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