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                 *to_ipoib_neigh(neigh->neighbour) = NULL;
 118                 neigh->neighbour->ops->destructor = NULL;
 119                 kfree(neigh);
 120         }
 121
 122         spin_unlock_irqrestore(&priv->lock, flags);
 123
 124         if (mcast->ah)
 125                 ipoib_put_ah(mcast->ah);
 126
 127         while (!skb_queue_empty(&mcast->pkt_queue)) {
 128                 ++tx_dropped;
 129                 dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
 130         }
 131
 132         spin_lock_irqsave(&priv->tx_lock, flags);
 133         priv->stats.tx_dropped += tx_dropped;
 134         spin_unlock_irqrestore(&priv->tx_lock, flags);
 135
 136         kfree(mcast);
 137 }
 138
 139 static struct ipoib_mcast *ipoib_mcast_alloc(struct net_device *dev,
 140                                              int can_sleep)
 141 {
 142         struct ipoib_mcast *mcast;
 143
 144         mcast = kzalloc(sizeof *mcast, can_sleep ? GFP_KERNEL : GFP_ATOMIC);
 145         if (!mcast)
 146                 return NULL;
 147
 148         mcast->dev = dev;
 149         mcast->created = 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, union ib_gid *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->raw, 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         int ret;
 217
 218         mcast->mcmember = *mcmember;
 219
 220         /* Set the cached Q_Key before we attach if it's the broadcast group */
 221         if (!memcmp(mcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
 222                     sizeof (union ib_gid))) {
 223                 priv->qkey = be32_to_cpu(priv->broadcast->mcmember.qkey);
 224                 priv->tx_wr.wr.ud.remote_qkey = priv->qkey;
 225         }
 226
 227         if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 228                 if (test_and_set_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 229                         ipoib_warn(priv, "multicast group " IPOIB_GID_FMT
 230                                    " already attached\n",
 231                                    IPOIB_GID_ARG(mcast->mcmember.mgid));
 232
 233                         return 0;
 234                 }
 235
 236                 ret = ipoib_mcast_attach(dev, be16_to_cpu(mcast->mcmember.mlid),
 237                                          &mcast->mcmember.mgid);
 238                 if (ret < 0) {
 239                         ipoib_warn(priv, "couldn't attach QP to multicast group "
 240                                    IPOIB_GID_FMT "\n",
 241                                    IPOIB_GID_ARG(mcast->mcmember.mgid));
 242
 243                         clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags);
 244                         return ret;
 245                 }
 246         }
 247
 248         {
 249                 struct ib_ah_attr av = {
 250                         .dlid          = be16_to_cpu(mcast->mcmember.mlid),
 251                         .port_num      = priv->port,
 252                         .sl            = mcast->mcmember.sl,
 253                         .ah_flags      = IB_AH_GRH,
 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                 int path_rate = ib_sa_rate_enum_to_int(mcast->mcmember.rate);
 262
 263                 av.grh.dgid = mcast->mcmember.mgid;
 264
 265                 if (path_rate > 0 && priv->local_rate > path_rate)
 266                         av.static_rate = (priv->local_rate - 1) / path_rate;
 267
 268                 ipoib_dbg_mcast(priv, "static_rate %d for local port %dX, mcmember %dX\n",
 269                                 av.static_rate, priv->local_rate,
 270                                 ib_sa_rate_enum_to_int(mcast->mcmember.rate));
 271
 272                 mcast->ah = ipoib_create_ah(dev, priv->pd, &av);
 273                 if (!mcast->ah) {
 274                         ipoib_warn(priv, "ib_address_create failed\n");
 275                 } else {
 276                         ipoib_dbg_mcast(priv, "MGID " IPOIB_GID_FMT
 277                                         " AV %p, LID 0x%04x, SL %d\n",
 278                                         IPOIB_GID_ARG(mcast->mcmember.mgid),
 279                                         mcast->ah->ah,
 280                                         be16_to_cpu(mcast->mcmember.mlid),
 281                                         mcast->mcmember.sl);
 282                 }
 283         }
 284
 285         /* actually send any queued packets */
 286         spin_lock_irq(&priv->tx_lock);
 287         while (!skb_queue_empty(&mcast->pkt_queue)) {
 288                 struct sk_buff *skb = skb_dequeue(&mcast->pkt_queue);
 289                 spin_unlock_irq(&priv->tx_lock);
 290
 291                 skb->dev = dev;
 292
 293                 if (!skb->dst || !skb->dst->neighbour) {
 294                         /* put pseudoheader back on for next time */
 295                         skb_push(skb, sizeof (struct ipoib_pseudoheader));
 296                 }
 297
 298                 if (dev_queue_xmit(skb))
 299                         ipoib_warn(priv, "dev_queue_xmit failed to requeue packet\n");
 300                 spin_lock_irq(&priv->tx_lock);
 301         }
 302         spin_unlock_irq(&priv->tx_lock);
 303
 304         return 0;
 305 }
 306
 307 static void
 308 ipoib_mcast_sendonly_join_complete(int status,
 309                                    struct ib_sa_mcmember_rec *mcmember,
 310                                    void *mcast_ptr)
 311 {
 312         struct ipoib_mcast *mcast = mcast_ptr;
 313         struct net_device *dev = mcast->dev;
 314         struct ipoib_dev_priv *priv = netdev_priv(dev);
 315
 316         if (!status)
 317                 ipoib_mcast_join_finish(mcast, mcmember);
 318         else {
 319                 if (mcast->logcount++ < 20)
 320                         ipoib_dbg_mcast(netdev_priv(dev), "multicast join failed for "
 321                                         IPOIB_GID_FMT ", status %d\n",
 322                                         IPOIB_GID_ARG(mcast->mcmember.mgid), status);
 323
 324                 /* Flush out any queued packets */
 325                 spin_lock_irq(&priv->tx_lock);
 326                 while (!skb_queue_empty(&mcast->pkt_queue)) {
 327                         ++priv->stats.tx_dropped;
 328                         dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
 329                 }
 330                 spin_unlock_irq(&priv->tx_lock);
 331
 332                 /* Clear the busy flag so we try again */
 333                 clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 334         }
 335
 336         complete(&mcast->done);
 337 }
 338
 339 static int ipoib_mcast_sendonly_join(struct ipoib_mcast *mcast)
 340 {
 341         struct net_device *dev = mcast->dev;
 342         struct ipoib_dev_priv *priv = netdev_priv(dev);
 343         struct ib_sa_mcmember_rec rec = {
 344 #if 0                           /* Some SMs don't support send-only yet */
 345                 .join_state = 4
 346 #else
 347                 .join_state = 1
 348 #endif
 349         };
 350         int ret = 0;
 351
 352         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
 353                 ipoib_dbg_mcast(priv, "device shutting down, no multicast joins\n");
 354                 return -ENODEV;
 355         }
 356
 357         if (test_and_set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)) {
 358                 ipoib_dbg_mcast(priv, "multicast entry busy, skipping\n");
 359                 return -EBUSY;
 360         }
 361
 362         rec.mgid     = mcast->mcmember.mgid;
 363         rec.port_gid = priv->local_gid;
 364         rec.pkey     = cpu_to_be16(priv->pkey);
 365
 366         init_completion(&mcast->done);
 367
 368         ret = ib_sa_mcmember_rec_set(priv->ca, priv->port, &rec,
 369                                      IB_SA_MCMEMBER_REC_MGID            |
 370                                      IB_SA_MCMEMBER_REC_PORT_GID        |
 371                                      IB_SA_MCMEMBER_REC_PKEY            |
 372                                      IB_SA_MCMEMBER_REC_JOIN_STATE,
 373                                      1000, GFP_ATOMIC,
 374                                      ipoib_mcast_sendonly_join_complete,
 375                                      mcast, &mcast->query);
 376         if (ret < 0) {
 377                 ipoib_warn(priv, "ib_sa_mcmember_rec_set failed (ret = %d)\n",
 378                            ret);
 379         } else {
 380                 ipoib_dbg_mcast(priv, "no multicast record for " IPOIB_GID_FMT
 381                                 ", starting join\n",
 382                                 IPOIB_GID_ARG(mcast->mcmember.mgid));
 383
 384                 mcast->query_id = ret;
 385         }
 386
 387         return ret;
 388 }
 389
 390 static void ipoib_mcast_join_complete(int status,
 391                                       struct ib_sa_mcmember_rec *mcmember,
 392                                       void *mcast_ptr)
 393 {
 394         struct ipoib_mcast *mcast = mcast_ptr;
 395         struct net_device *dev = mcast->dev;
 396         struct ipoib_dev_priv *priv = netdev_priv(dev);
 397
 398         ipoib_dbg_mcast(priv, "join completion for " IPOIB_GID_FMT
 399                         " (status %d)\n",
 400                         IPOIB_GID_ARG(mcast->mcmember.mgid), status);
 401
 402         if (!status && !ipoib_mcast_join_finish(mcast, mcmember)) {
 403                 mcast->backoff = 1;
 404                 mutex_lock(&mcast_mutex);
 405                 if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 406                         queue_work(ipoib_workqueue, &priv->mcast_task);
 407                 mutex_unlock(&mcast_mutex);
 408                 complete(&mcast->done);
 409                 return;
 410         }
 411
 412         if (status == -EINTR) {
 413                 complete(&mcast->done);
 414                 return;
 415         }
 416
 417         if (status && mcast->logcount++ < 20) {
 418                 if (status == -ETIMEDOUT || status == -EINTR) {
 419                         ipoib_dbg_mcast(priv, "multicast join failed for " IPOIB_GID_FMT
 420                                         ", status %d\n",
 421                                         IPOIB_GID_ARG(mcast->mcmember.mgid),
 422                                         status);
 423                 } else {
 424                         ipoib_warn(priv, "multicast join failed for "
 425                                    IPOIB_GID_FMT ", status %d\n",
 426                                    IPOIB_GID_ARG(mcast->mcmember.mgid),
 427                                    status);
 428                 }
 429         }
 430
 431         mcast->backoff *= 2;
 432         if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
 433                 mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
 434
 435         mcast->query = NULL;
 436
 437         mutex_lock(&mcast_mutex);
 438         if (test_bit(IPOIB_MCAST_RUN, &priv->flags)) {
 439                 if (status == -ETIMEDOUT)
 440                         queue_work(ipoib_workqueue, &priv->mcast_task);
 441                 else
 442                         queue_delayed_work(ipoib_workqueue, &priv->mcast_task,
 443                                            mcast->backoff * HZ);
 444         } else
 445                 complete(&mcast->done);
 446         mutex_unlock(&mcast_mutex);
 447
 448         return;
 449 }
 450
 451 static void ipoib_mcast_join(struct net_device *dev, struct ipoib_mcast *mcast,
 452                              int create)
 453 {
 454         struct ipoib_dev_priv *priv = netdev_priv(dev);
 455         struct ib_sa_mcmember_rec rec = {
 456                 .join_state = 1
 457         };
 458         ib_sa_comp_mask comp_mask;
 459         int ret = 0;
 460
 461         ipoib_dbg_mcast(priv, "joining MGID " IPOIB_GID_FMT "\n",
 462                         IPOIB_GID_ARG(mcast->mcmember.mgid));
 463
 464         rec.mgid     = mcast->mcmember.mgid;
 465         rec.port_gid = priv->local_gid;
 466         rec.pkey     = cpu_to_be16(priv->pkey);
 467
 468         comp_mask =
 469                 IB_SA_MCMEMBER_REC_MGID         |
 470                 IB_SA_MCMEMBER_REC_PORT_GID     |
 471                 IB_SA_MCMEMBER_REC_PKEY         |
 472                 IB_SA_MCMEMBER_REC_JOIN_STATE;
 473
 474         if (create) {
 475                 comp_mask |=
 476                         IB_SA_MCMEMBER_REC_QKEY         |
 477                         IB_SA_MCMEMBER_REC_SL           |
 478                         IB_SA_MCMEMBER_REC_FLOW_LABEL   |
 479                         IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
 480
 481                 rec.qkey          = priv->broadcast->mcmember.qkey;
 482                 rec.sl            = priv->broadcast->mcmember.sl;
 483                 rec.flow_label    = priv->broadcast->mcmember.flow_label;
 484                 rec.traffic_class = priv->broadcast->mcmember.traffic_class;
 485         }
 486
 487         init_completion(&mcast->done);
 488
 489         ret = ib_sa_mcmember_rec_set(priv->ca, priv->port, &rec, comp_mask,
 490                                      mcast->backoff * 1000, GFP_ATOMIC,
 491                                      ipoib_mcast_join_complete,
 492                                      mcast, &mcast->query);
 493
 494         if (ret < 0) {
 495                 ipoib_warn(priv, "ib_sa_mcmember_rec_set failed, status %d\n", ret);
 496
 497                 mcast->backoff *= 2;
 498                 if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
 499                         mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
 500
 501                 mutex_lock(&mcast_mutex);
 502                 if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 503                         queue_delayed_work(ipoib_workqueue,
 504                                            &priv->mcast_task,
 505                                            mcast->backoff * HZ);
 506                 mutex_unlock(&mcast_mutex);
 507         } else
 508                 mcast->query_id = ret;
 509 }
 510
 511 void ipoib_mcast_join_task(void *dev_ptr)
 512 {
 513         struct net_device *dev = dev_ptr;
 514         struct ipoib_dev_priv *priv = netdev_priv(dev);
 515
 516         if (!test_bit(IPOIB_MCAST_RUN, &priv->flags))
 517                 return;
 518
 519         if (ib_query_gid(priv->ca, priv->port, 0, &priv->local_gid))
 520                 ipoib_warn(priv, "ib_gid_entry_get() failed\n");
 521         else
 522                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
 523
 524         {
 525                 struct ib_port_attr attr;
 526
 527                 if (!ib_query_port(priv->ca, priv->port, &attr)) {
 528                         priv->local_lid  = attr.lid;
 529                         priv->local_rate = attr.active_speed *
 530                                 ib_width_enum_to_int(attr.active_width);
 531                 } else
 532                         ipoib_warn(priv, "ib_query_port failed\n");
 533         }
 534
 535         if (!priv->broadcast) {
 536                 struct ipoib_mcast *broadcast;
 537
 538                 broadcast = ipoib_mcast_alloc(dev, 1);
 539                 if (!broadcast) {
 540                         ipoib_warn(priv, "failed to allocate broadcast group\n");
 541                         mutex_lock(&mcast_mutex);
 542                         if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 543                                 queue_delayed_work(ipoib_workqueue,
 544                                                    &priv->mcast_task, HZ);
 545                         mutex_unlock(&mcast_mutex);
 546                         return;
 547                 }
 548
 549                 spin_lock_irq(&priv->lock);
 550                 memcpy(broadcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
 551                        sizeof (union ib_gid));
 552                 priv->broadcast = broadcast;
 553
 554                 __ipoib_mcast_add(dev, priv->broadcast);
 555                 spin_unlock_irq(&priv->lock);
 556         }
 557
 558         if (!test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
 559                 ipoib_mcast_join(dev, priv->broadcast, 0);
 560                 return;
 561         }
 562
 563         while (1) {
 564                 struct ipoib_mcast *mcast = NULL;
 565
 566                 spin_lock_irq(&priv->lock);
 567                 list_for_each_entry(mcast, &priv->multicast_list, list) {
 568                         if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)
 569                             && !test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)
 570                             && !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 571                                 /* Found the next unjoined group */
 572                                 break;
 573                         }
 574                 }
 575                 spin_unlock_irq(&priv->lock);
 576
 577                 if (&mcast->list == &priv->multicast_list) {
 578                         /* All done */
 579                         break;
 580                 }
 581
 582                 ipoib_mcast_join(dev, mcast, 1);
 583                 return;
 584         }
 585
 586         priv->mcast_mtu = ib_mtu_enum_to_int(priv->broadcast->mcmember.mtu) -
 587                 IPOIB_ENCAP_LEN;
 588         dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
 589
 590         ipoib_dbg_mcast(priv, "successfully joined all multicast groups\n");
 591
 592         clear_bit(IPOIB_MCAST_RUN, &priv->flags);
 593         netif_carrier_on(dev);
 594 }
 595
 596 int ipoib_mcast_start_thread(struct net_device *dev)
 597 {
 598         struct ipoib_dev_priv *priv = netdev_priv(dev);
 599
 600         ipoib_dbg_mcast(priv, "starting multicast thread\n");
 601
 602         mutex_lock(&mcast_mutex);
 603         if (!test_and_set_bit(IPOIB_MCAST_RUN, &priv->flags))
 604                 queue_work(ipoib_workqueue, &priv->mcast_task);
 605         mutex_unlock(&mcast_mutex);
 606
 607         spin_lock_irq(&priv->lock);
 608         set_bit(IPOIB_MCAST_STARTED, &priv->flags);
 609         spin_unlock_irq(&priv->lock);
 610
 611         return 0;
 612 }
 613
 614 int ipoib_mcast_stop_thread(struct net_device *dev, int flush)
 615 {
 616         struct ipoib_dev_priv *priv = netdev_priv(dev);
 617         struct ipoib_mcast *mcast;
 618
 619         ipoib_dbg_mcast(priv, "stopping multicast thread\n");
 620
 621         spin_lock_irq(&priv->lock);
 622         clear_bit(IPOIB_MCAST_STARTED, &priv->flags);
 623         spin_unlock_irq(&priv->lock);
 624
 625         mutex_lock(&mcast_mutex);
 626         clear_bit(IPOIB_MCAST_RUN, &priv->flags);
 627         cancel_delayed_work(&priv->mcast_task);
 628         mutex_unlock(&mcast_mutex);
 629
 630         if (flush)
 631                 flush_workqueue(ipoib_workqueue);
 632
 633         if (priv->broadcast && priv->broadcast->query) {
 634                 ib_sa_cancel_query(priv->broadcast->query_id, priv->broadcast->query);
 635                 priv->broadcast->query = NULL;
 636                 ipoib_dbg_mcast(priv, "waiting for bcast\n");
 637                 wait_for_completion(&priv->broadcast->done);
 638         }
 639
 640         list_for_each_entry(mcast, &priv->multicast_list, list) {
 641                 if (mcast->query) {
 642                         ib_sa_cancel_query(mcast->query_id, mcast->query);
 643                         mcast->query = NULL;
 644                         ipoib_dbg_mcast(priv, "waiting for MGID " IPOIB_GID_FMT "\n",
 645                                         IPOIB_GID_ARG(mcast->mcmember.mgid));
 646                         wait_for_completion(&mcast->done);
 647                 }
 648         }
 649
 650         return 0;
 651 }
 652
 653 static int ipoib_mcast_leave(struct net_device *dev, struct ipoib_mcast *mcast)
 654 {
 655         struct ipoib_dev_priv *priv = netdev_priv(dev);
 656         struct ib_sa_mcmember_rec rec = {
 657                 .join_state = 1
 658         };
 659         int ret = 0;
 660
 661         if (!test_and_clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags))
 662                 return 0;
 663
 664         ipoib_dbg_mcast(priv, "leaving MGID " IPOIB_GID_FMT "\n",
 665                         IPOIB_GID_ARG(mcast->mcmember.mgid));
 666
 667         rec.mgid     = mcast->mcmember.mgid;
 668         rec.port_gid = priv->local_gid;
 669         rec.pkey     = cpu_to_be16(priv->pkey);
 670
 671         /* Remove ourselves from the multicast group */
 672         ret = ipoib_mcast_detach(dev, be16_to_cpu(mcast->mcmember.mlid),
 673                                  &mcast->mcmember.mgid);
 674         if (ret)
 675                 ipoib_warn(priv, "ipoib_mcast_detach failed (result = %d)\n", ret);
 676
 677         /*
 678          * Just make one shot at leaving and don't wait for a reply;
 679          * if we fail, too bad.
 680          */
 681         ret = ib_sa_mcmember_rec_delete(priv->ca, priv->port, &rec,
 682                                         IB_SA_MCMEMBER_REC_MGID         |
 683                                         IB_SA_MCMEMBER_REC_PORT_GID     |
 684                                         IB_SA_MCMEMBER_REC_PKEY         |
 685                                         IB_SA_MCMEMBER_REC_JOIN_STATE,
 686                                         0, GFP_ATOMIC, NULL,
 687                                         mcast, &mcast->query);
 688         if (ret < 0)
 689                 ipoib_warn(priv, "ib_sa_mcmember_rec_delete failed "
 690                            "for leave (result = %d)\n", ret);
 691
 692         return 0;
 693 }
 694
 695 void ipoib_mcast_send(struct net_device *dev, union ib_gid *mgid,
 696                       struct sk_buff *skb)
 697 {
 698         struct ipoib_dev_priv *priv = netdev_priv(dev);
 699         struct ipoib_mcast *mcast;
 700
 701         /*
 702          * We can only be called from ipoib_start_xmit, so we're
 703          * inside tx_lock -- no need to save/restore flags.
 704          */
 705         spin_lock(&priv->lock);
 706
 707         if (!test_bit(IPOIB_MCAST_STARTED, &priv->flags)        ||
 708             !priv->broadcast                                    ||
 709             !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
 710                 ++priv->stats.tx_dropped;
 711                 dev_kfree_skb_any(skb);
 712                 goto unlock;
 713         }
 714
 715         mcast = __ipoib_mcast_find(dev, mgid);
 716         if (!mcast) {
 717                 /* Let's create a new send only group now */
 718                 ipoib_dbg_mcast(priv, "setting up send only multicast group for "
 719                                 IPOIB_GID_FMT "\n", IPOIB_GID_ARG(*mgid));
 720
 721                 mcast = ipoib_mcast_alloc(dev, 0);
 722                 if (!mcast) {
 723                         ipoib_warn(priv, "unable to allocate memory for "
 724                                    "multicast structure\n");
 725                         ++priv->stats.tx_dropped;
 726                         dev_kfree_skb_any(skb);
 727                         goto out;
 728                 }
 729
 730                 set_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags);
 731                 mcast->mcmember.mgid = *mgid;
 732                 __ipoib_mcast_add(dev, mcast);
 733                 list_add_tail(&mcast->list, &priv->multicast_list);
 734         }
 735
 736         if (!mcast->ah) {
 737                 if (skb_queue_len(&mcast->pkt_queue) < IPOIB_MAX_MCAST_QUEUE)
 738                         skb_queue_tail(&mcast->pkt_queue, skb);
 739                 else {
 740                         ++priv->stats.tx_dropped;
 741                         dev_kfree_skb_any(skb);
 742                 }
 743
 744                 if (mcast->query)
 745                         ipoib_dbg_mcast(priv, "no address vector, "
 746                                         "but multicast join already started\n");
 747                 else if (test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags))
 748                         ipoib_mcast_sendonly_join(mcast);
 749
 750                 /*
 751                  * If lookup completes between here and out:, don't
 752                  * want to send packet twice.
 753                  */
 754                 mcast = NULL;
 755         }
 756
 757 out:
 758         if (mcast && mcast->ah) {
 759                 if (skb->dst            &&
 760                     skb->dst->neighbour &&
 761                     !*to_ipoib_neigh(skb->dst->neighbour)) {
 762                         struct ipoib_neigh *neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
 763
 764                         if (neigh) {
 765                                 kref_get(&mcast->ah->ref);
 766                                 neigh->ah       = mcast->ah;
 767                                 neigh->neighbour = skb->dst->neighbour;
 768                                 *to_ipoib_neigh(skb->dst->neighbour) = neigh;
 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         spin_lock_irqsave(&dev->xmit_lock, flags);
 825         spin_lock(&priv->lock);
 826
 827         /*
 828          * Unfortunately, the networking core only gives us a list of all of
 829          * the multicast hardware addresses. We need to figure out which ones
 830          * are new and which ones have been removed
 831          */
 832
 833         /* Clear out the found flag */
 834         list_for_each_entry(mcast, &priv->multicast_list, list)
 835                 clear_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
 836
 837         /* Mark all of the entries that are found or don't exist */
 838         for (mclist = dev->mc_list; mclist; mclist = mclist->next) {
 839                 union ib_gid mgid;
 840
 841                 memcpy(mgid.raw, mclist->dmi_addr + 4, sizeof mgid);
 842
 843                 /* Add in the P_Key */
 844                 mgid.raw[4] = (priv->pkey >> 8) & 0xff;
 845                 mgid.raw[5] = priv->pkey & 0xff;
 846
 847                 mcast = __ipoib_mcast_find(dev, &mgid);
 848                 if (!mcast || test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 849                         struct ipoib_mcast *nmcast;
 850
 851                         /* Not found or send-only group, let's add a new entry */
 852                         ipoib_dbg_mcast(priv, "adding multicast entry for mgid "
 853                                         IPOIB_GID_FMT "\n", IPOIB_GID_ARG(mgid));
 854
 855                         nmcast = ipoib_mcast_alloc(dev, 0);
 856                         if (!nmcast) {
 857                                 ipoib_warn(priv, "unable to allocate memory for multicast structure\n");
 858                                 continue;
 859                         }
 860
 861                         set_bit(IPOIB_MCAST_FLAG_FOUND, &nmcast->flags);
 862
 863                         nmcast->mcmember.mgid = mgid;
 864
 865                         if (mcast) {
 866                                 /* Destroy the send only entry */
 867                                 list_del(&mcast->list);
 868                                 list_add_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_del(&mcast->list);
 894                         list_add_tail(&mcast->list, &remove_list);
 895                 }
 896         }
 897
 898         spin_unlock(&priv->lock);
 899         spin_unlock_irqrestore(&dev->xmit_lock, flags);
 900
 901         /* We have to cancel outside of the spinlock */
 902         list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
 903                 ipoib_mcast_leave(mcast->dev, mcast);
 904                 ipoib_mcast_free(mcast);
 905         }
 906
 907         if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 908                 ipoib_mcast_start_thread(dev);
 909 }
 910
 911 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 912
 913 struct ipoib_mcast_iter *ipoib_mcast_iter_init(struct net_device *dev)
 914 {
 915         struct ipoib_mcast_iter *iter;
 916
 917         iter = kmalloc(sizeof *iter, GFP_KERNEL);
 918         if (!iter)
 919                 return NULL;
 920
 921         iter->dev = dev;
 922         memset(iter->mgid.raw, 0, 16);
 923
 924         if (ipoib_mcast_iter_next(iter)) {
 925                 kfree(iter);
 926                 return NULL;
 927         }
 928
 929         return iter;
 930 }
 931
 932 int ipoib_mcast_iter_next(struct ipoib_mcast_iter *iter)
 933 {
 934         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
 935         struct rb_node *n;
 936         struct ipoib_mcast *mcast;
 937         int ret = 1;
 938
 939         spin_lock_irq(&priv->lock);
 940
 941         n = rb_first(&priv->multicast_tree);
 942
 943         while (n) {
 944                 mcast = rb_entry(n, struct ipoib_mcast, rb_node);
 945
 946                 if (memcmp(iter->mgid.raw, mcast->mcmember.mgid.raw,
 947                            sizeof (union ib_gid)) < 0) {
 948                         iter->mgid      = mcast->mcmember.mgid;
 949                         iter->created   = mcast->created;
 950                         iter->queuelen  = skb_queue_len(&mcast->pkt_queue);
 951                         iter->complete  = !!mcast->ah;
 952                         iter->send_only = !!(mcast->flags & (1 << IPOIB_MCAST_FLAG_SENDONLY));
 953
 954                         ret = 0;
 955
 956                         break;
 957                 }
 958
 959                 n = rb_next(n);
 960         }
 961
 962         spin_unlock_irq(&priv->lock);
 963
 964         return ret;
 965 }
 966
 967 void ipoib_mcast_iter_read(struct ipoib_mcast_iter *iter,
 968                            union ib_gid *mgid,
 969                            unsigned long *created,
 970                            unsigned int *queuelen,
 971                            unsigned int *complete,
 972                            unsigned int *send_only)
 973 {
 974         *mgid      = iter->mgid;
 975         *created   = iter->created;
 976         *queuelen  = iter->queuelen;
 977         *complete  = iter->complete;
 978         *send_only = iter->send_only;
 979 }
 980
 981 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */