drivers/infiniband/ulp/ipoib/ipoib_multicast.c

   1 /*
   2  * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
   3  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
   4  * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
   5  *
   6  * This software is available to you under a choice of one of two
   7  * licenses.  You may choose to be licensed under the terms of the GNU
   8  * General Public License (GPL) Version 2, available from the file
   9  * COPYING in the main directory of this source tree, or the
  10  * OpenIB.org BSD license below:
  11  *
  12  *     Redistribution and use in source and binary forms, with or
  13  *     without modification, are permitted provided that the following
  14  *     conditions are met:
  15  *
  16  *      - Redistributions of source code must retain the above
  17  *        copyright notice, this list of conditions and the following
  18  *        disclaimer.
  19  *
  20  *      - Redistributions in binary form must reproduce the above
  21  *        copyright notice, this list of conditions and the following
  22  *        disclaimer in the documentation and/or other materials
  23  *        provided with the distribution.
  24  *
  25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  32  * SOFTWARE.
  33  */
  34
  35 #include <linux/skbuff.h>
  36 #include <linux/rtnetlink.h>
  37 #include <linux/moduleparam.h>
  38 #include <linux/ip.h>
  39 #include <linux/in.h>
  40 #include <linux/igmp.h>
  41 #include <linux/inetdevice.h>
  42 #include <linux/delay.h>
  43 #include <linux/completion.h>
  44 #include <linux/slab.h>
  45
  46 #include <net/dst.h>
  47
  48 #include "ipoib.h"
  49
  50 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
  51 static int mcast_debug_level;
  52
  53 module_param(mcast_debug_level, int, 0644);
  54 MODULE_PARM_DESC(mcast_debug_level,
  55                  "Enable multicast debug tracing if > 0");
  56 #endif
  57
  58 struct ipoib_mcast_iter {
  59         struct net_device *dev;
  60         union ib_gid       mgid;
  61         unsigned long      created;
  62         unsigned int       queuelen;
  63         unsigned int       complete;
  64         unsigned int       send_only;
  65 };
  66
  67 /* join state that allows creating mcg with sendonly member request */
  68 #define SENDONLY_FULLMEMBER_JOIN        8
  69
  70 /*
  71  * This should be called with the priv->lock held
  72  */
  73 static void __ipoib_mcast_schedule_join_thread(struct ipoib_dev_priv *priv,
  74                                                struct ipoib_mcast *mcast,
  75                                                bool delay)
  76 {
  77         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
  78                 return;
  79
  80         /*
  81          * We will be scheduling *something*, so cancel whatever is
  82          * currently scheduled first
  83          */
  84         cancel_delayed_work(&priv->mcast_task);
  85         if (mcast && delay) {
  86                 /*
  87                  * We had a failure and want to schedule a retry later
  88                  */
  89                 mcast->backoff *= 2;
  90                 if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
  91                         mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
  92                 mcast->delay_until = jiffies + (mcast->backoff * HZ);
  93                 /*
  94                  * Mark this mcast for its delay, but restart the
  95                  * task immediately.  The join task will make sure to
  96                  * clear out all entries without delays, and then
  97                  * schedule itself to run again when the earliest
  98                  * delay expires
  99                  */
 100                 queue_delayed_work(priv->wq, &priv->mcast_task, 0);
 101         } else if (delay) {
 102                 /*
 103                  * Special case of retrying after a failure to
 104                  * allocate the broadcast multicast group, wait
 105                  * 1 second and try again
 106                  */
 107                 queue_delayed_work(priv->wq, &priv->mcast_task, HZ);
 108         } else
 109                 queue_delayed_work(priv->wq, &priv->mcast_task, 0);
 110 }
 111
 112 static void ipoib_mcast_free(struct ipoib_mcast *mcast)
 113 {
 114         struct net_device *dev = mcast->dev;
 115         int tx_dropped = 0;
 116
 117         ipoib_dbg_mcast(netdev_priv(dev), "deleting multicast group %pI6\n",
 118                         mcast->mcmember.mgid.raw);
 119
 120         /* remove all neigh connected to this mcast */
 121         ipoib_del_neighs_by_gid(dev, mcast->mcmember.mgid.raw);
 122
 123         if (mcast->ah)
 124                 ipoib_put_ah(mcast->ah);
 125
 126         while (!skb_queue_empty(&mcast->pkt_queue)) {
 127                 ++tx_dropped;
 128                 dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
 129         }
 130
 131         netif_tx_lock_bh(dev);
 132         dev->stats.tx_dropped += tx_dropped;
 133         netif_tx_unlock_bh(dev);
 134
 135         kfree(mcast);
 136 }
 137
 138 static struct ipoib_mcast *ipoib_mcast_alloc(struct net_device *dev,
 139                                              int can_sleep)
 140 {
 141         struct ipoib_mcast *mcast;
 142
 143         mcast = kzalloc(sizeof *mcast, can_sleep ? GFP_KERNEL : GFP_ATOMIC);
 144         if (!mcast)
 145                 return NULL;
 146
 147         mcast->dev = dev;
 148         mcast->created = jiffies;
 149         mcast->delay_until = jiffies;
 150         mcast->backoff = 1;
 151
 152         INIT_LIST_HEAD(&mcast->list);
 153         INIT_LIST_HEAD(&mcast->neigh_list);
 154         skb_queue_head_init(&mcast->pkt_queue);
 155
 156         return mcast;
 157 }
 158
 159 static struct ipoib_mcast *__ipoib_mcast_find(struct net_device *dev, void *mgid)
 160 {
 161         struct ipoib_dev_priv *priv = netdev_priv(dev);
 162         struct rb_node *n = priv->multicast_tree.rb_node;
 163
 164         while (n) {
 165                 struct ipoib_mcast *mcast;
 166                 int ret;
 167
 168                 mcast = rb_entry(n, struct ipoib_mcast, rb_node);
 169
 170                 ret = memcmp(mgid, mcast->mcmember.mgid.raw,
 171                              sizeof (union ib_gid));
 172                 if (ret < 0)
 173                         n = n->rb_left;
 174                 else if (ret > 0)
 175                         n = n->rb_right;
 176                 else
 177                         return mcast;
 178         }
 179
 180         return NULL;
 181 }
 182
 183 static int __ipoib_mcast_add(struct net_device *dev, struct ipoib_mcast *mcast)
 184 {
 185         struct ipoib_dev_priv *priv = netdev_priv(dev);
 186         struct rb_node **n = &priv->multicast_tree.rb_node, *pn = NULL;
 187
 188         while (*n) {
 189                 struct ipoib_mcast *tmcast;
 190                 int ret;
 191
 192                 pn = *n;
 193                 tmcast = rb_entry(pn, struct ipoib_mcast, rb_node);
 194
 195                 ret = memcmp(mcast->mcmember.mgid.raw, tmcast->mcmember.mgid.raw,
 196                              sizeof (union ib_gid));
 197                 if (ret < 0)
 198                         n = &pn->rb_left;
 199                 else if (ret > 0)
 200                         n = &pn->rb_right;
 201                 else
 202                         return -EEXIST;
 203         }
 204
 205         rb_link_node(&mcast->rb_node, pn, n);
 206         rb_insert_color(&mcast->rb_node, &priv->multicast_tree);
 207
 208         return 0;
 209 }
 210
 211 static int ipoib_mcast_join_finish(struct ipoib_mcast *mcast,
 212                                    struct ib_sa_mcmember_rec *mcmember)
 213 {
 214         struct net_device *dev = mcast->dev;
 215         struct ipoib_dev_priv *priv = netdev_priv(dev);
 216         struct ipoib_ah *ah;
 217         int ret;
 218         int set_qkey = 0;
 219
 220         mcast->mcmember = *mcmember;
 221
 222         /* Set the multicast MTU and cached Q_Key before we attach if it's
 223          * the broadcast group.
 224          */
 225         if (!memcmp(mcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
 226                     sizeof (union ib_gid))) {
 227                 spin_lock_irq(&priv->lock);
 228                 if (!priv->broadcast) {
 229                         spin_unlock_irq(&priv->lock);
 230                         return -EAGAIN;
 231                 }
 232                 /*update priv member according to the new mcast*/
 233                 priv->broadcast->mcmember.qkey = mcmember->qkey;
 234                 priv->broadcast->mcmember.mtu = mcmember->mtu;
 235                 priv->broadcast->mcmember.traffic_class = mcmember->traffic_class;
 236                 priv->broadcast->mcmember.rate = mcmember->rate;
 237                 priv->broadcast->mcmember.sl = mcmember->sl;
 238                 priv->broadcast->mcmember.flow_label = mcmember->flow_label;
 239                 priv->broadcast->mcmember.hop_limit = mcmember->hop_limit;
 240                 /* assume if the admin and the mcast are the same both can be changed */
 241                 if (priv->mcast_mtu == priv->admin_mtu)
 242                         priv->admin_mtu =
 243                         priv->mcast_mtu =
 244                         IPOIB_UD_MTU(ib_mtu_enum_to_int(priv->broadcast->mcmember.mtu));
 245                 else
 246                         priv->mcast_mtu =
 247                         IPOIB_UD_MTU(ib_mtu_enum_to_int(priv->broadcast->mcmember.mtu));
 248
 249                 priv->qkey = be32_to_cpu(priv->broadcast->mcmember.qkey);
 250                 spin_unlock_irq(&priv->lock);
 251                 priv->tx_wr.remote_qkey = priv->qkey;
 252                 set_qkey = 1;
 253         }
 254
 255         if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 256                 if (test_and_set_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 257                         ipoib_warn(priv, "multicast group %pI6 already attached\n",
 258                                    mcast->mcmember.mgid.raw);
 259
 260                         return 0;
 261                 }
 262
 263                 ret = ipoib_mcast_attach(dev, be16_to_cpu(mcast->mcmember.mlid),
 264                                          &mcast->mcmember.mgid, set_qkey);
 265                 if (ret < 0) {
 266                         ipoib_warn(priv, "couldn't attach QP to multicast group %pI6\n",
 267                                    mcast->mcmember.mgid.raw);
 268
 269                         clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags);
 270                         return ret;
 271                 }
 272         }
 273
 274         {
 275                 struct ib_ah_attr av = {
 276                         .dlid          = be16_to_cpu(mcast->mcmember.mlid),
 277                         .port_num      = priv->port,
 278                         .sl            = mcast->mcmember.sl,
 279                         .ah_flags      = IB_AH_GRH,
 280                         .static_rate   = mcast->mcmember.rate,
 281                         .grh           = {
 282                                 .flow_label    = be32_to_cpu(mcast->mcmember.flow_label),
 283                                 .hop_limit     = mcast->mcmember.hop_limit,
 284                                 .sgid_index    = 0,
 285                                 .traffic_class = mcast->mcmember.traffic_class
 286                         }
 287                 };
 288                 av.grh.dgid = mcast->mcmember.mgid;
 289
 290                 ah = ipoib_create_ah(dev, priv->pd, &av);
 291                 if (IS_ERR(ah)) {
 292                         ipoib_warn(priv, "ib_address_create failed %ld\n",
 293                                 -PTR_ERR(ah));
 294                         /* use original error */
 295                         return PTR_ERR(ah);
 296                 } else {
 297                         spin_lock_irq(&priv->lock);
 298                         mcast->ah = ah;
 299                         spin_unlock_irq(&priv->lock);
 300
 301                         ipoib_dbg_mcast(priv, "MGID %pI6 AV %p, LID 0x%04x, SL %d\n",
 302                                         mcast->mcmember.mgid.raw,
 303                                         mcast->ah->ah,
 304                                         be16_to_cpu(mcast->mcmember.mlid),
 305                                         mcast->mcmember.sl);
 306                 }
 307         }
 308
 309         /* actually send any queued packets */
 310         netif_tx_lock_bh(dev);
 311         while (!skb_queue_empty(&mcast->pkt_queue)) {
 312                 struct sk_buff *skb = skb_dequeue(&mcast->pkt_queue);
 313
 314                 netif_tx_unlock_bh(dev);
 315
 316                 skb->dev = dev;
 317                 if (dev_queue_xmit(skb))
 318                         ipoib_warn(priv, "dev_queue_xmit failed to requeue packet\n");
 319
 320                 netif_tx_lock_bh(dev);
 321         }
 322         netif_tx_unlock_bh(dev);
 323
 324         return 0;
 325 }
 326
 327 void ipoib_mcast_carrier_on_task(struct work_struct *work)
 328 {
 329         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
 330                                                    carrier_on_task);
 331         struct ib_port_attr attr;
 332         int ret;
 333
 334         if (ib_query_port(priv->ca, priv->port, &attr) ||
 335             attr.state != IB_PORT_ACTIVE) {
 336                 ipoib_dbg(priv, "Keeping carrier off until IB port is active\n");
 337                 return;
 338         }
 339         /*
 340          * Check if can send sendonly MCG's with sendonly-fullmember join state.
 341          * It done here after the successfully join to the broadcast group,
 342          * because the broadcast group must always be joined first and is always
 343          * re-joined if the SM changes substantially.
 344          */
 345         ret = ipoib_check_sm_sendonly_fullmember_support(priv);
 346         if (ret < 0)
 347                 pr_debug("%s failed query sm support for sendonly-fullmember (ret: %d)\n",
 348                          priv->dev->name, ret);
 349
 350         /*
 351          * Take rtnl_lock to avoid racing with ipoib_stop() and
 352          * turning the carrier back on while a device is being
 353          * removed.  However, ipoib_stop() will attempt to flush
 354          * the workqueue while holding the rtnl lock, so loop
 355          * on trylock until either we get the lock or we see
 356          * FLAG_OPER_UP go away as that signals that we are bailing
 357          * and can safely ignore the carrier on work.
 358          */
 359         while (!rtnl_trylock()) {
 360                 if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
 361                         return;
 362                 else
 363                         msleep(20);
 364         }
 365         if (!ipoib_cm_admin_enabled(priv->dev))
 366                 dev_set_mtu(priv->dev, min(priv->mcast_mtu, priv->admin_mtu));
 367         netif_carrier_on(priv->dev);
 368         rtnl_unlock();
 369 }
 370
 371 static int ipoib_mcast_join_complete(int status,
 372                                      struct ib_sa_multicast *multicast)
 373 {
 374         struct ipoib_mcast *mcast = multicast->context;
 375         struct net_device *dev = mcast->dev;
 376         struct ipoib_dev_priv *priv = netdev_priv(dev);
 377
 378         ipoib_dbg_mcast(priv, "%sjoin completion for %pI6 (status %d)\n",
 379                         test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) ?
 380                         "sendonly " : "",
 381                         mcast->mcmember.mgid.raw, status);
 382
 383         /* We trap for port events ourselves. */
 384         if (status == -ENETRESET) {
 385                 status = 0;
 386                 goto out;
 387         }
 388
 389         if (!status)
 390                 status = ipoib_mcast_join_finish(mcast, &multicast->rec);
 391
 392         if (!status) {
 393                 mcast->backoff = 1;
 394                 mcast->delay_until = jiffies;
 395
 396                 /*
 397                  * Defer carrier on work to priv->wq to avoid a
 398                  * deadlock on rtnl_lock here.  Requeue our multicast
 399                  * work too, which will end up happening right after
 400                  * our carrier on task work and will allow us to
 401                  * send out all of the non-broadcast joins
 402                  */
 403                 if (mcast == priv->broadcast) {
 404                         spin_lock_irq(&priv->lock);
 405                         queue_work(priv->wq, &priv->carrier_on_task);
 406                         __ipoib_mcast_schedule_join_thread(priv, NULL, 0);
 407                         goto out_locked;
 408                 }
 409         } else {
 410                 bool silent_fail =
 411                     test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) &&
 412                     status == -EINVAL;
 413
 414                 if (mcast->logcount < 20) {
 415                         if (status == -ETIMEDOUT || status == -EAGAIN ||
 416                             silent_fail) {
 417                                 ipoib_dbg_mcast(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                         } else {
 421                                 ipoib_warn(priv, "%smulticast join failed for %pI6, status %d\n",
 422                                                 test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) ? "sendonly " : "",
 423                                            mcast->mcmember.mgid.raw, status);
 424                         }
 425
 426                         if (!silent_fail)
 427                                 mcast->logcount++;
 428                 }
 429
 430                 if (test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) &&
 431                     mcast->backoff >= 2) {
 432                         /*
 433                          * We only retry sendonly joins once before we drop
 434                          * the packet and quit trying to deal with the
 435                          * group.  However, we leave the group in the
 436                          * mcast list as an unjoined group.  If we want to
 437                          * try joining again, we simply queue up a packet
 438                          * and restart the join thread.  The empty queue
 439                          * is why the join thread ignores this group.
 440                          */
 441                         mcast->backoff = 1;
 442                         netif_tx_lock_bh(dev);
 443                         while (!skb_queue_empty(&mcast->pkt_queue)) {
 444                                 ++dev->stats.tx_dropped;
 445                                 dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
 446                         }
 447                         netif_tx_unlock_bh(dev);
 448                 } else {
 449                         spin_lock_irq(&priv->lock);
 450                         /* Requeue this join task with a backoff delay */
 451                         __ipoib_mcast_schedule_join_thread(priv, mcast, 1);
 452                         goto out_locked;
 453                 }
 454         }
 455 out:
 456         spin_lock_irq(&priv->lock);
 457 out_locked:
 458         /*
 459          * Make sure to set mcast->mc before we clear the busy flag to avoid
 460          * racing with code that checks for BUSY before checking mcast->mc
 461          */
 462         if (status)
 463                 mcast->mc = NULL;
 464         else
 465                 mcast->mc = multicast;
 466         clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 467         spin_unlock_irq(&priv->lock);
 468         complete(&mcast->done);
 469
 470         return status;
 471 }
 472
 473 /*
 474  * Caller must hold 'priv->lock'
 475  */
 476 static int ipoib_mcast_join(struct net_device *dev, struct ipoib_mcast *mcast)
 477 {
 478         struct ipoib_dev_priv *priv = netdev_priv(dev);
 479         struct ib_sa_multicast *multicast;
 480         struct ib_sa_mcmember_rec rec = {
 481                 .join_state = 1
 482         };
 483         ib_sa_comp_mask comp_mask;
 484         int ret = 0;
 485
 486         if (!priv->broadcast ||
 487             !test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
 488                 return -EINVAL;
 489
 490         ipoib_dbg_mcast(priv, "joining MGID %pI6\n", mcast->mcmember.mgid.raw);
 491
 492         rec.mgid     = mcast->mcmember.mgid;
 493         rec.port_gid = priv->local_gid;
 494         rec.pkey     = cpu_to_be16(priv->pkey);
 495
 496         comp_mask =
 497                 IB_SA_MCMEMBER_REC_MGID         |
 498                 IB_SA_MCMEMBER_REC_PORT_GID     |
 499                 IB_SA_MCMEMBER_REC_PKEY         |
 500                 IB_SA_MCMEMBER_REC_JOIN_STATE;
 501
 502         if (mcast != priv->broadcast) {
 503                 /*
 504                  * RFC 4391:
 505                  *  The MGID MUST use the same P_Key, Q_Key, SL, MTU,
 506                  *  and HopLimit as those used in the broadcast-GID.  The rest
 507                  *  of attributes SHOULD follow the values used in the
 508                  *  broadcast-GID as well.
 509                  */
 510                 comp_mask |=
 511                         IB_SA_MCMEMBER_REC_QKEY                 |
 512                         IB_SA_MCMEMBER_REC_MTU_SELECTOR         |
 513                         IB_SA_MCMEMBER_REC_MTU                  |
 514                         IB_SA_MCMEMBER_REC_TRAFFIC_CLASS        |
 515                         IB_SA_MCMEMBER_REC_RATE_SELECTOR        |
 516                         IB_SA_MCMEMBER_REC_RATE                 |
 517                         IB_SA_MCMEMBER_REC_SL                   |
 518                         IB_SA_MCMEMBER_REC_FLOW_LABEL           |
 519                         IB_SA_MCMEMBER_REC_HOP_LIMIT;
 520
 521                 rec.qkey          = priv->broadcast->mcmember.qkey;
 522                 rec.mtu_selector  = IB_SA_EQ;
 523                 rec.mtu           = priv->broadcast->mcmember.mtu;
 524                 rec.traffic_class = priv->broadcast->mcmember.traffic_class;
 525                 rec.rate_selector = IB_SA_EQ;
 526                 rec.rate          = priv->broadcast->mcmember.rate;
 527                 rec.sl            = priv->broadcast->mcmember.sl;
 528                 rec.flow_label    = priv->broadcast->mcmember.flow_label;
 529                 rec.hop_limit     = priv->broadcast->mcmember.hop_limit;
 530
 531                 /*
 532                  * Send-only IB Multicast joins work at the core IB layer but
 533                  * require specific SM support.
 534                  * We can use such joins here only if the current SM supports that feature.
 535                  * However, if not, we emulate an Ethernet multicast send,
 536                  * which does not require a multicast subscription and will
 537                  * still send properly. The most appropriate thing to
 538                  * do is to create the group if it doesn't exist as that
 539                  * most closely emulates the behavior, from a user space
 540                  * application perspective, of Ethernet multicast operation.
 541                  */
 542                 if (test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) &&
 543                     priv->sm_fullmember_sendonly_support)
 544                         /* SM supports sendonly-fullmember, otherwise fallback to full-member */
 545                         rec.join_state = SENDONLY_FULLMEMBER_JOIN;
 546         }
 547         spin_unlock_irq(&priv->lock);
 548
 549         multicast = ib_sa_join_multicast(&ipoib_sa_client, priv->ca, priv->port,
 550                                          &rec, comp_mask, GFP_KERNEL,
 551                                          ipoib_mcast_join_complete, mcast);
 552         spin_lock_irq(&priv->lock);
 553         if (IS_ERR(multicast)) {
 554                 ret = PTR_ERR(multicast);
 555                 ipoib_warn(priv, "ib_sa_join_multicast failed, status %d\n", ret);
 556                 /* Requeue this join task with a backoff delay */
 557                 __ipoib_mcast_schedule_join_thread(priv, mcast, 1);
 558                 clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 559                 spin_unlock_irq(&priv->lock);
 560                 complete(&mcast->done);
 561                 spin_lock_irq(&priv->lock);
 562         }
 563         return 0;
 564 }
 565
 566 void ipoib_mcast_join_task(struct work_struct *work)
 567 {
 568         struct ipoib_dev_priv *priv =
 569                 container_of(work, struct ipoib_dev_priv, mcast_task.work);
 570         struct net_device *dev = priv->dev;
 571         struct ib_port_attr port_attr;
 572         unsigned long delay_until = 0;
 573         struct ipoib_mcast *mcast = NULL;
 574
 575         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
 576                 return;
 577
 578         if (ib_query_port(priv->ca, priv->port, &port_attr) ||
 579             port_attr.state != IB_PORT_ACTIVE) {
 580                 ipoib_dbg(priv, "port state is not ACTIVE (state = %d) suspending join task\n",
 581                           port_attr.state);
 582                 return;
 583         }
 584         priv->local_lid = port_attr.lid;
 585         netif_addr_lock_bh(dev);
 586
 587         if (!test_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags)) {
 588                 netif_addr_unlock_bh(dev);
 589                 return;
 590         }
 591         netif_addr_unlock_bh(dev);
 592
 593         spin_lock_irq(&priv->lock);
 594         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
 595                 goto out;
 596
 597         if (!priv->broadcast) {
 598                 struct ipoib_mcast *broadcast;
 599
 600                 broadcast = ipoib_mcast_alloc(dev, 0);
 601                 if (!broadcast) {
 602                         ipoib_warn(priv, "failed to allocate broadcast group\n");
 603                         /*
 604                          * Restart us after a 1 second delay to retry
 605                          * creating our broadcast group and attaching to
 606                          * it.  Until this succeeds, this ipoib dev is
 607                          * completely stalled (multicast wise).
 608                          */
 609                         __ipoib_mcast_schedule_join_thread(priv, NULL, 1);
 610                         goto out;
 611                 }
 612
 613                 memcpy(broadcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
 614                        sizeof (union ib_gid));
 615                 priv->broadcast = broadcast;
 616
 617                 __ipoib_mcast_add(dev, priv->broadcast);
 618         }
 619
 620         if (!test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
 621                 if (IS_ERR_OR_NULL(priv->broadcast->mc) &&
 622                     !test_bit(IPOIB_MCAST_FLAG_BUSY, &priv->broadcast->flags)) {
 623                         mcast = priv->broadcast;
 624                         if (mcast->backoff > 1 &&
 625                             time_before(jiffies, mcast->delay_until)) {
 626                                 delay_until = mcast->delay_until;
 627                                 mcast = NULL;
 628                         }
 629                 }
 630                 goto out;
 631         }
 632
 633         /*
 634          * We'll never get here until the broadcast group is both allocated
 635          * and attached
 636          */
 637         list_for_each_entry(mcast, &priv->multicast_list, list) {
 638                 if (IS_ERR_OR_NULL(mcast->mc) &&
 639                     !test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags) &&
 640                     (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) ||
 641                      !skb_queue_empty(&mcast->pkt_queue))) {
 642                         if (mcast->backoff == 1 ||
 643                             time_after_eq(jiffies, mcast->delay_until)) {
 644                                 /* Found the next unjoined group */
 645                                 init_completion(&mcast->done);
 646                                 set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 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                 init_completion(&mcast->done);
 668                 set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 669                 ipoib_mcast_join(dev, mcast);
 670         }
 671         spin_unlock_irq(&priv->lock);
 672 }
 673
 674 int ipoib_mcast_start_thread(struct net_device *dev)
 675 {
 676         struct ipoib_dev_priv *priv = netdev_priv(dev);
 677         unsigned long flags;
 678
 679         ipoib_dbg_mcast(priv, "starting multicast thread\n");
 680
 681         spin_lock_irqsave(&priv->lock, flags);
 682         __ipoib_mcast_schedule_join_thread(priv, NULL, 0);
 683         spin_unlock_irqrestore(&priv->lock, flags);
 684
 685         return 0;
 686 }
 687
 688 int ipoib_mcast_stop_thread(struct net_device *dev)
 689 {
 690         struct ipoib_dev_priv *priv = netdev_priv(dev);
 691         unsigned long flags;
 692
 693         ipoib_dbg_mcast(priv, "stopping multicast thread\n");
 694
 695         spin_lock_irqsave(&priv->lock, flags);
 696         cancel_delayed_work(&priv->mcast_task);
 697         spin_unlock_irqrestore(&priv->lock, flags);
 698
 699         flush_workqueue(priv->wq);
 700
 701         return 0;
 702 }
 703
 704 static int ipoib_mcast_leave(struct net_device *dev, struct ipoib_mcast *mcast)
 705 {
 706         struct ipoib_dev_priv *priv = netdev_priv(dev);
 707         int ret = 0;
 708
 709         if (test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
 710                 ipoib_warn(priv, "ipoib_mcast_leave on an in-flight join\n");
 711
 712         if (!IS_ERR_OR_NULL(mcast->mc))
 713                 ib_sa_free_multicast(mcast->mc);
 714
 715         if (test_and_clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 716                 ipoib_dbg_mcast(priv, "leaving MGID %pI6\n",
 717                                 mcast->mcmember.mgid.raw);
 718
 719                 /* Remove ourselves from the multicast group */
 720                 ret = ib_detach_mcast(priv->qp, &mcast->mcmember.mgid,
 721                                       be16_to_cpu(mcast->mcmember.mlid));
 722                 if (ret)
 723                         ipoib_warn(priv, "ib_detach_mcast failed (result = %d)\n", ret);
 724         } else if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags))
 725                 ipoib_dbg(priv, "leaving with no mcmember but not a "
 726                           "SENDONLY join\n");
 727
 728         return 0;
 729 }
 730
 731 /*
 732  * Check if the multicast group is sendonly. If so remove it from the maps
 733  * and add to the remove list
 734  */
 735 void ipoib_check_and_add_mcast_sendonly(struct ipoib_dev_priv *priv, u8 *mgid,
 736                                 struct list_head *remove_list)
 737 {
 738         /* Is this multicast ? */
 739         if (*mgid == 0xff) {
 740                 struct ipoib_mcast *mcast = __ipoib_mcast_find(priv->dev, mgid);
 741
 742                 if (mcast && test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 743                         list_del(&mcast->list);
 744                         rb_erase(&mcast->rb_node, &priv->multicast_tree);
 745                         list_add_tail(&mcast->list, remove_list);
 746                 }
 747         }
 748 }
 749
 750 void ipoib_mcast_remove_list(struct list_head *remove_list)
 751 {
 752         struct ipoib_mcast *mcast, *tmcast;
 753
 754         list_for_each_entry_safe(mcast, tmcast, remove_list, list) {
 755                 ipoib_mcast_leave(mcast->dev, mcast);
 756                 ipoib_mcast_free(mcast);
 757         }
 758 }
 759
 760 void ipoib_mcast_send(struct net_device *dev, u8 *daddr, struct sk_buff *skb)
 761 {
 762         struct ipoib_dev_priv *priv = netdev_priv(dev);
 763         struct ipoib_mcast *mcast;
 764         unsigned long flags;
 765         void *mgid = daddr + 4;
 766
 767         spin_lock_irqsave(&priv->lock, flags);
 768
 769         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)         ||
 770             !priv->broadcast                                    ||
 771             !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
 772                 ++dev->stats.tx_dropped;
 773                 dev_kfree_skb_any(skb);
 774                 goto unlock;
 775         }
 776
 777         mcast = __ipoib_mcast_find(dev, mgid);
 778         if (!mcast || !mcast->ah) {
 779                 if (!mcast) {
 780                         /* Let's create a new send only group now */
 781                         ipoib_dbg_mcast(priv, "setting up send only multicast group for %pI6\n",
 782                                         mgid);
 783
 784                         mcast = ipoib_mcast_alloc(dev, 0);
 785                         if (!mcast) {
 786                                 ipoib_warn(priv, "unable to allocate memory "
 787                                            "for multicast structure\n");
 788                                 ++dev->stats.tx_dropped;
 789                                 dev_kfree_skb_any(skb);
 790                                 goto unlock;
 791                         }
 792
 793                         set_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags);
 794                         memcpy(mcast->mcmember.mgid.raw, mgid,
 795                                sizeof (union ib_gid));
 796                         __ipoib_mcast_add(dev, mcast);
 797                         list_add_tail(&mcast->list, &priv->multicast_list);
 798                 }
 799                 if (skb_queue_len(&mcast->pkt_queue) < IPOIB_MAX_MCAST_QUEUE) {
 800                         /* put pseudoheader back on for next time */
 801                         skb_push(skb, sizeof(struct ipoib_pseudo_header));
 802                         skb_queue_tail(&mcast->pkt_queue, skb);
 803                 } else {
 804                         ++dev->stats.tx_dropped;
 805                         dev_kfree_skb_any(skb);
 806                 }
 807                 if (!test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)) {
 808                         __ipoib_mcast_schedule_join_thread(priv, NULL, 0);
 809                 }
 810         } else {
 811                 struct ipoib_neigh *neigh;
 812
 813                 spin_unlock_irqrestore(&priv->lock, flags);
 814                 neigh = ipoib_neigh_get(dev, daddr);
 815                 spin_lock_irqsave(&priv->lock, flags);
 816                 if (!neigh) {
 817                         neigh = ipoib_neigh_alloc(daddr, dev);
 818                         if (neigh) {
 819                                 kref_get(&mcast->ah->ref);
 820                                 neigh->ah       = mcast->ah;
 821                                 list_add_tail(&neigh->list, &mcast->neigh_list);
 822                         }
 823                 }
 824                 spin_unlock_irqrestore(&priv->lock, flags);
 825                 ipoib_send(dev, skb, mcast->ah, IB_MULTICAST_QPN);
 826                 if (neigh)
 827                         ipoib_neigh_put(neigh);
 828                 return;
 829         }
 830
 831 unlock:
 832         spin_unlock_irqrestore(&priv->lock, flags);
 833 }
 834
 835 void ipoib_mcast_dev_flush(struct net_device *dev)
 836 {
 837         struct ipoib_dev_priv *priv = netdev_priv(dev);
 838         LIST_HEAD(remove_list);
 839         struct ipoib_mcast *mcast, *tmcast;
 840         unsigned long flags;
 841
 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 }
 870
 871 static int ipoib_mcast_addr_is_valid(const u8 *addr, const u8 *broadcast)
 872 {
 873         /* reserved QPN, prefix, scope */
 874         if (memcmp(addr, broadcast, 6))
 875                 return 0;
 876         /* signature lower, pkey */
 877         if (memcmp(addr + 7, broadcast + 7, 3))
 878                 return 0;
 879         return 1;
 880 }
 881
 882 void ipoib_mcast_restart_task(struct work_struct *work)
 883 {
 884         struct ipoib_dev_priv *priv =
 885                 container_of(work, struct ipoib_dev_priv, restart_task);
 886         struct net_device *dev = priv->dev;
 887         struct netdev_hw_addr *ha;
 888         struct ipoib_mcast *mcast, *tmcast;
 889         LIST_HEAD(remove_list);
 890         unsigned long flags;
 891         struct ib_sa_mcmember_rec rec;
 892
 893         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
 894                 /*
 895                  * shortcut...on shutdown flush is called next, just
 896                  * let it do all the work
 897                  */
 898                 return;
 899
 900         ipoib_dbg_mcast(priv, "restarting multicast task\n");
 901
 902         local_irq_save(flags);
 903         netif_addr_lock(dev);
 904         spin_lock(&priv->lock);
 905
 906         /*
 907          * Unfortunately, the networking core only gives us a list of all of
 908          * the multicast hardware addresses. We need to figure out which ones
 909          * are new and which ones have been removed
 910          */
 911
 912         /* Clear out the found flag */
 913         list_for_each_entry(mcast, &priv->multicast_list, list)
 914                 clear_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
 915
 916         /* Mark all of the entries that are found or don't exist */
 917         netdev_for_each_mc_addr(ha, dev) {
 918                 union ib_gid mgid;
 919
 920                 if (!ipoib_mcast_addr_is_valid(ha->addr, dev->broadcast))
 921                         continue;
 922
 923                 memcpy(mgid.raw, ha->addr + 4, sizeof mgid);
 924
 925                 mcast = __ipoib_mcast_find(dev, &mgid);
 926                 if (!mcast || test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 927                         struct ipoib_mcast *nmcast;
 928
 929                         /* ignore group which is directly joined by userspace */
 930                         if (test_bit(IPOIB_FLAG_UMCAST, &priv->flags) &&
 931                             !ib_sa_get_mcmember_rec(priv->ca, priv->port, &mgid, &rec)) {
 932                                 ipoib_dbg_mcast(priv, "ignoring multicast entry for mgid %pI6\n",
 933                                                 mgid.raw);
 934                                 continue;
 935                         }
 936
 937                         /* Not found or send-only group, let's add a new entry */
 938                         ipoib_dbg_mcast(priv, "adding multicast entry for mgid %pI6\n",
 939                                         mgid.raw);
 940
 941                         nmcast = ipoib_mcast_alloc(dev, 0);
 942                         if (!nmcast) {
 943                                 ipoib_warn(priv, "unable to allocate memory for multicast structure\n");
 944                                 continue;
 945                         }
 946
 947                         set_bit(IPOIB_MCAST_FLAG_FOUND, &nmcast->flags);
 948
 949                         nmcast->mcmember.mgid = mgid;
 950
 951                         if (mcast) {
 952                                 /* Destroy the send only entry */
 953                                 list_move_tail(&mcast->list, &remove_list);
 954
 955                                 rb_replace_node(&mcast->rb_node,
 956                                                 &nmcast->rb_node,
 957                                                 &priv->multicast_tree);
 958                         } else
 959                                 __ipoib_mcast_add(dev, nmcast);
 960
 961                         list_add_tail(&nmcast->list, &priv->multicast_list);
 962                 }
 963
 964                 if (mcast)
 965                         set_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
 966         }
 967
 968         /* Remove all of the entries don't exist anymore */
 969         list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
 970                 if (!test_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags) &&
 971                     !test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 972                         ipoib_dbg_mcast(priv, "deleting multicast group %pI6\n",
 973                                         mcast->mcmember.mgid.raw);
 974
 975                         rb_erase(&mcast->rb_node, &priv->multicast_tree);
 976
 977                         /* Move to the remove list */
 978                         list_move_tail(&mcast->list, &remove_list);
 979                 }
 980         }
 981
 982         spin_unlock(&priv->lock);
 983         netif_addr_unlock(dev);
 984         local_irq_restore(flags);
 985
 986         /*
 987          * make sure the in-flight joins have finished before we attempt
 988          * to leave
 989          */
 990         list_for_each_entry_safe(mcast, tmcast, &remove_list, list)
 991                 if (test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
 992                         wait_for_completion(&mcast->done);
 993
 994         ipoib_mcast_remove_list(&remove_list);
 995
 996         /*
 997          * Double check that we are still up
 998          */
 999         if (test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
1000                 spin_lock_irqsave(&priv->lock, flags);
1001                 __ipoib_mcast_schedule_join_thread(priv, NULL, 0);
1002                 spin_unlock_irqrestore(&priv->lock, flags);
1003         }
1004 }
1005
1006 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
1007
1008 struct ipoib_mcast_iter *ipoib_mcast_iter_init(struct net_device *dev)
1009 {
1010         struct ipoib_mcast_iter *iter;
1011
1012         iter = kmalloc(sizeof *iter, GFP_KERNEL);
1013         if (!iter)
1014                 return NULL;
1015
1016         iter->dev = dev;
1017         memset(iter->mgid.raw, 0, 16);
1018
1019         if (ipoib_mcast_iter_next(iter)) {
1020                 kfree(iter);
1021                 return NULL;
1022         }
1023
1024         return iter;
1025 }
1026
1027 int ipoib_mcast_iter_next(struct ipoib_mcast_iter *iter)
1028 {
1029         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
1030         struct rb_node *n;
1031         struct ipoib_mcast *mcast;
1032         int ret = 1;
1033
1034         spin_lock_irq(&priv->lock);
1035
1036         n = rb_first(&priv->multicast_tree);
1037
1038         while (n) {
1039                 mcast = rb_entry(n, struct ipoib_mcast, rb_node);
1040
1041                 if (memcmp(iter->mgid.raw, mcast->mcmember.mgid.raw,
1042                            sizeof (union ib_gid)) < 0) {
1043                         iter->mgid      = mcast->mcmember.mgid;
1044                         iter->created   = mcast->created;
1045                         iter->queuelen  = skb_queue_len(&mcast->pkt_queue);
1046                         iter->complete  = !!mcast->ah;
1047                         iter->send_only = !!(mcast->flags & (1 << IPOIB_MCAST_FLAG_SENDONLY));
1048
1049                         ret = 0;
1050
1051                         break;
1052                 }
1053
1054                 n = rb_next(n);
1055         }
1056
1057         spin_unlock_irq(&priv->lock);
1058
1059         return ret;
1060 }
1061
1062 void ipoib_mcast_iter_read(struct ipoib_mcast_iter *iter,
1063                            union ib_gid *mgid,
1064                            unsigned long *created,
1065                            unsigned int *queuelen,
1066                            unsigned int *complete,
1067                            unsigned int *send_only)
1068 {
1069         *mgid      = iter->mgid;
1070         *created   = iter->created;
1071         *queuelen  = iter->queuelen;
1072         *complete  = iter->complete;
1073         *send_only = iter->send_only;
1074 }
1075
1076 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */