ib_sa_mcmember_rec rec = {
 328 #if 0                           /* Some SMs don't support send-only yet */
 329                 .join_state = 4
 330 #else
 331                 .join_state = 1
 332 #endif
 333         };
 334         int ret = 0;
 335
 336         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
 337                 ipoib_dbg_mcast(priv, "device shutting down, no multicast joins\n");
 338                 return -ENODEV;
 339         }
 340
 341         if (test_and_set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)) {
 342                 ipoib_dbg_mcast(priv, "multicast entry busy, skipping\n");
 343                 return -EBUSY;
 344         }
 345
 346         rec.mgid     = mcast->mcmember.mgid;
 347         rec.port_gid = priv->local_gid;
 348         rec.pkey     = cpu_to_be16(priv->pkey);
 349
 350         mcast->mc = ib_sa_join_multicast(&ipoib_sa_client, priv->ca,
 351                                          priv->port, &rec,
 352                                          IB_SA_MCMEMBER_REC_MGID        |
 353                                          IB_SA_MCMEMBER_REC_PORT_GID    |
 354                                          IB_SA_MCMEMBER_REC_PKEY        |
 355                                          IB_SA_MCMEMBER_REC_JOIN_STATE,
 356                                          GFP_ATOMIC,
 357                                          ipoib_mcast_sendonly_join_complete,
 358                                          mcast);
 359         if (IS_ERR(mcast->mc)) {
 360                 ret = PTR_ERR(mcast->mc);
 361                 clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 362                 ipoib_warn(priv, "ib_sa_join_multicast failed (ret = %d)\n",
 363                            ret);
 364         } else {
 365                 ipoib_dbg_mcast(priv, "no multicast record for %pI6, starting join\n",
 366                                 mcast->mcmember.mgid.raw);
 367         }
 368
 369         return ret;
 370 }
 371
 372 void ipoib_mcast_carrier_on_task(struct work_struct *work)
 373 {
 374         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
 375                                                    carrier_on_task);
 376         struct ib_port_attr attr;
 377
 378         /*
 379          * Take rtnl_lock to avoid racing with ipoib_stop() and
 380          * turning the carrier back on while a device is being
 381          * removed.
 382          */
 383         if (ib_query_port(priv->ca, priv->port, &attr) ||
 384             attr.state != IB_PORT_ACTIVE) {
 385                 ipoib_dbg(priv, "Keeping carrier off until IB port is active\n");
 386                 return;
 387         }
 388
 389         rtnl_lock();
 390         netif_carrier_on(priv->dev);
 391         rtnl_unlock();
 392 }
 393
 394 static int ipoib_mcast_join_complete(int status,
 395                                      struct ib_sa_multicast *multicast)
 396 {
 397         struct ipoib_mcast *mcast = multicast->context;
 398         struct net_device *dev = mcast->dev;
 399         struct ipoib_dev_priv *priv = netdev_priv(dev);
 400
 401         ipoib_dbg_mcast(priv, "join completion for %pI6 (status %d)\n",
 402                         mcast->mcmember.mgid.raw, status);
 403
 404         /* We trap for port events ourselves. */
 405         if (status == -ENETRESET)
 406                 return 0;
 407
 408         if (!status)
 409                 status = ipoib_mcast_join_finish(mcast, &multicast->rec);
 410
 411         if (!status) {
 412                 mcast->backoff = 1;
 413                 mutex_lock(&mcast_mutex);
 414                 if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 415                         queue_delayed_work(ipoib_workqueue,
 416                                            &priv->mcast_task, 0);
 417                 mutex_unlock(&mcast_mutex);
 418
 419                 /*
 420                  * Defer carrier on work to ipoib_workqueue to avoid a
 421                  * deadlock on rtnl_lock here.
 422                  */
 423                 if (mcast == priv->broadcast)
 424                         queue_work(ipoib_workqueue, &priv->carrier_on_task);
 425
 426                 return 0;
 427         }
 428
 429         if (mcast->logcount++ < 20) {
 430                 if (status == -ETIMEDOUT || status == -EAGAIN) {
 431                         ipoib_dbg_mcast(priv, "multicast join failed for %pI6, status %d\n",
 432                                         mcast->mcmember.mgid.raw, status);
 433                 } else {
 434                         ipoib_warn(priv, "multicast join failed for %pI6, status %d\n",
 435                                    mcast->mcmember.mgid.raw, status);
 436                 }
 437         }
 438
 439         mcast->backoff *= 2;
 440         if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
 441                 mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
 442
 443         /* Clear the busy flag so we try again */
 444         status = test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 445
 446         mutex_lock(&mcast_mutex);
 447         spin_lock_irq(&priv->lock);
 448         if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 449                 queue_delayed_work(ipoib_workqueue, &priv->mcast_task,
 450                                    mcast->backoff * HZ);
 451         spin_unlock_irq(&priv->lock);
 452         mutex_unlock(&mcast_mutex);
 453
 454         return status;
 455 }
 456
 457 static void ipoib_mcast_join(struct net_device *dev, struct ipoib_mcast *mcast,
 458                              int create)
 459 {
 460         struct ipoib_dev_priv *priv = netdev_priv(dev);
 461         struct ib_sa_mcmember_rec rec = {
 462                 .join_state = 1
 463         };
 464         ib_sa_comp_mask comp_mask;
 465         int ret = 0;
 466
 467         ipoib_dbg_mcast(priv, "joining MGID %pI6\n", mcast->mcmember.mgid.raw);
 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_MTU_SELECTOR         |
 483                         IB_SA_MCMEMBER_REC_MTU                  |
 484                         IB_SA_MCMEMBER_REC_TRAFFIC_CLASS        |
 485                         IB_SA_MCMEMBER_REC_RATE_SELECTOR        |
 486                         IB_SA_MCMEMBER_REC_RATE                 |
 487                         IB_SA_MCMEMBER_REC_SL                   |
 488                         IB_SA_MCMEMBER_REC_FLOW_LABEL           |
 489                         IB_SA_MCMEMBER_REC_HOP_LIMIT;
 490
 491                 rec.qkey          = priv->broadcast->mcmember.qkey;
 492                 rec.mtu_selector  = IB_SA_EQ;
 493                 rec.mtu           = priv->broadcast->mcmember.mtu;
 494                 rec.traffic_class = priv->broadcast->mcmember.traffic_class;
 495                 rec.rate_selector = IB_SA_EQ;
 496                 rec.rate          = priv->broadcast->mcmember.rate;
 497                 rec.sl            = priv->broadcast->mcmember.sl;
 498                 rec.flow_label    = priv->broadcast->mcmember.flow_label;
 499                 rec.hop_limit     = priv->broadcast->mcmember.hop_limit;
 500         }
 501
 502         set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 503         mcast->mc = ib_sa_join_multicast(&ipoib_sa_client, priv->ca, priv->port,
 504                                          &rec, comp_mask, GFP_KERNEL,
 505                                          ipoib_mcast_join_complete, mcast);
 506         if (IS_ERR(mcast->mc)) {
 507                 clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 508                 ret = PTR_ERR(mcast->mc);
 509                 ipoib_warn(priv, "ib_sa_join_multicast failed, status %d\n", ret);
 510
 511                 mcast->backoff *= 2;
 512                 if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
 513                         mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
 514
 515                 mutex_lock(&mcast_mutex);
 516                 if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 517                         queue_delayed_work(ipoib_workqueue,
 518                                            &priv->mcast_task,
 519                                            mcast->backoff * HZ);
 520                 mutex_unlock(&mcast_mutex);
 521         }
 522 }
 523
 524 void ipoib_mcast_join_task(struct work_struct *work)
 525 {
 526         struct ipoib_dev_priv *priv =
 527                 container_of(work, struct ipoib_dev_priv, mcast_task.work);
 528         struct net_device *dev = priv->dev;
 529
 530         if (!test_bit(IPOIB_MCAST_RUN, &priv->flags))
 531                 return;
 532
 533         if (ib_query_gid(priv->ca, priv->port, 0, &priv->local_gid))
 534                 ipoib_warn(priv, "ib_query_gid() failed\n");
 535         else
 536                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
 537
 538         {
 539                 struct ib_port_attr attr;
 540
 541                 if (!ib_query_port(priv->ca, priv->port, &attr))
 542                         priv->local_lid = attr.lid;
 543                 else
 544                         ipoib_warn(priv, "ib_query_port failed\n");
 545         }
 546
 547         if (!priv->broadcast) {
 548                 struct ipoib_mcast *broadcast;
 549
 550                 if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 551                         return;
 552
 553                 broadcast = ipoib_mcast_alloc(dev, 1);
 554                 if (!broadcast) {
 555                         ipoib_warn(priv, "failed to allocate broadcast group\n");
 556                         mutex_lock(&mcast_mutex);
 557                         if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 558                                 queue_delayed_work(ipoib_workqueue,
 559                                                    &priv->mcast_task, HZ);
 560                         mutex_unlock(&mcast_mutex);
 561                         return;
 562                 }
 563
 564                 spin_lock_irq(&priv->lock);
 565                 memcpy(broadcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
 566                        sizeof (union ib_gid));
 567                 priv->broadcast = broadcast;
 568
 569                 __ipoib_mcast_add(dev, priv->broadcast);
 570                 spin_unlock_irq(&priv->lock);
 571         }
 572
 573         if (!test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
 574                 if (!test_bit(IPOIB_MCAST_FLAG_BUSY, &priv->broadcast->flags))
 575                         ipoib_mcast_join(dev, priv->broadcast, 0);
 576                 return;
 577         }
 578
 579         while (1) {
 580                 struct ipoib_mcast *mcast = NULL;
 581
 582                 spin_lock_irq(&priv->lock);
 583                 list_for_each_entry(mcast, &priv->multicast_list, list) {
 584                         if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)
 585                             && !test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)
 586                             && !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 587                                 /* Found the next unjoined group */
 588                                 break;
 589                         }
 590                 }
 591                 spin_unlock_irq(&priv->lock);
 592
 593                 if (&mcast->list == &priv->multicast_list) {
 594                         /* All done */
 595                         break;
 596                 }
 597
 598                 ipoib_mcast_join(dev, mcast, 1);
 599                 return;
 600         }
 601
 602         ipoib_dbg_mcast(priv, "successfully joined all multicast groups\n");
 603
 604         clear_bit(IPOIB_MCAST_RUN, &priv->flags);
 605 }
 606
 607 int ipoib_mcast_start_thread(struct net_device *dev)
 608 {
 609         struct ipoib_dev_priv *priv = netdev_priv(dev);
 610
 611         ipoib_dbg_mcast(priv, "starting multicast thread\n");
 612
 613         mutex_lock(&mcast_mutex);
 614         if (!test_and_set_bit(IPOIB_MCAST_RUN, &priv->flags))
 615                 queue_delayed_work(ipoib_workqueue, &priv->mcast_task, 0);
 616         mutex_unlock(&mcast_mutex);
 617
 618         return 0;
 619 }
 620
 621 int ipoib_mcast_stop_thread(struct net_device *dev, int flush)
 622 {
 623         struct ipoib_dev_priv *priv = netdev_priv(dev);
 624
 625         ipoib_dbg_mcast(priv, "stopping multicast thread\n");
 626
 627         mutex_lock(&mcast_mutex);
 628         clear_bit(IPOIB_MCAST_RUN, &priv->flags);
 629         cancel_delayed_work(&priv->mcast_task);
 630         mutex_unlock(&mcast_mutex);
 631
 632         if (flush)
 633                 flush_workqueue(ipoib_workqueue);
 634
 635         return 0;
 636 }
 637
 638 static int ipoib_mcast_leave(struct net_device *dev, struct ipoib_mcast *mcast)
 639 {
 640         struct ipoib_dev_priv *priv = netdev_priv(dev);
 641         int ret = 0;
 642
 643         if (test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
 644                 ib_sa_free_multicast(mcast->mc);
 645
 646         if (test_and_clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 647                 ipoib_dbg_mcast(priv, "leaving MGID %pI6\n",
 648                                 mcast->mcmember.mgid.raw);
 649
 650                 /* Remove ourselves from the multicast group */
 651                 ret = ib_detach_mcast(priv->qp, &mcast->mcmember.mgid,
 652                                       be16_to_cpu(mcast->mcmember.mlid));
 653                 if (ret)
 654                         ipoib_warn(priv, "ib_detach_mcast failed (result = %d)\n", ret);
 655         }
 656
 657         return 0;
 658 }
 659
 660 void ipoib_mcast_send(struct net_device *dev, void *mgid, struct sk_buff *skb)
 661 {
 662         struct ipoib_dev_priv *priv = netdev_priv(dev);
 663         struct ipoib_mcast *mcast;
 664         unsigned long flags;
 665
 666         spin_lock_irqsave(&priv->lock, flags);
 667
 668         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)         ||
 669             !priv->broadcast                                    ||
 670             !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
 671                 ++dev->stats.tx_dropped;
 672                 dev_kfree_skb_any(skb);
 673                 goto unlock;
 674         }
 675
 676         mcast = __ipoib_mcast_find(dev, mgid);
 677         if (!mcast) {
 678                 /* Let's create a new send only group now */
 679                 ipoib_dbg_mcast(priv, "setting up send only multicast group for %pI6\n",
 680                                 mgid);
 681
 682                 mcast = ipoib_mcast_alloc(dev, 0);
 683                 if (!mcast) {
 684                         ipoib_warn(priv, "unable to allocate memory for "
 685                                    "multicast structure\n");
 686                         ++dev->stats.tx_dropped;
 687                         dev_kfree_skb_any(skb);
 688                         goto out;
 689                 }
 690
 691                 set_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags);
 692                 memcpy(mcast->mcmember.mgid.raw, mgid, sizeof (union ib_gid));
 693                 __ipoib_mcast_add(dev, mcast);
 694                 list_add_tail(&mcast->list, &priv->multicast_list);
 695         }
 696
 697         if (!mcast->ah) {
 698                 if (skb_queue_len(&mcast->pkt_queue) < IPOIB_MAX_MCAST_QUEUE)
 699                         skb_queue_tail(&mcast->pkt_queue, skb);
 700                 else {
 701                         ++dev->stats.tx_dropped;
 702                         dev_kfree_skb_any(skb);
 703                 }
 704
 705                 if (test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
 706                         ipoib_dbg_mcast(priv, "no address vector, "
 707                                         "but multicast join already started\n");
 708                 else if (test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags))
 709                         ipoib_mcast_sendonly_join(mcast);
 710
 711                 /*
 712                  * If lookup completes between here and out:, don't
 713                  * want to send packet twice.
 714                  */
 715                 mcast = NULL;
 716         }
 717
 718 out:
 719         if (mcast && mcast->ah) {
 720                 struct dst_entry *dst = skb_dst(skb);
 721                 struct neighbour *n = NULL;
 722
 723                 rcu_read_lock();
 724                 if (dst)
 725                         n = dst_get_neighbour(dst);
 726                 if (n && !*to_ipoib_neigh(n)) {
 727                         struct ipoib_neigh *neigh = ipoib_neigh_alloc(n,
 728                                                                       skb->dev);
 729
 730                         if (neigh) {
 731                                 kref_get(&mcast->ah->ref);
 732                                 neigh->ah       = mcast->ah;
 733                                 list_add_tail(&neigh->list, &mcast->neigh_list);
 734                         }
 735                 }
 736                 rcu_read_unlock();
 737                 spin_unlock_irqrestore(&priv->lock, flags);
 738                 ipoib_send(dev, skb, mcast->ah, IB_MULTICAST_QPN);
 739                 return;
 740         }
 741
 742 unlock:
 743         spin_unlock_irqrestore(&priv->lock, flags);
 744 }
 745
 746 void ipoib_mcast_dev_flush(struct net_device *dev)
 747 {
 748         struct ipoib_dev_priv *priv = netdev_priv(dev);
 749         LIST_HEAD(remove_list);
 750         struct ipoib_mcast *mcast, *tmcast;
 751         unsigned long flags;
 752
 753         ipoib_dbg_mcast(priv, "flushing multicast list\n");
 754
 755         spin_lock_irqsave(&priv->lock, flags);
 756
 757         list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
 758                 list_del(&mcast->list);
 759                 rb_erase(&mcast->rb_node, &priv->multicast_tree);
 760                 list_add_tail(&mcast->list, &remove_list);
 761         }
 762
 763         if (priv->broadcast) {
 764                 rb_erase(&priv->broadcast->rb_node, &priv->multicast_tree);
 765                 list_add_tail(&priv->broadcast->list, &remove_list);
 766                 priv->broadcast = NULL;
 767         }
 768
 769         spin_unlock_irqrestore(&priv->lock, flags);
 770
 771         list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
 772                 ipoib_mcast_leave(dev, mcast);
 773                 ipoib_mcast_free(mcast);
 774         }
 775 }
 776
 777 static int ipoib_mcast_addr_is_valid(const u8 *addr, const u8 *broadcast)
 778 {
 779         /* reserved QPN, prefix, scope */
 780         if (memcmp(addr, broadcast, 6))
 781                 return 0;
 782         /* signature lower, pkey */
 783         if (memcmp(addr + 7, broadcast + 7, 3))
 784                 return 0;
 785         return 1;
 786 }
 787
 788 void ipoib_mcast_restart_task(struct work_struct *work)
 789 {
 790         struct ipoib_dev_priv *priv =
 791                 container_of(work, struct ipoib_dev_priv, restart_task);
 792         struct net_device *dev = priv->dev;
 793         struct netdev_hw_addr *ha;
 794         struct ipoib_mcast *mcast, *tmcast;
 795         LIST_HEAD(remove_list);
 796         unsigned long flags;
 797         struct ib_sa_mcmember_rec rec;
 798
 799         ipoib_dbg_mcast(priv, "restarting multicast task\n");
 800
 801         ipoib_mcast_stop_thread(dev, 0);
 802
 803         local_irq_save(flags);
 804         netif_addr_lock(dev);
 805         spin_lock(&priv->lock);
 806
 807         /*
 808          * Unfortunately, the networking core only gives us a list of all of
 809          * the multicast hardware addresses. We need to figure out which ones
 810          * are new and which ones have been removed
 811          */
 812
 813         /* Clear out the found flag */
 814         list_for_each_entry(mcast, &priv->multicast_list, list)
 815                 clear_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
 816
 817         /* Mark all of the entries that are found or don't exist */
 818         netdev_for_each_mc_addr(ha, dev) {
 819                 union ib_gid mgid;
 820
 821                 if (!ipoib_mcast_addr_is_valid(ha->addr, dev->broadcast))
 822                         continue;
 823
 824                 memcpy(mgid.raw, ha->addr + 4, sizeof mgid);
 825
 826                 mcast = __ipoib_mcast_find(dev, &mgid);
 827                 if (!mcast || test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 828                         struct ipoib_mcast *nmcast;
 829
 830                         /* ignore group which is directly joined by userspace */
 831                         if (test_bit(IPOIB_FLAG_UMCAST, &priv->flags) &&
 832                             !ib_sa_get_mcmember_rec(priv->ca, priv->port, &mgid, &rec)) {
 833                                 ipoib_dbg_mcast(priv, "ignoring multicast entry for mgid %pI6\n",
 834                                                 mgid.raw);
 835                                 continue;
 836                         }
 837
 838                         /* Not found or send-only group, let's add a new entry */
 839                         ipoib_dbg_mcast(priv, "adding multicast entry for mgid %pI6\n",
 840                                         mgid.raw);
 841
 842                         nmcast = ipoib_mcast_alloc(dev, 0);
 843                         if (!nmcast) {
 844                                 ipoib_warn(priv, "unable to allocate memory for multicast structure\n");
 845                                 continue;
 846                         }
 847
 848                         set_bit(IPOIB_MCAST_FLAG_FOUND, &nmcast->flags);
 849
 850                         nmcast->mcmember.mgid = mgid;
 851
 852                         if (mcast) {
 853                                 /* Destroy the send only entry */
 854                                 list_move_tail(&mcast->list, &remove_list);
 855
 856                                 rb_replace_node(&mcast->rb_node,
 857                                                 &nmcast->rb_node,
 858                                                 &priv->multicast_tree);
 859                         } else
 860                                 __ipoib_mcast_add(dev, nmcast);
 861
 862                         list_add_tail(&nmcast->list, &priv->multicast_list);
 863                 }
 864
 865                 if (mcast)
 866                         set_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
 867         }
 868
 869         /* Remove all of the entries don't exist anymore */
 870         list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
 871                 if (!test_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags) &&
 872                     !test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 873                         ipoib_dbg_mcast(priv, "deleting multicast group %pI6\n",
 874                                         mcast->mcmember.mgid.raw);
 875
 876                         rb_erase(&mcast->rb_node, &priv->multicast_tree);
 877
 878                         /* Move to the remove list */
 879                         list_move_tail(&mcast->list, &remove_list);
 880                 }
 881         }
 882
 883         spin_unlock(&priv->lock);
 884         netif_addr_unlock(dev);
 885         local_irq_restore(flags);
 886
 887         /* We have to cancel outside of the spinlock */
 888         list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
 889                 ipoib_mcast_leave(mcast->dev, mcast);
 890                 ipoib_mcast_free(mcast);
 891         }
 892
 893         if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 894                 ipoib_mcast_start_thread(dev);
 895 }
 896
 897 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 898
 899 struct ipoib_mcast_iter *ipoib_mcast_iter_init(struct net_device *dev)
 900 {
 901         struct ipoib_mcast_iter *iter;
 902
 903         iter = kmalloc(sizeof *iter, GFP_KERNEL);
 904         if (!iter)
 905                 return NULL;
 906
 907         iter->dev = dev;
 908         memset(iter->mgid.raw, 0, 16);
 909
 910         if (ipoib_mcast_iter_next(iter)) {
 911                 kfree(iter);
 912                 return NULL;
 913         }
 914
 915         return iter;
 916 }
 917
 918 int ipoib_mcast_iter_next(struct ipoib_mcast_iter *iter)
 919 {
 920         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
 921         struct rb_node *n;
 922         struct ipoib_mcast *mcast;
 923         int ret = 1;
 924
 925         spin_lock_irq(&priv->lock);
 926
 927         n = rb_first(&priv->multicast_tree);
 928
 929         while (n) {
 930                 mcast = rb_entry(n, struct ipoib_mcast, rb_node);
 931
 932                 if (memcmp(iter->mgid.raw, mcast->mcmember.mgid.raw,
 933                            sizeof (union ib_gid)) < 0) {
 934                         iter->mgid      = mcast->mcmember.mgid;
 935                         iter->created   = mcast->created;
 936                         iter->queuelen  = skb_queue_len(&mcast->pkt_queue);
 937                         iter->complete  = !!mcast->ah;
 938                         iter->send_only = !!(mcast->flags & (1 << IPOIB_MCAST_FLAG_SENDONLY));
 939
 940                         ret = 0;
 941
 942                         break;
 943                 }
 944
 945                 n = rb_next(n);
 946         }
 947
 948         spin_unlock_irq(&priv->lock);
 949
 950         return ret;
 951 }
 952
 953 void ipoib_mcast_iter_read(struct ipoib_mcast_iter *iter,
 954                            union ib_gid *mgid,
 955                            unsigned long *created,
 956                            unsigned int *queuelen,
 957                            unsigned int *complete,
 958                            unsigned int *send_only)
 959 {
 960         *mgid      = iter->mgid;
 961         *created   = iter->created;
 962         *queuelen  = iter->queuelen;
 963         *complete  = iter->complete;
 964         *send_only = iter->send_only;
 965 }
 966
 967 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */