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 /* join state that allows creating mcg with sendonly member request */
  68 #define SENDONLY_FULLMEMBER_JOIN        8
  69
  70 /*
  71  * This should be called with the priv->lock held
  72  */
  73 static void __ipoib_mcast_schedule_join_thread(struct ipoib_dev_priv *priv,
  74                                                struct ipoib_mcast *mcast,
  75                                                bool delay)
  76 {
  77         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
  78                 return;
  79
  80         /*
  81          * We will be scheduling *something*, so cancel whatever is
  82          * currently scheduled first
  83          */
  84         cancel_delayed_work(&priv->mcast_task);
  85         if (mcast && delay) {
  86                 /*
  87                  * We had a failure and want to schedule a retry later
  88                  */
  89                 mcast->backoff *= 2;
  90                 if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
  91                         mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
  92                 mcast->delay_until = jiffies + (mcast->backoff * HZ);
  93                 /*
  94                  * Mark this mcast for its delay, but restart the
  95                  * task immediately.  The join task will make sure to
  96                  * clear out all entries without delays, and then
  97                  * schedule itself to run again when the earliest
  98                  * delay expires
  99                  */
 100                 queue_delayed_work(priv->wq, &priv->mcast_task, 0);
 101         } else if (delay) {
 102                 /*
 103                  * Special case of retrying after a failure to
 104                  * allocate the broadcast multicast group, wait
 105                  * 1 second and try again
 106                  */
 107                 queue_delayed_work(priv->wq, &priv->mcast_task, HZ);
 108         } else
 109                 queue_delayed_work(priv->wq, &priv->mcast_task, 0);
 110 }
 111
 112 static void ipoib_mcast_free(struct ipoib_mcast *mcast)
 113 {
 114         struct net_device *dev = mcast->dev;
 115         int tx_dropped = 0;
 116
 117         ipoib_dbg_mcast(netdev_priv(dev), "deleting multicast group %pI6\n",
 118                         mcast->mcmember.mgid.raw);
 119
 120         /* remove all neigh connected to this mcast */
 121         ipoib_del_neighs_by_gid(dev, mcast->mcmember.mgid.raw);
 122
 123         if (mcast->ah)
 124                 ipoib_put_ah(mcast->ah);
 125
 126         while (!skb_queue_empty(&mcast->pkt_queue)) {
 127                 ++tx_dropped;
 128                 dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
 129         }
 130
 131         netif_tx_lock_bh(dev);
 132         dev->stats.tx_dropped += tx_dropped;
 133         netif_tx_unlock_bh(dev);
 134
 135         kfree(mcast);
 136 }
 137
 138 static struct ipoib_mcast *ipoib_mcast_alloc(struct net_device *dev,
 139                                              int can_sleep)
 140 {
 141         struct ipoib_mcast *mcast;
 142
 143         mcast = kzalloc(sizeof *mcast, can_sleep ? GFP_KERNEL : GFP_ATOMIC);
 144         if (!mcast)
 145                 return NULL;
 146
 147         mcast->dev = dev;
 148         mcast->created = jiffies;
 149         mcast->delay_until = jiffies;
 150         mcast->backoff = 1;
 151
 152         INIT_LIST_HEAD(&mcast->list);
 153         INIT_LIST_HEAD(&mcast->neigh_list);
 154         skb_queue_head_init(&mcast->pkt_queue);
 155
 156         return mcast;
 157 }
 158
 159 static struct ipoib_mcast *__ipoib_mcast_find(struct net_device *dev, void *mgid)
 160 {
 161         struct ipoib_dev_priv *priv = netdev_priv(dev);
 162         struct rb_node *n = priv->multicast_tree.rb_node;
 163
 164         while (n) {
 165                 struct ipoib_mcast *mcast;
 166                 int ret;
 167
 168                 mcast = rb_entry(n, struct ipoib_mcast, rb_node);
 169
 170                 ret = memcmp(mgid, mcast->mcmember.mgid.raw,
 171                              sizeof (union ib_gid));
 172                 if (ret < 0)
 173                         n = n->rb_left;
 174                 else if (ret > 0)
 175                         n = n->rb_right;
 176                 else
 177                         return mcast;
 178         }
 179
 180         return NULL;
 181 }
 182
 183 static int __ipoib_mcast_add(struct net_device *dev, struct ipoib_mcast *mcast)
 184 {
 185         struct ipoib_dev_priv *priv = netdev_priv(dev);
 186         struct rb_node **n = &priv->multicast_tree.rb_node, *pn = NULL;
 187
 188         while (*n) {
 189                 struct ipoib_mcast *tmcast;
 190                 int ret;
 191
 192                 pn = *n;
 193                 tmcast = rb_entry(pn, struct ipoib_mcast, rb_node);
 194
 195                 ret = memcmp(mcast->mcmember.mgid.raw, tmcast->mcmember.mgid.raw,
 196                              sizeof (union ib_gid));
 197                 if (ret < 0)
 198                         n = &pn->rb_left;
 199                 else if (ret > 0)
 200                         n = &pn->rb_right;
 201                 else
 202                         return -EEXIST;
 203         }
 204
 205         rb_link_node(&mcast->rb_node, pn, n);
 206         rb_insert_color(&mcast->rb_node, &priv->multicast_tree);
 207
 208         return 0;
 209 }
 210
 211 static int ipoib_mcast_join_finish(struct ipoib_mcast *mcast,
 212                                    struct ib_sa_mcmember_rec *mcmember)
 213 {
 214         struct net_device *dev = mcast->dev;
 215         struct ipoib_dev_priv *priv = netdev_priv(dev);
 216         struct ipoib_ah *ah;
 217         int ret;
 218         int set_qkey = 0;
 219
 220         mcast->mcmember = *mcmember;
 221
 222         /* Set the multicast MTU and cached Q_Key before we attach if it's
 223          * the broadcast group.
 224          */
 225         if (!memcmp(mcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
 226                     sizeof (union ib_gid))) {
 227                 spin_lock_irq(&priv->lock);
 228                 if (!priv->broadcast) {
 229                         spin_unlock_irq(&priv->lock);
 230                         return -EAGAIN;
 231                 }
 232                 /*update priv member according to the new mcast*/
 233                 priv->broadcast->mcmember.qkey = mcmember->qkey;
 234                 priv->broadcast->mcmember.mtu = mcmember->mtu;
 235                 priv->broadcast->mcmember.traffic_class = mcmember->traffic_class;
 236                 priv->broadcast->mcmember.rate = mcmember->rate;
 237                 priv->broadcast->mcmember.sl = mcmember->sl;
 238                 priv->broadcast->mcmember.flow_label = mcmember->flow_label;
 239                 priv->broadcast->mcmember.hop_limit = mcmember->hop_limit;
 240                 /* assume if the admin and the mcast are the same both can be changed */
 241                 if (priv->mcast_mtu == priv->admin_mtu)
 242                         priv->admin_mtu =
 243                         priv->mcast_mtu =
 244                         IPOIB_UD_MTU(ib_mtu_enum_to_int(priv->broadcast->mcmember.mtu));
 245                 else
 246                         priv->mcast_mtu =
 247                         IPOIB_UD_MTU(ib_mtu_enum_to_int(priv->broadcast->mcmember.mtu));
 248
 249                 priv->qkey = be32_to_cpu(priv->broadcast->mcmember.qkey);
 250                 spin_unlock_irq(&priv->lock);
 251                 priv->tx_wr.remote_qkey = priv->qkey;
 252                 set_qkey = 1;
 253         }
 254
 255         if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 256                 if (test_and_set_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 257                         ipoib_warn(priv, "multicast group %pI6 already attached\n",
 258                                    mcast->mcmember.mgid.raw);
 259
 260                         return 0;
 261                 }
 262
 263                 ret = ipoib_mcast_attach(dev, be16_to_cpu(mcast->mcmember.mlid),
 264                                          &mcast->mcmember.mgid, set_qkey);
 265                 if (ret < 0) {
 266                         ipoib_warn(priv, "couldn't attach QP to multicast group %pI6\n",
 267                                    mcast->mcmember.mgid.raw);
 268
 269                         clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags);
 270                         return ret;
 271                 }
 272         }
 273
 274         {
 275                 struct ib_ah_attr av = {
 276                         .dlid          = be16_to_cpu(mcast->mcmember.mlid),
 277                         .port_num      = priv->port,
 278                         .sl            = mcast->mcmember.sl,
 279                         .ah_flags      = IB_AH_GRH,
 280                         .static_rate   = mcast->mcmember.rate,
 281                         .grh           = {
 282                                 .flow_label    = be32_to_cpu(mcast->mcmember.flow_label),
 283                                 .hop_limit     = mcast->mcmember.hop_limit,
 284                                 .sgid_index    = 0,
 285                                 .traffic_class = mcast->mcmember.traffic_class
 286                         }
 287                 };
 288                 av.grh.dgid = mcast->mcmember.mgid;
 289
 290                 ah = ipoib_create_ah(dev, priv->pd, &av);
 291                 if (IS_ERR(ah)) {
 292                         ipoib_warn(priv, "ib_address_create failed %ld\n",
 293                                 -PTR_ERR(ah));
 294                         /* use original error */
 295                         return PTR_ERR(ah);
 296                 } else {
 297                         spin_lock_irq(&priv->lock);
 298                         mcast->ah = ah;
 299                         spin_unlock_irq(&priv->lock);
 300
 301                         ipoib_dbg_mcast(priv, "MGID %pI6 AV %p, LID 0x%04x, SL %d\n",
 302                                         mcast->mcmember.mgid.raw,
 303                                         mcast->ah->ah,
 304                                         be16_to_cpu(mcast->mcmember.mlid),
 305                                         mcast->mcmember.sl);
 306                 }
 307         }
 308
 309         /* actually send any queued packets */
 310         netif_tx_lock_bh(dev);
 311         while (!skb_queue_empty(&mcast->pkt_queue)) {
 312                 struct sk_buff *skb = skb_dequeue(&mcast->pkt_queue);
 313
 314                 netif_tx_unlock_bh(dev);
 315
 316                 skb->dev = dev;
 317
 318                 ret = dev_queue_xmit(skb);
 319                 if (ret)
 320                         ipoib_warn(priv, "%s:dev_queue_xmit failed to re-queue packet, ret:%d\n",
 321                                    __func__, ret);
 322                 netif_tx_lock_bh(dev);
 323         }
 324         netif_tx_unlock_bh(dev);
 325
 326         return 0;
 327 }
 328
 329 void ipoib_mcast_carrier_on_task(struct work_struct *work)
 330 {
 331         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
 332                                                    carrier_on_task);
 333         struct ib_port_attr attr;
 334         int ret;
 335
 336         if (ib_query_port(priv->ca, priv->port, &attr) ||
 337             attr.state != IB_PORT_ACTIVE) {
 338                 ipoib_dbg(priv, "Keeping carrier off until IB port is active\n");
 339                 return;
 340         }
 341         /*
 342          * Check if can send sendonly MCG's with sendonly-fullmember join state.
 343          * It done here after the successfully join to the broadcast group,
 344          * because the broadcast group must always be joined first and is always
 345          * re-joined if the SM changes substantially.
 346          */
 347         ret = ipoib_check_sm_sendonly_fullmember_support(priv);
 348         if (ret < 0)
 349                 pr_debug("%s failed query sm support for sendonly-fullmember (ret: %d)\n",
 350                          priv->dev->name, ret);
 351
 352         /*
 353          * Take rtnl_lock to avoid racing with ipoib_stop() and
 354          * turning the carrier back on while a device is being
 355          * removed.  However, ipoib_stop() will attempt to flush
 356          * the workqueue while holding the rtnl lock, so loop
 357          * on trylock until either we get the lock or we see
 358          * FLAG_OPER_UP go away as that signals that we are bailing
 359          * and can safely ignore the carrier on work.
 360          */
 361         while (!rtnl_trylock()) {
 362                 if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
 363                         return;
 364                 else
 365                         msleep(20);
 366         }
 367         if (!ipoib_cm_admin_enabled(priv->dev))
 368                 dev_set_mtu(priv->dev, min(priv->mcast_mtu, priv->admin_mtu));
 369         netif_carrier_on(priv->dev);
 370         rtnl_unlock();
 371 }
 372
 373 static int ipoib_mcast_join_complete(int status,
 374                                      struct ib_sa_multicast *multicast)
 375 {
 376         struct ipoib_mcast *mcast = multicast->context;
 377         struct net_device *dev = mcast->dev;
 378         struct ipoib_dev_priv *priv = netdev_priv(dev);
 379
 380         ipoib_dbg_mcast(priv, "%sjoin completion for %pI6 (status %d)\n",
 381                         test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) ?
 382                         "sendonly " : "",
 383                         mcast->mcmember.mgid.raw, status);
 384
 385         /* We trap for port events ourselves. */
 386         if (status == -ENETRESET) {
 387                 status = 0;
 388                 goto out;
 389         }
 390
 391         if (!status)
 392                 status = ipoib_mcast_join_finish(mcast, &multicast->rec);
 393
 394         if (!status) {
 395                 mcast->backoff = 1;
 396                 mcast->delay_until = jiffies;
 397
 398                 /*
 399                  * Defer carrier on work to priv->wq to avoid a
 400                  * deadlock on rtnl_lock here.  Requeue our multicast
 401                  * work too, which will end up happening right after
 402                  * our carrier on task work and will allow us to
 403                  * send out all of the non-broadcast joins
 404                  */
 405                 if (mcast == priv->broadcast) {
 406                         spin_lock_irq(&priv->lock);
 407                         queue_work(priv->wq, &priv->carrier_on_task);
 408                         __ipoib_mcast_schedule_join_thread(priv, NULL, 0);
 409                         goto out_locked;
 410                 }
 411         } else {
 412                 bool silent_fail =
 413                     test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) &&
 414                     status == -EINVAL;
 415
 416                 if (mcast->logcount < 20) {
 417                         if (status == -ETIMEDOUT || status == -EAGAIN ||
 418                             silent_fail) {
 419                                 ipoib_dbg_mcast(priv, "%smulticast join failed for %pI6, status %d\n",
 420                                                 test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) ? "sendonly " : "",
 421                                                 mcast->mcmember.mgid.raw, status);
 422                         } else {
 423                                 ipoib_warn(priv, "%smulticast join failed for %pI6, status %d\n",
 424                                                 test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) ? "sendonly " : "",
 425                                            mcast->mcmember.mgid.raw, status);
 426                         }
 427
 428                         if (!silent_fail)
 429                                 mcast->logcount++;
 430                 }
 431
 432                 if (test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) &&
 433                     mcast->backoff >= 2) {
 434                         /*
 435                          * We only retry sendonly joins once before we drop
 436                          * the packet and quit trying to deal with the
 437                          * group.  However, we leave the group in the
 438                          * mcast list as an unjoined group.  If we want to
 439                          * try joining again, we simply queue up a packet
 440                          * and restart the join thread.  The empty queue
 441                          * is why the join thread ignores this group.
 442                          */
 443                         mcast->backoff = 1;
 444                         netif_tx_lock_bh(dev);
 445                         while (!skb_queue_empty(&mcast->pkt_queue)) {
 446                                 ++dev->stats.tx_dropped;
 447                                 dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
 448                         }
 449                         netif_tx_unlock_bh(dev);
 450                 } else {
 451                         spin_lock_irq(&priv->lock);
 452                         /* Requeue this join task with a backoff delay */
 453                         __ipoib_mcast_schedule_join_thread(priv, mcast, 1);
 454                         goto out_locked;
 455                 }
 456         }
 457 out:
 458         spin_lock_irq(&priv->lock);
 459 out_locked:
 460         /*
 461          * Make sure to set mcast->mc before we clear the busy flag to avoid
 462          * racing with code that checks for BUSY before checking mcast->mc
 463          */
 464         if (status)
 465                 mcast->mc = NULL;
 466         else
 467                 mcast->mc = multicast;
 468         clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 469         spin_unlock_irq(&priv->lock);
 470         complete(&mcast->done);
 471
 472         return status;
 473 }
 474
 475 /*
 476  * Caller must hold 'priv->lock'
 477  */
 478 static int ipoib_mcast_join(struct net_device *dev, struct ipoib_mcast *mcast)
 479 {
 480         struct ipoib_dev_priv *priv = netdev_priv(dev);
 481         struct ib_sa_multicast *multicast;
 482         struct ib_sa_mcmember_rec rec = {
 483                 .join_state = 1
 484         };
 485         ib_sa_comp_mask comp_mask;
 486         int ret = 0;
 487
 488         if (!priv->broadcast ||
 489             !test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
 490                 return -EINVAL;
 491
 492         ipoib_dbg_mcast(priv, "joining MGID %pI6\n", mcast->mcmember.mgid.raw);
 493
 494         rec.mgid     = mcast->mcmember.mgid;
 495         rec.port_gid = priv->local_gid;
 496         rec.pkey     = cpu_to_be16(priv->pkey);
 497
 498         comp_mask =
 499                 IB_SA_MCMEMBER_REC_MGID         |
 500                 IB_SA_MCMEMBER_REC_PORT_GID     |
 501                 IB_SA_MCMEMBER_REC_PKEY         |
 502                 IB_SA_MCMEMBER_REC_JOIN_STATE;
 503
 504         if (mcast != priv->broadcast) {
 505                 /*
 506                  * RFC 4391:
 507                  *  The MGID MUST use the same P_Key, Q_Key, SL, MTU,
 508                  *  and HopLimit as those used in the broadcast-GID.  The rest
 509                  *  of attributes SHOULD follow the values used in the
 510                  *  broadcast-GID as well.
 511                  */
 512                 comp_mask |=
 513                         IB_SA_MCMEMBER_REC_QKEY                 |
 514                         IB_SA_MCMEMBER_REC_MTU_SELECTOR         |
 515                         IB_SA_MCMEMBER_REC_MTU                  |
 516                         IB_SA_MCMEMBER_REC_TRAFFIC_CLASS        |
 517                         IB_SA_MCMEMBER_REC_RATE_SELECTOR        |
 518                         IB_SA_MCMEMBER_REC_RATE                 |
 519                         IB_SA_MCMEMBER_REC_SL                   |
 520                         IB_SA_MCMEMBER_REC_FLOW_LABEL           |
 521                         IB_SA_MCMEMBER_REC_HOP_LIMIT;
 522
 523                 rec.qkey          = priv->broadcast->mcmember.qkey;
 524                 rec.mtu_selector  = IB_SA_EQ;
 525                 rec.mtu           = priv->broadcast->mcmember.mtu;
 526                 rec.traffic_class = priv->broadcast->mcmember.traffic_class;
 527                 rec.rate_selector = IB_SA_EQ;
 528                 rec.rate          = priv->broadcast->mcmember.rate;
 529                 rec.sl            = priv->broadcast->mcmember.sl;
 530                 rec.flow_label    = priv->broadcast->mcmember.flow_label;
 531                 rec.hop_limit     = priv->broadcast->mcmember.hop_limit;
 532
 533                 /*
 534                  * Send-only IB Multicast joins work at the core IB layer but
 535                  * require specific SM support.
 536                  * We can use such joins here only if the current SM supports that feature.
 537                  * However, if not, we emulate an Ethernet multicast send,
 538                  * which does not require a multicast subscription and will
 539                  * still send properly. The most appropriate thing to
 540                  * do is to create the group if it doesn't exist as that
 541                  * most closely emulates the behavior, from a user space
 542                  * application perspective, of Ethernet multicast operation.
 543                  */
 544                 if (test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) &&
 545                     priv->sm_fullmember_sendonly_support)
 546                         /* SM supports sendonly-fullmember, otherwise fallback to full-member */
 547                         rec.join_state = SENDONLY_FULLMEMBER_JOIN;
 548         }
 549         spin_unlock_irq(&priv->lock);
 550
 551         multicast = ib_sa_join_multicast(&ipoib_sa_client, priv->ca, priv->port,
 552                                          &rec, comp_mask, GFP_KERNEL,
 553                                          ipoib_mcast_join_complete, mcast);
 554         spin_lock_irq(&priv->lock);
 555         if (IS_ERR(multicast)) {
 556                 ret = PTR_ERR(multicast);
 557                 ipoib_warn(priv, "ib_sa_join_multicast failed, status %d\n", ret);
 558                 /* Requeue this join task with a backoff delay */
 559                 __ipoib_mcast_schedule_join_thread(priv, mcast, 1);
 560                 clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 561                 spin_unlock_irq(&priv->lock);
 562                 complete(&mcast->done);
 563                 spin_lock_irq(&priv->lock);
 564         }
 565         return 0;
 566 }
 567
 568 void ipoib_mcast_join_task(struct work_struct *work)
 569 {
 570         struct ipoib_dev_priv *priv =
 571                 container_of(work, struct ipoib_dev_priv, mcast_task.work);
 572         struct net_device *dev = priv->dev;
 573         struct ib_port_attr port_attr;
 574         unsigned long delay_until = 0;
 575         struct ipoib_mcast *mcast = NULL;
 576
 577         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
 578                 return;
 579
 580         if (ib_query_port(priv->ca, priv->port, &port_attr)) {
 581                 ipoib_dbg(priv, "ib_query_port() failed\n");
 582                 return;
 583         }
 584         if (port_attr.state != IB_PORT_ACTIVE) {
 585                 ipoib_dbg(priv, "port state is not ACTIVE (state = %d) suspending join task\n",
 586                           port_attr.state);
 587                 return;
 588         }
 589         priv->local_lid = port_attr.lid;
 590         netif_addr_lock_bh(dev);
 591
 592         if (!test_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags)) {
 593                 netif_addr_unlock_bh(dev);
 594                 return;
 595         }
 596         netif_addr_unlock_bh(dev);
 597
 598         spin_lock_irq(&priv->lock);
 599         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
 600                 goto out;
 601
 602         if (!priv->broadcast) {
 603                 struct ipoib_mcast *broadcast;
 604
 605                 broadcast = ipoib_mcast_alloc(dev, 0);
 606                 if (!broadcast) {
 607                         ipoib_warn(priv, "failed to allocate broadcast group\n");
 608                         /*
 609                          * Restart us after a 1 second delay to retry
 610                          * creating our broadcast group and attaching to
 611                          * it.  Until this succeeds, this ipoib dev is
 612                          * completely stalled (multicast wise).
 613                          */
 614                         __ipoib_mcast_schedule_join_thread(priv, NULL, 1);
 615                         goto out;
 616                 }
 617
 618                 memcpy(broadcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
 619                        sizeof (union ib_gid));
 620                 priv->broadcast = broadcast;
 621
 622                 __ipoib_mcast_add(dev, priv->broadcast);
 623         }
 624
 625         if (!test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
 626                 if (IS_ERR_OR_NULL(priv->broadcast->mc) &&
 627                     !test_bit(IPOIB_MCAST_FLAG_BUSY, &priv->broadcast->flags)) {
 628                         mcast = priv->broadcast;
 629                         if (mcast->backoff > 1 &&
 630                             time_before(jiffies, mcast->delay_until)) {
 631                                 delay_until = mcast->delay_until;
 632                                 mcast = NULL;
 633                         }
 634                 }
 635                 goto out;
 636         }
 637
 638         /*
 639          * We'll never get here until the broadcast group is both allocated
 640          * and attached
 641          */
 642         list_for_each_entry(mcast, &priv->multicast_list, list) {
 643                 if (IS_ERR_OR_NULL(mcast->mc) &&
 644                     !test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags) &&
 645                     (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) ||
 646                      !skb_queue_empty(&mcast->pkt_queue))) {
 647                         if (mcast->backoff == 1 ||
 648                             time_after_eq(jiffies, mcast->delay_until)) {
 649                                 /* Found the next unjoined group */
 650                                 init_completion(&mcast->done);
 651                                 set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 652                                 if (ipoib_mcast_join(dev, mcast)) {
 653                                         spin_unlock_irq(&priv->lock);
 654                                         return;
 655                                 }
 656                         } else if (!delay_until ||
 657                                  time_before(mcast->delay_until, delay_until))
 658                                 delay_until = mcast->delay_until;
 659                 }
 660         }
 661
 662         mcast = NULL;
 663         ipoib_dbg_mcast(priv, "successfully started all multicast joins\n");
 664
 665 out:
 666         if (delay_until) {
 667                 cancel_delayed_work(&priv->mcast_task);
 668                 queue_delayed_work(priv->wq, &priv->mcast_task,
 669                                    delay_until - jiffies);
 670         }
 671         if (mcast) {
 672                 init_completion(&mcast->done);
 673                 set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 674                 ipoib_mcast_join(dev, mcast);
 675         }
 676         spin_unlock_irq(&priv->lock);
 677 }
 678
 679 void ipoib_mcast_start_thread(struct net_device *dev)
 680 {
 681         struct ipoib_dev_priv *priv = netdev_priv(dev);
 682         unsigned long flags;
 683
 684         ipoib_dbg_mcast(priv, "starting multicast thread\n");
 685
 686         spin_lock_irqsave(&priv->lock, flags);
 687         __ipoib_mcast_schedule_join_thread(priv, NULL, 0);
 688         spin_unlock_irqrestore(&priv->lock, flags);
 689 }
 690
 691 int ipoib_mcast_stop_thread(struct net_device *dev)
 692 {
 693         struct ipoib_dev_priv *priv = netdev_priv(dev);
 694         unsigned long flags;
 695
 696         ipoib_dbg_mcast(priv, "stopping multicast thread\n");
 697
 698         spin_lock_irqsave(&priv->lock, flags);
 699         cancel_delayed_work(&priv->mcast_task);
 700         spin_unlock_irqrestore(&priv->lock, flags);
 701
 702         flush_workqueue(priv->wq);
 703
 704         return 0;
 705 }
 706
 707 static int ipoib_mcast_leave(struct net_device *dev, struct ipoib_mcast *mcast)
 708 {
 709         struct ipoib_dev_priv *priv = netdev_priv(dev);
 710         int ret = 0;
 711
 712         if (test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
 713                 ipoib_warn(priv, "ipoib_mcast_leave on an in-flight join\n");
 714
 715         if (!IS_ERR_OR_NULL(mcast->mc))
 716                 ib_sa_free_multicast(mcast->mc);
 717
 718         if (test_and_clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 719                 ipoib_dbg_mcast(priv, "leaving MGID %pI6\n",
 720                                 mcast->mcmember.mgid.raw);
 721
 722                 /* Remove ourselves from the multicast group */
 723                 ret = ib_detach_mcast(priv->qp, &mcast->mcmember.mgid,
 724                                       be16_to_cpu(mcast->mcmember.mlid));
 725                 if (ret)
 726                         ipoib_warn(priv, "ib_detach_mcast failed (result = %d)\n", ret);
 727         } else if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags))
 728                 ipoib_dbg(priv, "leaving with no mcmember but not a "
 729                           "SENDONLY join\n");
 730
 731         return 0;
 732 }
 733
 734 /*
 735  * Check if the multicast group is sendonly. If so remove it from the maps
 736  * and add to the remove list
 737  */
 738 void ipoib_check_and_add_mcast_sendonly(struct ipoib_dev_priv *priv, u8 *mgid,
 739                                 struct list_head *remove_list)
 740 {
 741         /* Is this multicast ? */
 742         if (*mgid == 0xff) {
 743                 struct ipoib_mcast *mcast = __ipoib_mcast_find(priv->dev, mgid);
 744
 745                 if (mcast && test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 746                         list_del(&mcast->list);
 747                         rb_erase(&mcast->rb_node, &priv->multicast_tree);
 748                         list_add_tail(&mcast->list, remove_list);
 749                 }
 750         }
 751 }
 752
 753 void ipoib_mcast_remove_list(struct list_head *remove_list)
 754 {
 755         struct ipoib_mcast *mcast, *tmcast;
 756
 757         list_for_each_entry_safe(mcast, tmcast, remove_list, list) {
 758                 ipoib_mcast_leave(mcast->dev, mcast);
 759                 ipoib_mcast_free(mcast);
 760         }
 761 }
 762
 763 void ipoib_mcast_send(struct net_device *dev, u8 *daddr, struct sk_buff *skb)
 764 {
 765         struct ipoib_dev_priv *priv = netdev_priv(dev);
 766         struct ipoib_mcast *mcast;
 767         unsigned long flags;
 768         void *mgid = daddr + 4;
 769
 770         spin_lock_irqsave(&priv->lock, flags);
 771
 772         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)         ||
 773             !priv->broadcast                                    ||
 774             !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
 775                 ++dev->stats.tx_dropped;
 776                 dev_kfree_skb_any(skb);
 777                 goto unlock;
 778         }
 779
 780         mcast = __ipoib_mcast_find(dev, mgid);
 781         if (!mcast || !mcast->ah) {
 782                 if (!mcast) {
 783                         /* Let's create a new send only group now */
 784                         ipoib_dbg_mcast(priv, "setting up send only multicast group for %pI6\n",
 785                                         mgid);
 786
 787                         mcast = ipoib_mcast_alloc(dev, 0);
 788                         if (!mcast) {
 789                                 ipoib_warn(priv, "unable to allocate memory "
 790                                            "for multicast structure\n");
 791                                 ++dev->stats.tx_dropped;
 792                                 dev_kfree_skb_any(skb);
 793                                 goto unlock;
 794                         }
 795
 796                         set_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags);
 797                         memcpy(mcast->mcmember.mgid.raw, mgid,
 798                                sizeof (union ib_gid));
 799                         __ipoib_mcast_add(dev, mcast);
 800                         list_add_tail(&mcast->list, &priv->multicast_list);
 801                 }
 802                 if (skb_queue_len(&mcast->pkt_queue) < IPOIB_MAX_MCAST_QUEUE) {
 803                         /* put pseudoheader back on for next time */
 804                         skb_push(skb, sizeof(struct ipoib_pseudo_header));
 805                         skb_queue_tail(&mcast->pkt_queue, skb);
 806                 } else {
 807                         ++dev->stats.tx_dropped;
 808                         dev_kfree_skb_any(skb);
 809                 }
 810                 if (!test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)) {
 811                         __ipoib_mcast_schedule_join_thread(priv, NULL, 0);
 812                 }
 813         } else {
 814                 struct ipoib_neigh *neigh;
 815
 816                 spin_unlock_irqrestore(&priv->lock, flags);
 817                 neigh = ipoib_neigh_get(dev, daddr);
 818                 spin_lock_irqsave(&priv->lock, flags);
 819                 if (!neigh) {
 820                         neigh = ipoib_neigh_alloc(daddr, dev);
 821                         if (neigh) {
 822                                 kref_get(&mcast->ah->ref);
 823                                 neigh->ah       = mcast->ah;
 824                                 list_add_tail(&neigh->list, &mcast->neigh_list);
 825                         }
 826                 }
 827                 spin_unlock_irqrestore(&priv->lock, flags);
 828                 ipoib_send(dev, skb, mcast->ah, IB_MULTICAST_QPN);
 829                 if (neigh)
 830                         ipoib_neigh_put(neigh);
 831                 return;
 832         }
 833
 834 unlock:
 835         spin_unlock_irqrestore(&priv->lock, flags);
 836 }
 837
 838 void ipoib_mcast_dev_flush(struct net_device *dev)
 839 {
 840         struct ipoib_dev_priv *priv = netdev_priv(dev);
 841         LIST_HEAD(remove_list);
 842         struct ipoib_mcast *mcast, *tmcast;
 843         unsigned long flags;
 844
 845         ipoib_dbg_mcast(priv, "flushing multicast list\n");
 846
 847         spin_lock_irqsave(&priv->lock, flags);
 848
 849         list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
 850                 list_del(&mcast->list);
 851                 rb_erase(&mcast->rb_node, &priv->multicast_tree);
 852                 list_add_tail(&mcast->list, &remove_list);
 853         }
 854
 855         if (priv->broadcast) {
 856                 rb_erase(&priv->broadcast->rb_node, &priv->multicast_tree);
 857                 list_add_tail(&priv->broadcast->list, &remove_list);
 858                 priv->broadcast = NULL;
 859         }
 860
 861         spin_unlock_irqrestore(&priv->lock, flags);
 862
 863         /*
 864          * make sure the in-flight joins have finished before we attempt
 865          * to leave
 866          */
 867         list_for_each_entry_safe(mcast, tmcast, &remove_list, list)
 868                 if (test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
 869                         wait_for_completion(&mcast->done);
 870
 871         ipoib_mcast_remove_list(&remove_list);
 872 }
 873
 874 static int ipoib_mcast_addr_is_valid(const u8 *addr, const u8 *broadcast)
 875 {
 876         /* reserved QPN, prefix, scope */
 877         if (memcmp(addr, broadcast, 6))
 878                 return 0;
 879         /* signature lower, pkey */
 880         if (memcmp(addr + 7, broadcast + 7, 3))
 881                 return 0;
 882         return 1;
 883 }
 884
 885 void ipoib_mcast_restart_task(struct work_struct *work)
 886 {
 887         struct ipoib_dev_priv *priv =
 888                 container_of(work, struct ipoib_dev_priv, restart_task);
 889         struct net_device *dev = priv->dev;
 890         struct netdev_hw_addr *ha;
 891         struct ipoib_mcast *mcast, *tmcast;
 892         LIST_HEAD(remove_list);
 893         unsigned long flags;
 894         struct ib_sa_mcmember_rec rec;
 895
 896         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
 897                 /*
 898                  * shortcut...on shutdown flush is called next, just
 899                  * let it do all the work
 900                  */
 901                 return;
 902
 903         ipoib_dbg_mcast(priv, "restarting multicast task\n");
 904
 905         local_irq_save(flags);
 906         netif_addr_lock(dev);
 907         spin_lock(&priv->lock);
 908
 909         /*
 910          * Unfortunately, the networking core only gives us a list of all of
 911          * the multicast hardware addresses. We need to figure out which ones
 912          * are new and which ones have been removed
 913          */
 914
 915         /* Clear out the found flag */
 916         list_for_each_entry(mcast, &priv->multicast_list, list)
 917                 clear_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
 918
 919         /* Mark all of the entries that are found or don't exist */
 920         netdev_for_each_mc_addr(ha, dev) {
 921                 union ib_gid mgid;
 922
 923                 if (!ipoib_mcast_addr_is_valid(ha->addr, dev->broadcast))
 924                         continue;
 925
 926                 memcpy(mgid.raw, ha->addr + 4, sizeof mgid);
 927
 928                 mcast = __ipoib_mcast_find(dev, &mgid);
 929                 if (!mcast || test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 930                         struct ipoib_mcast *nmcast;
 931
 932                         /* ignore group which is directly joined by userspace */
 933                         if (test_bit(IPOIB_FLAG_UMCAST, &priv->flags) &&
 934                             !ib_sa_get_mcmember_rec(priv->ca, priv->port, &mgid, &rec)) {
 935                                 ipoib_dbg_mcast(priv, "ignoring multicast entry for mgid %pI6\n",
 936                                                 mgid.raw);
 937                                 continue;
 938                         }
 939
 940                         /* Not found or send-only group, let's add a new entry */
 941                         ipoib_dbg_mcast(priv, "adding multicast entry for mgid %pI6\n",
 942                                         mgid.raw);
 943
 944                         nmcast = ipoib_mcast_alloc(dev, 0);
 945                         if (!nmcast) {
 946                                 ipoib_warn(priv, "unable to allocate memory for multicast structure\n");
 947                                 continue;
 948                         }
 949
 950                         set_bit(IPOIB_MCAST_FLAG_FOUND, &nmcast->flags);
 951
 952                         nmcast->mcmember.mgid = mgid;
 953
 954                         if (mcast) {
 955                                 /* Destroy the send only entry */
 956                                 list_move_tail(&mcast->list, &remove_list);
 957
 958                                 rb_replace_node(&mcast->rb_node,
 959                                                 &nmcast->rb_node,
 960                                                 &priv->multicast_tree);
 961                         } else
 962                                 __ipoib_mcast_add(dev, nmcast);
 963
 964                         list_add_tail(&nmcast->list, &priv->multicast_list);
 965                 }
 966
 967                 if (mcast)
 968                         set_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
 969         }
 970
 971         /* Remove all of the entries don't exist anymore */
 972         list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
 973                 if (!test_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags) &&
 974                     !test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 975                         ipoib_dbg_mcast(priv, "deleting multicast group %pI6\n",
 976                                         mcast->mcmember.mgid.raw);
 977
 978                         rb_erase(&mcast->rb_node, &priv->multicast_tree);
 979
 980                         /* Move to the remove list */
 981                         list_move_tail(&mcast->list, &remove_list);
 982                 }
 983         }
 984
 985         spin_unlock(&priv->lock);
 986         netif_addr_unlock(dev);
 987         local_irq_restore(flags);
 988
 989         /*
 990          * make sure the in-flight joins have finished before we attempt
 991          * to leave
 992          */
 993         list_for_each_entry_safe(mcast, tmcast, &remove_list, list)
 994                 if (test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
 995                         wait_for_completion(&mcast->done);
 996
 997         ipoib_mcast_remove_list(&remove_list);
 998
 999         /*
1000          * Double check that we are still up
1001          */
1002         if (test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
1003                 spin_lock_irqsave(&priv->lock, flags);
1004                 __ipoib_mcast_schedule_join_thread(priv, NULL, 0);
1005                 spin_unlock_irqrestore(&priv->lock, flags);
1006         }
1007 }
1008
1009 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
1010
1011 struct ipoib_mcast_iter *ipoib_mcast_iter_init(struct net_device *dev)
1012 {
1013         struct ipoib_mcast_iter *iter;
1014
1015         iter = kmalloc(sizeof *iter, GFP_KERNEL);
1016         if (!iter)
1017                 return NULL;
1018
1019         iter->dev = dev;
1020         memset(iter->mgid.raw, 0, 16);
1021
1022         if (ipoib_mcast_iter_next(iter)) {
1023                 kfree(iter);
1024                 return NULL;
1025         }
1026
1027         return iter;
1028 }
1029
1030 int ipoib_mcast_iter_next(struct ipoib_mcast_iter *iter)
1031 {
1032         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
1033         struct rb_node *n;
1034         struct ipoib_mcast *mcast;
1035         int ret = 1;
1036
1037         spin_lock_irq(&priv->lock);
1038
1039         n = rb_first(&priv->multicast_tree);
1040
1041         while (n) {
1042                 mcast = rb_entry(n, struct ipoib_mcast, rb_node);
1043
1044                 if (memcmp(iter->mgid.raw, mcast->mcmember.mgid.raw,
1045                            sizeof (union ib_gid)) < 0) {
1046                         iter->mgid      = mcast->mcmember.mgid;
1047                         iter->created   = mcast->created;
1048                         iter->queuelen  = skb_queue_len(&mcast->pkt_queue);
1049                         iter->complete  = !!mcast->ah;
1050                         iter->send_only = !!(mcast->flags & (1 << IPOIB_MCAST_FLAG_SENDONLY));
1051
1052                         ret = 0;
1053
1054                         break;
1055                 }
1056
1057                 n = rb_next(n);
1058         }
1059
1060         spin_unlock_irq(&priv->lock);
1061
1062         return ret;
1063 }
1064
1065 void ipoib_mcast_iter_read(struct ipoib_mcast_iter *iter,
1066                            union ib_gid *mgid,
1067                            unsigned long *created,
1068                            unsigned int *queuelen,
1069                            unsigned int *complete,
1070                            unsigned int *send_only)
1071 {
1072         *mgid      = iter->mgid;
1073         *created   = iter->created;
1074         *queuelen  = iter->queuelen;
1075         *complete  = iter->complete;
1076         *send_only = iter->send_only;
1077 }
1078
1079 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */