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(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_SL           |
 477                         IB_SA_MCMEMBER_REC_FLOW_LABEL   |
 478                         IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
 479
 480                 rec.qkey          = priv->broadcast->mcmember.qkey;
 481                 rec.sl            = priv->broadcast->mcmember.sl;
 482                 rec.flow_label    = priv->broadcast->mcmember.flow_label;
 483                 rec.traffic_class = priv->broadcast->mcmember.traffic_class;
 484         }
 485
 486         init_completion(&mcast->done);
 487
 488         ret = ib_sa_mcmember_rec_set(priv->ca, priv->port, &rec, comp_mask,
 489                                      mcast->backoff * 1000, GFP_ATOMIC,
 490                                      ipoib_mcast_join_complete,
 491                                      mcast, &mcast->query);
 492
 493         if (ret < 0) {
 494                 ipoib_warn(priv, "ib_sa_mcmember_rec_set failed, status %d\n", ret);
 495
 496                 mcast->backoff *= 2;
 497                 if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
 498                         mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
 499
 500                 mutex_lock(&mcast_mutex);
 501                 if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 502                         queue_delayed_work(ipoib_workqueue,
 503                                            &priv->mcast_task,
 504                                            mcast->backoff * HZ);
 505                 mutex_unlock(&mcast_mutex);
 506         } else
 507                 mcast->query_id = ret;
 508 }
 509
 510 void ipoib_mcast_join_task(void *dev_ptr)
 511 {
 512         struct net_device *dev = dev_ptr;
 513         struct ipoib_dev_priv *priv = netdev_priv(dev);
 514
 515         if (!test_bit(IPOIB_MCAST_RUN, &priv->flags))
 516                 return;
 517
 518         if (ib_query_gid(priv->ca, priv->port, 0, &priv->local_gid))
 519                 ipoib_warn(priv, "ib_gid_entry_get() failed\n");
 520         else
 521                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
 522
 523         {
 524                 struct ib_port_attr attr;
 525
 526                 if (!ib_query_port(priv->ca, priv->port, &attr)) {
 527                         priv->local_lid  = attr.lid;
 528                         priv->local_rate = attr.active_speed *
 529                                 ib_width_enum_to_int(attr.active_width);
 530                 } else
 531                         ipoib_warn(priv, "ib_query_port failed\n");
 532         }
 533
 534         if (!priv->broadcast) {
 535                 struct ipoib_mcast *broadcast;
 536
 537                 broadcast = ipoib_mcast_alloc(dev, 1);
 538                 if (!broadcast) {
 539                         ipoib_warn(priv, "failed to allocate broadcast group\n");
 540                         mutex_lock(&mcast_mutex);
 541                         if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 542                                 queue_delayed_work(ipoib_workqueue,
 543                                                    &priv->mcast_task, HZ);
 544                         mutex_unlock(&mcast_mutex);
 545                         return;
 546                 }
 547
 548                 spin_lock_irq(&priv->lock);
 549                 memcpy(broadcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
 550                        sizeof (union ib_gid));
 551                 priv->broadcast = broadcast;
 552
 553                 __ipoib_mcast_add(dev, priv->broadcast);
 554                 spin_unlock_irq(&priv->lock);
 555         }
 556
 557         if (!test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
 558                 ipoib_mcast_join(dev, priv->broadcast, 0);
 559                 return;
 560         }
 561
 562         while (1) {
 563                 struct ipoib_mcast *mcast = NULL;
 564
 565                 spin_lock_irq(&priv->lock);
 566                 list_for_each_entry(mcast, &priv->multicast_list, list) {
 567                         if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)
 568                             && !test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)
 569                             && !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 570                                 /* Found the next unjoined group */
 571                                 break;
 572                         }
 573                 }
 574                 spin_unlock_irq(&priv->lock);
 575
 576                 if (&mcast->list == &priv->multicast_list) {
 577                         /* All done */
 578                         break;
 579                 }
 580
 581                 ipoib_mcast_join(dev, mcast, 1);
 582                 return;
 583         }
 584
 585         priv->mcast_mtu = ib_mtu_enum_to_int(priv->broadcast->mcmember.mtu) -
 586                 IPOIB_ENCAP_LEN;
 587         dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
 588
 589         ipoib_dbg_mcast(priv, "successfully joined all multicast groups\n");
 590
 591         clear_bit(IPOIB_MCAST_RUN, &priv->flags);
 592         netif_carrier_on(dev);
 593 }
 594
 595 int ipoib_mcast_start_thread(struct net_device *dev)
 596 {
 597         struct ipoib_dev_priv *priv = netdev_priv(dev);
 598
 599         ipoib_dbg_mcast(priv, "starting multicast thread\n");
 600
 601         mutex_lock(&mcast_mutex);
 602         if (!test_and_set_bit(IPOIB_MCAST_RUN, &priv->flags))
 603                 queue_work(ipoib_workqueue, &priv->mcast_task);
 604         mutex_unlock(&mcast_mutex);
 605
 606         spin_lock_irq(&priv->lock);
 607         set_bit(IPOIB_MCAST_STARTED, &priv->flags);
 608         spin_unlock_irq(&priv->lock);
 609
 610         return 0;
 611 }
 612
 613 static void wait_for_mcast_join(struct ipoib_dev_priv *priv,
 614                                 struct ipoib_mcast *mcast)
 615 {
 616         spin_lock_irq(&priv->lock);
 617         if (mcast && mcast->query) {
 618                 ib_sa_cancel_query(mcast->query_id, mcast->query);
 619                 mcast->query = NULL;
 620                 spin_unlock_irq(&priv->lock);
 621                 ipoib_dbg_mcast(priv, "waiting for MGID " IPOIB_GID_FMT "\n",
 622                                 IPOIB_GID_ARG(mcast->mcmember.mgid));
 623                 wait_for_completion(&mcast->done);
 624         }
 625         else
 626                 spin_unlock_irq(&priv->lock);
 627 }
 628
 629 int ipoib_mcast_stop_thread(struct net_device *dev, int flush)
 630 {
 631         struct ipoib_dev_priv *priv = netdev_priv(dev);
 632         struct ipoib_mcast *mcast;
 633
 634         ipoib_dbg_mcast(priv, "stopping multicast thread\n");
 635
 636         spin_lock_irq(&priv->lock);
 637         clear_bit(IPOIB_MCAST_STARTED, &priv->flags);
 638         spin_unlock_irq(&priv->lock);
 639
 640         mutex_lock(&mcast_mutex);
 641         clear_bit(IPOIB_MCAST_RUN, &priv->flags);
 642         cancel_delayed_work(&priv->mcast_task);
 643         mutex_unlock(&mcast_mutex);
 644
 645         if (flush)
 646                 flush_workqueue(ipoib_workqueue);
 647
 648         wait_for_mcast_join(priv, priv->broadcast);
 649
 650         list_for_each_entry(mcast, &priv->multicast_list, list)
 651                 wait_for_mcast_join(priv, mcast);
 652
 653         return 0;
 654 }
 655
 656 static int ipoib_mcast_leave(struct net_device *dev, struct ipoib_mcast *mcast)
 657 {
 658         struct ipoib_dev_priv *priv = netdev_priv(dev);
 659         struct ib_sa_mcmember_rec rec = {
 660                 .join_state = 1
 661         };
 662         int ret = 0;
 663
 664         if (!test_and_clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags))
 665                 return 0;
 666
 667         ipoib_dbg_mcast(priv, "leaving MGID " IPOIB_GID_FMT "\n",
 668                         IPOIB_GID_ARG(mcast->mcmember.mgid));
 669
 670         rec.mgid     = mcast->mcmember.mgid;
 671         rec.port_gid = priv->local_gid;
 672         rec.pkey     = cpu_to_be16(priv->pkey);
 673
 674         /* Remove ourselves from the multicast group */
 675         ret = ipoib_mcast_detach(dev, be16_to_cpu(mcast->mcmember.mlid),
 676                                  &mcast->mcmember.mgid);
 677         if (ret)
 678                 ipoib_warn(priv, "ipoib_mcast_detach failed (result = %d)\n", ret);
 679
 680         /*
 681          * Just make one shot at leaving and don't wait for a reply;
 682          * if we fail, too bad.
 683          */
 684         ret = ib_sa_mcmember_rec_delete(priv->ca, priv->port, &rec,
 685                                         IB_SA_MCMEMBER_REC_MGID         |
 686                                         IB_SA_MCMEMBER_REC_PORT_GID     |
 687                                         IB_SA_MCMEMBER_REC_PKEY         |
 688                                         IB_SA_MCMEMBER_REC_JOIN_STATE,
 689                                         0, GFP_ATOMIC, NULL,
 690                                         mcast, &mcast->query);
 691         if (ret < 0)
 692                 ipoib_warn(priv, "ib_sa_mcmember_rec_delete failed "
 693                            "for leave (result = %d)\n", ret);
 694
 695         return 0;
 696 }
 697
 698 void ipoib_mcast_send(struct net_device *dev, void *mgid, struct sk_buff *skb)
 699 {
 700         struct ipoib_dev_priv *priv = netdev_priv(dev);
 701         struct ipoib_mcast *mcast;
 702
 703         /*
 704          * We can only be called from ipoib_start_xmit, so we're
 705          * inside tx_lock -- no need to save/restore flags.
 706          */
 707         spin_lock(&priv->lock);
 708
 709         if (!test_bit(IPOIB_MCAST_STARTED, &priv->flags)        ||
 710             !priv->broadcast                                    ||
 711             !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
 712                 ++priv->stats.tx_dropped;
 713                 dev_kfree_skb_any(skb);
 714                 goto unlock;
 715         }
 716
 717         mcast = __ipoib_mcast_find(dev, mgid);
 718         if (!mcast) {
 719                 /* Let's create a new send only group now */
 720                 ipoib_dbg_mcast(priv, "setting up send only multicast group for "
 721                                 IPOIB_GID_FMT "\n", IPOIB_GID_RAW_ARG(mgid));
 722
 723                 mcast = ipoib_mcast_alloc(dev, 0);
 724                 if (!mcast) {
 725                         ipoib_warn(priv, "unable to allocate memory for "
 726                                    "multicast structure\n");
 727                         ++priv->stats.tx_dropped;
 728                         dev_kfree_skb_any(skb);
 729                         goto out;
 730                 }
 731
 732                 set_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags);
 733                 memcpy(mcast->mcmember.mgid.raw, mgid, sizeof (union ib_gid));
 734                 __ipoib_mcast_add(dev, mcast);
 735                 list_add_tail(&mcast->list, &priv->multicast_list);
 736         }
 737
 738         if (!mcast->ah) {
 739                 if (skb_queue_len(&mcast->pkt_queue) < IPOIB_MAX_MCAST_QUEUE)
 740                         skb_queue_tail(&mcast->pkt_queue, skb);
 741                 else {
 742                         ++priv->stats.tx_dropped;
 743                         dev_kfree_skb_any(skb);
 744                 }
 745
 746                 if (mcast->query)
 747                         ipoib_dbg_mcast(priv, "no address vector, "
 748                                         "but multicast join already started\n");
 749                 else if (test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags))
 750                         ipoib_mcast_sendonly_join(mcast);
 751
 752                 /*
 753                  * If lookup completes between here and out:, don't
 754                  * want to send packet twice.
 755                  */
 756                 mcast = NULL;
 757         }
 758
 759 out:
 760         if (mcast && mcast->ah) {
 761                 if (skb->dst            &&
 762                     skb->dst->neighbour &&
 763                     !*to_ipoib_neigh(skb->dst->neighbour)) {
 764                         struct ipoib_neigh *neigh = ipoib_neigh_alloc(skb->dst->neighbour);
 765
 766                         if (neigh) {
 767                                 kref_get(&mcast->ah->ref);
 768                                 neigh->ah       = mcast->ah;
 769                                 list_add_tail(&neigh->list, &mcast->neigh_list);
 770                         }
 771                 }
 772
 773                 ipoib_send(dev, skb, mcast->ah, IB_MULTICAST_QPN);
 774         }
 775
 776 unlock:
 777         spin_unlock(&priv->lock);
 778 }
 779
 780 void ipoib_mcast_dev_flush(struct net_device *dev)
 781 {
 782         struct ipoib_dev_priv *priv = netdev_priv(dev);
 783         LIST_HEAD(remove_list);
 784         struct ipoib_mcast *mcast, *tmcast;
 785         unsigned long flags;
 786
 787         ipoib_dbg_mcast(priv, "flushing multicast list\n");
 788
 789         spin_lock_irqsave(&priv->lock, flags);
 790
 791         list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
 792                 list_del(&mcast->list);
 793                 rb_erase(&mcast->rb_node, &priv->multicast_tree);
 794                 list_add_tail(&mcast->list, &remove_list);
 795         }
 796
 797         if (priv->broadcast) {
 798                 rb_erase(&priv->broadcast->rb_node, &priv->multicast_tree);
 799                 list_add_tail(&priv->broadcast->list, &remove_list);
 800                 priv->broadcast = NULL;
 801         }
 802
 803         spin_unlock_irqrestore(&priv->lock, flags);
 804
 805         list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
 806                 ipoib_mcast_leave(dev, mcast);
 807                 ipoib_mcast_free(mcast);
 808         }
 809 }
 810
 811 void ipoib_mcast_restart_task(void *dev_ptr)
 812 {
 813         struct net_device *dev = dev_ptr;
 814         struct ipoib_dev_priv *priv = netdev_priv(dev);
 815         struct dev_mc_list *mclist;
 816         struct ipoib_mcast *mcast, *tmcast;
 817         LIST_HEAD(remove_list);
 818         unsigned long flags;
 819
 820         ipoib_dbg_mcast(priv, "restarting multicast task\n");
 821
 822         ipoib_mcast_stop_thread(dev, 0);
 823
 824         local_irq_save(flags);
 825         netif_tx_lock(dev);
 826         spin_lock(&priv->lock);
 827
 828         /*
 829          * Unfortunately, the networking core only gives us a list of all of
 830          * the multicast hardware addresses. We need to figure out which ones
 831          * are new and which ones have been removed
 832          */
 833
 834         /* Clear out the found flag */
 835         list_for_each_entry(mcast, &priv->multicast_list, list)
 836                 clear_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
 837
 838         /* Mark all of the entries that are found or don't exist */
 839         for (mclist = dev->mc_list; mclist; mclist = mclist->next) {
 840                 union ib_gid mgid;
 841
 842                 memcpy(mgid.raw, mclist->dmi_addr + 4, sizeof mgid);
 843
 844                 /* Add in the P_Key */
 845                 mgid.raw[4] = (priv->pkey >> 8) & 0xff;
 846                 mgid.raw[5] = priv->pkey & 0xff;
 847
 848                 mcast = __ipoib_mcast_find(dev, &mgid);
 849                 if (!mcast || test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 850                         struct ipoib_mcast *nmcast;
 851
 852                         /* Not found or send-only group, let's add a new entry */
 853                         ipoib_dbg_mcast(priv, "adding multicast entry for mgid "
 854                                         IPOIB_GID_FMT "\n", IPOIB_GID_ARG(mgid));
 855
 856                         nmcast = ipoib_mcast_alloc(dev, 0);
 857                         if (!nmcast) {
 858                                 ipoib_warn(priv, "unable to allocate memory for multicast structure\n");
 859                                 continue;
 860                         }
 861
 862                         set_bit(IPOIB_MCAST_FLAG_FOUND, &nmcast->flags);
 863
 864                         nmcast->mcmember.mgid = mgid;
 865
 866                         if (mcast) {
 867                                 /* Destroy the send only entry */
 868                                 list_move_tail(&mcast->list, &remove_list);
 869
 870                                 rb_replace_node(&mcast->rb_node,
 871                                                 &nmcast->rb_node,
 872                                                 &priv->multicast_tree);
 873                         } else
 874                                 __ipoib_mcast_add(dev, nmcast);
 875
 876                         list_add_tail(&nmcast->list, &priv->multicast_list);
 877                 }
 878
 879                 if (mcast)
 880                         set_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
 881         }
 882
 883         /* Remove all of the entries don't exist anymore */
 884         list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
 885                 if (!test_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags) &&
 886                     !test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 887                         ipoib_dbg_mcast(priv, "deleting multicast group " IPOIB_GID_FMT "\n",
 888                                         IPOIB_GID_ARG(mcast->mcmember.mgid));
 889
 890                         rb_erase(&mcast->rb_node, &priv->multicast_tree);
 891
 892                         /* Move to the remove list */
 893                         list_move_tail(&mcast->list, &remove_list);
 894                 }
 895         }
 896
 897         spin_unlock(&priv->lock);
 898         netif_tx_unlock(dev);
 899         local_irq_restore(flags);
 900
 901         /* We have to cancel outside of the spinlock */
 902         list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
 903                 wait_for_mcast_join(priv, mcast);
 904                 ipoib_mcast_leave(mcast->dev, mcast);
 905                 ipoib_mcast_free(mcast);
 906         }
 907
 908         if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 909                 ipoib_mcast_start_thread(dev);
 910 }
 911
 912 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 913
 914 struct ipoib_mcast_iter *ipoib_mcast_iter_init(struct net_device *dev)
 915 {
 916         struct ipoib_mcast_iter *iter;
 917
 918         iter = kmalloc(sizeof *iter, GFP_KERNEL);
 919         if (!iter)
 920                 return NULL;
 921
 922         iter->dev = dev;
 923         memset(iter->mgid.raw, 0, 16);
 924
 925         if (ipoib_mcast_iter_next(iter)) {
 926                 kfree(iter);
 927                 return NULL;
 928         }
 929
 930         return iter;
 931 }
 932
 933 int ipoib_mcast_iter_next(struct ipoib_mcast_iter *iter)
 934 {
 935         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
 936         struct rb_node *n;
 937         struct ipoib_mcast *mcast;
 938         int ret = 1;
 939
 940         spin_lock_irq(&priv->lock);
 941
 942         n = rb_first(&priv->multicast_tree);
 943
 944         while (n) {
 945                 mcast = rb_entry(n, struct ipoib_mcast, rb_node);
 946
 947                 if (memcmp(iter->mgid.raw, mcast->mcmember.mgid.raw,
 948                            sizeof (union ib_gid)) < 0) {
 949                         iter->mgid      = mcast->mcmember.mgid;
 950                         iter->created   = mcast->created;
 951                         iter->queuelen  = skb_queue_len(&mcast->pkt_queue);
 952                         iter->complete  = !!mcast->ah;
 953                         iter->send_only = !!(mcast->flags & (1 << IPOIB_MCAST_FLAG_SENDONLY));
 954
 955                         ret = 0;
 956
 957                         break;
 958                 }
 959
 960                 n = rb_next(n);
 961         }
 962
 963         spin_unlock_irq(&priv->lock);
 964
 965         return ret;
 966 }
 967
 968 void ipoib_mcast_iter_read(struct ipoib_mcast_iter *iter,
 969                            union ib_gid *mgid,
 970                            unsigned long *created,
 971                            unsigned int *queuelen,
 972                            unsigned int *complete,
 973                            unsigned int *send_only)
 974 {
 975         *mgid      = iter->mgid;
 976         *created   = iter->created;
 977         *queuelen  = iter->queuelen;
 978         *complete  = iter->complete;
 979         *send_only = iter->send_only;
 980 }
 981
 982 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */