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(ipoib_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 = ipoib_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 = ipoib_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 = ipoib_priv(dev);
 216         struct rdma_netdev *rn = netdev_priv(dev);
 217         struct ipoib_ah *ah;
 218         int ret;
 219         int set_qkey = 0;
 220
 221         mcast->mcmember = *mcmember;
 222
 223         /* Set the multicast MTU and cached Q_Key before we attach if it's
 224          * the broadcast group.
 225          */
 226         if (!memcmp(mcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
 227                     sizeof (union ib_gid))) {
 228                 spin_lock_irq(&priv->lock);
 229                 if (!priv->broadcast) {
 230                         spin_unlock_irq(&priv->lock);
 231                         return -EAGAIN;
 232                 }
 233                 /*update priv member according to the new mcast*/
 234                 priv->broadcast->mcmember.qkey = mcmember->qkey;
 235                 priv->broadcast->mcmember.mtu = mcmember->mtu;
 236                 priv->broadcast->mcmember.traffic_class = mcmember->traffic_class;
 237                 priv->broadcast->mcmember.rate = mcmember->rate;
 238                 priv->broadcast->mcmember.sl = mcmember->sl;
 239                 priv->broadcast->mcmember.flow_label = mcmember->flow_label;
 240                 priv->broadcast->mcmember.hop_limit = mcmember->hop_limit;
 241                 /* assume if the admin and the mcast are the same both can be changed */
 242                 if (priv->mcast_mtu == priv->admin_mtu)
 243                         priv->admin_mtu =
 244                         priv->mcast_mtu =
 245                         IPOIB_UD_MTU(ib_mtu_enum_to_int(priv->broadcast->mcmember.mtu));
 246                 else
 247                         priv->mcast_mtu =
 248                         IPOIB_UD_MTU(ib_mtu_enum_to_int(priv->broadcast->mcmember.mtu));
 249
 250                 priv->qkey = be32_to_cpu(priv->broadcast->mcmember.qkey);
 251                 spin_unlock_irq(&priv->lock);
 252                 priv->tx_wr.remote_qkey = priv->qkey;
 253                 set_qkey = 1;
 254         }
 255
 256         if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 257                 if (test_and_set_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 258                         ipoib_warn(priv, "multicast group %pI6 already attached\n",
 259                                    mcast->mcmember.mgid.raw);
 260
 261                         return 0;
 262                 }
 263
 264                 ret = rn->attach_mcast(dev, priv->ca, &mcast->mcmember.mgid,
 265                                        be16_to_cpu(mcast->mcmember.mlid),
 266                                        set_qkey, priv->qkey);
 267                 if (ret < 0) {
 268                         ipoib_warn(priv, "couldn't attach QP to multicast group %pI6\n",
 269                                    mcast->mcmember.mgid.raw);
 270
 271                         clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags);
 272                         return ret;
 273                 }
 274         }
 275
 276         {
 277                 struct ib_ah_attr av = {
 278                         .dlid          = be16_to_cpu(mcast->mcmember.mlid),
 279                         .port_num      = priv->port,
 280                         .sl            = mcast->mcmember.sl,
 281                         .ah_flags      = IB_AH_GRH,
 282                         .static_rate   = mcast->mcmember.rate,
 283                         .grh           = {
 284                                 .flow_label    = be32_to_cpu(mcast->mcmember.flow_label),
 285                                 .hop_limit     = mcast->mcmember.hop_limit,
 286                                 .sgid_index    = 0,
 287                                 .traffic_class = mcast->mcmember.traffic_class
 288                         }
 289                 };
 290                 av.grh.dgid = mcast->mcmember.mgid;
 291
 292                 ah = ipoib_create_ah(dev, priv->pd, &av);
 293                 if (IS_ERR(ah)) {
 294                         ipoib_warn(priv, "ib_address_create failed %ld\n",
 295                                 -PTR_ERR(ah));
 296                         /* use original error */
 297                         return PTR_ERR(ah);
 298                 } else {
 299                         spin_lock_irq(&priv->lock);
 300                         mcast->ah = ah;
 301                         spin_unlock_irq(&priv->lock);
 302
 303                         ipoib_dbg_mcast(priv, "MGID %pI6 AV %p, LID 0x%04x, SL %d\n",
 304                                         mcast->mcmember.mgid.raw,
 305                                         mcast->ah->ah,
 306                                         be16_to_cpu(mcast->mcmember.mlid),
 307                                         mcast->mcmember.sl);
 308                 }
 309         }
 310
 311         /* actually send any queued packets */
 312         netif_tx_lock_bh(dev);
 313         while (!skb_queue_empty(&mcast->pkt_queue)) {
 314                 struct sk_buff *skb = skb_dequeue(&mcast->pkt_queue);
 315
 316                 netif_tx_unlock_bh(dev);
 317
 318                 skb->dev = dev;
 319
 320                 ret = dev_queue_xmit(skb);
 321                 if (ret)
 322                         ipoib_warn(priv, "%s:dev_queue_xmit failed to re-queue packet, ret:%d\n",
 323                                    __func__, ret);
 324                 netif_tx_lock_bh(dev);
 325         }
 326         netif_tx_unlock_bh(dev);
 327
 328         return 0;
 329 }
 330
 331 void ipoib_mcast_carrier_on_task(struct work_struct *work)
 332 {
 333         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
 334                                                    carrier_on_task);
 335         struct ib_port_attr attr;
 336         int ret;
 337
 338         if (ib_query_port(priv->ca, priv->port, &attr) ||
 339             attr.state != IB_PORT_ACTIVE) {
 340                 ipoib_dbg(priv, "Keeping carrier off until IB port is active\n");
 341                 return;
 342         }
 343         /*
 344          * Check if can send sendonly MCG's with sendonly-fullmember join state.
 345          * It done here after the successfully join to the broadcast group,
 346          * because the broadcast group must always be joined first and is always
 347          * re-joined if the SM changes substantially.
 348          */
 349         ret = ipoib_check_sm_sendonly_fullmember_support(priv);
 350         if (ret < 0)
 351                 pr_debug("%s failed query sm support for sendonly-fullmember (ret: %d)\n",
 352                          priv->dev->name, ret);
 353
 354         /*
 355          * Take rtnl_lock to avoid racing with ipoib_stop() and
 356          * turning the carrier back on while a device is being
 357          * removed.  However, ipoib_stop() will attempt to flush
 358          * the workqueue while holding the rtnl lock, so loop
 359          * on trylock until either we get the lock or we see
 360          * FLAG_OPER_UP go away as that signals that we are bailing
 361          * and can safely ignore the carrier on work.
 362          */
 363         while (!rtnl_trylock()) {
 364                 if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
 365                         return;
 366                 else
 367                         msleep(20);
 368         }
 369         if (!ipoib_cm_admin_enabled(priv->dev))
 370                 dev_set_mtu(priv->dev, min(priv->mcast_mtu, priv->admin_mtu));
 371         netif_carrier_on(priv->dev);
 372         rtnl_unlock();
 373 }
 374
 375 static int ipoib_mcast_join_complete(int status,
 376                                      struct ib_sa_multicast *multicast)
 377 {
 378         struct ipoib_mcast *mcast = multicast->context;
 379         struct net_device *dev = mcast->dev;
 380         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 381
 382         ipoib_dbg_mcast(priv, "%sjoin completion for %pI6 (status %d)\n",
 383                         test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) ?
 384                         "sendonly " : "",
 385                         mcast->mcmember.mgid.raw, status);
 386
 387         /* We trap for port events ourselves. */
 388         if (status == -ENETRESET) {
 389                 status = 0;
 390                 goto out;
 391         }
 392
 393         if (!status)
 394                 status = ipoib_mcast_join_finish(mcast, &multicast->rec);
 395
 396         if (!status) {
 397                 mcast->backoff = 1;
 398                 mcast->delay_until = jiffies;
 399
 400                 /*
 401                  * Defer carrier on work to priv->wq to avoid a
 402                  * deadlock on rtnl_lock here.  Requeue our multicast
 403                  * work too, which will end up happening right after
 404                  * our carrier on task work and will allow us to
 405                  * send out all of the non-broadcast joins
 406                  */
 407                 if (mcast == priv->broadcast) {
 408                         spin_lock_irq(&priv->lock);
 409                         queue_work(priv->wq, &priv->carrier_on_task);
 410                         __ipoib_mcast_schedule_join_thread(priv, NULL, 0);
 411                         goto out_locked;
 412                 }
 413         } else {
 414                 bool silent_fail =
 415                     test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) &&
 416                     status == -EINVAL;
 417
 418                 if (mcast->logcount < 20) {
 419                         if (status == -ETIMEDOUT || status == -EAGAIN ||
 420                             silent_fail) {
 421                                 ipoib_dbg_mcast(priv, "%smulticast join failed for %pI6, status %d\n",
 422                                                 test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) ? "sendonly " : "",
 423                                                 mcast->mcmember.mgid.raw, status);
 424                         } else {
 425                                 ipoib_warn(priv, "%smulticast join failed for %pI6, status %d\n",
 426                                                 test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) ? "sendonly " : "",
 427                                            mcast->mcmember.mgid.raw, status);
 428                         }
 429
 430                         if (!silent_fail)
 431                                 mcast->logcount++;
 432                 }
 433
 434                 if (test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) &&
 435                     mcast->backoff >= 2) {
 436                         /*
 437                          * We only retry sendonly joins once before we drop
 438                          * the packet and quit trying to deal with the
 439                          * group.  However, we leave the group in the
 440                          * mcast list as an unjoined group.  If we want to
 441                          * try joining again, we simply queue up a packet
 442                          * and restart the join thread.  The empty queue
 443                          * is why the join thread ignores this group.
 444                          */
 445                         mcast->backoff = 1;
 446                         netif_tx_lock_bh(dev);
 447                         while (!skb_queue_empty(&mcast->pkt_queue)) {
 448                                 ++dev->stats.tx_dropped;
 449                                 dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
 450                         }
 451                         netif_tx_unlock_bh(dev);
 452                 } else {
 453                         spin_lock_irq(&priv->lock);
 454                         /* Requeue this join task with a backoff delay */
 455                         __ipoib_mcast_schedule_join_thread(priv, mcast, 1);
 456                         goto out_locked;
 457                 }
 458         }
 459 out:
 460         spin_lock_irq(&priv->lock);
 461 out_locked:
 462         /*
 463          * Make sure to set mcast->mc before we clear the busy flag to avoid
 464          * racing with code that checks for BUSY before checking mcast->mc
 465          */
 466         if (status)
 467                 mcast->mc = NULL;
 468         else
 469                 mcast->mc = multicast;
 470         clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 471         spin_unlock_irq(&priv->lock);
 472         complete(&mcast->done);
 473
 474         return status;
 475 }
 476
 477 /*
 478  * Caller must hold 'priv->lock'
 479  */
 480 static int ipoib_mcast_join(struct net_device *dev, struct ipoib_mcast *mcast)
 481 {
 482         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 483         struct ib_sa_multicast *multicast;
 484         struct ib_sa_mcmember_rec rec = {
 485                 .join_state = 1
 486         };
 487         ib_sa_comp_mask comp_mask;
 488         int ret = 0;
 489
 490         if (!priv->broadcast ||
 491             !test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
 492                 return -EINVAL;
 493
 494         ipoib_dbg_mcast(priv, "joining MGID %pI6\n", mcast->mcmember.mgid.raw);
 495
 496         rec.mgid     = mcast->mcmember.mgid;
 497         rec.port_gid = priv->local_gid;
 498         rec.pkey     = cpu_to_be16(priv->pkey);
 499
 500         comp_mask =
 501                 IB_SA_MCMEMBER_REC_MGID         |
 502                 IB_SA_MCMEMBER_REC_PORT_GID     |
 503                 IB_SA_MCMEMBER_REC_PKEY         |
 504                 IB_SA_MCMEMBER_REC_JOIN_STATE;
 505
 506         if (mcast != priv->broadcast) {
 507                 /*
 508                  * RFC 4391:
 509                  *  The MGID MUST use the same P_Key, Q_Key, SL, MTU,
 510                  *  and HopLimit as those used in the broadcast-GID.  The rest
 511                  *  of attributes SHOULD follow the values used in the
 512                  *  broadcast-GID as well.
 513                  */
 514                 comp_mask |=
 515                         IB_SA_MCMEMBER_REC_QKEY                 |
 516                         IB_SA_MCMEMBER_REC_MTU_SELECTOR         |
 517                         IB_SA_MCMEMBER_REC_MTU                  |
 518                         IB_SA_MCMEMBER_REC_TRAFFIC_CLASS        |
 519                         IB_SA_MCMEMBER_REC_RATE_SELECTOR        |
 520                         IB_SA_MCMEMBER_REC_RATE                 |
 521                         IB_SA_MCMEMBER_REC_SL                   |
 522                         IB_SA_MCMEMBER_REC_FLOW_LABEL           |
 523                         IB_SA_MCMEMBER_REC_HOP_LIMIT;
 524
 525                 rec.qkey          = priv->broadcast->mcmember.qkey;
 526                 rec.mtu_selector  = IB_SA_EQ;
 527                 rec.mtu           = priv->broadcast->mcmember.mtu;
 528                 rec.traffic_class = priv->broadcast->mcmember.traffic_class;
 529                 rec.rate_selector = IB_SA_EQ;
 530                 rec.rate          = priv->broadcast->mcmember.rate;
 531                 rec.sl            = priv->broadcast->mcmember.sl;
 532                 rec.flow_label    = priv->broadcast->mcmember.flow_label;
 533                 rec.hop_limit     = priv->broadcast->mcmember.hop_limit;
 534
 535                 /*
 536                  * Send-only IB Multicast joins work at the core IB layer but
 537                  * require specific SM support.
 538                  * We can use such joins here only if the current SM supports that feature.
 539                  * However, if not, we emulate an Ethernet multicast send,
 540                  * which does not require a multicast subscription and will
 541                  * still send properly. The most appropriate thing to
 542                  * do is to create the group if it doesn't exist as that
 543                  * most closely emulates the behavior, from a user space
 544                  * application perspective, of Ethernet multicast operation.
 545                  */
 546                 if (test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) &&
 547                     priv->sm_fullmember_sendonly_support)
 548                         /* SM supports sendonly-fullmember, otherwise fallback to full-member */
 549                         rec.join_state = SENDONLY_FULLMEMBER_JOIN;
 550         }
 551         spin_unlock_irq(&priv->lock);
 552
 553         multicast = ib_sa_join_multicast(&ipoib_sa_client, priv->ca, priv->port,
 554                                          &rec, comp_mask, GFP_KERNEL,
 555                                          ipoib_mcast_join_complete, mcast);
 556         spin_lock_irq(&priv->lock);
 557         if (IS_ERR(multicast)) {
 558                 ret = PTR_ERR(multicast);
 559                 ipoib_warn(priv, "ib_sa_join_multicast failed, status %d\n", ret);
 560                 /* Requeue this join task with a backoff delay */
 561                 __ipoib_mcast_schedule_join_thread(priv, mcast, 1);
 562                 clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 563                 spin_unlock_irq(&priv->lock);
 564                 complete(&mcast->done);
 565                 spin_lock_irq(&priv->lock);
 566         }
 567         return 0;
 568 }
 569
 570 void ipoib_mcast_join_task(struct work_struct *work)
 571 {
 572         struct ipoib_dev_priv *priv =
 573                 container_of(work, struct ipoib_dev_priv, mcast_task.work);
 574         struct net_device *dev = priv->dev;
 575         struct ib_port_attr port_attr;
 576         unsigned long delay_until = 0;
 577         struct ipoib_mcast *mcast = NULL;
 578
 579         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
 580                 return;
 581
 582         if (ib_query_port(priv->ca, priv->port, &port_attr)) {
 583                 ipoib_dbg(priv, "ib_query_port() failed\n");
 584                 return;
 585         }
 586         if (port_attr.state != IB_PORT_ACTIVE) {
 587                 ipoib_dbg(priv, "port state is not ACTIVE (state = %d) suspending join task\n",
 588                           port_attr.state);
 589                 return;
 590         }
 591         priv->local_lid = port_attr.lid;
 592         netif_addr_lock_bh(dev);
 593
 594         if (!test_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags)) {
 595                 netif_addr_unlock_bh(dev);
 596                 return;
 597         }
 598         netif_addr_unlock_bh(dev);
 599
 600         spin_lock_irq(&priv->lock);
 601         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
 602                 goto out;
 603
 604         if (!priv->broadcast) {
 605                 struct ipoib_mcast *broadcast;
 606
 607                 broadcast = ipoib_mcast_alloc(dev, 0);
 608                 if (!broadcast) {
 609                         ipoib_warn(priv, "failed to allocate broadcast group\n");
 610                         /*
 611                          * Restart us after a 1 second delay to retry
 612                          * creating our broadcast group and attaching to
 613                          * it.  Until this succeeds, this ipoib dev is
 614                          * completely stalled (multicast wise).
 615                          */
 616                         __ipoib_mcast_schedule_join_thread(priv, NULL, 1);
 617                         goto out;
 618                 }
 619
 620                 memcpy(broadcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
 621                        sizeof (union ib_gid));
 622                 priv->broadcast = broadcast;
 623
 624                 __ipoib_mcast_add(dev, priv->broadcast);
 625         }
 626
 627         if (!test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
 628                 if (IS_ERR_OR_NULL(priv->broadcast->mc) &&
 629                     !test_bit(IPOIB_MCAST_FLAG_BUSY, &priv->broadcast->flags)) {
 630                         mcast = priv->broadcast;
 631                         if (mcast->backoff > 1 &&
 632                             time_before(jiffies, mcast->delay_until)) {
 633                                 delay_until = mcast->delay_until;
 634                                 mcast = NULL;
 635                         }
 636                 }
 637                 goto out;
 638         }
 639
 640         /*
 641          * We'll never get here until the broadcast group is both allocated
 642          * and attached
 643          */
 644         list_for_each_entry(mcast, &priv->multicast_list, list) {
 645                 if (IS_ERR_OR_NULL(mcast->mc) &&
 646                     !test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags) &&
 647                     (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) ||
 648                      !skb_queue_empty(&mcast->pkt_queue))) {
 649                         if (mcast->backoff == 1 ||
 650                             time_after_eq(jiffies, mcast->delay_until)) {
 651                                 /* Found the next unjoined group */
 652                                 init_completion(&mcast->done);
 653                                 set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 654                                 if (ipoib_mcast_join(dev, mcast)) {
 655                                         spin_unlock_irq(&priv->lock);
 656                                         return;
 657                                 }
 658                         } else if (!delay_until ||
 659                                  time_before(mcast->delay_until, delay_until))
 660                                 delay_until = mcast->delay_until;
 661                 }
 662         }
 663
 664         mcast = NULL;
 665         ipoib_dbg_mcast(priv, "successfully started all multicast joins\n");
 666
 667 out:
 668         if (delay_until) {
 669                 cancel_delayed_work(&priv->mcast_task);
 670                 queue_delayed_work(priv->wq, &priv->mcast_task,
 671                                    delay_until - jiffies);
 672         }
 673         if (mcast) {
 674                 init_completion(&mcast->done);
 675                 set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 676                 ipoib_mcast_join(dev, mcast);
 677         }
 678         spin_unlock_irq(&priv->lock);
 679 }
 680
 681 void ipoib_mcast_start_thread(struct net_device *dev)
 682 {
 683         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 684         unsigned long flags;
 685
 686         ipoib_dbg_mcast(priv, "starting multicast thread\n");
 687
 688         spin_lock_irqsave(&priv->lock, flags);
 689         __ipoib_mcast_schedule_join_thread(priv, NULL, 0);
 690         spin_unlock_irqrestore(&priv->lock, flags);
 691 }
 692
 693 int ipoib_mcast_stop_thread(struct net_device *dev)
 694 {
 695         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 696         unsigned long flags;
 697
 698         ipoib_dbg_mcast(priv, "stopping multicast thread\n");
 699
 700         spin_lock_irqsave(&priv->lock, flags);
 701         cancel_delayed_work(&priv->mcast_task);
 702         spin_unlock_irqrestore(&priv->lock, flags);
 703
 704         flush_workqueue(priv->wq);
 705
 706         return 0;
 707 }
 708
 709 static int ipoib_mcast_leave(struct net_device *dev, struct ipoib_mcast *mcast)
 710 {
 711         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 712         struct rdma_netdev *rn = netdev_priv(dev);
 713         int ret = 0;
 714
 715         if (test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
 716                 ipoib_warn(priv, "ipoib_mcast_leave on an in-flight join\n");
 717
 718         if (!IS_ERR_OR_NULL(mcast->mc))
 719                 ib_sa_free_multicast(mcast->mc);
 720
 721         if (test_and_clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 722                 ipoib_dbg_mcast(priv, "leaving MGID %pI6\n",
 723                                 mcast->mcmember.mgid.raw);
 724
 725                 /* Remove ourselves from the multicast group */
 726                 ret = rn->detach_mcast(dev, priv->ca, &mcast->mcmember.mgid,
 727                                        be16_to_cpu(mcast->mcmember.mlid));
 728                 if (ret)
 729                         ipoib_warn(priv, "ib_detach_mcast failed (result = %d)\n", ret);
 730         } else if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags))
 731                 ipoib_dbg(priv, "leaving with no mcmember but not a "
 732                           "SENDONLY join\n");
 733
 734         return 0;
 735 }
 736
 737 /*
 738  * Check if the multicast group is sendonly. If so remove it from the maps
 739  * and add to the remove list
 740  */
 741 void ipoib_check_and_add_mcast_sendonly(struct ipoib_dev_priv *priv, u8 *mgid,
 742                                 struct list_head *remove_list)
 743 {
 744         /* Is this multicast ? */
 745         if (*mgid == 0xff) {
 746                 struct ipoib_mcast *mcast = __ipoib_mcast_find(priv->dev, mgid);
 747
 748                 if (mcast && test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 749                         list_del(&mcast->list);
 750                         rb_erase(&mcast->rb_node, &priv->multicast_tree);
 751                         list_add_tail(&mcast->list, remove_list);
 752                 }
 753         }
 754 }
 755
 756 void ipoib_mcast_remove_list(struct list_head *remove_list)
 757 {
 758         struct ipoib_mcast *mcast, *tmcast;
 759
 760         list_for_each_entry_safe(mcast, tmcast, remove_list, list) {
 761                 ipoib_mcast_leave(mcast->dev, mcast);
 762                 ipoib_mcast_free(mcast);
 763         }
 764 }
 765
 766 void ipoib_mcast_send(struct net_device *dev, u8 *daddr, struct sk_buff *skb)
 767 {
 768         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 769         struct rdma_netdev *rn = netdev_priv(dev);
 770         struct ipoib_mcast *mcast;
 771         unsigned long flags;
 772         void *mgid = daddr + 4;
 773
 774         spin_lock_irqsave(&priv->lock, flags);
 775
 776         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)         ||
 777             !priv->broadcast                                    ||
 778             !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
 779                 ++dev->stats.tx_dropped;
 780                 dev_kfree_skb_any(skb);
 781                 goto unlock;
 782         }
 783
 784         mcast = __ipoib_mcast_find(dev, mgid);
 785         if (!mcast || !mcast->ah) {
 786                 if (!mcast) {
 787                         /* Let's create a new send only group now */
 788                         ipoib_dbg_mcast(priv, "setting up send only multicast group for %pI6\n",
 789                                         mgid);
 790
 791                         mcast = ipoib_mcast_alloc(dev, 0);
 792                         if (!mcast) {
 793                                 ipoib_warn(priv, "unable to allocate memory "
 794                                            "for multicast structure\n");
 795                                 ++dev->stats.tx_dropped;
 796                                 dev_kfree_skb_any(skb);
 797                                 goto unlock;
 798                         }
 799
 800                         set_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags);
 801                         memcpy(mcast->mcmember.mgid.raw, mgid,
 802                                sizeof (union ib_gid));
 803                         __ipoib_mcast_add(dev, mcast);
 804                         list_add_tail(&mcast->list, &priv->multicast_list);
 805                 }
 806                 if (skb_queue_len(&mcast->pkt_queue) < IPOIB_MAX_MCAST_QUEUE) {
 807                         /* put pseudoheader back on for next time */
 808                         skb_push(skb, sizeof(struct ipoib_pseudo_header));
 809                         skb_queue_tail(&mcast->pkt_queue, skb);
 810                 } else {
 811                         ++dev->stats.tx_dropped;
 812                         dev_kfree_skb_any(skb);
 813                 }
 814                 if (!test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)) {
 815                         __ipoib_mcast_schedule_join_thread(priv, NULL, 0);
 816                 }
 817         } else {
 818                 struct ipoib_neigh *neigh;
 819
 820                 spin_unlock_irqrestore(&priv->lock, flags);
 821                 neigh = ipoib_neigh_get(dev, daddr);
 822                 spin_lock_irqsave(&priv->lock, flags);
 823                 if (!neigh) {
 824                         neigh = ipoib_neigh_alloc(daddr, dev);
 825                         if (neigh) {
 826                                 kref_get(&mcast->ah->ref);
 827                                 neigh->ah       = mcast->ah;
 828                                 list_add_tail(&neigh->list, &mcast->neigh_list);
 829                         }
 830                 }
 831                 spin_unlock_irqrestore(&priv->lock, flags);
 832                 mcast->ah->last_send = rn->send(dev, skb, mcast->ah->ah,
 833                                                 IB_MULTICAST_QPN);
 834                 if (neigh)
 835                         ipoib_neigh_put(neigh);
 836                 return;
 837         }
 838
 839 unlock:
 840         spin_unlock_irqrestore(&priv->lock, flags);
 841 }
 842
 843 void ipoib_mcast_dev_flush(struct net_device *dev)
 844 {
 845         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 846         LIST_HEAD(remove_list);
 847         struct ipoib_mcast *mcast, *tmcast;
 848         unsigned long flags;
 849
 850         ipoib_dbg_mcast(priv, "flushing multicast list\n");
 851
 852         spin_lock_irqsave(&priv->lock, flags);
 853
 854         list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
 855                 list_del(&mcast->list);
 856                 rb_erase(&mcast->rb_node, &priv->multicast_tree);
 857                 list_add_tail(&mcast->list, &remove_list);
 858         }
 859
 860         if (priv->broadcast) {
 861                 rb_erase(&priv->broadcast->rb_node, &priv->multicast_tree);
 862                 list_add_tail(&priv->broadcast->list, &remove_list);
 863                 priv->broadcast = NULL;
 864         }
 865
 866         spin_unlock_irqrestore(&priv->lock, flags);
 867
 868         /*
 869          * make sure the in-flight joins have finished before we attempt
 870          * to leave
 871          */
 872         list_for_each_entry_safe(mcast, tmcast, &remove_list, list)
 873                 if (test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
 874                         wait_for_completion(&mcast->done);
 875
 876         ipoib_mcast_remove_list(&remove_list);
 877 }
 878
 879 static int ipoib_mcast_addr_is_valid(const u8 *addr, const u8 *broadcast)
 880 {
 881         /* reserved QPN, prefix, scope */
 882         if (memcmp(addr, broadcast, 6))
 883                 return 0;
 884         /* signature lower, pkey */
 885         if (memcmp(addr + 7, broadcast + 7, 3))
 886                 return 0;
 887         return 1;
 888 }
 889
 890 void ipoib_mcast_restart_task(struct work_struct *work)
 891 {
 892         struct ipoib_dev_priv *priv =
 893                 container_of(work, struct ipoib_dev_priv, restart_task);
 894         struct net_device *dev = priv->dev;
 895         struct netdev_hw_addr *ha;
 896         struct ipoib_mcast *mcast, *tmcast;
 897         LIST_HEAD(remove_list);
 898         unsigned long flags;
 899         struct ib_sa_mcmember_rec rec;
 900
 901         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
 902                 /*
 903                  * shortcut...on shutdown flush is called next, just
 904                  * let it do all the work
 905                  */
 906                 return;
 907
 908         ipoib_dbg_mcast(priv, "restarting multicast task\n");
 909
 910         local_irq_save(flags);
 911         netif_addr_lock(dev);
 912         spin_lock(&priv->lock);
 913
 914         /*
 915          * Unfortunately, the networking core only gives us a list of all of
 916          * the multicast hardware addresses. We need to figure out which ones
 917          * are new and which ones have been removed
 918          */
 919
 920         /* Clear out the found flag */
 921         list_for_each_entry(mcast, &priv->multicast_list, list)
 922                 clear_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
 923
 924         /* Mark all of the entries that are found or don't exist */
 925         netdev_for_each_mc_addr(ha, dev) {
 926                 union ib_gid mgid;
 927
 928                 if (!ipoib_mcast_addr_is_valid(ha->addr, dev->broadcast))
 929                         continue;
 930
 931                 memcpy(mgid.raw, ha->addr + 4, sizeof mgid);
 932
 933                 mcast = __ipoib_mcast_find(dev, &mgid);
 934                 if (!mcast || test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 935                         struct ipoib_mcast *nmcast;
 936
 937                         /* ignore group which is directly joined by userspace */
 938                         if (test_bit(IPOIB_FLAG_UMCAST, &priv->flags) &&
 939                             !ib_sa_get_mcmember_rec(priv->ca, priv->port, &mgid, &rec)) {
 940                                 ipoib_dbg_mcast(priv, "ignoring multicast entry for mgid %pI6\n",
 941                                                 mgid.raw);
 942                                 continue;
 943                         }
 944
 945                         /* Not found or send-only group, let's add a new entry */
 946                         ipoib_dbg_mcast(priv, "adding multicast entry for mgid %pI6\n",
 947                                         mgid.raw);
 948
 949                         nmcast = ipoib_mcast_alloc(dev, 0);
 950                         if (!nmcast) {
 951                                 ipoib_warn(priv, "unable to allocate memory for multicast structure\n");
 952                                 continue;
 953                         }
 954
 955                         set_bit(IPOIB_MCAST_FLAG_FOUND, &nmcast->flags);
 956
 957                         nmcast->mcmember.mgid = mgid;
 958
 959                         if (mcast) {
 960                                 /* Destroy the send only entry */
 961                                 list_move_tail(&mcast->list, &remove_list);
 962
 963                                 rb_replace_node(&mcast->rb_node,
 964                                                 &nmcast->rb_node,
 965                                                 &priv->multicast_tree);
 966                         } else
 967                                 __ipoib_mcast_add(dev, nmcast);
 968
 969                         list_add_tail(&nmcast->list, &priv->multicast_list);
 970                 }
 971
 972                 if (mcast)
 973                         set_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
 974         }
 975
 976         /* Remove all of the entries don't exist anymore */
 977         list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
 978                 if (!test_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags) &&
 979                     !test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 980                         ipoib_dbg_mcast(priv, "deleting multicast group %pI6\n",
 981                                         mcast->mcmember.mgid.raw);
 982
 983                         rb_erase(&mcast->rb_node, &priv->multicast_tree);
 984
 985                         /* Move to the remove list */
 986                         list_move_tail(&mcast->list, &remove_list);
 987                 }
 988         }
 989
 990         spin_unlock(&priv->lock);
 991         netif_addr_unlock(dev);
 992         local_irq_restore(flags);
 993
 994         /*
 995          * make sure the in-flight joins have finished before we attempt
 996          * to leave
 997          */
 998         list_for_each_entry_safe(mcast, tmcast, &remove_list, list)
 999                 if (test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
1000                         wait_for_completion(&mcast->done);
1001
1002         ipoib_mcast_remove_list(&remove_list);
1003
1004         /*
1005          * Double check that we are still up
1006          */
1007         if (test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
1008                 spin_lock_irqsave(&priv->lock, flags);
1009                 __ipoib_mcast_schedule_join_thread(priv, NULL, 0);
1010                 spin_unlock_irqrestore(&priv->lock, flags);
1011         }
1012 }
1013
1014 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
1015
1016 struct ipoib_mcast_iter *ipoib_mcast_iter_init(struct net_device *dev)
1017 {
1018         struct ipoib_mcast_iter *iter;
1019
1020         iter = kmalloc(sizeof *iter, GFP_KERNEL);
1021         if (!iter)
1022                 return NULL;
1023
1024         iter->dev = dev;
1025         memset(iter->mgid.raw, 0, 16);
1026
1027         if (ipoib_mcast_iter_next(iter)) {
1028                 kfree(iter);
1029                 return NULL;
1030         }
1031
1032         return iter;
1033 }
1034
1035 int ipoib_mcast_iter_next(struct ipoib_mcast_iter *iter)
1036 {
1037         struct ipoib_dev_priv *priv = ipoib_priv(iter->dev);
1038         struct rb_node *n;
1039         struct ipoib_mcast *mcast;
1040         int ret = 1;
1041
1042         spin_lock_irq(&priv->lock);
1043
1044         n = rb_first(&priv->multicast_tree);
1045
1046         while (n) {
1047                 mcast = rb_entry(n, struct ipoib_mcast, rb_node);
1048
1049                 if (memcmp(iter->mgid.raw, mcast->mcmember.mgid.raw,
1050                            sizeof (union ib_gid)) < 0) {
1051                         iter->mgid      = mcast->mcmember.mgid;
1052                         iter->created   = mcast->created;
1053                         iter->queuelen  = skb_queue_len(&mcast->pkt_queue);
1054                         iter->complete  = !!mcast->ah;
1055                         iter->send_only = !!(mcast->flags & (1 << IPOIB_MCAST_FLAG_SENDONLY));
1056
1057                         ret = 0;
1058
1059                         break;
1060                 }
1061
1062                 n = rb_next(n);
1063         }
1064
1065         spin_unlock_irq(&priv->lock);
1066
1067         return ret;
1068 }
1069
1070 void ipoib_mcast_iter_read(struct ipoib_mcast_iter *iter,
1071                            union ib_gid *mgid,
1072                            unsigned long *created,
1073                            unsigned int *queuelen,
1074                            unsigned int *complete,
1075                            unsigned int *send_only)
1076 {
1077         *mgid      = iter->mgid;
1078         *created   = iter->created;
1079         *queuelen  = iter->queuelen;
1080         *complete  = iter->complete;
1081         *send_only = iter->send_only;
1082 }
1083
1084 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */