SAFE public projects git trees. - safe/jmp/linux-2.6/blob - drivers/infiniband/ulp/ipoib/ipoib_multicast.c

   1 /*
   2  * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
   3  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
   4  * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
   5  *
   6  * This software is available to you under a choice of one of two
   7  * licenses.  You may choose to be licensed under the terms of the GNU
   8  * General Public License (GPL) Version 2, available from the file
   9  * COPYING in the main directory of this source tree, or the
  10  * OpenIB.org BSD license below:
  11  *
  12  *     Redistribution and use in source and binary forms, with or
  13  *     without modification, are permitted provided that the following
  14  *     conditions are met:
  15  *
  16  *      - Redistributions of source code must retain the above
  17  *        copyright notice, this list of conditions and the following
  18  *        disclaimer.
  19  *
  20  *      - Redistributions in binary form must reproduce the above
  21  *        copyright notice, this list of conditions and the following
  22  *        disclaimer in the documentation and/or other materials
  23  *        provided with the distribution.
  24  *
  25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  32  * SOFTWARE.
  33  */
  34
  35 #include <linux/skbuff.h>
  36 #include <linux/rtnetlink.h>
  37 #include <linux/ip.h>
  38 #include <linux/in.h>
  39 #include <linux/igmp.h>
  40 #include <linux/inetdevice.h>
  41 #include <linux/delay.h>
  42 #include <linux/completion.h>
  43
  44 #include <net/dst.h>
  45
  46 #include "ipoib.h"
  47
  48 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
  49 static int mcast_debug_level;
  50
  51 module_param(mcast_debug_level, int, 0644);
  52 MODULE_PARM_DESC(mcast_debug_level,
  53                  "Enable multicast debug tracing if > 0");
  54 #endif
  55
  56 static DEFINE_MUTEX(mcast_mutex);
  57
  58 struct ipoib_mcast_iter {
  59         struct net_device *dev;
  60         union ib_gid       mgid;
  61         unsigned long      created;
  62         unsigned int       queuelen;
  63         unsigned int       complete;
  64         unsigned int       send_only;
  65 };
  66
  67 static void ipoib_mcast_free(struct ipoib_mcast *mcast)
  68 {
  69         struct net_device *dev = mcast->dev;
  70         struct ipoib_dev_priv *priv = netdev_priv(dev);
  71         struct ipoib_neigh *neigh, *tmp;
  72         unsigned long flags;
  73         int tx_dropped = 0;
  74
  75         ipoib_dbg_mcast(netdev_priv(dev),
  76                         "deleting multicast group " IPOIB_GID_FMT "\n",
  77                         IPOIB_GID_ARG(mcast->mcmember.mgid));
  78
  79         spin_lock_irqsave(&priv->lock, flags);
  80
  81         list_for_each_entry_safe(neigh, tmp, &mcast->neigh_list, list) {
  82                 /*
  83                  * It's safe to call ipoib_put_ah() inside priv->lock
  84                  * here, because we know that mcast->ah will always
  85                  * hold one more reference, so ipoib_put_ah() will
  86                  * never do more than decrement the ref count.
  87                  */
  88                 if (neigh->ah)
  89                         ipoib_put_ah(neigh->ah);
  90                 ipoib_neigh_free(dev, neigh);
  91         }
  92
  93         spin_unlock_irqrestore(&priv->lock, flags);
  94
  95         if (mcast->ah)
  96                 ipoib_put_ah(mcast->ah);
  97
  98         while (!skb_queue_empty(&mcast->pkt_queue)) {
  99                 ++tx_dropped;
 100                 dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
 101         }
 102
 103         spin_lock_irqsave(&priv->tx_lock, flags);
 104         dev->stats.tx_dropped += tx_dropped;
 105         spin_unlock_irqrestore(&priv->tx_lock, flags);
 106
 107         kfree(mcast);
 108 }
 109
 110 static struct ipoib_mcast *ipoib_mcast_alloc(struct net_device *dev,
 111                                              int can_sleep)
 112 {
 113         struct ipoib_mcast *mcast;
 114
 115         mcast = kzalloc(sizeof *mcast, can_sleep ? GFP_KERNEL : GFP_ATOMIC);
 116         if (!mcast)
 117                 return NULL;
 118
 119         mcast->dev = dev;
 120         mcast->created = jiffies;
 121         mcast->backoff = 1;
 122
 123         INIT_LIST_HEAD(&mcast->list);
 124         INIT_LIST_HEAD(&mcast->neigh_list);
 125         skb_queue_head_init(&mcast->pkt_queue);
 126
 127         return mcast;
 128 }
 129
 130 static struct ipoib_mcast *__ipoib_mcast_find(struct net_device *dev, void *mgid)
 131 {
 132         struct ipoib_dev_priv *priv = netdev_priv(dev);
 133         struct rb_node *n = priv->multicast_tree.rb_node;
 134
 135         while (n) {
 136                 struct ipoib_mcast *mcast;
 137                 int ret;
 138
 139                 mcast = rb_entry(n, struct ipoib_mcast, rb_node);
 140
 141                 ret = memcmp(mgid, mcast->mcmember.mgid.raw,
 142                              sizeof (union ib_gid));
 143                 if (ret < 0)
 144                         n = n->rb_left;
 145                 else if (ret > 0)
 146                         n = n->rb_right;
 147                 else
 148                         return mcast;
 149         }
 150
 151         return NULL;
 152 }
 153
 154 static int __ipoib_mcast_add(struct net_device *dev, struct ipoib_mcast *mcast)
 155 {
 156         struct ipoib_dev_priv *priv = netdev_priv(dev);
 157         struct rb_node **n = &priv->multicast_tree.rb_node, *pn = NULL;
 158
 159         while (*n) {
 160                 struct ipoib_mcast *tmcast;
 161                 int ret;
 162
 163                 pn = *n;
 164                 tmcast = rb_entry(pn, struct ipoib_mcast, rb_node);
 165
 166                 ret = memcmp(mcast->mcmember.mgid.raw, tmcast->mcmember.mgid.raw,
 167                              sizeof (union ib_gid));
 168                 if (ret < 0)
 169                         n = &pn->rb_left;
 170                 else if (ret > 0)
 171                         n = &pn->rb_right;
 172                 else
 173                         return -EEXIST;
 174         }
 175
 176         rb_link_node(&mcast->rb_node, pn, n);
 177         rb_insert_color(&mcast->rb_node, &priv->multicast_tree);
 178
 179         return 0;
 180 }
 181
 182 static int ipoib_mcast_join_finish(struct ipoib_mcast *mcast,
 183                                    struct ib_sa_mcmember_rec *mcmember)
 184 {
 185         struct net_device *dev = mcast->dev;
 186         struct ipoib_dev_priv *priv = netdev_priv(dev);
 187         struct ipoib_ah *ah;
 188         int ret;
 189
 190         mcast->mcmember = *mcmember;
 191
 192         /* Set the cached Q_Key before we attach if it's the broadcast group */
 193         if (!memcmp(mcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
 194                     sizeof (union ib_gid))) {
 195                 spin_lock_irq(&priv->lock);
 196                 if (!priv->broadcast) {
 197                         spin_unlock_irq(&priv->lock);
 198                         return -EAGAIN;
 199                 }
 200                 priv->qkey = be32_to_cpu(priv->broadcast->mcmember.qkey);
 201                 spin_unlock_irq(&priv->lock);
 202                 priv->tx_wr.wr.ud.remote_qkey = priv->qkey;
 203         }
 204
 205         if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 206                 if (test_and_set_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 207                         ipoib_warn(priv, "multicast group " IPOIB_GID_FMT
 208                                    " already attached\n",
 209                                    IPOIB_GID_ARG(mcast->mcmember.mgid));
 210
 211                         return 0;
 212                 }
 213
 214                 ret = ipoib_mcast_attach(dev, be16_to_cpu(mcast->mcmember.mlid),
 215                                          &mcast->mcmember.mgid);
 216                 if (ret < 0) {
 217                         ipoib_warn(priv, "couldn't attach QP to multicast group "
 218                                    IPOIB_GID_FMT "\n",
 219                                    IPOIB_GID_ARG(mcast->mcmember.mgid));
 220
 221                         clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags);
 222                         return ret;
 223                 }
 224         }
 225
 226         {
 227                 struct ib_ah_attr av = {
 228                         .dlid          = be16_to_cpu(mcast->mcmember.mlid),
 229                         .port_num      = priv->port,
 230                         .sl            = mcast->mcmember.sl,
 231                         .ah_flags      = IB_AH_GRH,
 232                         .static_rate   = mcast->mcmember.rate,
 233                         .grh           = {
 234                                 .flow_label    = be32_to_cpu(mcast->mcmember.flow_label),
 235                                 .hop_limit     = mcast->mcmember.hop_limit,
 236                                 .sgid_index    = 0,
 237                                 .traffic_class = mcast->mcmember.traffic_class
 238                         }
 239                 };
 240                 av.grh.dgid = mcast->mcmember.mgid;
 241
 242                 ah = ipoib_create_ah(dev, priv->pd, &av);
 243                 if (!ah) {
 244                         ipoib_warn(priv, "ib_address_create failed\n");
 245                 } else {
 246                         spin_lock_irq(&priv->lock);
 247                         mcast->ah = ah;
 248                         spin_unlock_irq(&priv->lock);
 249
 250                         ipoib_dbg_mcast(priv, "MGID " IPOIB_GID_FMT
 251                                         " AV %p, LID 0x%04x, SL %d\n",
 252                                         IPOIB_GID_ARG(mcast->mcmember.mgid),
 253                                         mcast->ah->ah,
 254                                         be16_to_cpu(mcast->mcmember.mlid),
 255                                         mcast->mcmember.sl);
 256                 }
 257         }
 258
 259         /* actually send any queued packets */
 260         spin_lock_irq(&priv->tx_lock);
 261         while (!skb_queue_empty(&mcast->pkt_queue)) {
 262                 struct sk_buff *skb = skb_dequeue(&mcast->pkt_queue);
 263                 spin_unlock_irq(&priv->tx_lock);
 264
 265                 skb->dev = dev;
 266
 267                 if (!skb->dst || !skb->dst->neighbour) {
 268                         /* put pseudoheader back on for next time */
 269                         skb_push(skb, sizeof (struct ipoib_pseudoheader));
 270                 }
 271
 272                 if (dev_queue_xmit(skb))
 273                         ipoib_warn(priv, "dev_queue_xmit failed to requeue packet\n");
 274                 spin_lock_irq(&priv->tx_lock);
 275         }
 276         spin_unlock_irq(&priv->tx_lock);
 277
 278         return 0;
 279 }
 280
 281 static int
 282 ipoib_mcast_sendonly_join_complete(int status,
 283                                    struct ib_sa_multicast *multicast)
 284 {
 285         struct ipoib_mcast *mcast = multicast->context;
 286         struct net_device *dev = mcast->dev;
 287         struct ipoib_dev_priv *priv = netdev_priv(dev);
 288
 289         /* We trap for port events ourselves. */
 290         if (status == -ENETRESET)
 291                 return 0;
 292
 293         if (!status)
 294                 status = ipoib_mcast_join_finish(mcast, &multicast->rec);
 295
 296         if (status) {
 297                 if (mcast->logcount++ < 20)
 298                         ipoib_dbg_mcast(netdev_priv(dev), "multicast join failed for "
 299                                         IPOIB_GID_FMT ", status %d\n",
 300                                         IPOIB_GID_ARG(mcast->mcmember.mgid), status);
 301
 302                 /* Flush out any queued packets */
 303                 spin_lock_irq(&priv->tx_lock);
 304                 while (!skb_queue_empty(&mcast->pkt_queue)) {
 305                         ++dev->stats.tx_dropped;
 306                         dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
 307                 }
 308                 spin_unlock_irq(&priv->tx_lock);
 309
 310                 /* Clear the busy flag so we try again */
 311                 status = test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY,
 312                                             &mcast->flags);
 313         }
 314         return status;
 315 }
 316
 317 static int ipoib_mcast_sendonly_join(struct ipoib_mcast *mcast)
 318 {
 319         struct net_device *dev = mcast->dev;
 320         struct ipoib_dev_priv *priv = netdev_priv(dev);
 321         struct ib_sa_mcmember_rec rec = {
 322 #if 0                           /* Some SMs don't support send-only yet */
 323                 .join_state = 4
 324 #else
 325                 .join_state = 1
 326 #endif
 327         };
 328         int ret = 0;
 329
 330         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
 331                 ipoib_dbg_mcast(priv, "device shutting down, no multicast joins\n");
 332                 return -ENODEV;
 333         }
 334
 335         if (test_and_set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)) {
 336                 ipoib_dbg_mcast(priv, "multicast entry busy, skipping\n");
 337                 return -EBUSY;
 338         }
 339
 340         rec.mgid     = mcast->mcmember.mgid;
 341         rec.port_gid = priv->local_gid;
 342         rec.pkey     = cpu_to_be16(priv->pkey);
 343
 344         mcast->mc = ib_sa_join_multicast(&ipoib_sa_client, priv->ca,
 345                                          priv->port, &rec,
 346                                          IB_SA_MCMEMBER_REC_MGID        |
 347                                          IB_SA_MCMEMBER_REC_PORT_GID    |
 348                                          IB_SA_MCMEMBER_REC_PKEY        |
 349                                          IB_SA_MCMEMBER_REC_JOIN_STATE,
 350                                          GFP_ATOMIC,
 351                                          ipoib_mcast_sendonly_join_complete,
 352                                          mcast);
 353         if (IS_ERR(mcast->mc)) {
 354                 ret = PTR_ERR(mcast->mc);
 355                 clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 356                 ipoib_warn(priv, "ib_sa_join_multicast failed (ret = %d)\n",
 357                            ret);
 358         } else {
 359                 ipoib_dbg_mcast(priv, "no multicast record for " IPOIB_GID_FMT
 360                                 ", starting join\n",
 361                                 IPOIB_GID_ARG(mcast->mcmember.mgid));
 362         }
 363
 364         return ret;
 365 }
 366
 367 static int ipoib_mcast_join_complete(int status,
 368                                      struct ib_sa_multicast *multicast)
 369 {
 370         struct ipoib_mcast *mcast = multicast->context;
 371         struct net_device *dev = mcast->dev;
 372         struct ipoib_dev_priv *priv = netdev_priv(dev);
 373
 374         ipoib_dbg_mcast(priv, "join completion for " IPOIB_GID_FMT
 375                         " (status %d)\n",
 376                         IPOIB_GID_ARG(mcast->mcmember.mgid), status);
 377
 378         /* We trap for port events ourselves. */
 379         if (status == -ENETRESET)
 380                 return 0;
 381
 382         if (!status)
 383                 status = ipoib_mcast_join_finish(mcast, &multicast->rec);
 384
 385         if (!status) {
 386                 mcast->backoff = 1;
 387                 mutex_lock(&mcast_mutex);
 388                 if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 389                         queue_delayed_work(ipoib_workqueue,
 390                                            &priv->mcast_task, 0);
 391                 mutex_unlock(&mcast_mutex);
 392
 393                 if (mcast == priv->broadcast)
 394                         netif_carrier_on(dev);
 395
 396                 return 0;
 397         }
 398
 399         if (mcast->logcount++ < 20) {
 400                 if (status == -ETIMEDOUT) {
 401                         ipoib_dbg_mcast(priv, "multicast join failed for " IPOIB_GID_FMT
 402                                         ", status %d\n",
 403                                         IPOIB_GID_ARG(mcast->mcmember.mgid),
 404                                         status);
 405                 } else {
 406                         ipoib_warn(priv, "multicast join failed for "
 407                                    IPOIB_GID_FMT ", status %d\n",
 408                                    IPOIB_GID_ARG(mcast->mcmember.mgid),
 409                                    status);
 410                 }
 411         }
 412
 413         mcast->backoff *= 2;
 414         if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
 415                 mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
 416
 417         /* Clear the busy flag so we try again */
 418         status = test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 419
 420         mutex_lock(&mcast_mutex);
 421         spin_lock_irq(&priv->lock);
 422         if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 423                 queue_delayed_work(ipoib_workqueue, &priv->mcast_task,
 424                                    mcast->backoff * HZ);
 425         spin_unlock_irq(&priv->lock);
 426         mutex_unlock(&mcast_mutex);
 427
 428         return status;
 429 }
 430
 431 static void ipoib_mcast_join(struct net_device *dev, struct ipoib_mcast *mcast,
 432                              int create)
 433 {
 434         struct ipoib_dev_priv *priv = netdev_priv(dev);
 435         struct ib_sa_mcmember_rec rec = {
 436                 .join_state = 1
 437         };
 438         ib_sa_comp_mask comp_mask;
 439         int ret = 0;
 440
 441         ipoib_dbg_mcast(priv, "joining MGID " IPOIB_GID_FMT "\n",
 442                         IPOIB_GID_ARG(mcast->mcmember.mgid));
 443
 444         rec.mgid     = mcast->mcmember.mgid;
 445         rec.port_gid = priv->local_gid;
 446         rec.pkey     = cpu_to_be16(priv->pkey);
 447
 448         comp_mask =
 449                 IB_SA_MCMEMBER_REC_MGID         |
 450                 IB_SA_MCMEMBER_REC_PORT_GID     |
 451                 IB_SA_MCMEMBER_REC_PKEY         |
 452                 IB_SA_MCMEMBER_REC_JOIN_STATE;
 453
 454         if (create) {
 455                 comp_mask |=
 456                         IB_SA_MCMEMBER_REC_QKEY                 |
 457                         IB_SA_MCMEMBER_REC_MTU_SELECTOR         |
 458                         IB_SA_MCMEMBER_REC_MTU                  |
 459                         IB_SA_MCMEMBER_REC_TRAFFIC_CLASS        |
 460                         IB_SA_MCMEMBER_REC_RATE_SELECTOR        |
 461                         IB_SA_MCMEMBER_REC_RATE                 |
 462                         IB_SA_MCMEMBER_REC_SL                   |
 463                         IB_SA_MCMEMBER_REC_FLOW_LABEL           |
 464                         IB_SA_MCMEMBER_REC_HOP_LIMIT;
 465
 466                 rec.qkey          = priv->broadcast->mcmember.qkey;
 467                 rec.mtu_selector  = IB_SA_EQ;
 468                 rec.mtu           = priv->broadcast->mcmember.mtu;
 469                 rec.traffic_class = priv->broadcast->mcmember.traffic_class;
 470                 rec.rate_selector = IB_SA_EQ;
 471                 rec.rate          = priv->broadcast->mcmember.rate;
 472                 rec.sl            = priv->broadcast->mcmember.sl;
 473                 rec.flow_label    = priv->broadcast->mcmember.flow_label;
 474                 rec.hop_limit     = priv->broadcast->mcmember.hop_limit;
 475         }
 476
 477         set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 478         mcast->mc = ib_sa_join_multicast(&ipoib_sa_client, priv->ca, priv->port,
 479                                          &rec, comp_mask, GFP_KERNEL,
 480                                          ipoib_mcast_join_complete, mcast);
 481         if (IS_ERR(mcast->mc)) {
 482                 clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 483                 ret = PTR_ERR(mcast->mc);
 484                 ipoib_warn(priv, "ib_sa_join_multicast failed, status %d\n", ret);
 485
 486                 mcast->backoff *= 2;
 487                 if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
 488                         mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
 489
 490                 mutex_lock(&mcast_mutex);
 491                 if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 492                         queue_delayed_work(ipoib_workqueue,
 493                                            &priv->mcast_task,
 494                                            mcast->backoff * HZ);
 495                 mutex_unlock(&mcast_mutex);
 496         }
 497 }
 498
 499 void ipoib_mcast_join_task(struct work_struct *work)
 500 {
 501         struct ipoib_dev_priv *priv =
 502                 container_of(work, struct ipoib_dev_priv, mcast_task.work);
 503         struct net_device *dev = priv->dev;
 504
 505         if (!test_bit(IPOIB_MCAST_RUN, &priv->flags))
 506                 return;
 507
 508         if (ib_query_gid(priv->ca, priv->port, 0, &priv->local_gid))
 509                 ipoib_warn(priv, "ib_query_gid() failed\n");
 510         else
 511                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
 512
 513         {
 514                 struct ib_port_attr attr;
 515
 516                 if (!ib_query_port(priv->ca, priv->port, &attr))
 517                         priv->local_lid = attr.lid;
 518                 else
 519                         ipoib_warn(priv, "ib_query_port failed\n");
 520         }
 521
 522         if (!priv->broadcast) {
 523                 struct ipoib_mcast *broadcast;
 524
 525                 broadcast = ipoib_mcast_alloc(dev, 1);
 526                 if (!broadcast) {
 527                         ipoib_warn(priv, "failed to allocate broadcast group\n");
 528                         mutex_lock(&mcast_mutex);
 529                         if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 530                                 queue_delayed_work(ipoib_workqueue,
 531                                                    &priv->mcast_task, HZ);
 532                         mutex_unlock(&mcast_mutex);
 533                         return;
 534                 }
 535
 536                 spin_lock_irq(&priv->lock);
 537                 memcpy(broadcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
 538                        sizeof (union ib_gid));
 539                 priv->broadcast = broadcast;
 540
 541                 __ipoib_mcast_add(dev, priv->broadcast);
 542                 spin_unlock_irq(&priv->lock);
 543         }
 544
 545         if (!test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
 546                 if (!test_bit(IPOIB_MCAST_FLAG_BUSY, &priv->broadcast->flags))
 547                         ipoib_mcast_join(dev, priv->broadcast, 0);
 548                 return;
 549         }
 550
 551         while (1) {
 552                 struct ipoib_mcast *mcast = NULL;
 553
 554                 spin_lock_irq(&priv->lock);
 555                 list_for_each_entry(mcast, &priv->multicast_list, list) {
 556                         if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)
 557                             && !test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)
 558                             && !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 559                                 /* Found the next unjoined group */
 560                                 break;
 561                         }
 562                 }
 563                 spin_unlock_irq(&priv->lock);
 564
 565                 if (&mcast->list == &priv->multicast_list) {
 566                         /* All done */
 567                         break;
 568                 }
 569
 570                 ipoib_mcast_join(dev, mcast, 1);
 571                 return;
 572         }
 573
 574         priv->mcast_mtu = IPOIB_UD_MTU(ib_mtu_enum_to_int(priv->broadcast->mcmember.mtu));
 575
 576         if (!ipoib_cm_admin_enabled(dev))
 577                 dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
 578
 579         ipoib_dbg_mcast(priv, "successfully joined all multicast groups\n");
 580
 581         clear_bit(IPOIB_MCAST_RUN, &priv->flags);
 582 }
 583
 584 int ipoib_mcast_start_thread(struct net_device *dev)
 585 {
 586         struct ipoib_dev_priv *priv = netdev_priv(dev);
 587
 588         ipoib_dbg_mcast(priv, "starting multicast thread\n");
 589
 590         mutex_lock(&mcast_mutex);
 591         if (!test_and_set_bit(IPOIB_MCAST_RUN, &priv->flags))
 592                 queue_delayed_work(ipoib_workqueue, &priv->mcast_task, 0);
 593         mutex_unlock(&mcast_mutex);
 594
 595         spin_lock_irq(&priv->lock);
 596         set_bit(IPOIB_MCAST_STARTED, &priv->flags);
 597         spin_unlock_irq(&priv->lock);
 598
 599         return 0;
 600 }
 601
 602 int ipoib_mcast_stop_thread(struct net_device *dev, int flush)
 603 {
 604         struct ipoib_dev_priv *priv = netdev_priv(dev);
 605
 606         ipoib_dbg_mcast(priv, "stopping multicast thread\n");
 607
 608         spin_lock_irq(&priv->lock);
 609         clear_bit(IPOIB_MCAST_STARTED, &priv->flags);
 610         spin_unlock_irq(&priv->lock);
 611
 612         mutex_lock(&mcast_mutex);
 613         clear_bit(IPOIB_MCAST_RUN, &priv->flags);
 614         cancel_delayed_work(&priv->mcast_task);
 615         mutex_unlock(&mcast_mutex);
 616
 617         if (flush)
 618                 flush_workqueue(ipoib_workqueue);
 619
 620         return 0;
 621 }
 622
 623 static int ipoib_mcast_leave(struct net_device *dev, struct ipoib_mcast *mcast)
 624 {
 625         struct ipoib_dev_priv *priv = netdev_priv(dev);
 626         int ret = 0;
 627
 628         if (test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
 629                 ib_sa_free_multicast(mcast->mc);
 630
 631         if (test_and_clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 632                 ipoib_dbg_mcast(priv, "leaving MGID " IPOIB_GID_FMT "\n",
 633                                 IPOIB_GID_ARG(mcast->mcmember.mgid));
 634
 635                 /* Remove ourselves from the multicast group */
 636                 ret = ipoib_mcast_detach(dev, be16_to_cpu(mcast->mcmember.mlid),
 637                                          &mcast->mcmember.mgid);
 638                 if (ret)
 639                         ipoib_warn(priv, "ipoib_mcast_detach failed (result = %d)\n", ret);
 640         }
 641
 642         return 0;
 643 }
 644
 645 void ipoib_mcast_send(struct net_device *dev, void *mgid, struct sk_buff *skb)
 646 {
 647         struct ipoib_dev_priv *priv = netdev_priv(dev);
 648         struct ipoib_mcast *mcast;
 649
 650         /*
 651          * We can only be called from ipoib_start_xmit, so we're
 652          * inside tx_lock -- no need to save/restore flags.
 653          */
 654         spin_lock(&priv->lock);
 655
 656         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)         ||
 657             !priv->broadcast                                    ||
 658             !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
 659                 ++dev->stats.tx_dropped;
 660                 dev_kfree_skb_any(skb);
 661                 goto unlock;
 662         }
 663
 664         mcast = __ipoib_mcast_find(dev, mgid);
 665         if (!mcast) {
 666                 /* Let's create a new send only group now */
 667                 ipoib_dbg_mcast(priv, "setting up send only multicast group for "
 668                                 IPOIB_GID_FMT "\n", IPOIB_GID_RAW_ARG(mgid));
 669
 670                 mcast = ipoib_mcast_alloc(dev, 0);
 671                 if (!mcast) {
 672                         ipoib_warn(priv, "unable to allocate memory for "
 673                                    "multicast structure\n");
 674                         ++dev->stats.tx_dropped;
 675                         dev_kfree_skb_any(skb);
 676                         goto out;
 677                 }
 678
 679                 set_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags);
 680                 memcpy(mcast->mcmember.mgid.raw, mgid, sizeof (union ib_gid));
 681                 __ipoib_mcast_add(dev, mcast);
 682                 list_add_tail(&mcast->list, &priv->multicast_list);
 683         }
 684
 685         if (!mcast->ah) {
 686                 if (skb_queue_len(&mcast->pkt_queue) < IPOIB_MAX_MCAST_QUEUE)
 687                         skb_queue_tail(&mcast->pkt_queue, skb);
 688                 else {
 689                         ++dev->stats.tx_dropped;
 690                         dev_kfree_skb_any(skb);
 691                 }
 692
 693                 if (test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
 694                         ipoib_dbg_mcast(priv, "no address vector, "
 695                                         "but multicast join already started\n");
 696                 else if (test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags))
 697                         ipoib_mcast_sendonly_join(mcast);
 698
 699                 /*
 700                  * If lookup completes between here and out:, don't
 701                  * want to send packet twice.
 702                  */
 703                 mcast = NULL;
 704         }
 705
 706 out:
 707         if (mcast && mcast->ah) {
 708                 if (skb->dst            &&
 709                     skb->dst->neighbour &&
 710                     !*to_ipoib_neigh(skb->dst->neighbour)) {
 711                         struct ipoib_neigh *neigh = ipoib_neigh_alloc(skb->dst->neighbour,
 712                                                                         skb->dev);
 713
 714                         if (neigh) {
 715                                 kref_get(&mcast->ah->ref);
 716                                 neigh->ah       = mcast->ah;
 717                                 list_add_tail(&neigh->list, &mcast->neigh_list);
 718                         }
 719                 }
 720
 721                 ipoib_send(dev, skb, mcast->ah, IB_MULTICAST_QPN);
 722         }
 723
 724 unlock:
 725         spin_unlock(&priv->lock);
 726 }
 727
 728 void ipoib_mcast_dev_flush(struct net_device *dev)
 729 {
 730         struct ipoib_dev_priv *priv = netdev_priv(dev);
 731         LIST_HEAD(remove_list);
 732         struct ipoib_mcast *mcast, *tmcast;
 733         unsigned long flags;
 734
 735         ipoib_dbg_mcast(priv, "flushing multicast list\n");
 736
 737         spin_lock_irqsave(&priv->lock, flags);
 738
 739         list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
 740                 list_del(&mcast->list);
 741                 rb_erase(&mcast->rb_node, &priv->multicast_tree);
 742                 list_add_tail(&mcast->list, &remove_list);
 743         }
 744
 745         if (priv->broadcast) {
 746                 rb_erase(&priv->broadcast->rb_node, &priv->multicast_tree);
 747                 list_add_tail(&priv->broadcast->list, &remove_list);
 748                 priv->broadcast = NULL;
 749         }
 750
 751         spin_unlock_irqrestore(&priv->lock, flags);
 752
 753         list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
 754                 ipoib_mcast_leave(dev, mcast);
 755                 ipoib_mcast_free(mcast);
 756         }
 757 }
 758
 759 void ipoib_mcast_restart_task(struct work_struct *work)
 760 {
 761         struct ipoib_dev_priv *priv =
 762                 container_of(work, struct ipoib_dev_priv, restart_task);
 763         struct net_device *dev = priv->dev;
 764         struct dev_mc_list *mclist;
 765         struct ipoib_mcast *mcast, *tmcast;
 766         LIST_HEAD(remove_list);
 767         unsigned long flags;
 768         struct ib_sa_mcmember_rec rec;
 769
 770         ipoib_dbg_mcast(priv, "restarting multicast task\n");
 771
 772         ipoib_mcast_stop_thread(dev, 0);
 773
 774         local_irq_save(flags);
 775         netif_tx_lock(dev);
 776         spin_lock(&priv->lock);
 777
 778         /*
 779          * Unfortunately, the networking core only gives us a list of all of
 780          * the multicast hardware addresses. We need to figure out which ones
 781          * are new and which ones have been removed
 782          */
 783
 784         /* Clear out the found flag */
 785         list_for_each_entry(mcast, &priv->multicast_list, list)
 786                 clear_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
 787
 788         /* Mark all of the entries that are found or don't exist */
 789         for (mclist = dev->mc_list; mclist; mclist = mclist->next) {
 790                 union ib_gid mgid;
 791
 792                 memcpy(mgid.raw, mclist->dmi_addr + 4, sizeof mgid);
 793
 794                 mcast = __ipoib_mcast_find(dev, &mgid);
 795                 if (!mcast || test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 796                         struct ipoib_mcast *nmcast;
 797
 798                         /* ignore group which is directly joined by userspace */
 799                         if (test_bit(IPOIB_FLAG_UMCAST, &priv->flags) &&
 800                             !ib_sa_get_mcmember_rec(priv->ca, priv->port, &mgid, &rec)) {
 801                                 ipoib_dbg_mcast(priv, "ignoring multicast entry for mgid "
 802                                                 IPOIB_GID_FMT "\n", IPOIB_GID_ARG(mgid));
 803                                 continue;
 804                         }
 805
 806                         /* Not found or send-only group, let's add a new entry */
 807                         ipoib_dbg_mcast(priv, "adding multicast entry for mgid "
 808                                         IPOIB_GID_FMT "\n", IPOIB_GID_ARG(mgid));
 809
 810                         nmcast = ipoib_mcast_alloc(dev, 0);
 811                         if (!nmcast) {
 812                                 ipoib_warn(priv, "unable to allocate memory for multicast structure\n");
 813                                 continue;
 814                         }
 815
 816                         set_bit(IPOIB_MCAST_FLAG_FOUND, &nmcast->flags);
 817
 818                         nmcast->mcmember.mgid = mgid;
 819
 820                         if (mcast) {
 821                                 /* Destroy the send only entry */
 822                                 list_move_tail(&mcast->list, &remove_list);
 823
 824                                 rb_replace_node(&mcast->rb_node,
 825                                                 &nmcast->rb_node,
 826                                                 &priv->multicast_tree);
 827                         } else
 828                                 __ipoib_mcast_add(dev, nmcast);
 829
 830                         list_add_tail(&nmcast->list, &priv->multicast_list);
 831                 }
 832
 833                 if (mcast)
 834                         set_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
 835         }
 836
 837         /* Remove all of the entries don't exist anymore */
 838         list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
 839                 if (!test_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags) &&
 840                     !test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 841                         ipoib_dbg_mcast(priv, "deleting multicast group " IPOIB_GID_FMT "\n",
 842                                         IPOIB_GID_ARG(mcast->mcmember.mgid));
 843
 844                         rb_erase(&mcast->rb_node, &priv->multicast_tree);
 845
 846                         /* Move to the remove list */
 847                         list_move_tail(&mcast->list, &remove_list);
 848                 }
 849         }
 850
 851         spin_unlock(&priv->lock);
 852         netif_tx_unlock(dev);
 853         local_irq_restore(flags);
 854
 855         /* We have to cancel outside of the spinlock */
 856         list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
 857                 ipoib_mcast_leave(mcast->dev, mcast);
 858                 ipoib_mcast_free(mcast);
 859         }
 860
 861         if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 862                 ipoib_mcast_start_thread(dev);
 863 }
 864
 865 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 866
 867 struct ipoib_mcast_iter *ipoib_mcast_iter_init(struct net_device *dev)
 868 {
 869         struct ipoib_mcast_iter *iter;
 870
 871         iter = kmalloc(sizeof *iter, GFP_KERNEL);
 872         if (!iter)
 873                 return NULL;
 874
 875         iter->dev = dev;
 876         memset(iter->mgid.raw, 0, 16);
 877
 878         if (ipoib_mcast_iter_next(iter)) {
 879                 kfree(iter);
 880                 return NULL;
 881         }
 882
 883         return iter;
 884 }
 885
 886 int ipoib_mcast_iter_next(struct ipoib_mcast_iter *iter)
 887 {
 888         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
 889         struct rb_node *n;
 890         struct ipoib_mcast *mcast;
 891         int ret = 1;
 892
 893         spin_lock_irq(&priv->lock);
 894
 895         n = rb_first(&priv->multicast_tree);
 896
 897         while (n) {
 898                 mcast = rb_entry(n, struct ipoib_mcast, rb_node);
 899
 900                 if (memcmp(iter->mgid.raw, mcast->mcmember.mgid.raw,
 901                            sizeof (union ib_gid)) < 0) {
 902                         iter->mgid      = mcast->mcmember.mgid;
 903                         iter->created   = mcast->created;
 904                         iter->queuelen  = skb_queue_len(&mcast->pkt_queue);
 905                         iter->complete  = !!mcast->ah;
 906                         iter->send_only = !!(mcast->flags & (1 << IPOIB_MCAST_FLAG_SENDONLY));
 907
 908                         ret = 0;
 909
 910                         break;
 911                 }
 912
 913                 n = rb_next(n);
 914         }
 915
 916         spin_unlock_irq(&priv->lock);
 917
 918         return ret;
 919 }
 920
 921 void ipoib_mcast_iter_read(struct ipoib_mcast_iter *iter,
 922                            union ib_gid *mgid,
 923                            unsigned long *created,
 924                            unsigned int *queuelen,
 925                            unsigned int *complete,
 926                            unsigned int *send_only)
 927 {
 928         *mgid      = iter->mgid;
 929         *created   = iter->created;
 930         *queuelen  = iter->queuelen;
 931         *complete  = iter->complete;
 932         *send_only = iter->send_only;
 933 }
 934
 935 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */