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