kernel/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_FLAG_OPER_UP, &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 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 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                 /*update priv member according to the new mcast*/
 230                 priv->broadcast->mcmember.qkey = mcmember->qkey;
 231                 priv->broadcast->mcmember.mtu = mcmember->mtu;
 232                 priv->broadcast->mcmember.traffic_class = mcmember->traffic_class;
 233                 priv->broadcast->mcmember.rate = mcmember->rate;
 234                 priv->broadcast->mcmember.sl = mcmember->sl;
 235                 priv->broadcast->mcmember.flow_label = mcmember->flow_label;
 236                 priv->broadcast->mcmember.hop_limit = mcmember->hop_limit;
 237                 /* assume if the admin and the mcast are the same both can be changed */
 238                 if (priv->mcast_mtu == priv->admin_mtu)
 239                         priv->admin_mtu =
 240                         priv->mcast_mtu =
 241                         IPOIB_UD_MTU(ib_mtu_enum_to_int(priv->broadcast->mcmember.mtu));
 242                 else
 243                         priv->mcast_mtu =
 244                         IPOIB_UD_MTU(ib_mtu_enum_to_int(priv->broadcast->mcmember.mtu));
 245
 246                 priv->qkey = be32_to_cpu(priv->broadcast->mcmember.qkey);
 247                 spin_unlock_irq(&priv->lock);
 248                 priv->tx_wr.remote_qkey = priv->qkey;
 249                 set_qkey = 1;
 250         }
 251
 252         if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 253                 if (test_and_set_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 254                         ipoib_warn(priv, "multicast group %pI6 already attached\n",
 255                                    mcast->mcmember.mgid.raw);
 256
 257                         return 0;
 258                 }
 259
 260                 ret = ipoib_mcast_attach(dev, be16_to_cpu(mcast->mcmember.mlid),
 261                                          &mcast->mcmember.mgid, set_qkey);
 262                 if (ret < 0) {
 263                         ipoib_warn(priv, "couldn't attach QP to multicast group %pI6\n",
 264                                    mcast->mcmember.mgid.raw);
 265
 266                         clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags);
 267                         return ret;
 268                 }
 269         }
 270
 271         {
 272                 struct ib_ah_attr av = {
 273                         .dlid          = be16_to_cpu(mcast->mcmember.mlid),
 274                         .port_num      = priv->port,
 275                         .sl            = mcast->mcmember.sl,
 276                         .ah_flags      = IB_AH_GRH,
 277                         .static_rate   = mcast->mcmember.rate,
 278                         .grh           = {
 279                                 .flow_label    = be32_to_cpu(mcast->mcmember.flow_label),
 280                                 .hop_limit     = mcast->mcmember.hop_limit,
 281                                 .sgid_index    = 0,
 282                                 .traffic_class = mcast->mcmember.traffic_class
 283                         }
 284                 };
 285                 av.grh.dgid = mcast->mcmember.mgid;
 286
 287                 ah = ipoib_create_ah(dev, priv->pd, &av);
 288                 if (IS_ERR(ah)) {
 289                         ipoib_warn(priv, "ib_address_create failed %ld\n",
 290                                 -PTR_ERR(ah));
 291                         /* use original error */
 292                         return PTR_ERR(ah);
 293                 } else {
 294                         spin_lock_irq(&priv->lock);
 295                         mcast->ah = ah;
 296                         spin_unlock_irq(&priv->lock);
 297
 298                         ipoib_dbg_mcast(priv, "MGID %pI6 AV %p, LID 0x%04x, SL %d\n",
 299                                         mcast->mcmember.mgid.raw,
 300                                         mcast->ah->ah,
 301                                         be16_to_cpu(mcast->mcmember.mlid),
 302                                         mcast->mcmember.sl);
 303                 }
 304         }
 305
 306         /* actually send any queued packets */
 307         netif_tx_lock_bh(dev);
 308         while (!skb_queue_empty(&mcast->pkt_queue)) {
 309                 struct sk_buff *skb = skb_dequeue(&mcast->pkt_queue);
 310
 311                 netif_tx_unlock_bh(dev);
 312
 313                 skb->dev = dev;
 314                 if (dev_queue_xmit(skb))
 315                         ipoib_warn(priv, "dev_queue_xmit failed to requeue packet\n");
 316
 317                 netif_tx_lock_bh(dev);
 318         }
 319         netif_tx_unlock_bh(dev);
 320
 321         return 0;
 322 }
 323
 324 void ipoib_mcast_carrier_on_task(struct work_struct *work)
 325 {
 326         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
 327                                                    carrier_on_task);
 328         struct ib_port_attr attr;
 329
 330         if (ib_query_port(priv->ca, priv->port, &attr) ||
 331             attr.state != IB_PORT_ACTIVE) {
 332                 ipoib_dbg(priv, "Keeping carrier off until IB port is active\n");
 333                 return;
 334         }
 335
 336         /*
 337          * Take rtnl_lock to avoid racing with ipoib_stop() and
 338          * turning the carrier back on while a device is being
 339          * removed.  However, ipoib_stop() will attempt to flush
 340          * the workqueue while holding the rtnl lock, so loop
 341          * on trylock until either we get the lock or we see
 342          * FLAG_OPER_UP go away as that signals that we are bailing
 343          * and can safely ignore the carrier on work.
 344          */
 345         while (!rtnl_trylock()) {
 346                 if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
 347                         return;
 348                 else
 349                         msleep(20);
 350         }
 351         if (!ipoib_cm_admin_enabled(priv->dev))
 352                 dev_set_mtu(priv->dev, min(priv->mcast_mtu, priv->admin_mtu));
 353         netif_carrier_on(priv->dev);
 354         rtnl_unlock();
 355 }
 356
 357 static int ipoib_mcast_join_complete(int status,
 358                                      struct ib_sa_multicast *multicast)
 359 {
 360         struct ipoib_mcast *mcast = multicast->context;
 361         struct net_device *dev = mcast->dev;
 362         struct ipoib_dev_priv *priv = netdev_priv(dev);
 363
 364         ipoib_dbg_mcast(priv, "%sjoin completion for %pI6 (status %d)\n",
 365                         test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) ?
 366                         "sendonly " : "",
 367                         mcast->mcmember.mgid.raw, status);
 368
 369         /* We trap for port events ourselves. */
 370         if (status == -ENETRESET) {
 371                 status = 0;
 372                 goto out;
 373         }
 374
 375         if (!status)
 376                 status = ipoib_mcast_join_finish(mcast, &multicast->rec);
 377
 378         if (!status) {
 379                 mcast->backoff = 1;
 380                 mcast->delay_until = jiffies;
 381
 382                 /*
 383                  * Defer carrier on work to priv->wq to avoid a
 384                  * deadlock on rtnl_lock here.  Requeue our multicast
 385                  * work too, which will end up happening right after
 386                  * our carrier on task work and will allow us to
 387                  * send out all of the non-broadcast joins
 388                  */
 389                 if (mcast == priv->broadcast) {
 390                         spin_lock_irq(&priv->lock);
 391                         queue_work(priv->wq, &priv->carrier_on_task);
 392                         __ipoib_mcast_schedule_join_thread(priv, NULL, 0);
 393                         goto out_locked;
 394                 }
 395         } else {
 396                 bool silent_fail =
 397                     test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) &&
 398                     status == -EINVAL;
 399
 400                 if (mcast->logcount < 20) {
 401                         if (status == -ETIMEDOUT || status == -EAGAIN ||
 402                             silent_fail) {
 403                                 ipoib_dbg_mcast(priv, "%smulticast join failed for %pI6, status %d\n",
 404                                                 test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) ? "sendonly " : "",
 405                                                 mcast->mcmember.mgid.raw, status);
 406                         } else {
 407                                 ipoib_warn(priv, "%smulticast join failed for %pI6, status %d\n",
 408                                                 test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) ? "sendonly " : "",
 409                                            mcast->mcmember.mgid.raw, status);
 410                         }
 411
 412                         if (!silent_fail)
 413                                 mcast->logcount++;
 414                 }
 415
 416                 if (test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) &&
 417                     mcast->backoff >= 2) {
 418                         /*
 419                          * We only retry sendonly joins once before we drop
 420                          * the packet and quit trying to deal with the
 421                          * group.  However, we leave the group in the
 422                          * mcast list as an unjoined group.  If we want to
 423                          * try joining again, we simply queue up a packet
 424                          * and restart the join thread.  The empty queue
 425                          * is why the join thread ignores this group.
 426                          */
 427                         mcast->backoff = 1;
 428                         netif_tx_lock_bh(dev);
 429                         while (!skb_queue_empty(&mcast->pkt_queue)) {
 430                                 ++dev->stats.tx_dropped;
 431                                 dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
 432                         }
 433                         netif_tx_unlock_bh(dev);
 434                 } else {
 435                         spin_lock_irq(&priv->lock);
 436                         /* Requeue this join task with a backoff delay */
 437                         __ipoib_mcast_schedule_join_thread(priv, mcast, 1);
 438                         goto out_locked;
 439                 }
 440         }
 441 out:
 442         spin_lock_irq(&priv->lock);
 443 out_locked:
 444         /*
 445          * Make sure to set mcast->mc before we clear the busy flag to avoid
 446          * racing with code that checks for BUSY before checking mcast->mc
 447          */
 448         if (status)
 449                 mcast->mc = NULL;
 450         else
 451                 mcast->mc = multicast;
 452         clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 453         spin_unlock_irq(&priv->lock);
 454         complete(&mcast->done);
 455
 456         return status;
 457 }
 458
 459 /*
 460  * Caller must hold 'priv->lock'
 461  */
 462 static int ipoib_mcast_join(struct net_device *dev, struct ipoib_mcast *mcast)
 463 {
 464         struct ipoib_dev_priv *priv = netdev_priv(dev);
 465         struct ib_sa_multicast *multicast;
 466         struct ib_sa_mcmember_rec rec = {
 467                 .join_state = 1
 468         };
 469         ib_sa_comp_mask comp_mask;
 470         int ret = 0;
 471
 472         if (!priv->broadcast ||
 473             !test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
 474                 return -EINVAL;
 475
 476         ipoib_dbg_mcast(priv, "joining MGID %pI6\n", mcast->mcmember.mgid.raw);
 477
 478         rec.mgid     = mcast->mcmember.mgid;
 479         rec.port_gid = priv->local_gid;
 480         rec.pkey     = cpu_to_be16(priv->pkey);
 481
 482         comp_mask =
 483                 IB_SA_MCMEMBER_REC_MGID         |
 484                 IB_SA_MCMEMBER_REC_PORT_GID     |
 485                 IB_SA_MCMEMBER_REC_PKEY         |
 486                 IB_SA_MCMEMBER_REC_JOIN_STATE;
 487
 488         if (mcast != priv->broadcast) {
 489                 /*
 490                  * RFC 4391:
 491                  *  The MGID MUST use the same P_Key, Q_Key, SL, MTU,
 492                  *  and HopLimit as those used in the broadcast-GID.  The rest
 493                  *  of attributes SHOULD follow the values used in the
 494                  *  broadcast-GID as well.
 495                  */
 496                 comp_mask |=
 497                         IB_SA_MCMEMBER_REC_QKEY                 |
 498                         IB_SA_MCMEMBER_REC_MTU_SELECTOR         |
 499                         IB_SA_MCMEMBER_REC_MTU                  |
 500                         IB_SA_MCMEMBER_REC_TRAFFIC_CLASS        |
 501                         IB_SA_MCMEMBER_REC_RATE_SELECTOR        |
 502                         IB_SA_MCMEMBER_REC_RATE                 |
 503                         IB_SA_MCMEMBER_REC_SL                   |
 504                         IB_SA_MCMEMBER_REC_FLOW_LABEL           |
 505                         IB_SA_MCMEMBER_REC_HOP_LIMIT;
 506
 507                 rec.qkey          = priv->broadcast->mcmember.qkey;
 508                 rec.mtu_selector  = IB_SA_EQ;
 509                 rec.mtu           = priv->broadcast->mcmember.mtu;
 510                 rec.traffic_class = priv->broadcast->mcmember.traffic_class;
 511                 rec.rate_selector = IB_SA_EQ;
 512                 rec.rate          = priv->broadcast->mcmember.rate;
 513                 rec.sl            = priv->broadcast->mcmember.sl;
 514                 rec.flow_label    = priv->broadcast->mcmember.flow_label;
 515                 rec.hop_limit     = priv->broadcast->mcmember.hop_limit;
 516
 517                 /*
 518                  * Send-only IB Multicast joins do not work at the core
 519                  * IB layer yet, so we can't use them here.  However,
 520                  * we are emulating an Ethernet multicast send, which
 521                  * does not require a multicast subscription and will
 522                  * still send properly.  The most appropriate thing to
 523                  * do is to create the group if it doesn't exist as that
 524                  * most closely emulates the behavior, from a user space
 525                  * application perspecitive, of Ethernet multicast
 526                  * operation.  For now, we do a full join, maybe later
 527                  * when the core IB layers support send only joins we
 528                  * will use them.
 529                  */
 530 #if 0
 531                 if (test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags))
 532                         rec.join_state = 4;
 533 #endif
 534         }
 535         spin_unlock_irq(&priv->lock);
 536
 537         multicast = ib_sa_join_multicast(&ipoib_sa_client, priv->ca, priv->port,
 538                                          &rec, comp_mask, GFP_KERNEL,
 539                                          ipoib_mcast_join_complete, mcast);
 540         spin_lock_irq(&priv->lock);
 541         if (IS_ERR(multicast)) {
 542                 ret = PTR_ERR(multicast);
 543                 ipoib_warn(priv, "ib_sa_join_multicast failed, status %d\n", ret);
 544                 /* Requeue this join task with a backoff delay */
 545                 __ipoib_mcast_schedule_join_thread(priv, mcast, 1);
 546                 clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 547                 spin_unlock_irq(&priv->lock);
 548                 complete(&mcast->done);
 549                 spin_lock_irq(&priv->lock);
 550         }
 551         return 0;
 552 }
 553
 554 void ipoib_mcast_join_task(struct work_struct *work)
 555 {
 556         struct ipoib_dev_priv *priv =
 557                 container_of(work, struct ipoib_dev_priv, mcast_task.work);
 558         struct net_device *dev = priv->dev;
 559         struct ib_port_attr port_attr;
 560         unsigned long delay_until = 0;
 561         struct ipoib_mcast *mcast = NULL;
 562
 563         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
 564                 return;
 565
 566         if (ib_query_port(priv->ca, priv->port, &port_attr)) {
 567                 ipoib_dbg(priv, "ib_query_port() failed\n");
 568                 return;
 569         }
 570         if (port_attr.state != IB_PORT_ACTIVE) {
 571                 ipoib_dbg(priv, "port state is not ACTIVE (state = %d) suspending join task\n",
 572                           port_attr.state);
 573                 return;
 574         }
 575         priv->local_lid = port_attr.lid;
 576
 577         if (ib_query_gid(priv->ca, priv->port, 0, &priv->local_gid, NULL))
 578                 ipoib_warn(priv, "ib_query_gid() failed\n");
 579         else
 580                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
 581
 582         spin_lock_irq(&priv->lock);
 583         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
 584                 goto out;
 585
 586         if (!priv->broadcast) {
 587                 struct ipoib_mcast *broadcast;
 588
 589                 broadcast = ipoib_mcast_alloc(dev, 0);
 590                 if (!broadcast) {
 591                         ipoib_warn(priv, "failed to allocate broadcast group\n");
 592                         /*
 593                          * Restart us after a 1 second delay to retry
 594                          * creating our broadcast group and attaching to
 595                          * it.  Until this succeeds, this ipoib dev is
 596                          * completely stalled (multicast wise).
 597                          */
 598                         __ipoib_mcast_schedule_join_thread(priv, NULL, 1);
 599                         goto out;
 600                 }
 601
 602                 memcpy(broadcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
 603                        sizeof (union ib_gid));
 604                 priv->broadcast = broadcast;
 605
 606                 __ipoib_mcast_add(dev, priv->broadcast);
 607         }
 608
 609         if (!test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
 610                 if (IS_ERR_OR_NULL(priv->broadcast->mc) &&
 611                     !test_bit(IPOIB_MCAST_FLAG_BUSY, &priv->broadcast->flags)) {
 612                         mcast = priv->broadcast;
 613                         if (mcast->backoff > 1 &&
 614                             time_before(jiffies, mcast->delay_until)) {
 615                                 delay_until = mcast->delay_until;
 616                                 mcast = NULL;
 617                         }
 618                 }
 619                 goto out;
 620         }
 621
 622         /*
 623          * We'll never get here until the broadcast group is both allocated
 624          * and attached
 625          */
 626         list_for_each_entry(mcast, &priv->multicast_list, list) {
 627                 if (IS_ERR_OR_NULL(mcast->mc) &&
 628                     !test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags) &&
 629                     (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) ||
 630                      !skb_queue_empty(&mcast->pkt_queue))) {
 631                         if (mcast->backoff == 1 ||
 632                             time_after_eq(jiffies, mcast->delay_until)) {
 633                                 /* Found the next unjoined group */
 634                                 init_completion(&mcast->done);
 635                                 set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 636                                 if (ipoib_mcast_join(dev, mcast)) {
 637                                         spin_unlock_irq(&priv->lock);
 638                                         return;
 639                                 }
 640                         } else if (!delay_until ||
 641                                  time_before(mcast->delay_until, delay_until))
 642                                 delay_until = mcast->delay_until;
 643                 }
 644         }
 645
 646         mcast = NULL;
 647         ipoib_dbg_mcast(priv, "successfully started all multicast joins\n");
 648
 649 out:
 650         if (delay_until) {
 651                 cancel_delayed_work(&priv->mcast_task);
 652                 queue_delayed_work(priv->wq, &priv->mcast_task,
 653                                    delay_until - jiffies);
 654         }
 655         if (mcast) {
 656                 init_completion(&mcast->done);
 657                 set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 658                 ipoib_mcast_join(dev, mcast);
 659         }
 660         spin_unlock_irq(&priv->lock);
 661 }
 662
 663 int ipoib_mcast_start_thread(struct net_device *dev)
 664 {
 665         struct ipoib_dev_priv *priv = netdev_priv(dev);
 666         unsigned long flags;
 667
 668         ipoib_dbg_mcast(priv, "starting multicast thread\n");
 669
 670         spin_lock_irqsave(&priv->lock, flags);
 671         __ipoib_mcast_schedule_join_thread(priv, NULL, 0);
 672         spin_unlock_irqrestore(&priv->lock, flags);
 673
 674         return 0;
 675 }
 676
 677 int ipoib_mcast_stop_thread(struct net_device *dev)
 678 {
 679         struct ipoib_dev_priv *priv = netdev_priv(dev);
 680         unsigned long flags;
 681
 682         ipoib_dbg_mcast(priv, "stopping multicast thread\n");
 683
 684         spin_lock_irqsave(&priv->lock, flags);
 685         cancel_delayed_work(&priv->mcast_task);
 686         spin_unlock_irqrestore(&priv->lock, flags);
 687
 688         flush_workqueue(priv->wq);
 689
 690         return 0;
 691 }
 692
 693 int ipoib_mcast_leave(struct net_device *dev, struct ipoib_mcast *mcast)
 694 {
 695         struct ipoib_dev_priv *priv = netdev_priv(dev);
 696         int ret = 0;
 697
 698         if (test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
 699                 ipoib_warn(priv, "ipoib_mcast_leave on an in-flight join\n");
 700
 701         if (!IS_ERR_OR_NULL(mcast->mc))
 702                 ib_sa_free_multicast(mcast->mc);
 703
 704         if (test_and_clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 705                 ipoib_dbg_mcast(priv, "leaving MGID %pI6\n",
 706                                 mcast->mcmember.mgid.raw);
 707
 708                 /* Remove ourselves from the multicast group */
 709                 ret = ib_detach_mcast(priv->qp, &mcast->mcmember.mgid,
 710                                       be16_to_cpu(mcast->mcmember.mlid));
 711                 if (ret)
 712                         ipoib_warn(priv, "ib_detach_mcast failed (result = %d)\n", ret);
 713         } else if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags))
 714                 ipoib_dbg(priv, "leaving with no mcmember but not a "
 715                           "SENDONLY join\n");
 716
 717         return 0;
 718 }
 719
 720 void ipoib_mcast_send(struct net_device *dev, u8 *daddr, struct sk_buff *skb)
 721 {
 722         struct ipoib_dev_priv *priv = netdev_priv(dev);
 723         struct ipoib_mcast *mcast;
 724         unsigned long flags;
 725         void *mgid = daddr + 4;
 726
 727         spin_lock_irqsave(&priv->lock, flags);
 728
 729         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)         ||
 730             !priv->broadcast                                    ||
 731             !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
 732                 ++dev->stats.tx_dropped;
 733                 dev_kfree_skb_any(skb);
 734                 goto unlock;
 735         }
 736
 737         mcast = __ipoib_mcast_find(dev, mgid);
 738         if (!mcast || !mcast->ah) {
 739                 if (!mcast) {
 740                         /* Let's create a new send only group now */
 741                         ipoib_dbg_mcast(priv, "setting up send only multicast group for %pI6\n",
 742                                         mgid);
 743
 744                         mcast = ipoib_mcast_alloc(dev, 0);
 745                         if (!mcast) {
 746                                 ipoib_warn(priv, "unable to allocate memory "
 747                                            "for multicast structure\n");
 748                                 ++dev->stats.tx_dropped;
 749                                 dev_kfree_skb_any(skb);
 750                                 goto unlock;
 751                         }
 752
 753                         set_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags);
 754                         memcpy(mcast->mcmember.mgid.raw, mgid,
 755                                sizeof (union ib_gid));
 756                         __ipoib_mcast_add(dev, mcast);
 757                         list_add_tail(&mcast->list, &priv->multicast_list);
 758                 }
 759                 if (skb_queue_len(&mcast->pkt_queue) < IPOIB_MAX_MCAST_QUEUE) {
 760                         /* put pseudoheader back on for next time */
 761                         skb_push(skb, sizeof(struct ipoib_pseudo_header));
 762                         skb_queue_tail(&mcast->pkt_queue, skb);
 763                 } else {
 764                         ++dev->stats.tx_dropped;
 765                         dev_kfree_skb_any(skb);
 766                 }
 767                 if (!test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)) {
 768                         __ipoib_mcast_schedule_join_thread(priv, NULL, 0);
 769                 }
 770         } else {
 771                 struct ipoib_neigh *neigh;
 772
 773                 spin_unlock_irqrestore(&priv->lock, flags);
 774                 neigh = ipoib_neigh_get(dev, daddr);
 775                 spin_lock_irqsave(&priv->lock, flags);
 776                 if (!neigh) {
 777                         neigh = ipoib_neigh_alloc(daddr, dev);
 778                         if (neigh) {
 779                                 kref_get(&mcast->ah->ref);
 780                                 neigh->ah       = mcast->ah;
 781                                 list_add_tail(&neigh->list, &mcast->neigh_list);
 782                         }
 783                 }
 784                 spin_unlock_irqrestore(&priv->lock, flags);
 785                 ipoib_send(dev, skb, mcast->ah, IB_MULTICAST_QPN);
 786                 if (neigh)
 787                         ipoib_neigh_put(neigh);
 788                 return;
 789         }
 790
 791 unlock:
 792         spin_unlock_irqrestore(&priv->lock, flags);
 793 }
 794
 795 void ipoib_mcast_dev_flush(struct net_device *dev)
 796 {
 797         struct ipoib_dev_priv *priv = netdev_priv(dev);
 798         LIST_HEAD(remove_list);
 799         struct ipoib_mcast *mcast, *tmcast;
 800         unsigned long flags;
 801
 802         ipoib_dbg_mcast(priv, "flushing multicast list\n");
 803
 804         spin_lock_irqsave(&priv->lock, flags);
 805
 806         list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
 807                 list_del(&mcast->list);
 808                 rb_erase(&mcast->rb_node, &priv->multicast_tree);
 809                 list_add_tail(&mcast->list, &remove_list);
 810         }
 811
 812         if (priv->broadcast) {
 813                 rb_erase(&priv->broadcast->rb_node, &priv->multicast_tree);
 814                 list_add_tail(&priv->broadcast->list, &remove_list);
 815                 priv->broadcast = NULL;
 816         }
 817
 818         spin_unlock_irqrestore(&priv->lock, flags);
 819
 820         /*
 821          * make sure the in-flight joins have finished before we attempt
 822          * to leave
 823          */
 824         list_for_each_entry_safe(mcast, tmcast, &remove_list, list)
 825                 if (test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
 826                         wait_for_completion(&mcast->done);
 827
 828         list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
 829                 ipoib_mcast_leave(dev, mcast);
 830                 ipoib_mcast_free(mcast);
 831         }
 832 }
 833
 834 static int ipoib_mcast_addr_is_valid(const u8 *addr, const u8 *broadcast)
 835 {
 836         /* reserved QPN, prefix, scope */
 837         if (memcmp(addr, broadcast, 6))
 838                 return 0;
 839         /* signature lower, pkey */
 840         if (memcmp(addr + 7, broadcast + 7, 3))
 841                 return 0;
 842         return 1;
 843 }
 844
 845 void ipoib_mcast_restart_task(struct work_struct *work)
 846 {
 847         struct ipoib_dev_priv *priv =
 848                 container_of(work, struct ipoib_dev_priv, restart_task);
 849         struct net_device *dev = priv->dev;
 850         struct netdev_hw_addr *ha;
 851         struct ipoib_mcast *mcast, *tmcast;
 852         LIST_HEAD(remove_list);
 853         unsigned long flags;
 854         struct ib_sa_mcmember_rec rec;
 855
 856         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
 857                 /*
 858                  * shortcut...on shutdown flush is called next, just
 859                  * let it do all the work
 860                  */
 861                 return;
 862
 863         ipoib_dbg_mcast(priv, "restarting multicast task\n");
 864
 865         local_irq_save_nort(flags);
 866         netif_addr_lock(dev);
 867         spin_lock(&priv->lock);
 868
 869         /*
 870          * Unfortunately, the networking core only gives us a list of all of
 871          * the multicast hardware addresses. We need to figure out which ones
 872          * are new and which ones have been removed
 873          */
 874
 875         /* Clear out the found flag */
 876         list_for_each_entry(mcast, &priv->multicast_list, list)
 877                 clear_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
 878
 879         /* Mark all of the entries that are found or don't exist */
 880         netdev_for_each_mc_addr(ha, dev) {
 881                 union ib_gid mgid;
 882
 883                 if (!ipoib_mcast_addr_is_valid(ha->addr, dev->broadcast))
 884                         continue;
 885
 886                 memcpy(mgid.raw, ha->addr + 4, sizeof mgid);
 887
 888                 mcast = __ipoib_mcast_find(dev, &mgid);
 889                 if (!mcast || test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 890                         struct ipoib_mcast *nmcast;
 891
 892                         /* ignore group which is directly joined by userspace */
 893                         if (test_bit(IPOIB_FLAG_UMCAST, &priv->flags) &&
 894                             !ib_sa_get_mcmember_rec(priv->ca, priv->port, &mgid, &rec)) {
 895                                 ipoib_dbg_mcast(priv, "ignoring multicast entry for mgid %pI6\n",
 896                                                 mgid.raw);
 897                                 continue;
 898                         }
 899
 900                         /* Not found or send-only group, let's add a new entry */
 901                         ipoib_dbg_mcast(priv, "adding multicast entry for mgid %pI6\n",
 902                                         mgid.raw);
 903
 904                         nmcast = ipoib_mcast_alloc(dev, 0);
 905                         if (!nmcast) {
 906                                 ipoib_warn(priv, "unable to allocate memory for multicast structure\n");
 907                                 continue;
 908                         }
 909
 910                         set_bit(IPOIB_MCAST_FLAG_FOUND, &nmcast->flags);
 911
 912                         nmcast->mcmember.mgid = mgid;
 913
 914                         if (mcast) {
 915                                 /* Destroy the send only entry */
 916                                 list_move_tail(&mcast->list, &remove_list);
 917
 918                                 rb_replace_node(&mcast->rb_node,
 919                                                 &nmcast->rb_node,
 920                                                 &priv->multicast_tree);
 921                         } else
 922                                 __ipoib_mcast_add(dev, nmcast);
 923
 924                         list_add_tail(&nmcast->list, &priv->multicast_list);
 925                 }
 926
 927                 if (mcast)
 928                         set_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
 929         }
 930
 931         /* Remove all of the entries don't exist anymore */
 932         list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
 933                 if (!test_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags) &&
 934                     !test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 935                         ipoib_dbg_mcast(priv, "deleting multicast group %pI6\n",
 936                                         mcast->mcmember.mgid.raw);
 937
 938                         rb_erase(&mcast->rb_node, &priv->multicast_tree);
 939
 940                         /* Move to the remove list */
 941                         list_move_tail(&mcast->list, &remove_list);
 942                 }
 943         }
 944
 945         spin_unlock(&priv->lock);
 946         netif_addr_unlock(dev);
 947         local_irq_restore_nort(flags);
 948
 949         /*
 950          * make sure the in-flight joins have finished before we attempt
 951          * to leave
 952          */
 953         list_for_each_entry_safe(mcast, tmcast, &remove_list, list)
 954                 if (test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
 955                         wait_for_completion(&mcast->done);
 956
 957         list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
 958                 ipoib_mcast_leave(mcast->dev, mcast);
 959                 ipoib_mcast_free(mcast);
 960         }
 961
 962         /*
 963          * Double check that we are still up
 964          */
 965         if (test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
 966                 spin_lock_irqsave(&priv->lock, flags);
 967                 __ipoib_mcast_schedule_join_thread(priv, NULL, 0);
 968                 spin_unlock_irqrestore(&priv->lock, flags);
 969         }
 970 }
 971
 972 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 973
 974 struct ipoib_mcast_iter *ipoib_mcast_iter_init(struct net_device *dev)
 975 {
 976         struct ipoib_mcast_iter *iter;
 977
 978         iter = kmalloc(sizeof *iter, GFP_KERNEL);
 979         if (!iter)
 980                 return NULL;
 981
 982         iter->dev = dev;
 983         memset(iter->mgid.raw, 0, 16);
 984
 985         if (ipoib_mcast_iter_next(iter)) {
 986                 kfree(iter);
 987                 return NULL;
 988         }
 989
 990         return iter;
 991 }
 992
 993 int ipoib_mcast_iter_next(struct ipoib_mcast_iter *iter)
 994 {
 995         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
 996         struct rb_node *n;
 997         struct ipoib_mcast *mcast;
 998         int ret = 1;
 999
1000         spin_lock_irq(&priv->lock);
1001
1002         n = rb_first(&priv->multicast_tree);
1003
1004         while (n) {
1005                 mcast = rb_entry(n, struct ipoib_mcast, rb_node);
1006
1007                 if (memcmp(iter->mgid.raw, mcast->mcmember.mgid.raw,
1008                            sizeof (union ib_gid)) < 0) {
1009                         iter->mgid      = mcast->mcmember.mgid;
1010                         iter->created   = mcast->created;
1011                         iter->queuelen  = skb_queue_len(&mcast->pkt_queue);
1012                         iter->complete  = !!mcast->ah;
1013                         iter->send_only = !!(mcast->flags & (1 << IPOIB_MCAST_FLAG_SENDONLY));
1014
1015                         ret = 0;
1016
1017                         break;
1018                 }
1019
1020                 n = rb_next(n);
1021         }
1022
1023         spin_unlock_irq(&priv->lock);
1024
1025         return ret;
1026 }
1027
1028 void ipoib_mcast_iter_read(struct ipoib_mcast_iter *iter,
1029                            union ib_gid *mgid,
1030                            unsigned long *created,
1031                            unsigned int *queuelen,
1032                            unsigned int *complete,
1033                            unsigned int *send_only)
1034 {
1035         *mgid      = iter->mgid;
1036         *created   = iter->created;
1037         *queuelen  = iter->queuelen;
1038         *complete  = iter->complete;
1039         *send_only = iter->send_only;
1040 }
1041
1042 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */