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  * $Id: ipoib_multicast.c 1362 2004-12-18 15:56:29Z roland $
  35  */
  36
  37 #include <linux/skbuff.h>
  38 #include <linux/rtnetlink.h>
  39 #include <linux/ip.h>
  40 #include <linux/in.h>
  41 #include <linux/igmp.h>
  42 #include <linux/inetdevice.h>
  43 #include <linux/delay.h>
  44 #include <linux/completion.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 static DEFINE_MUTEX(mcast_mutex);
  59
  60 /* Used for all multicast joins (broadcast, IPv4 mcast and IPv6 mcast) */
  61 struct ipoib_mcast {
  62         struct ib_sa_mcmember_rec mcmember;
  63         struct ipoib_ah          *ah;
  64
  65         struct rb_node    rb_node;
  66         struct list_head  list;
  67         struct completion done;
  68
  69         int                 query_id;
  70         struct ib_sa_query *query;
  71
  72         unsigned long created;
  73         unsigned long backoff;
  74
  75         unsigned long flags;
  76         unsigned char logcount;
  77
  78         struct list_head  neigh_list;
  79
  80         struct sk_buff_head pkt_queue;
  81
  82         struct net_device *dev;
  83 };
  84
  85 struct ipoib_mcast_iter {
  86         struct net_device *dev;
  87         union ib_gid       mgid;
  88         unsigned long      created;
  89         unsigned int       queuelen;
  90         unsigned int       complete;
  91         unsigned int       send_only;
  92 };
  93
  94 static void ipoib_mcast_free(struct ipoib_mcast *mcast)
  95 {
  96         struct net_device *dev = mcast->dev;
  97         struct ipoib_dev_priv *priv = netdev_priv(dev);
  98         struct ipoib_neigh *neigh, *tmp;
  99         unsigned long flags;
 100         int tx_dropped = 0;
 101
 102         ipoib_dbg_mcast(netdev_priv(dev),
 103                         "deleting multicast group " IPOIB_GID_FMT "\n",
 104                         IPOIB_GID_ARG(mcast->mcmember.mgid));
 105
 106         spin_lock_irqsave(&priv->lock, flags);
 107
 108         list_for_each_entry_safe(neigh, tmp, &mcast->neigh_list, list) {
 109                 /*
 110                  * It's safe to call ipoib_put_ah() inside priv->lock
 111                  * here, because we know that mcast->ah will always
 112                  * hold one more reference, so ipoib_put_ah() will
 113                  * never do more than decrement the ref count.
 114                  */
 115                 if (neigh->ah)
 116                         ipoib_put_ah(neigh->ah);
 117                 ipoib_neigh_free(neigh);
 118         }
 119
 120         spin_unlock_irqrestore(&priv->lock, flags);
 121
 122         if (mcast->ah)
 123                 ipoib_put_ah(mcast->ah);
 124
 125         while (!skb_queue_empty(&mcast->pkt_queue)) {
 126                 ++tx_dropped;
 127                 dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
 128         }
 129
 130         spin_lock_irqsave(&priv->tx_lock, flags);
 131         priv->stats.tx_dropped += tx_dropped;
 132         spin_unlock_irqrestore(&priv->tx_lock, flags);
 133
 134         kfree(mcast);
 135 }
 136
 137 static struct ipoib_mcast *ipoib_mcast_alloc(struct net_device *dev,
 138                                              int can_sleep)
 139 {
 140         struct ipoib_mcast *mcast;
 141
 142         mcast = kzalloc(sizeof *mcast, can_sleep ? GFP_KERNEL : GFP_ATOMIC);
 143         if (!mcast)
 144                 return NULL;
 145
 146         mcast->dev = dev;
 147         mcast->created = jiffies;
 148         mcast->backoff = 1;
 149
 150         INIT_LIST_HEAD(&mcast->list);
 151         INIT_LIST_HEAD(&mcast->neigh_list);
 152         skb_queue_head_init(&mcast->pkt_queue);
 153
 154         return mcast;
 155 }
 156
 157 static struct ipoib_mcast *__ipoib_mcast_find(struct net_device *dev, void *mgid)
 158 {
 159         struct ipoib_dev_priv *priv = netdev_priv(dev);
 160         struct rb_node *n = priv->multicast_tree.rb_node;
 161
 162         while (n) {
 163                 struct ipoib_mcast *mcast;
 164                 int ret;
 165
 166                 mcast = rb_entry(n, struct ipoib_mcast, rb_node);
 167
 168                 ret = memcmp(mgid, mcast->mcmember.mgid.raw,
 169                              sizeof (union ib_gid));
 170                 if (ret < 0)
 171                         n = n->rb_left;
 172                 else if (ret > 0)
 173                         n = n->rb_right;
 174                 else
 175                         return mcast;
 176         }
 177
 178         return NULL;
 179 }
 180
 181 static int __ipoib_mcast_add(struct net_device *dev, struct ipoib_mcast *mcast)
 182 {
 183         struct ipoib_dev_priv *priv = netdev_priv(dev);
 184         struct rb_node **n = &priv->multicast_tree.rb_node, *pn = NULL;
 185
 186         while (*n) {
 187                 struct ipoib_mcast *tmcast;
 188                 int ret;
 189
 190                 pn = *n;
 191                 tmcast = rb_entry(pn, struct ipoib_mcast, rb_node);
 192
 193                 ret = memcmp(mcast->mcmember.mgid.raw, tmcast->mcmember.mgid.raw,
 194                              sizeof (union ib_gid));
 195                 if (ret < 0)
 196                         n = &pn->rb_left;
 197                 else if (ret > 0)
 198                         n = &pn->rb_right;
 199                 else
 200                         return -EEXIST;
 201         }
 202
 203         rb_link_node(&mcast->rb_node, pn, n);
 204         rb_insert_color(&mcast->rb_node, &priv->multicast_tree);
 205
 206         return 0;
 207 }
 208
 209 static int ipoib_mcast_join_finish(struct ipoib_mcast *mcast,
 210                                    struct ib_sa_mcmember_rec *mcmember)
 211 {
 212         struct net_device *dev = mcast->dev;
 213         struct ipoib_dev_priv *priv = netdev_priv(dev);
 214         struct ipoib_ah *ah;
 215         int ret;
 216
 217         mcast->mcmember = *mcmember;
 218
 219         /* Set the cached Q_Key before we attach if it's the broadcast group */
 220         if (!memcmp(mcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
 221                     sizeof (union ib_gid))) {
 222                 priv->qkey = be32_to_cpu(priv->broadcast->mcmember.qkey);
 223                 priv->tx_wr.wr.ud.remote_qkey = priv->qkey;
 224         }
 225
 226         if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 227                 if (test_and_set_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 228                         ipoib_warn(priv, "multicast group " IPOIB_GID_FMT
 229                                    " already attached\n",
 230                                    IPOIB_GID_ARG(mcast->mcmember.mgid));
 231
 232                         return 0;
 233                 }
 234
 235                 ret = ipoib_mcast_attach(dev, be16_to_cpu(mcast->mcmember.mlid),
 236                                          &mcast->mcmember.mgid);
 237                 if (ret < 0) {
 238                         ipoib_warn(priv, "couldn't attach QP to multicast group "
 239                                    IPOIB_GID_FMT "\n",
 240                                    IPOIB_GID_ARG(mcast->mcmember.mgid));
 241
 242                         clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags);
 243                         return ret;
 244                 }
 245         }
 246
 247         {
 248                 struct ib_ah_attr av = {
 249                         .dlid          = be16_to_cpu(mcast->mcmember.mlid),
 250                         .port_num      = priv->port,
 251                         .sl            = mcast->mcmember.sl,
 252                         .ah_flags      = IB_AH_GRH,
 253                         .static_rate   = mcast->mcmember.rate,
 254                         .grh           = {
 255                                 .flow_label    = be32_to_cpu(mcast->mcmember.flow_label),
 256                                 .hop_limit     = mcast->mcmember.hop_limit,
 257                                 .sgid_index    = 0,
 258                                 .traffic_class = mcast->mcmember.traffic_class
 259                         }
 260                 };
 261                 av.grh.dgid = mcast->mcmember.mgid;
 262
 263                 ah = ipoib_create_ah(dev, priv->pd, &av);
 264                 if (!ah) {
 265                         ipoib_warn(priv, "ib_address_create failed\n");
 266                 } else {
 267                         spin_lock_irq(&priv->lock);
 268                         mcast->ah = ah;
 269                         spin_unlock_irq(&priv->lock);
 270
 271                         ipoib_dbg_mcast(priv, "MGID " IPOIB_GID_FMT
 272                                         " AV %p, LID 0x%04x, SL %d\n",
 273                                         IPOIB_GID_ARG(mcast->mcmember.mgid),
 274                                         mcast->ah->ah,
 275                                         be16_to_cpu(mcast->mcmember.mlid),
 276                                         mcast->mcmember.sl);
 277                 }
 278         }
 279
 280         /* actually send any queued packets */
 281         spin_lock_irq(&priv->tx_lock);
 282         while (!skb_queue_empty(&mcast->pkt_queue)) {
 283                 struct sk_buff *skb = skb_dequeue(&mcast->pkt_queue);
 284                 spin_unlock_irq(&priv->tx_lock);
 285
 286                 skb->dev = dev;
 287
 288                 if (!skb->dst || !skb->dst->neighbour) {
 289                         /* put pseudoheader back on for next time */
 290                         skb_push(skb, sizeof (struct ipoib_pseudoheader));
 291                 }
 292
 293                 if (dev_queue_xmit(skb))
 294                         ipoib_warn(priv, "dev_queue_xmit failed to requeue packet\n");
 295                 spin_lock_irq(&priv->tx_lock);
 296         }
 297         spin_unlock_irq(&priv->tx_lock);
 298
 299         return 0;
 300 }
 301
 302 static void
 303 ipoib_mcast_sendonly_join_complete(int status,
 304                                    struct ib_sa_mcmember_rec *mcmember,
 305                                    void *mcast_ptr)
 306 {
 307         struct ipoib_mcast *mcast = mcast_ptr;
 308         struct net_device *dev = mcast->dev;
 309         struct ipoib_dev_priv *priv = netdev_priv(dev);
 310
 311         if (!status)
 312                 ipoib_mcast_join_finish(mcast, mcmember);
 313         else {
 314                 if (mcast->logcount++ < 20)
 315                         ipoib_dbg_mcast(netdev_priv(dev), "multicast join failed for "
 316                                         IPOIB_GID_FMT ", status %d\n",
 317                                         IPOIB_GID_ARG(mcast->mcmember.mgid), status);
 318
 319                 /* Flush out any queued packets */
 320                 spin_lock_irq(&priv->tx_lock);
 321                 while (!skb_queue_empty(&mcast->pkt_queue)) {
 322                         ++priv->stats.tx_dropped;
 323                         dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
 324                 }
 325                 spin_unlock_irq(&priv->tx_lock);
 326
 327                 /* Clear the busy flag so we try again */
 328                 clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 329                 mcast->query = NULL;
 330         }
 331
 332         complete(&mcast->done);
 333 }
 334
 335 static int ipoib_mcast_sendonly_join(struct ipoib_mcast *mcast)
 336 {
 337         struct net_device *dev = mcast->dev;
 338         struct ipoib_dev_priv *priv = netdev_priv(dev);
 339         struct ib_sa_mcmember_rec rec = {
 340 #if 0                           /* Some SMs don't support send-only yet */
 341                 .join_state = 4
 342 #else
 343                 .join_state = 1
 344 #endif
 345         };
 346         int ret = 0;
 347
 348         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
 349                 ipoib_dbg_mcast(priv, "device shutting down, no multicast joins\n");
 350                 return -ENODEV;
 351         }
 352
 353         if (test_and_set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)) {
 354                 ipoib_dbg_mcast(priv, "multicast entry busy, skipping\n");
 355                 return -EBUSY;
 356         }
 357
 358         rec.mgid     = mcast->mcmember.mgid;
 359         rec.port_gid = priv->local_gid;
 360         rec.pkey     = cpu_to_be16(priv->pkey);
 361
 362         init_completion(&mcast->done);
 363
 364         ret = ib_sa_mcmember_rec_set(&ipoib_sa_client, priv->ca, priv->port, &rec,
 365                                      IB_SA_MCMEMBER_REC_MGID            |
 366                                      IB_SA_MCMEMBER_REC_PORT_GID        |
 367                                      IB_SA_MCMEMBER_REC_PKEY            |
 368                                      IB_SA_MCMEMBER_REC_JOIN_STATE,
 369                                      1000, GFP_ATOMIC,
 370                                      ipoib_mcast_sendonly_join_complete,
 371                                      mcast, &mcast->query);
 372         if (ret < 0) {
 373                 ipoib_warn(priv, "ib_sa_mcmember_rec_set failed (ret = %d)\n",
 374                            ret);
 375         } else {
 376                 ipoib_dbg_mcast(priv, "no multicast record for " IPOIB_GID_FMT
 377                                 ", starting join\n",
 378                                 IPOIB_GID_ARG(mcast->mcmember.mgid));
 379
 380                 mcast->query_id = ret;
 381         }
 382
 383         return ret;
 384 }
 385
 386 static void ipoib_mcast_join_complete(int status,
 387                                       struct ib_sa_mcmember_rec *mcmember,
 388                                       void *mcast_ptr)
 389 {
 390         struct ipoib_mcast *mcast = mcast_ptr;
 391         struct net_device *dev = mcast->dev;
 392         struct ipoib_dev_priv *priv = netdev_priv(dev);
 393
 394         ipoib_dbg_mcast(priv, "join completion for " IPOIB_GID_FMT
 395                         " (status %d)\n",
 396                         IPOIB_GID_ARG(mcast->mcmember.mgid), status);
 397
 398         if (!status && !ipoib_mcast_join_finish(mcast, mcmember)) {
 399                 mcast->backoff = 1;
 400                 mutex_lock(&mcast_mutex);
 401                 if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 402                         queue_work(ipoib_workqueue, &priv->mcast_task);
 403                 mutex_unlock(&mcast_mutex);
 404                 complete(&mcast->done);
 405                 return;
 406         }
 407
 408         if (status == -EINTR) {
 409                 complete(&mcast->done);
 410                 return;
 411         }
 412
 413         if (status && mcast->logcount++ < 20) {
 414                 if (status == -ETIMEDOUT || status == -EINTR) {
 415                         ipoib_dbg_mcast(priv, "multicast join failed for " IPOIB_GID_FMT
 416                                         ", status %d\n",
 417                                         IPOIB_GID_ARG(mcast->mcmember.mgid),
 418                                         status);
 419                 } else {
 420                         ipoib_warn(priv, "multicast join failed for "
 421                                    IPOIB_GID_FMT ", status %d\n",
 422                                    IPOIB_GID_ARG(mcast->mcmember.mgid),
 423                                    status);
 424                 }
 425         }
 426
 427         mcast->backoff *= 2;
 428         if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
 429                 mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
 430
 431         mutex_lock(&mcast_mutex);
 432
 433         spin_lock_irq(&priv->lock);
 434         mcast->query = NULL;
 435
 436         if (test_bit(IPOIB_MCAST_RUN, &priv->flags)) {
 437                 if (status == -ETIMEDOUT)
 438                         queue_work(ipoib_workqueue, &priv->mcast_task);
 439                 else
 440                         queue_delayed_work(ipoib_workqueue, &priv->mcast_task,
 441                                            mcast->backoff * HZ);
 442         } else
 443                 complete(&mcast->done);
 444         spin_unlock_irq(&priv->lock);
 445         mutex_unlock(&mcast_mutex);
 446
 447         return;
 448 }
 449
 450 static void ipoib_mcast_join(struct net_device *dev, struct ipoib_mcast *mcast,
 451                              int create)
 452 {
 453         struct ipoib_dev_priv *priv = netdev_priv(dev);
 454         struct ib_sa_mcmember_rec rec = {
 455                 .join_state = 1
 456         };
 457         ib_sa_comp_mask comp_mask;
 458         int ret = 0;
 459
 460         ipoib_dbg_mcast(priv, "joining MGID " IPOIB_GID_FMT "\n",
 461                         IPOIB_GID_ARG(mcast->mcmember.mgid));
 462
 463         rec.mgid     = mcast->mcmember.mgid;
 464         rec.port_gid = priv->local_gid;
 465         rec.pkey     = cpu_to_be16(priv->pkey);
 466
 467         comp_mask =
 468                 IB_SA_MCMEMBER_REC_MGID         |
 469                 IB_SA_MCMEMBER_REC_PORT_GID     |
 470                 IB_SA_MCMEMBER_REC_PKEY         |
 471                 IB_SA_MCMEMBER_REC_JOIN_STATE;
 472
 473         if (create) {
 474                 comp_mask |=
 475                         IB_SA_MCMEMBER_REC_QKEY                 |
 476                         IB_SA_MCMEMBER_REC_MTU_SELECTOR         |
 477                         IB_SA_MCMEMBER_REC_MTU                  |
 478                         IB_SA_MCMEMBER_REC_TRAFFIC_CLASS        |
 479                         IB_SA_MCMEMBER_REC_RATE_SELECTOR        |
 480                         IB_SA_MCMEMBER_REC_RATE                 |
 481                         IB_SA_MCMEMBER_REC_SL                   |
 482                         IB_SA_MCMEMBER_REC_FLOW_LABEL           |
 483                         IB_SA_MCMEMBER_REC_HOP_LIMIT;
 484
 485                 rec.qkey          = priv->broadcast->mcmember.qkey;
 486                 rec.mtu_selector  = IB_SA_EQ;
 487                 rec.mtu           = priv->broadcast->mcmember.mtu;
 488                 rec.traffic_class = priv->broadcast->mcmember.traffic_class;
 489                 rec.rate_selector = IB_SA_EQ;
 490                 rec.rate          = priv->broadcast->mcmember.rate;
 491                 rec.sl            = priv->broadcast->mcmember.sl;
 492                 rec.flow_label    = priv->broadcast->mcmember.flow_label;
 493                 rec.hop_limit     = priv->broadcast->mcmember.hop_limit;
 494         }
 495
 496         init_completion(&mcast->done);
 497
 498         ret = ib_sa_mcmember_rec_set(&ipoib_sa_client, priv->ca, priv->port,
 499                                      &rec, comp_mask, mcast->backoff * 1000,
 500                                      GFP_ATOMIC, ipoib_mcast_join_complete,
 501                                      mcast, &mcast->query);
 502
 503         if (ret < 0) {
 504                 ipoib_warn(priv, "ib_sa_mcmember_rec_set failed, status %d\n", ret);
 505
 506                 mcast->backoff *= 2;
 507                 if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
 508                         mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
 509
 510                 mutex_lock(&mcast_mutex);
 511                 if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 512                         queue_delayed_work(ipoib_workqueue,
 513                                            &priv->mcast_task,
 514                                            mcast->backoff * HZ);
 515                 mutex_unlock(&mcast_mutex);
 516         } else
 517                 mcast->query_id = ret;
 518 }
 519
 520 void ipoib_mcast_join_task(void *dev_ptr)
 521 {
 522         struct net_device *dev = dev_ptr;
 523         struct ipoib_dev_priv *priv = netdev_priv(dev);
 524
 525         if (!test_bit(IPOIB_MCAST_RUN, &priv->flags))
 526                 return;
 527
 528         if (ib_query_gid(priv->ca, priv->port, 0, &priv->local_gid))
 529                 ipoib_warn(priv, "ib_gid_entry_get() failed\n");
 530         else
 531                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
 532
 533         {
 534                 struct ib_port_attr attr;
 535
 536                 if (!ib_query_port(priv->ca, priv->port, &attr)) {
 537                         priv->local_lid  = attr.lid;
 538                         priv->local_rate = attr.active_speed *
 539                                 ib_width_enum_to_int(attr.active_width);
 540                 } else
 541                 ipoib_warn(priv, "ib_query_port failed\n");
 542         }
 543
 544         if (!priv->broadcast) {
 545                 struct ipoib_mcast *broadcast;
 546
 547                 broadcast = ipoib_mcast_alloc(dev, 1);
 548                 if (!broadcast) {
 549                         ipoib_warn(priv, "failed to allocate broadcast group\n");
 550                         mutex_lock(&mcast_mutex);
 551                         if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 552                                 queue_delayed_work(ipoib_workqueue,
 553                                                    &priv->mcast_task, HZ);
 554                         mutex_unlock(&mcast_mutex);
 555                         return;
 556                 }
 557
 558                 spin_lock_irq(&priv->lock);
 559                 memcpy(broadcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
 560                        sizeof (union ib_gid));
 561                 priv->broadcast = broadcast;
 562
 563                 __ipoib_mcast_add(dev, priv->broadcast);
 564                 spin_unlock_irq(&priv->lock);
 565         }
 566
 567         if (!test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
 568                 ipoib_mcast_join(dev, priv->broadcast, 0);
 569                 return;
 570         }
 571
 572         while (1) {
 573                 struct ipoib_mcast *mcast = NULL;
 574
 575                 spin_lock_irq(&priv->lock);
 576                 list_for_each_entry(mcast, &priv->multicast_list, list) {
 577                         if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)
 578                             && !test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)
 579                             && !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 580                                 /* Found the next unjoined group */
 581                                 break;
 582                         }
 583                 }
 584                 spin_unlock_irq(&priv->lock);
 585
 586                 if (&mcast->list == &priv->multicast_list) {
 587                         /* All done */
 588                         break;
 589                 }
 590
 591                 ipoib_mcast_join(dev, mcast, 1);
 592                 return;
 593         }
 594
 595         priv->mcast_mtu = ib_mtu_enum_to_int(priv->broadcast->mcmember.mtu) -
 596                 IPOIB_ENCAP_LEN;
 597         dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
 598
 599         ipoib_dbg_mcast(priv, "successfully joined all multicast groups\n");
 600
 601         clear_bit(IPOIB_MCAST_RUN, &priv->flags);
 602         netif_carrier_on(dev);
 603 }
 604
 605 int ipoib_mcast_start_thread(struct net_device *dev)
 606 {
 607         struct ipoib_dev_priv *priv = netdev_priv(dev);
 608
 609         ipoib_dbg_mcast(priv, "starting multicast thread\n");
 610
 611         mutex_lock(&mcast_mutex);
 612         if (!test_and_set_bit(IPOIB_MCAST_RUN, &priv->flags))
 613                 queue_work(ipoib_workqueue, &priv->mcast_task);
 614         mutex_unlock(&mcast_mutex);
 615
 616         spin_lock_irq(&priv->lock);
 617         set_bit(IPOIB_MCAST_STARTED, &priv->flags);
 618         spin_unlock_irq(&priv->lock);
 619
 620         return 0;
 621 }
 622
 623 static void wait_for_mcast_join(struct ipoib_dev_priv *priv,
 624                                 struct ipoib_mcast *mcast)
 625 {
 626         spin_lock_irq(&priv->lock);
 627         if (mcast && mcast->query) {
 628                 ib_sa_cancel_query(mcast->query_id, mcast->query);
 629                 mcast->query = NULL;
 630                 spin_unlock_irq(&priv->lock);
 631                 ipoib_dbg_mcast(priv, "waiting for MGID " IPOIB_GID_FMT "\n",
 632                                 IPOIB_GID_ARG(mcast->mcmember.mgid));
 633                 wait_for_completion(&mcast->done);
 634         }
 635         else
 636                 spin_unlock_irq(&priv->lock);
 637 }
 638
 639 int ipoib_mcast_stop_thread(struct net_device *dev, int flush)
 640 {
 641         struct ipoib_dev_priv *priv = netdev_priv(dev);
 642         struct ipoib_mcast *mcast;
 643
 644         ipoib_dbg_mcast(priv, "stopping multicast thread\n");
 645
 646         spin_lock_irq(&priv->lock);
 647         clear_bit(IPOIB_MCAST_STARTED, &priv->flags);
 648         spin_unlock_irq(&priv->lock);
 649
 650         mutex_lock(&mcast_mutex);
 651         clear_bit(IPOIB_MCAST_RUN, &priv->flags);
 652         cancel_delayed_work(&priv->mcast_task);
 653         mutex_unlock(&mcast_mutex);
 654
 655         if (flush)
 656                 flush_workqueue(ipoib_workqueue);
 657
 658         wait_for_mcast_join(priv, priv->broadcast);
 659
 660         list_for_each_entry(mcast, &priv->multicast_list, list)
 661                 wait_for_mcast_join(priv, mcast);
 662
 663         return 0;
 664 }
 665
 666 static int ipoib_mcast_leave(struct net_device *dev, struct ipoib_mcast *mcast)
 667 {
 668         struct ipoib_dev_priv *priv = netdev_priv(dev);
 669         struct ib_sa_mcmember_rec rec = {
 670                 .join_state = 1
 671         };
 672         int ret = 0;
 673
 674         if (!test_and_clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags))
 675                 return 0;
 676
 677         ipoib_dbg_mcast(priv, "leaving MGID " IPOIB_GID_FMT "\n",
 678                         IPOIB_GID_ARG(mcast->mcmember.mgid));
 679
 680         rec.mgid     = mcast->mcmember.mgid;
 681         rec.port_gid = priv->local_gid;
 682         rec.pkey     = cpu_to_be16(priv->pkey);
 683
 684         /* Remove ourselves from the multicast group */
 685         ret = ipoib_mcast_detach(dev, be16_to_cpu(mcast->mcmember.mlid),
 686                                  &mcast->mcmember.mgid);
 687         if (ret)
 688                 ipoib_warn(priv, "ipoib_mcast_detach failed (result = %d)\n", ret);
 689
 690         /*
 691          * Just make one shot at leaving and don't wait for a reply;
 692          * if we fail, too bad.
 693          */
 694         ret = ib_sa_mcmember_rec_delete(&ipoib_sa_client, priv->ca, priv->port, &rec,
 695                                         IB_SA_MCMEMBER_REC_MGID         |
 696                                         IB_SA_MCMEMBER_REC_PORT_GID     |
 697                                         IB_SA_MCMEMBER_REC_PKEY         |
 698                                         IB_SA_MCMEMBER_REC_JOIN_STATE,
 699                                         0, GFP_ATOMIC, NULL,
 700                                         mcast, &mcast->query);
 701         if (ret < 0)
 702                 ipoib_warn(priv, "ib_sa_mcmember_rec_delete failed "
 703                            "for leave (result = %d)\n", ret);
 704
 705         return 0;
 706 }
 707
 708 void ipoib_mcast_send(struct net_device *dev, void *mgid, struct sk_buff *skb)
 709 {
 710         struct ipoib_dev_priv *priv = netdev_priv(dev);
 711         struct ipoib_mcast *mcast;
 712
 713         /*
 714          * We can only be called from ipoib_start_xmit, so we're
 715          * inside tx_lock -- no need to save/restore flags.
 716          */
 717         spin_lock(&priv->lock);
 718
 719         if (!test_bit(IPOIB_MCAST_STARTED, &priv->flags)        ||
 720             !priv->broadcast                                    ||
 721             !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
 722                 ++priv->stats.tx_dropped;
 723                 dev_kfree_skb_any(skb);
 724                 goto unlock;
 725         }
 726
 727         mcast = __ipoib_mcast_find(dev, mgid);
 728         if (!mcast) {
 729                 /* Let's create a new send only group now */
 730                 ipoib_dbg_mcast(priv, "setting up send only multicast group for "
 731                                 IPOIB_GID_FMT "\n", IPOIB_GID_RAW_ARG(mgid));
 732
 733                 mcast = ipoib_mcast_alloc(dev, 0);
 734                 if (!mcast) {
 735                         ipoib_warn(priv, "unable to allocate memory for "
 736                                    "multicast structure\n");
 737                         ++priv->stats.tx_dropped;
 738                         dev_kfree_skb_any(skb);
 739                         goto out;
 740                 }
 741
 742                 set_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags);
 743                 memcpy(mcast->mcmember.mgid.raw, mgid, sizeof (union ib_gid));
 744                 __ipoib_mcast_add(dev, mcast);
 745                 list_add_tail(&mcast->list, &priv->multicast_list);
 746         }
 747
 748         if (!mcast->ah) {
 749                 if (skb_queue_len(&mcast->pkt_queue) < IPOIB_MAX_MCAST_QUEUE)
 750                         skb_queue_tail(&mcast->pkt_queue, skb);
 751                 else {
 752                         ++priv->stats.tx_dropped;
 753                         dev_kfree_skb_any(skb);
 754                 }
 755
 756                 if (mcast->query)
 757                         ipoib_dbg_mcast(priv, "no address vector, "
 758                                         "but multicast join already started\n");
 759                 else if (test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags))
 760                         ipoib_mcast_sendonly_join(mcast);
 761
 762                 /*
 763                  * If lookup completes between here and out:, don't
 764                  * want to send packet twice.
 765                  */
 766                 mcast = NULL;
 767         }
 768
 769 out:
 770         if (mcast && mcast->ah) {
 771                 if (skb->dst            &&
 772                     skb->dst->neighbour &&
 773                     !*to_ipoib_neigh(skb->dst->neighbour)) {
 774                         struct ipoib_neigh *neigh = ipoib_neigh_alloc(skb->dst->neighbour);
 775
 776                         if (neigh) {
 777                                 kref_get(&mcast->ah->ref);
 778                                 neigh->ah       = mcast->ah;
 779                                 list_add_tail(&neigh->list, &mcast->neigh_list);
 780                         }
 781                 }
 782
 783                 ipoib_send(dev, skb, mcast->ah, IB_MULTICAST_QPN);
 784         }
 785
 786 unlock:
 787         spin_unlock(&priv->lock);
 788 }
 789
 790 void ipoib_mcast_dev_flush(struct net_device *dev)
 791 {
 792         struct ipoib_dev_priv *priv = netdev_priv(dev);
 793         LIST_HEAD(remove_list);
 794         struct ipoib_mcast *mcast, *tmcast;
 795         unsigned long flags;
 796
 797         ipoib_dbg_mcast(priv, "flushing multicast list\n");
 798
 799         spin_lock_irqsave(&priv->lock, flags);
 800
 801         list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
 802                 list_del(&mcast->list);
 803                 rb_erase(&mcast->rb_node, &priv->multicast_tree);
 804                 list_add_tail(&mcast->list, &remove_list);
 805         }
 806
 807         if (priv->broadcast) {
 808                 rb_erase(&priv->broadcast->rb_node, &priv->multicast_tree);
 809                 list_add_tail(&priv->broadcast->list, &remove_list);
 810                 priv->broadcast = NULL;
 811         }
 812
 813         spin_unlock_irqrestore(&priv->lock, flags);
 814
 815         list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
 816                 ipoib_mcast_leave(dev, mcast);
 817                 ipoib_mcast_free(mcast);
 818         }
 819 }
 820
 821 void ipoib_mcast_restart_task(void *dev_ptr)
 822 {
 823         struct net_device *dev = dev_ptr;
 824         struct ipoib_dev_priv *priv = netdev_priv(dev);
 825         struct dev_mc_list *mclist;
 826         struct ipoib_mcast *mcast, *tmcast;
 827         LIST_HEAD(remove_list);
 828         unsigned long flags;
 829
 830         ipoib_dbg_mcast(priv, "restarting multicast task\n");
 831
 832         ipoib_mcast_stop_thread(dev, 0);
 833
 834         local_irq_save(flags);
 835         netif_tx_lock(dev);
 836         spin_lock(&priv->lock);
 837
 838         /*
 839          * Unfortunately, the networking core only gives us a list of all of
 840          * the multicast hardware addresses. We need to figure out which ones
 841          * are new and which ones have been removed
 842          */
 843
 844         /* Clear out the found flag */
 845         list_for_each_entry(mcast, &priv->multicast_list, list)
 846                 clear_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
 847
 848         /* Mark all of the entries that are found or don't exist */
 849         for (mclist = dev->mc_list; mclist; mclist = mclist->next) {
 850                 union ib_gid mgid;
 851
 852                 memcpy(mgid.raw, mclist->dmi_addr + 4, sizeof mgid);
 853
 854                 /* Add in the P_Key */
 855                 mgid.raw[4] = (priv->pkey >> 8) & 0xff;
 856                 mgid.raw[5] = priv->pkey & 0xff;
 857
 858                 mcast = __ipoib_mcast_find(dev, &mgid);
 859                 if (!mcast || test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 860                         struct ipoib_mcast *nmcast;
 861
 862                         /* Not found or send-only group, let's add a new entry */
 863                         ipoib_dbg_mcast(priv, "adding multicast entry for mgid "
 864                                         IPOIB_GID_FMT "\n", IPOIB_GID_ARG(mgid));
 865
 866                         nmcast = ipoib_mcast_alloc(dev, 0);
 867                         if (!nmcast) {
 868                                 ipoib_warn(priv, "unable to allocate memory for multicast structure\n");
 869                                 continue;
 870                         }
 871
 872                         set_bit(IPOIB_MCAST_FLAG_FOUND, &nmcast->flags);
 873
 874                         nmcast->mcmember.mgid = mgid;
 875
 876                         if (mcast) {
 877                                 /* Destroy the send only entry */
 878                                 list_move_tail(&mcast->list, &remove_list);
 879
 880                                 rb_replace_node(&mcast->rb_node,
 881                                                 &nmcast->rb_node,
 882                                                 &priv->multicast_tree);
 883                         } else
 884                                 __ipoib_mcast_add(dev, nmcast);
 885
 886                         list_add_tail(&nmcast->list, &priv->multicast_list);
 887                 }
 888
 889                 if (mcast)
 890                         set_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
 891         }
 892
 893         /* Remove all of the entries don't exist anymore */
 894         list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
 895                 if (!test_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags) &&
 896                     !test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 897                         ipoib_dbg_mcast(priv, "deleting multicast group " IPOIB_GID_FMT "\n",
 898                                         IPOIB_GID_ARG(mcast->mcmember.mgid));
 899
 900                         rb_erase(&mcast->rb_node, &priv->multicast_tree);
 901
 902                         /* Move to the remove list */
 903                         list_move_tail(&mcast->list, &remove_list);
 904                 }
 905         }
 906
 907         spin_unlock(&priv->lock);
 908         netif_tx_unlock(dev);
 909         local_irq_restore(flags);
 910
 911         /* We have to cancel outside of the spinlock */
 912         list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
 913                 wait_for_mcast_join(priv, mcast);
 914                 ipoib_mcast_leave(mcast->dev, mcast);
 915                 ipoib_mcast_free(mcast);
 916         }
 917
 918         if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 919                 ipoib_mcast_start_thread(dev);
 920 }
 921
 922 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 923
 924 struct ipoib_mcast_iter *ipoib_mcast_iter_init(struct net_device *dev)
 925 {
 926         struct ipoib_mcast_iter *iter;
 927
 928         iter = kmalloc(sizeof *iter, GFP_KERNEL);
 929         if (!iter)
 930                 return NULL;
 931
 932         iter->dev = dev;
 933         memset(iter->mgid.raw, 0, 16);
 934
 935         if (ipoib_mcast_iter_next(iter)) {
 936                 kfree(iter);
 937                 return NULL;
 938         }
 939
 940         return iter;
 941 }
 942
 943 int ipoib_mcast_iter_next(struct ipoib_mcast_iter *iter)
 944 {
 945         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
 946         struct rb_node *n;
 947         struct ipoib_mcast *mcast;
 948         int ret = 1;
 949
 950         spin_lock_irq(&priv->lock);
 951
 952         n = rb_first(&priv->multicast_tree);
 953
 954         while (n) {
 955                 mcast = rb_entry(n, struct ipoib_mcast, rb_node);
 956
 957                 if (memcmp(iter->mgid.raw, mcast->mcmember.mgid.raw,
 958                            sizeof (union ib_gid)) < 0) {
 959                         iter->mgid      = mcast->mcmember.mgid;
 960                         iter->created   = mcast->created;
 961                         iter->queuelen  = skb_queue_len(&mcast->pkt_queue);
 962                         iter->complete  = !!mcast->ah;
 963                         iter->send_only = !!(mcast->flags & (1 << IPOIB_MCAST_FLAG_SENDONLY));
 964
 965                         ret = 0;
 966
 967                         break;
 968                 }
 969
 970                 n = rb_next(n);
 971         }
 972
 973         spin_unlock_irq(&priv->lock);
 974
 975         return ret;
 976 }
 977
 978 void ipoib_mcast_iter_read(struct ipoib_mcast_iter *iter,
 979                            union ib_gid *mgid,
 980                            unsigned long *created,
 981                            unsigned int *queuelen,
 982                            unsigned int *complete,
 983                            unsigned int *send_only)
 984 {
 985         *mgid      = iter->mgid;
 986         *created   = iter->created;
 987         *queuelen  = iter->queuelen;
 988         *complete  = iter->complete;
 989         *send_only = iter->send_only;
 990 }
 991
 992 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */