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 struct ipoib_mcast_iter {
  59         struct net_device *dev;
  60         union ib_gid       mgid;
  61         unsigned long      created;
  62         unsigned int       queuelen;
  63         unsigned int       complete;
  64         unsigned int       send_only;
  65 };
  66
  67 /*
  68  * This should be called with the priv->lock held
  69  */
  70 static void __ipoib_mcast_schedule_join_thread(struct ipoib_dev_priv *priv,
  71                                                struct ipoib_mcast *mcast,
  72                                                bool delay)
  73 {
  74         if (!test_bit(IPOIB_MCAST_RUN, &priv->flags))
  75                 return;
  76
  77         /*
  78          * We will be scheduling *something*, so cancel whatever is
  79          * currently scheduled first
  80          */
  81         cancel_delayed_work(&priv->mcast_task);
  82         if (mcast && delay) {
  83                 /*
  84                  * We had a failure and want to schedule a retry later
  85                  */
  86                 mcast->backoff *= 2;
  87                 if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
  88                         mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
  89                 mcast->delay_until = jiffies + (mcast->backoff * HZ);
  90                 /*
  91                  * Mark this mcast for its delay, but restart the
  92                  * task immediately.  The join task will make sure to
  93                  * clear out all entries without delays, and then
  94                  * schedule itself to run again when the earliest
  95                  * delay expires
  96                  */
  97                 queue_delayed_work(priv->wq, &priv->mcast_task, 0);
  98         } else if (delay) {
  99                 /*
 100                  * Special case of retrying after a failure to
 101                  * allocate the broadcast multicast group, wait
 102                  * 1 second and try again
 103                  */
 104                 queue_delayed_work(priv->wq, &priv->mcast_task, HZ);
 105         } else
 106                 queue_delayed_work(priv->wq, &priv->mcast_task, 0);
 107 }
 108
 109 static void ipoib_mcast_free(struct ipoib_mcast *mcast)
 110 {
 111         struct net_device *dev = mcast->dev;
 112         int tx_dropped = 0;
 113
 114         ipoib_dbg_mcast(netdev_priv(dev), "deleting multicast group %pI6\n",
 115                         mcast->mcmember.mgid.raw);
 116
 117         /* remove all neigh connected to this mcast */
 118         ipoib_del_neighs_by_gid(dev, mcast->mcmember.mgid.raw);
 119
 120         if (mcast->ah)
 121                 ipoib_put_ah(mcast->ah);
 122
 123         while (!skb_queue_empty(&mcast->pkt_queue)) {
 124                 ++tx_dropped;
 125                 dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
 126         }
 127
 128         netif_tx_lock_bh(dev);
 129         dev->stats.tx_dropped += tx_dropped;
 130         netif_tx_unlock_bh(dev);
 131
 132         kfree(mcast);
 133 }
 134
 135 static struct ipoib_mcast *ipoib_mcast_alloc(struct net_device *dev,
 136                                              int can_sleep)
 137 {
 138         struct ipoib_mcast *mcast;
 139
 140         mcast = kzalloc(sizeof *mcast, can_sleep ? GFP_KERNEL : GFP_ATOMIC);
 141         if (!mcast)
 142                 return NULL;
 143
 144         mcast->dev = dev;
 145         mcast->created = jiffies;
 146         mcast->delay_until = jiffies;
 147         mcast->backoff = 1;
 148
 149         INIT_LIST_HEAD(&mcast->list);
 150         INIT_LIST_HEAD(&mcast->neigh_list);
 151         skb_queue_head_init(&mcast->pkt_queue);
 152
 153         return mcast;
 154 }
 155
 156 static struct ipoib_mcast *__ipoib_mcast_find(struct net_device *dev, void *mgid)
 157 {
 158         struct ipoib_dev_priv *priv = netdev_priv(dev);
 159         struct rb_node *n = priv->multicast_tree.rb_node;
 160
 161         while (n) {
 162                 struct ipoib_mcast *mcast;
 163                 int ret;
 164
 165                 mcast = rb_entry(n, struct ipoib_mcast, rb_node);
 166
 167                 ret = memcmp(mgid, mcast->mcmember.mgid.raw,
 168                              sizeof (union ib_gid));
 169                 if (ret < 0)
 170                         n = n->rb_left;
 171                 else if (ret > 0)
 172                         n = n->rb_right;
 173                 else
 174                         return mcast;
 175         }
 176
 177         return NULL;
 178 }
 179
 180 static int __ipoib_mcast_add(struct net_device *dev, struct ipoib_mcast *mcast)
 181 {
 182         struct ipoib_dev_priv *priv = netdev_priv(dev);
 183         struct rb_node **n = &priv->multicast_tree.rb_node, *pn = NULL;
 184
 185         while (*n) {
 186                 struct ipoib_mcast *tmcast;
 187                 int ret;
 188
 189                 pn = *n;
 190                 tmcast = rb_entry(pn, struct ipoib_mcast, rb_node);
 191
 192                 ret = memcmp(mcast->mcmember.mgid.raw, tmcast->mcmember.mgid.raw,
 193                              sizeof (union ib_gid));
 194                 if (ret < 0)
 195                         n = &pn->rb_left;
 196                 else if (ret > 0)
 197                         n = &pn->rb_right;
 198                 else
 199                         return -EEXIST;
 200         }
 201
 202         rb_link_node(&mcast->rb_node, pn, n);
 203         rb_insert_color(&mcast->rb_node, &priv->multicast_tree);
 204
 205         return 0;
 206 }
 207
 208 static int ipoib_mcast_join_finish(struct ipoib_mcast *mcast,
 209                                    struct ib_sa_mcmember_rec *mcmember)
 210 {
 211         struct net_device *dev = mcast->dev;
 212         struct ipoib_dev_priv *priv = netdev_priv(dev);
 213         struct ipoib_ah *ah;
 214         int ret;
 215         int set_qkey = 0;
 216
 217         mcast->mcmember = *mcmember;
 218
 219         /* Set the multicast MTU and cached Q_Key before we attach if it's
 220          * the broadcast group.
 221          */
 222         if (!memcmp(mcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
 223                     sizeof (union ib_gid))) {
 224                 spin_lock_irq(&priv->lock);
 225                 if (!priv->broadcast) {
 226                         spin_unlock_irq(&priv->lock);
 227                         return -EAGAIN;
 228                 }
 229                 priv->mcast_mtu = IPOIB_UD_MTU(ib_mtu_enum_to_int(priv->broadcast->mcmember.mtu));
 230                 priv->qkey = be32_to_cpu(priv->broadcast->mcmember.qkey);
 231                 spin_unlock_irq(&priv->lock);
 232                 priv->tx_wr.wr.ud.remote_qkey = priv->qkey;
 233                 set_qkey = 1;
 234         }
 235
 236         if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 237                 if (test_and_set_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 238                         ipoib_warn(priv, "multicast group %pI6 already attached\n",
 239                                    mcast->mcmember.mgid.raw);
 240
 241                         return 0;
 242                 }
 243
 244                 ret = ipoib_mcast_attach(dev, be16_to_cpu(mcast->mcmember.mlid),
 245                                          &mcast->mcmember.mgid, set_qkey);
 246                 if (ret < 0) {
 247                         ipoib_warn(priv, "couldn't attach QP to multicast group %pI6\n",
 248                                    mcast->mcmember.mgid.raw);
 249
 250                         clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags);
 251                         return ret;
 252                 }
 253         }
 254
 255         {
 256                 struct ib_ah_attr av = {
 257                         .dlid          = be16_to_cpu(mcast->mcmember.mlid),
 258                         .port_num      = priv->port,
 259                         .sl            = mcast->mcmember.sl,
 260                         .ah_flags      = IB_AH_GRH,
 261                         .static_rate   = mcast->mcmember.rate,
 262                         .grh           = {
 263                                 .flow_label    = be32_to_cpu(mcast->mcmember.flow_label),
 264                                 .hop_limit     = mcast->mcmember.hop_limit,
 265                                 .sgid_index    = 0,
 266                                 .traffic_class = mcast->mcmember.traffic_class
 267                         }
 268                 };
 269                 av.grh.dgid = mcast->mcmember.mgid;
 270
 271                 ah = ipoib_create_ah(dev, priv->pd, &av);
 272                 if (IS_ERR(ah)) {
 273                         ipoib_warn(priv, "ib_address_create failed %ld\n",
 274                                 -PTR_ERR(ah));
 275                         /* use original error */
 276                         return PTR_ERR(ah);
 277                 } else {
 278                         spin_lock_irq(&priv->lock);
 279                         mcast->ah = ah;
 280                         spin_unlock_irq(&priv->lock);
 281
 282                         ipoib_dbg_mcast(priv, "MGID %pI6 AV %p, LID 0x%04x, SL %d\n",
 283                                         mcast->mcmember.mgid.raw,
 284                                         mcast->ah->ah,
 285                                         be16_to_cpu(mcast->mcmember.mlid),
 286                                         mcast->mcmember.sl);
 287                 }
 288         }
 289
 290         /* actually send any queued packets */
 291         netif_tx_lock_bh(dev);
 292         while (!skb_queue_empty(&mcast->pkt_queue)) {
 293                 struct sk_buff *skb = skb_dequeue(&mcast->pkt_queue);
 294
 295                 netif_tx_unlock_bh(dev);
 296
 297                 skb->dev = dev;
 298                 if (dev_queue_xmit(skb))
 299                         ipoib_warn(priv, "dev_queue_xmit failed to requeue packet\n");
 300
 301                 netif_tx_lock_bh(dev);
 302         }
 303         netif_tx_unlock_bh(dev);
 304
 305         return 0;
 306 }
 307
 308 void ipoib_mcast_carrier_on_task(struct work_struct *work)
 309 {
 310         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
 311                                                    carrier_on_task);
 312         struct ib_port_attr attr;
 313
 314         if (ib_query_port(priv->ca, priv->port, &attr) ||
 315             attr.state != IB_PORT_ACTIVE) {
 316                 ipoib_dbg(priv, "Keeping carrier off until IB port is active\n");
 317                 return;
 318         }
 319
 320         /*
 321          * Take rtnl_lock to avoid racing with ipoib_stop() and
 322          * turning the carrier back on while a device is being
 323          * removed.  However, ipoib_stop() will attempt to flush
 324          * the workqueue while holding the rtnl lock, so loop
 325          * on trylock until either we get the lock or we see
 326          * FLAG_OPER_UP go away as that signals that we are bailing
 327          * and can safely ignore the carrier on work.
 328          */
 329         while (!rtnl_trylock()) {
 330                 if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
 331                         return;
 332                 else
 333                         msleep(20);
 334         }
 335         if (!ipoib_cm_admin_enabled(priv->dev))
 336                 dev_set_mtu(priv->dev, min(priv->mcast_mtu, priv->admin_mtu));
 337         netif_carrier_on(priv->dev);
 338         rtnl_unlock();
 339 }
 340
 341 static int ipoib_mcast_join_complete(int status,
 342                                      struct ib_sa_multicast *multicast)
 343 {
 344         struct ipoib_mcast *mcast = multicast->context;
 345         struct net_device *dev = mcast->dev;
 346         struct ipoib_dev_priv *priv = netdev_priv(dev);
 347
 348         ipoib_dbg_mcast(priv, "%sjoin completion for %pI6 (status %d)\n",
 349                         test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) ?
 350                         "sendonly " : "",
 351                         mcast->mcmember.mgid.raw, status);
 352
 353         /* We trap for port events ourselves. */
 354         if (status == -ENETRESET) {
 355                 status = 0;
 356                 goto out;
 357         }
 358
 359         if (!status)
 360                 status = ipoib_mcast_join_finish(mcast, &multicast->rec);
 361
 362         if (!status) {
 363                 mcast->backoff = 1;
 364                 mcast->delay_until = jiffies;
 365
 366                 /*
 367                  * Defer carrier on work to priv->wq to avoid a
 368                  * deadlock on rtnl_lock here.  Requeue our multicast
 369                  * work too, which will end up happening right after
 370                  * our carrier on task work and will allow us to
 371                  * send out all of the non-broadcast joins
 372                  */
 373                 if (mcast == priv->broadcast) {
 374                         spin_lock_irq(&priv->lock);
 375                         queue_work(priv->wq, &priv->carrier_on_task);
 376                         __ipoib_mcast_schedule_join_thread(priv, NULL, 0);
 377                         goto out_locked;
 378                 }
 379         } else {
 380                 if (mcast->logcount++ < 20) {
 381                         if (status == -ETIMEDOUT || status == -EAGAIN) {
 382                                 ipoib_dbg_mcast(priv, "%smulticast join failed for %pI6, status %d\n",
 383                                                 test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) ? "sendonly " : "",
 384                                                 mcast->mcmember.mgid.raw, status);
 385                         } else {
 386                                 ipoib_warn(priv, "%smulticast join failed for %pI6, status %d\n",
 387                                                 test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) ? "sendonly " : "",
 388                                            mcast->mcmember.mgid.raw, status);
 389                         }
 390                 }
 391
 392                 if (test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) &&
 393                     mcast->backoff >= 2) {
 394                         /*
 395                          * We only retry sendonly joins once before we drop
 396                          * the packet and quit trying to deal with the
 397                          * group.  However, we leave the group in the
 398                          * mcast list as an unjoined group.  If we want to
 399                          * try joining again, we simply queue up a packet
 400                          * and restart the join thread.  The empty queue
 401                          * is why the join thread ignores this group.
 402                          */
 403                         mcast->backoff = 1;
 404                         netif_tx_lock_bh(dev);
 405                         while (!skb_queue_empty(&mcast->pkt_queue)) {
 406                                 ++dev->stats.tx_dropped;
 407                                 dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
 408                         }
 409                         netif_tx_unlock_bh(dev);
 410                 } else {
 411                         spin_lock_irq(&priv->lock);
 412                         /* Requeue this join task with a backoff delay */
 413                         __ipoib_mcast_schedule_join_thread(priv, mcast, 1);
 414                         goto out_locked;
 415                 }
 416         }
 417 out:
 418         spin_lock_irq(&priv->lock);
 419 out_locked:
 420         /*
 421          * Make sure to set mcast->mc before we clear the busy flag to avoid
 422          * racing with code that checks for BUSY before checking mcast->mc
 423          */
 424         if (status)
 425                 mcast->mc = NULL;
 426         else
 427                 mcast->mc = multicast;
 428         clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 429         spin_unlock_irq(&priv->lock);
 430         complete(&mcast->done);
 431
 432         return status;
 433 }
 434
 435 static void ipoib_mcast_join(struct net_device *dev, struct ipoib_mcast *mcast,
 436                              int create)
 437 {
 438         struct ipoib_dev_priv *priv = netdev_priv(dev);
 439         struct ib_sa_multicast *multicast;
 440         struct ib_sa_mcmember_rec rec = {
 441                 .join_state = 1
 442         };
 443         ib_sa_comp_mask comp_mask;
 444         int ret = 0;
 445
 446         ipoib_dbg_mcast(priv, "joining MGID %pI6\n", mcast->mcmember.mgid.raw);
 447
 448         rec.mgid     = mcast->mcmember.mgid;
 449         rec.port_gid = priv->local_gid;
 450         rec.pkey     = cpu_to_be16(priv->pkey);
 451
 452         comp_mask =
 453                 IB_SA_MCMEMBER_REC_MGID         |
 454                 IB_SA_MCMEMBER_REC_PORT_GID     |
 455                 IB_SA_MCMEMBER_REC_PKEY         |
 456                 IB_SA_MCMEMBER_REC_JOIN_STATE;
 457
 458         if (create) {
 459                 comp_mask |=
 460                         IB_SA_MCMEMBER_REC_QKEY                 |
 461                         IB_SA_MCMEMBER_REC_MTU_SELECTOR         |
 462                         IB_SA_MCMEMBER_REC_MTU                  |
 463                         IB_SA_MCMEMBER_REC_TRAFFIC_CLASS        |
 464                         IB_SA_MCMEMBER_REC_RATE_SELECTOR        |
 465                         IB_SA_MCMEMBER_REC_RATE                 |
 466                         IB_SA_MCMEMBER_REC_SL                   |
 467                         IB_SA_MCMEMBER_REC_FLOW_LABEL           |
 468                         IB_SA_MCMEMBER_REC_HOP_LIMIT;
 469
 470                 rec.qkey          = priv->broadcast->mcmember.qkey;
 471                 rec.mtu_selector  = IB_SA_EQ;
 472                 rec.mtu           = priv->broadcast->mcmember.mtu;
 473                 rec.traffic_class = priv->broadcast->mcmember.traffic_class;
 474                 rec.rate_selector = IB_SA_EQ;
 475                 rec.rate          = priv->broadcast->mcmember.rate;
 476                 rec.sl            = priv->broadcast->mcmember.sl;
 477                 rec.flow_label    = priv->broadcast->mcmember.flow_label;
 478                 rec.hop_limit     = priv->broadcast->mcmember.hop_limit;
 479         }
 480
 481         multicast = ib_sa_join_multicast(&ipoib_sa_client, priv->ca, priv->port,
 482                                          &rec, comp_mask, GFP_KERNEL,
 483                                          ipoib_mcast_join_complete, mcast);
 484         if (IS_ERR(multicast)) {
 485                 ret = PTR_ERR(multicast);
 486                 ipoib_warn(priv, "ib_sa_join_multicast failed, status %d\n", ret);
 487                 spin_lock_irq(&priv->lock);
 488                 /* Requeue this join task with a backoff delay */
 489                 __ipoib_mcast_schedule_join_thread(priv, mcast, 1);
 490                 clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 491                 spin_unlock_irq(&priv->lock);
 492                 complete(&mcast->done);
 493         }
 494 }
 495
 496 void ipoib_mcast_join_task(struct work_struct *work)
 497 {
 498         struct ipoib_dev_priv *priv =
 499                 container_of(work, struct ipoib_dev_priv, mcast_task.work);
 500         struct net_device *dev = priv->dev;
 501         struct ib_port_attr port_attr;
 502         unsigned long delay_until = 0;
 503         struct ipoib_mcast *mcast = NULL;
 504         int create = 1;
 505
 506         if (!test_bit(IPOIB_MCAST_RUN, &priv->flags))
 507                 return;
 508
 509         if (ib_query_port(priv->ca, priv->port, &port_attr) ||
 510             port_attr.state != IB_PORT_ACTIVE) {
 511                 ipoib_dbg(priv, "port state is not ACTIVE (state = %d) suspending join task\n",
 512                           port_attr.state);
 513                 return;
 514         }
 515         priv->local_lid = port_attr.lid;
 516
 517         if (ib_query_gid(priv->ca, priv->port, 0, &priv->local_gid))
 518                 ipoib_warn(priv, "ib_query_gid() failed\n");
 519         else
 520                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
 521
 522         spin_lock_irq(&priv->lock);
 523         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
 524                 goto out;
 525
 526         if (!priv->broadcast) {
 527                 struct ipoib_mcast *broadcast;
 528
 529                 broadcast = ipoib_mcast_alloc(dev, 0);
 530                 if (!broadcast) {
 531                         ipoib_warn(priv, "failed to allocate broadcast group\n");
 532                         /*
 533                          * Restart us after a 1 second delay to retry
 534                          * creating our broadcast group and attaching to
 535                          * it.  Until this succeeds, this ipoib dev is
 536                          * completely stalled (multicast wise).
 537                          */
 538                         __ipoib_mcast_schedule_join_thread(priv, NULL, 1);
 539                         goto out;
 540                 }
 541
 542                 memcpy(broadcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
 543                        sizeof (union ib_gid));
 544                 priv->broadcast = broadcast;
 545
 546                 __ipoib_mcast_add(dev, priv->broadcast);
 547         }
 548
 549         if (!test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
 550                 if (IS_ERR_OR_NULL(priv->broadcast->mc) &&
 551                     !test_bit(IPOIB_MCAST_FLAG_BUSY, &priv->broadcast->flags)) {
 552                         mcast = priv->broadcast;
 553                         create = 0;
 554                         if (mcast->backoff > 1 &&
 555                             time_before(jiffies, mcast->delay_until)) {
 556                                 delay_until = mcast->delay_until;
 557                                 mcast = NULL;
 558                         }
 559                 }
 560                 goto out;
 561         }
 562
 563         /*
 564          * We'll never get here until the broadcast group is both allocated
 565          * and attached
 566          */
 567         list_for_each_entry(mcast, &priv->multicast_list, list) {
 568                 if (IS_ERR_OR_NULL(mcast->mc) &&
 569                     !test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags) &&
 570                     (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) ||
 571                      !skb_queue_empty(&mcast->pkt_queue))) {
 572                         if (mcast->backoff == 1 ||
 573                             time_after_eq(jiffies, mcast->delay_until)) {
 574                                 /* Found the next unjoined group */
 575                                 init_completion(&mcast->done);
 576                                 set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 577                                 if (test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags))
 578                                         create = 0;
 579                                 else
 580                                         create = 1;
 581                                 spin_unlock_irq(&priv->lock);
 582                                 ipoib_mcast_join(dev, mcast, create);
 583                                 spin_lock_irq(&priv->lock);
 584                         } else if (!delay_until ||
 585                                  time_before(mcast->delay_until, delay_until))
 586                                 delay_until = mcast->delay_until;
 587                 }
 588         }
 589
 590         mcast = NULL;
 591         ipoib_dbg_mcast(priv, "successfully started all multicast joins\n");
 592
 593 out:
 594         if (delay_until) {
 595                 cancel_delayed_work(&priv->mcast_task);
 596                 queue_delayed_work(priv->wq, &priv->mcast_task,
 597                                    delay_until - jiffies);
 598         }
 599         if (mcast) {
 600                 init_completion(&mcast->done);
 601                 set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 602         }
 603         spin_unlock_irq(&priv->lock);
 604         if (mcast)
 605                 ipoib_mcast_join(dev, mcast, create);
 606 }
 607
 608 int ipoib_mcast_start_thread(struct net_device *dev)
 609 {
 610         struct ipoib_dev_priv *priv = netdev_priv(dev);
 611         unsigned long flags;
 612
 613         ipoib_dbg_mcast(priv, "starting multicast thread\n");
 614
 615         spin_lock_irqsave(&priv->lock, flags);
 616         set_bit(IPOIB_MCAST_RUN, &priv->flags);
 617         __ipoib_mcast_schedule_join_thread(priv, NULL, 0);
 618         spin_unlock_irqrestore(&priv->lock, flags);
 619
 620         return 0;
 621 }
 622
 623 int ipoib_mcast_stop_thread(struct net_device *dev)
 624 {
 625         struct ipoib_dev_priv *priv = netdev_priv(dev);
 626         unsigned long flags;
 627
 628         ipoib_dbg_mcast(priv, "stopping multicast thread\n");
 629
 630         spin_lock_irqsave(&priv->lock, flags);
 631         clear_bit(IPOIB_MCAST_RUN, &priv->flags);
 632         cancel_delayed_work(&priv->mcast_task);
 633         spin_unlock_irqrestore(&priv->lock, flags);
 634
 635         flush_workqueue(priv->wq);
 636
 637         return 0;
 638 }
 639
 640 static int ipoib_mcast_leave(struct net_device *dev, struct ipoib_mcast *mcast)
 641 {
 642         struct ipoib_dev_priv *priv = netdev_priv(dev);
 643         int ret = 0;
 644
 645         if (test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
 646                 ipoib_warn(priv, "ipoib_mcast_leave on an in-flight join\n");
 647
 648         if (!IS_ERR_OR_NULL(mcast->mc))
 649                 ib_sa_free_multicast(mcast->mc);
 650
 651         if (test_and_clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 652                 ipoib_dbg_mcast(priv, "leaving MGID %pI6\n",
 653                                 mcast->mcmember.mgid.raw);
 654
 655                 /* Remove ourselves from the multicast group */
 656                 ret = ib_detach_mcast(priv->qp, &mcast->mcmember.mgid,
 657                                       be16_to_cpu(mcast->mcmember.mlid));
 658                 if (ret)
 659                         ipoib_warn(priv, "ib_detach_mcast failed (result = %d)\n", ret);
 660         } else if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags))
 661                 ipoib_dbg(priv, "leaving with no mcmember but not a "
 662                           "SENDONLY join\n");
 663
 664         return 0;
 665 }
 666
 667 void ipoib_mcast_send(struct net_device *dev, u8 *daddr, struct sk_buff *skb)
 668 {
 669         struct ipoib_dev_priv *priv = netdev_priv(dev);
 670         struct ipoib_mcast *mcast;
 671         unsigned long flags;
 672         void *mgid = daddr + 4;
 673
 674         spin_lock_irqsave(&priv->lock, flags);
 675
 676         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)         ||
 677             !priv->broadcast                                    ||
 678             !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
 679                 ++dev->stats.tx_dropped;
 680                 dev_kfree_skb_any(skb);
 681                 goto unlock;
 682         }
 683
 684         mcast = __ipoib_mcast_find(dev, mgid);
 685         if (!mcast || !mcast->ah) {
 686                 if (!mcast) {
 687                         /* Let's create a new send only group now */
 688                         ipoib_dbg_mcast(priv, "setting up send only multicast group for %pI6\n",
 689                                         mgid);
 690
 691                         mcast = ipoib_mcast_alloc(dev, 0);
 692                         if (!mcast) {
 693                                 ipoib_warn(priv, "unable to allocate memory "
 694                                            "for multicast structure\n");
 695                                 ++dev->stats.tx_dropped;
 696                                 dev_kfree_skb_any(skb);
 697                                 goto unlock;
 698                         }
 699
 700                         set_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags);
 701                         memcpy(mcast->mcmember.mgid.raw, mgid,
 702                                sizeof (union ib_gid));
 703                         __ipoib_mcast_add(dev, mcast);
 704                         list_add_tail(&mcast->list, &priv->multicast_list);
 705                 }
 706                 if (skb_queue_len(&mcast->pkt_queue) < IPOIB_MAX_MCAST_QUEUE)
 707                         skb_queue_tail(&mcast->pkt_queue, skb);
 708                 else {
 709                         ++dev->stats.tx_dropped;
 710                         dev_kfree_skb_any(skb);
 711                 }
 712                 if (!test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)) {
 713                         __ipoib_mcast_schedule_join_thread(priv, NULL, 0);
 714                 }
 715         } else {
 716                 struct ipoib_neigh *neigh;
 717
 718                 spin_unlock_irqrestore(&priv->lock, flags);
 719                 neigh = ipoib_neigh_get(dev, daddr);
 720                 spin_lock_irqsave(&priv->lock, flags);
 721                 if (!neigh) {
 722                         neigh = ipoib_neigh_alloc(daddr, dev);
 723                         if (neigh) {
 724                                 kref_get(&mcast->ah->ref);
 725                                 neigh->ah       = mcast->ah;
 726                                 list_add_tail(&neigh->list, &mcast->neigh_list);
 727                         }
 728                 }
 729                 spin_unlock_irqrestore(&priv->lock, flags);
 730                 ipoib_send(dev, skb, mcast->ah, IB_MULTICAST_QPN);
 731                 if (neigh)
 732                         ipoib_neigh_put(neigh);
 733                 return;
 734         }
 735
 736 unlock:
 737         spin_unlock_irqrestore(&priv->lock, flags);
 738 }
 739
 740 void ipoib_mcast_dev_flush(struct net_device *dev)
 741 {
 742         struct ipoib_dev_priv *priv = netdev_priv(dev);
 743         LIST_HEAD(remove_list);
 744         struct ipoib_mcast *mcast, *tmcast;
 745         unsigned long flags;
 746
 747         ipoib_dbg_mcast(priv, "flushing multicast list\n");
 748
 749         spin_lock_irqsave(&priv->lock, flags);
 750
 751         list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
 752                 list_del(&mcast->list);
 753                 rb_erase(&mcast->rb_node, &priv->multicast_tree);
 754                 list_add_tail(&mcast->list, &remove_list);
 755         }
 756
 757         if (priv->broadcast) {
 758                 rb_erase(&priv->broadcast->rb_node, &priv->multicast_tree);
 759                 list_add_tail(&priv->broadcast->list, &remove_list);
 760                 priv->broadcast = NULL;
 761         }
 762
 763         spin_unlock_irqrestore(&priv->lock, flags);
 764
 765         /*
 766          * make sure the in-flight joins have finished before we attempt
 767          * to leave
 768          */
 769         list_for_each_entry_safe(mcast, tmcast, &remove_list, list)
 770                 if (test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
 771                         wait_for_completion(&mcast->done);
 772
 773         list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
 774                 ipoib_mcast_leave(dev, mcast);
 775                 ipoib_mcast_free(mcast);
 776         }
 777 }
 778
 779 static int ipoib_mcast_addr_is_valid(const u8 *addr, const u8 *broadcast)
 780 {
 781         /* reserved QPN, prefix, scope */
 782         if (memcmp(addr, broadcast, 6))
 783                 return 0;
 784         /* signature lower, pkey */
 785         if (memcmp(addr + 7, broadcast + 7, 3))
 786                 return 0;
 787         return 1;
 788 }
 789
 790 void ipoib_mcast_restart_task(struct work_struct *work)
 791 {
 792         struct ipoib_dev_priv *priv =
 793                 container_of(work, struct ipoib_dev_priv, restart_task);
 794         struct net_device *dev = priv->dev;
 795         struct netdev_hw_addr *ha;
 796         struct ipoib_mcast *mcast, *tmcast;
 797         LIST_HEAD(remove_list);
 798         unsigned long flags;
 799         struct ib_sa_mcmember_rec rec;
 800
 801         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
 802                 /*
 803                  * shortcut...on shutdown flush is called next, just
 804                  * let it do all the work
 805                  */
 806                 return;
 807
 808         ipoib_dbg_mcast(priv, "restarting multicast task\n");
 809
 810         local_irq_save(flags);
 811         netif_addr_lock(dev);
 812         spin_lock(&priv->lock);
 813
 814         /*
 815          * Unfortunately, the networking core only gives us a list of all of
 816          * the multicast hardware addresses. We need to figure out which ones
 817          * are new and which ones have been removed
 818          */
 819
 820         /* Clear out the found flag */
 821         list_for_each_entry(mcast, &priv->multicast_list, list)
 822                 clear_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
 823
 824         /* Mark all of the entries that are found or don't exist */
 825         netdev_for_each_mc_addr(ha, dev) {
 826                 union ib_gid mgid;
 827
 828                 if (!ipoib_mcast_addr_is_valid(ha->addr, dev->broadcast))
 829                         continue;
 830
 831                 memcpy(mgid.raw, ha->addr + 4, sizeof mgid);
 832
 833                 mcast = __ipoib_mcast_find(dev, &mgid);
 834                 if (!mcast || test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 835                         struct ipoib_mcast *nmcast;
 836
 837                         /* ignore group which is directly joined by userspace */
 838                         if (test_bit(IPOIB_FLAG_UMCAST, &priv->flags) &&
 839                             !ib_sa_get_mcmember_rec(priv->ca, priv->port, &mgid, &rec)) {
 840                                 ipoib_dbg_mcast(priv, "ignoring multicast entry for mgid %pI6\n",
 841                                                 mgid.raw);
 842                                 continue;
 843                         }
 844
 845                         /* Not found or send-only group, let's add a new entry */
 846                         ipoib_dbg_mcast(priv, "adding multicast entry for mgid %pI6\n",
 847                                         mgid.raw);
 848
 849                         nmcast = ipoib_mcast_alloc(dev, 0);
 850                         if (!nmcast) {
 851                                 ipoib_warn(priv, "unable to allocate memory for multicast structure\n");
 852                                 continue;
 853                         }
 854
 855                         set_bit(IPOIB_MCAST_FLAG_FOUND, &nmcast->flags);
 856
 857                         nmcast->mcmember.mgid = mgid;
 858
 859                         if (mcast) {
 860                                 /* Destroy the send only entry */
 861                                 list_move_tail(&mcast->list, &remove_list);
 862
 863                                 rb_replace_node(&mcast->rb_node,
 864                                                 &nmcast->rb_node,
 865                                                 &priv->multicast_tree);
 866                         } else
 867                                 __ipoib_mcast_add(dev, nmcast);
 868
 869                         list_add_tail(&nmcast->list, &priv->multicast_list);
 870                 }
 871
 872                 if (mcast)
 873                         set_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
 874         }
 875
 876         /* Remove all of the entries don't exist anymore */
 877         list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
 878                 if (!test_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags) &&
 879                     !test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 880                         ipoib_dbg_mcast(priv, "deleting multicast group %pI6\n",
 881                                         mcast->mcmember.mgid.raw);
 882
 883                         rb_erase(&mcast->rb_node, &priv->multicast_tree);
 884
 885                         /* Move to the remove list */
 886                         list_move_tail(&mcast->list, &remove_list);
 887                 }
 888         }
 889
 890         spin_unlock(&priv->lock);
 891         netif_addr_unlock(dev);
 892         local_irq_restore(flags);
 893
 894         /*
 895          * make sure the in-flight joins have finished before we attempt
 896          * to leave
 897          */
 898         list_for_each_entry_safe(mcast, tmcast, &remove_list, list)
 899                 if (test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
 900                         wait_for_completion(&mcast->done);
 901
 902         list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
 903                 ipoib_mcast_leave(mcast->dev, mcast);
 904                 ipoib_mcast_free(mcast);
 905         }
 906
 907         /*
 908          * Double check that we are still up
 909          */
 910         if (test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
 911                 spin_lock_irqsave(&priv->lock, flags);
 912                 __ipoib_mcast_schedule_join_thread(priv, NULL, 0);
 913                 spin_unlock_irqrestore(&priv->lock, flags);
 914         }
 915 }
 916
 917 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 918
 919 struct ipoib_mcast_iter *ipoib_mcast_iter_init(struct net_device *dev)
 920 {
 921         struct ipoib_mcast_iter *iter;
 922
 923         iter = kmalloc(sizeof *iter, GFP_KERNEL);
 924         if (!iter)
 925                 return NULL;
 926
 927         iter->dev = dev;
 928         memset(iter->mgid.raw, 0, 16);
 929
 930         if (ipoib_mcast_iter_next(iter)) {
 931                 kfree(iter);
 932                 return NULL;
 933         }
 934
 935         return iter;
 936 }
 937
 938 int ipoib_mcast_iter_next(struct ipoib_mcast_iter *iter)
 939 {
 940         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
 941         struct rb_node *n;
 942         struct ipoib_mcast *mcast;
 943         int ret = 1;
 944
 945         spin_lock_irq(&priv->lock);
 946
 947         n = rb_first(&priv->multicast_tree);
 948
 949         while (n) {
 950                 mcast = rb_entry(n, struct ipoib_mcast, rb_node);
 951
 952                 if (memcmp(iter->mgid.raw, mcast->mcmember.mgid.raw,
 953                            sizeof (union ib_gid)) < 0) {
 954                         iter->mgid      = mcast->mcmember.mgid;
 955                         iter->created   = mcast->created;
 956                         iter->queuelen  = skb_queue_len(&mcast->pkt_queue);
 957                         iter->complete  = !!mcast->ah;
 958                         iter->send_only = !!(mcast->flags & (1 << IPOIB_MCAST_FLAG_SENDONLY));
 959
 960                         ret = 0;
 961
 962                         break;
 963                 }
 964
 965                 n = rb_next(n);
 966         }
 967
 968         spin_unlock_irq(&priv->lock);
 969
 970         return ret;
 971 }
 972
 973 void ipoib_mcast_iter_read(struct ipoib_mcast_iter *iter,
 974                            union ib_gid *mgid,
 975                            unsigned long *created,
 976                            unsigned int *queuelen,
 977                            unsigned int *complete,
 978                            unsigned int *send_only)
 979 {
 980         *mgid      = iter->mgid;
 981         *created   = iter->created;
 982         *queuelen  = iter->queuelen;
 983         *complete  = iter->complete;
 984         *send_only = iter->send_only;
 985 }
 986
 987 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */