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 #include <linux/slab.h>
  44
  45 #include <net/dst.h>
  46
  47 #include "ipoib.h"
  48
  49 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
  50 static int mcast_debug_level;
  51
  52 module_param(mcast_debug_level, int, 0644);
  53 MODULE_PARM_DESC(mcast_debug_level,
  54                  "Enable multicast debug tracing if > 0");
  55 #endif
  56
  57 static DEFINE_MUTEX(mcast_mutex);
  58
  59 struct ipoib_mcast_iter {
  60         struct net_device *dev;
  61         union ib_gid       mgid;
  62         unsigned long      created;
  63         unsigned int       queuelen;
  64         unsigned int       complete;
  65         unsigned int       send_only;
  66 };
  67
  68 static void ipoib_mcast_free(struct ipoib_mcast *mcast)
  69 {
  70         struct net_device *dev = mcast->dev;
  71         struct ipoib_dev_priv *priv = netdev_priv(dev);
  72         struct ipoib_neigh *neigh, *tmp;
  73         int tx_dropped = 0;
  74
  75         ipoib_dbg_mcast(netdev_priv(dev), "deleting multicast group %pI6\n",
  76                         mcast->mcmember.mgid.raw);
  77
  78         spin_lock_irq(&priv->lock);
  79
  80         list_for_each_entry_safe(neigh, tmp, &mcast->neigh_list, list) {
  81                 /*
  82                  * It's safe to call ipoib_put_ah() inside priv->lock
  83                  * here, because we know that mcast->ah will always
  84                  * hold one more reference, so ipoib_put_ah() will
  85                  * never do more than decrement the ref count.
  86                  */
  87                 if (neigh->ah)
  88                         ipoib_put_ah(neigh->ah);
  89                 ipoib_neigh_free(dev, neigh);
  90         }
  91
  92         spin_unlock_irq(&priv->lock);
  93
  94         if (mcast->ah)
  95                 ipoib_put_ah(mcast->ah);
  96
  97         while (!skb_queue_empty(&mcast->pkt_queue)) {
  98                 ++tx_dropped;
  99                 dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
 100         }
 101
 102         netif_tx_lock_bh(dev);
 103         dev->stats.tx_dropped += tx_dropped;
 104         netif_tx_unlock_bh(dev);
 105
 106         kfree(mcast);
 107 }
 108
 109 static struct ipoib_mcast *ipoib_mcast_alloc(struct net_device *dev,
 110                                              int can_sleep)
 111 {
 112         struct ipoib_mcast *mcast;
 113
 114         mcast = kzalloc(sizeof *mcast, can_sleep ? GFP_KERNEL : GFP_ATOMIC);
 115         if (!mcast)
 116                 return NULL;
 117
 118         mcast->dev = dev;
 119         mcast->created = jiffies;
 120         mcast->backoff = 1;
 121
 122         INIT_LIST_HEAD(&mcast->list);
 123         INIT_LIST_HEAD(&mcast->neigh_list);
 124         skb_queue_head_init(&mcast->pkt_queue);
 125
 126         return mcast;
 127 }
 128
 129 static struct ipoib_mcast *__ipoib_mcast_find(struct net_device *dev, void *mgid)
 130 {
 131         struct ipoib_dev_priv *priv = netdev_priv(dev);
 132         struct rb_node *n = priv->multicast_tree.rb_node;
 133
 134         while (n) {
 135                 struct ipoib_mcast *mcast;
 136                 int ret;
 137
 138                 mcast = rb_entry(n, struct ipoib_mcast, rb_node);
 139
 140                 ret = memcmp(mgid, mcast->mcmember.mgid.raw,
 141                              sizeof (union ib_gid));
 142                 if (ret < 0)
 143                         n = n->rb_left;
 144                 else if (ret > 0)
 145                         n = n->rb_right;
 146                 else
 147                         return mcast;
 148         }
 149
 150         return NULL;
 151 }
 152
 153 static int __ipoib_mcast_add(struct net_device *dev, struct ipoib_mcast *mcast)
 154 {
 155         struct ipoib_dev_priv *priv = netdev_priv(dev);
 156         struct rb_node **n = &priv->multicast_tree.rb_node, *pn = NULL;
 157
 158         while (*n) {
 159                 struct ipoib_mcast *tmcast;
 160                 int ret;
 161
 162                 pn = *n;
 163                 tmcast = rb_entry(pn, struct ipoib_mcast, rb_node);
 164
 165                 ret = memcmp(mcast->mcmember.mgid.raw, tmcast->mcmember.mgid.raw,
 166                              sizeof (union ib_gid));
 167                 if (ret < 0)
 168                         n = &pn->rb_left;
 169                 else if (ret > 0)
 170                         n = &pn->rb_right;
 171                 else
 172                         return -EEXIST;
 173         }
 174
 175         rb_link_node(&mcast->rb_node, pn, n);
 176         rb_insert_color(&mcast->rb_node, &priv->multicast_tree);
 177
 178         return 0;
 179 }
 180
 181 static int ipoib_mcast_join_finish(struct ipoib_mcast *mcast,
 182                                    struct ib_sa_mcmember_rec *mcmember)
 183 {
 184         struct net_device *dev = mcast->dev;
 185         struct ipoib_dev_priv *priv = netdev_priv(dev);
 186         struct ipoib_ah *ah;
 187         int ret;
 188         int set_qkey = 0;
 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                 set_qkey = 1;
 204         }
 205
 206         if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 207                 if (test_and_set_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 208                         ipoib_warn(priv, "multicast group %pI6 already attached\n",
 209                                    mcast->mcmember.mgid.raw);
 210
 211                         return 0;
 212                 }
 213
 214                 ret = ipoib_mcast_attach(dev, be16_to_cpu(mcast->mcmember.mlid),
 215                                          &mcast->mcmember.mgid, set_qkey);
 216                 if (ret < 0) {
 217                         ipoib_warn(priv, "couldn't attach QP to multicast group %pI6\n",
 218                                    mcast->mcmember.mgid.raw);
 219
 220                         clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags);
 221                         return ret;
 222                 }
 223         }
 224
 225         {
 226                 struct ib_ah_attr av = {
 227                         .dlid          = be16_to_cpu(mcast->mcmember.mlid),
 228                         .port_num      = priv->port,
 229                         .sl            = mcast->mcmember.sl,
 230                         .ah_flags      = IB_AH_GRH,
 231                         .static_rate   = mcast->mcmember.rate,
 232                         .grh           = {
 233                                 .flow_label    = be32_to_cpu(mcast->mcmember.flow_label),
 234                                 .hop_limit     = mcast->mcmember.hop_limit,
 235                                 .sgid_index    = 0,
 236                                 .traffic_class = mcast->mcmember.traffic_class
 237                         }
 238                 };
 239                 av.grh.dgid = mcast->mcmember.mgid;
 240
 241                 ah = ipoib_create_ah(dev, priv->pd, &av);
 242                 if (!ah) {
 243                         ipoib_warn(priv, "ib_address_create failed\n");
 244                 } else {
 245                         spin_lock_irq(&priv->lock);
 246                         mcast->ah = ah;
 247                         spin_unlock_irq(&priv->lock);
 248
 249                         ipoib_dbg_mcast(priv, "MGID %pI6 AV %p, LID 0x%04x, SL %d\n",
 250                                         mcast->mcmember.mgid.raw,
 251                                         mcast->ah->ah,
 252                                         be16_to_cpu(mcast->mcmember.mlid),
 253                                         mcast->mcmember.sl);
 254                 }
 255         }
 256
 257         /* actually send any queued packets */
 258         netif_tx_lock_bh(dev);
 259         while (!skb_queue_empty(&mcast->pkt_queue)) {
 260                 struct sk_buff *skb = skb_dequeue(&mcast->pkt_queue);
 261                 netif_tx_unlock_bh(dev);
 262
 263                 skb->dev = dev;
 264
 265                 if (!skb_dst(skb) || !skb_dst(skb)->neighbour) {
 266                         /* put pseudoheader back on for next time */
 267                         skb_push(skb, sizeof (struct ipoib_pseudoheader));
 268                 }
 269
 270                 if (dev_queue_xmit(skb))
 271                         ipoib_warn(priv, "dev_queue_xmit failed to requeue packet\n");
 272                 netif_tx_lock_bh(dev);
 273         }
 274         netif_tx_unlock_bh(dev);
 275
 276         return 0;
 277 }
 278
 279 static int
 280 ipoib_mcast_sendonly_join_complete(int status,
 281                                    struct ib_sa_multicast *multicast)
 282 {
 283         struct ipoib_mcast *mcast = multicast->context;
 284         struct net_device *dev = mcast->dev;
 285
 286         /* We trap for port events ourselves. */
 287         if (status == -ENETRESET)
 288                 return 0;
 289
 290         if (!status)
 291                 status = ipoib_mcast_join_finish(mcast, &multicast->rec);
 292
 293         if (status) {
 294                 if (mcast->logcount++ < 20)
 295                         ipoib_dbg_mcast(netdev_priv(dev), "multicast join failed for %pI6, status %d\n",
 296                                         mcast->mcmember.mgid.raw, status);
 297
 298                 /* Flush out any queued packets */
 299                 netif_tx_lock_bh(dev);
 300                 while (!skb_queue_empty(&mcast->pkt_queue)) {
 301                         ++dev->stats.tx_dropped;
 302                         dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
 303                 }
 304                 netif_tx_unlock_bh(dev);
 305
 306                 /* Clear the busy flag so we try again */
 307                 status = test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY,
 308                                             &mcast->flags);
 309         }
 310         return status;
 311 }
 312
 313 static int ipoib_mcast_sendonly_join(struct ipoib_mcast *mcast)
 314 {
 315         struct net_device *dev = mcast->dev;
 316         struct ipoib_dev_priv *priv = netdev_priv(dev);
 317         struct ib_sa_mcmember_rec rec = {
 318 #if 0                           /* Some SMs don't support send-only yet */
 319                 .join_state = 4
 320 #else
 321                 .join_state = 1
 322 #endif
 323         };
 324         int ret = 0;
 325
 326         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
 327                 ipoib_dbg_mcast(priv, "device shutting down, no multicast joins\n");
 328                 return -ENODEV;
 329         }
 330
 331         if (test_and_set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)) {
 332                 ipoib_dbg_mcast(priv, "multicast entry busy, skipping\n");
 333                 return -EBUSY;
 334         }
 335
 336         rec.mgid     = mcast->mcmember.mgid;
 337         rec.port_gid = priv->local_gid;
 338         rec.pkey     = cpu_to_be16(priv->pkey);
 339
 340         mcast->mc = ib_sa_join_multicast(&ipoib_sa_client, priv->ca,
 341                                          priv->port, &rec,
 342                                          IB_SA_MCMEMBER_REC_MGID        |
 343                                          IB_SA_MCMEMBER_REC_PORT_GID    |
 344                                          IB_SA_MCMEMBER_REC_PKEY        |
 345                                          IB_SA_MCMEMBER_REC_JOIN_STATE,
 346                                          GFP_ATOMIC,
 347                                          ipoib_mcast_sendonly_join_complete,
 348                                          mcast);
 349         if (IS_ERR(mcast->mc)) {
 350                 ret = PTR_ERR(mcast->mc);
 351                 clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 352                 ipoib_warn(priv, "ib_sa_join_multicast failed (ret = %d)\n",
 353                            ret);
 354         } else {
 355                 ipoib_dbg_mcast(priv, "no multicast record for %pI6, starting join\n",
 356                                 mcast->mcmember.mgid.raw);
 357         }
 358
 359         return ret;
 360 }
 361
 362 void ipoib_mcast_carrier_on_task(struct work_struct *work)
 363 {
 364         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
 365                                                    carrier_on_task);
 366         struct ib_port_attr attr;
 367
 368         /*
 369          * Take rtnl_lock to avoid racing with ipoib_stop() and
 370          * turning the carrier back on while a device is being
 371          * removed.
 372          */
 373         if (ib_query_port(priv->ca, priv->port, &attr) ||
 374             attr.state != IB_PORT_ACTIVE) {
 375                 ipoib_dbg(priv, "Keeping carrier off until IB port is active\n");
 376                 return;
 377         }
 378
 379         rtnl_lock();
 380         netif_carrier_on(priv->dev);
 381         rtnl_unlock();
 382 }
 383
 384 static int ipoib_mcast_join_complete(int status,
 385                                      struct ib_sa_multicast *multicast)
 386 {
 387         struct ipoib_mcast *mcast = multicast->context;
 388         struct net_device *dev = mcast->dev;
 389         struct ipoib_dev_priv *priv = netdev_priv(dev);
 390
 391         ipoib_dbg_mcast(priv, "join completion for %pI6 (status %d)\n",
 392                         mcast->mcmember.mgid.raw, status);
 393
 394         /* We trap for port events ourselves. */
 395         if (status == -ENETRESET)
 396                 return 0;
 397
 398         if (!status)
 399                 status = ipoib_mcast_join_finish(mcast, &multicast->rec);
 400
 401         if (!status) {
 402                 mcast->backoff = 1;
 403                 mutex_lock(&mcast_mutex);
 404                 if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 405                         queue_delayed_work(ipoib_workqueue,
 406                                            &priv->mcast_task, 0);
 407                 mutex_unlock(&mcast_mutex);
 408
 409                 /*
 410                  * Defer carrier on work to ipoib_workqueue to avoid a
 411                  * deadlock on rtnl_lock here.
 412                  */
 413                 if (mcast == priv->broadcast)
 414                         queue_work(ipoib_workqueue, &priv->carrier_on_task);
 415
 416                 return 0;
 417         }
 418
 419         if (mcast->logcount++ < 20) {
 420                 if (status == -ETIMEDOUT || status == -EAGAIN) {
 421                         ipoib_dbg_mcast(priv, "multicast join failed for %pI6, status %d\n",
 422                                         mcast->mcmember.mgid.raw, status);
 423                 } else {
 424                         ipoib_warn(priv, "multicast join failed for %pI6, status %d\n",
 425                                    mcast->mcmember.mgid.raw, status);
 426                 }
 427         }
 428
 429         mcast->backoff *= 2;
 430         if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
 431                 mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
 432
 433         /* Clear the busy flag so we try again */
 434         status = test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 435
 436         mutex_lock(&mcast_mutex);
 437         spin_lock_irq(&priv->lock);
 438         if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 439                 queue_delayed_work(ipoib_workqueue, &priv->mcast_task,
 440                                    mcast->backoff * HZ);
 441         spin_unlock_irq(&priv->lock);
 442         mutex_unlock(&mcast_mutex);
 443
 444         return status;
 445 }
 446
 447 static void ipoib_mcast_join(struct net_device *dev, struct ipoib_mcast *mcast,
 448                              int create)
 449 {
 450         struct ipoib_dev_priv *priv = netdev_priv(dev);
 451         struct ib_sa_mcmember_rec rec = {
 452                 .join_state = 1
 453         };
 454         ib_sa_comp_mask comp_mask;
 455         int ret = 0;
 456
 457         ipoib_dbg_mcast(priv, "joining MGID %pI6\n", mcast->mcmember.mgid.raw);
 458
 459         rec.mgid     = mcast->mcmember.mgid;
 460         rec.port_gid = priv->local_gid;
 461         rec.pkey     = cpu_to_be16(priv->pkey);
 462
 463         comp_mask =
 464                 IB_SA_MCMEMBER_REC_MGID         |
 465                 IB_SA_MCMEMBER_REC_PORT_GID     |
 466                 IB_SA_MCMEMBER_REC_PKEY         |
 467                 IB_SA_MCMEMBER_REC_JOIN_STATE;
 468
 469         if (create) {
 470                 comp_mask |=
 471                         IB_SA_MCMEMBER_REC_QKEY                 |
 472                         IB_SA_MCMEMBER_REC_MTU_SELECTOR         |
 473                         IB_SA_MCMEMBER_REC_MTU                  |
 474                         IB_SA_MCMEMBER_REC_TRAFFIC_CLASS        |
 475                         IB_SA_MCMEMBER_REC_RATE_SELECTOR        |
 476                         IB_SA_MCMEMBER_REC_RATE                 |
 477                         IB_SA_MCMEMBER_REC_SL                   |
 478                         IB_SA_MCMEMBER_REC_FLOW_LABEL           |
 479                         IB_SA_MCMEMBER_REC_HOP_LIMIT;
 480
 481                 rec.qkey          = priv->broadcast->mcmember.qkey;
 482                 rec.mtu_selector  = IB_SA_EQ;
 483                 rec.mtu           = priv->broadcast->mcmember.mtu;
 484                 rec.traffic_class = priv->broadcast->mcmember.traffic_class;
 485                 rec.rate_selector = IB_SA_EQ;
 486                 rec.rate          = priv->broadcast->mcmember.rate;
 487                 rec.sl            = priv->broadcast->mcmember.sl;
 488                 rec.flow_label    = priv->broadcast->mcmember.flow_label;
 489                 rec.hop_limit     = priv->broadcast->mcmember.hop_limit;
 490         }
 491
 492         set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 493         mcast->mc = ib_sa_join_multicast(&ipoib_sa_client, priv->ca, priv->port,
 494                                          &rec, comp_mask, GFP_KERNEL,
 495                                          ipoib_mcast_join_complete, mcast);
 496         if (IS_ERR(mcast->mc)) {
 497                 clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 498                 ret = PTR_ERR(mcast->mc);
 499                 ipoib_warn(priv, "ib_sa_join_multicast failed, status %d\n", ret);
 500
 501                 mcast->backoff *= 2;
 502                 if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
 503                         mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
 504
 505                 mutex_lock(&mcast_mutex);
 506                 if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 507                         queue_delayed_work(ipoib_workqueue,
 508                                            &priv->mcast_task,
 509                                            mcast->backoff * HZ);
 510                 mutex_unlock(&mcast_mutex);
 511         }
 512 }
 513
 514 void ipoib_mcast_join_task(struct work_struct *work)
 515 {
 516         struct ipoib_dev_priv *priv =
 517                 container_of(work, struct ipoib_dev_priv, mcast_task.work);
 518         struct net_device *dev = priv->dev;
 519
 520         if (!test_bit(IPOIB_MCAST_RUN, &priv->flags))
 521                 return;
 522
 523         if (ib_query_gid(priv->ca, priv->port, 0, &priv->local_gid))
 524                 ipoib_warn(priv, "ib_query_gid() failed\n");
 525         else
 526                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
 527
 528         {
 529                 struct ib_port_attr attr;
 530
 531                 if (!ib_query_port(priv->ca, priv->port, &attr))
 532                         priv->local_lid = attr.lid;
 533                 else
 534                         ipoib_warn(priv, "ib_query_port failed\n");
 535         }
 536
 537         if (!priv->broadcast) {
 538                 struct ipoib_mcast *broadcast;
 539
 540                 if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 541                         return;
 542
 543                 broadcast = ipoib_mcast_alloc(dev, 1);
 544                 if (!broadcast) {
 545                         ipoib_warn(priv, "failed to allocate broadcast group\n");
 546                         mutex_lock(&mcast_mutex);
 547                         if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 548                                 queue_delayed_work(ipoib_workqueue,
 549                                                    &priv->mcast_task, HZ);
 550                         mutex_unlock(&mcast_mutex);
 551                         return;
 552                 }
 553
 554                 spin_lock_irq(&priv->lock);
 555                 memcpy(broadcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
 556                        sizeof (union ib_gid));
 557                 priv->broadcast = broadcast;
 558
 559                 __ipoib_mcast_add(dev, priv->broadcast);
 560                 spin_unlock_irq(&priv->lock);
 561         }
 562
 563         if (!test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
 564                 if (!test_bit(IPOIB_MCAST_FLAG_BUSY, &priv->broadcast->flags))
 565                         ipoib_mcast_join(dev, priv->broadcast, 0);
 566                 return;
 567         }
 568
 569         while (1) {
 570                 struct ipoib_mcast *mcast = NULL;
 571
 572                 spin_lock_irq(&priv->lock);
 573                 list_for_each_entry(mcast, &priv->multicast_list, list) {
 574                         if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)
 575                             && !test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)
 576                             && !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 577                                 /* Found the next unjoined group */
 578                                 break;
 579                         }
 580                 }
 581                 spin_unlock_irq(&priv->lock);
 582
 583                 if (&mcast->list == &priv->multicast_list) {
 584                         /* All done */
 585                         break;
 586                 }
 587
 588                 ipoib_mcast_join(dev, mcast, 1);
 589                 return;
 590         }
 591
 592         priv->mcast_mtu = IPOIB_UD_MTU(ib_mtu_enum_to_int(priv->broadcast->mcmember.mtu));
 593
 594         if (!ipoib_cm_admin_enabled(dev)) {
 595                 rtnl_lock();
 596                 dev_set_mtu(dev, min(priv->mcast_mtu, priv->admin_mtu));
 597                 rtnl_unlock();
 598         }
 599
 600         ipoib_dbg_mcast(priv, "successfully joined all multicast groups\n");
 601
 602         clear_bit(IPOIB_MCAST_RUN, &priv->flags);
 603 }
 604
 605 int ipoib_mcast_start_thread(struct net_device *dev)
 606 {
 607         struct ipoib_dev_priv *priv = netdev_priv(dev);
 608
 609         ipoib_dbg_mcast(priv, "starting multicast thread\n");
 610
 611         mutex_lock(&mcast_mutex);
 612         if (!test_and_set_bit(IPOIB_MCAST_RUN, &priv->flags))
 613                 queue_delayed_work(ipoib_workqueue, &priv->mcast_task, 0);
 614         mutex_unlock(&mcast_mutex);
 615
 616         return 0;
 617 }
 618
 619 int ipoib_mcast_stop_thread(struct net_device *dev, int flush)
 620 {
 621         struct ipoib_dev_priv *priv = netdev_priv(dev);
 622
 623         ipoib_dbg_mcast(priv, "stopping multicast thread\n");
 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         return 0;
 634 }
 635
 636 static int ipoib_mcast_leave(struct net_device *dev, struct ipoib_mcast *mcast)
 637 {
 638         struct ipoib_dev_priv *priv = netdev_priv(dev);
 639         int ret = 0;
 640
 641         if (test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
 642                 ib_sa_free_multicast(mcast->mc);
 643
 644         if (test_and_clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 645                 ipoib_dbg_mcast(priv, "leaving MGID %pI6\n",
 646                                 mcast->mcmember.mgid.raw);
 647
 648                 /* Remove ourselves from the multicast group */
 649                 ret = ib_detach_mcast(priv->qp, &mcast->mcmember.mgid,
 650                                       be16_to_cpu(mcast->mcmember.mlid));
 651                 if (ret)
 652                         ipoib_warn(priv, "ib_detach_mcast failed (result = %d)\n", ret);
 653         }
 654
 655         return 0;
 656 }
 657
 658 void ipoib_mcast_send(struct net_device *dev, void *mgid, struct sk_buff *skb)
 659 {
 660         struct ipoib_dev_priv *priv = netdev_priv(dev);
 661         struct ipoib_mcast *mcast;
 662         unsigned long flags;
 663
 664         spin_lock_irqsave(&priv->lock, flags);
 665
 666         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)         ||
 667             !priv->broadcast                                    ||
 668             !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
 669                 ++dev->stats.tx_dropped;
 670                 dev_kfree_skb_any(skb);
 671                 goto unlock;
 672         }
 673
 674         mcast = __ipoib_mcast_find(dev, mgid);
 675         if (!mcast) {
 676                 /* Let's create a new send only group now */
 677                 ipoib_dbg_mcast(priv, "setting up send only multicast group for %pI6\n",
 678                                 mgid);
 679
 680                 mcast = ipoib_mcast_alloc(dev, 0);
 681                 if (!mcast) {
 682                         ipoib_warn(priv, "unable to allocate memory for "
 683                                    "multicast structure\n");
 684                         ++dev->stats.tx_dropped;
 685                         dev_kfree_skb_any(skb);
 686                         goto out;
 687                 }
 688
 689                 set_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags);
 690                 memcpy(mcast->mcmember.mgid.raw, mgid, sizeof (union ib_gid));
 691                 __ipoib_mcast_add(dev, mcast);
 692                 list_add_tail(&mcast->list, &priv->multicast_list);
 693         }
 694
 695         if (!mcast->ah) {
 696                 if (skb_queue_len(&mcast->pkt_queue) < IPOIB_MAX_MCAST_QUEUE)
 697                         skb_queue_tail(&mcast->pkt_queue, skb);
 698                 else {
 699                         ++dev->stats.tx_dropped;
 700                         dev_kfree_skb_any(skb);
 701                 }
 702
 703                 if (test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
 704                         ipoib_dbg_mcast(priv, "no address vector, "
 705                                         "but multicast join already started\n");
 706                 else if (test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags))
 707                         ipoib_mcast_sendonly_join(mcast);
 708
 709                 /*
 710                  * If lookup completes between here and out:, don't
 711                  * want to send packet twice.
 712                  */
 713                 mcast = NULL;
 714         }
 715
 716 out:
 717         if (mcast && mcast->ah) {
 718                 if (skb_dst(skb)                &&
 719                     skb_dst(skb)->neighbour &&
 720                     !*to_ipoib_neigh(skb_dst(skb)->neighbour)) {
 721                         struct ipoib_neigh *neigh = ipoib_neigh_alloc(skb_dst(skb)->neighbour,
 722                                                                         skb->dev);
 723
 724                         if (neigh) {
 725                                 kref_get(&mcast->ah->ref);
 726                                 neigh->ah       = mcast->ah;
 727                                 list_add_tail(&neigh->list, &mcast->neigh_list);
 728                         }
 729                 }
 730
 731                 spin_unlock_irqrestore(&priv->lock, flags);
 732                 ipoib_send(dev, skb, mcast->ah, IB_MULTICAST_QPN);
 733                 return;
 734         }
 735
 736 unlock:
 737         spin_unlock_irqrestore(&priv->lock, flags);
 738 }
 739
 740 void ipoib_mcast_dev_flush(struct net_device *dev)
 741 {
 742         struct ipoib_dev_priv *priv = netdev_priv(dev);
 743         LIST_HEAD(remove_list);
 744         struct ipoib_mcast *mcast, *tmcast;
 745         unsigned long flags;
 746
 747         ipoib_dbg_mcast(priv, "flushing multicast list\n");
 748
 749         spin_lock_irqsave(&priv->lock, flags);
 750
 751         list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
 752                 list_del(&mcast->list);
 753                 rb_erase(&mcast->rb_node, &priv->multicast_tree);
 754                 list_add_tail(&mcast->list, &remove_list);
 755         }
 756
 757         if (priv->broadcast) {
 758                 rb_erase(&priv->broadcast->rb_node, &priv->multicast_tree);
 759                 list_add_tail(&priv->broadcast->list, &remove_list);
 760                 priv->broadcast = NULL;
 761         }
 762
 763         spin_unlock_irqrestore(&priv->lock, flags);
 764
 765         list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
 766                 ipoib_mcast_leave(dev, mcast);
 767                 ipoib_mcast_free(mcast);
 768         }
 769 }
 770
 771 static int ipoib_mcast_addr_is_valid(const u8 *addr, const u8 *broadcast)
 772 {
 773         /* reserved QPN, prefix, scope */
 774         if (memcmp(addr, broadcast, 6))
 775                 return 0;
 776         /* signature lower, pkey */
 777         if (memcmp(addr + 7, broadcast + 7, 3))
 778                 return 0;
 779         return 1;
 780 }
 781
 782 void ipoib_mcast_restart_task(struct work_struct *work)
 783 {
 784         struct ipoib_dev_priv *priv =
 785                 container_of(work, struct ipoib_dev_priv, restart_task);
 786         struct net_device *dev = priv->dev;
 787         struct netdev_hw_addr *ha;
 788         struct ipoib_mcast *mcast, *tmcast;
 789         LIST_HEAD(remove_list);
 790         unsigned long flags;
 791         struct ib_sa_mcmember_rec rec;
 792
 793         ipoib_dbg_mcast(priv, "restarting multicast task\n");
 794
 795         ipoib_mcast_stop_thread(dev, 0);
 796
 797         local_irq_save(flags);
 798         netif_addr_lock(dev);
 799         spin_lock(&priv->lock);
 800
 801         /*
 802          * Unfortunately, the networking core only gives us a list of all of
 803          * the multicast hardware addresses. We need to figure out which ones
 804          * are new and which ones have been removed
 805          */
 806
 807         /* Clear out the found flag */
 808         list_for_each_entry(mcast, &priv->multicast_list, list)
 809                 clear_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
 810
 811         /* Mark all of the entries that are found or don't exist */
 812         netdev_for_each_mc_addr(ha, dev) {
 813                 union ib_gid mgid;
 814
 815                 if (!ipoib_mcast_addr_is_valid(ha->addr, dev->broadcast))
 816                         continue;
 817
 818                 memcpy(mgid.raw, ha->addr + 4, sizeof mgid);
 819
 820                 mcast = __ipoib_mcast_find(dev, &mgid);
 821                 if (!mcast || test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 822                         struct ipoib_mcast *nmcast;
 823
 824                         /* ignore group which is directly joined by userspace */
 825                         if (test_bit(IPOIB_FLAG_UMCAST, &priv->flags) &&
 826                             !ib_sa_get_mcmember_rec(priv->ca, priv->port, &mgid, &rec)) {
 827                                 ipoib_dbg_mcast(priv, "ignoring multicast entry for mgid %pI6\n",
 828                                                 mgid.raw);
 829                                 continue;
 830                         }
 831
 832                         /* Not found or send-only group, let's add a new entry */
 833                         ipoib_dbg_mcast(priv, "adding multicast entry for mgid %pI6\n",
 834                                         mgid.raw);
 835
 836                         nmcast = ipoib_mcast_alloc(dev, 0);
 837                         if (!nmcast) {
 838                                 ipoib_warn(priv, "unable to allocate memory for multicast structure\n");
 839                                 continue;
 840                         }
 841
 842                         set_bit(IPOIB_MCAST_FLAG_FOUND, &nmcast->flags);
 843
 844                         nmcast->mcmember.mgid = mgid;
 845
 846                         if (mcast) {
 847                                 /* Destroy the send only entry */
 848                                 list_move_tail(&mcast->list, &remove_list);
 849
 850                                 rb_replace_node(&mcast->rb_node,
 851                                                 &nmcast->rb_node,
 852                                                 &priv->multicast_tree);
 853                         } else
 854                                 __ipoib_mcast_add(dev, nmcast);
 855
 856                         list_add_tail(&nmcast->list, &priv->multicast_list);
 857                 }
 858
 859                 if (mcast)
 860                         set_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
 861         }
 862
 863         /* Remove all of the entries don't exist anymore */
 864         list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
 865                 if (!test_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags) &&
 866                     !test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 867                         ipoib_dbg_mcast(priv, "deleting multicast group %pI6\n",
 868                                         mcast->mcmember.mgid.raw);
 869
 870                         rb_erase(&mcast->rb_node, &priv->multicast_tree);
 871
 872                         /* Move to the remove list */
 873                         list_move_tail(&mcast->list, &remove_list);
 874                 }
 875         }
 876
 877         spin_unlock(&priv->lock);
 878         netif_addr_unlock(dev);
 879         local_irq_restore(flags);
 880
 881         /* We have to cancel outside of the spinlock */
 882         list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
 883                 ipoib_mcast_leave(mcast->dev, mcast);
 884                 ipoib_mcast_free(mcast);
 885         }
 886
 887         if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 888                 ipoib_mcast_start_thread(dev);
 889 }
 890
 891 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 892
 893 struct ipoib_mcast_iter *ipoib_mcast_iter_init(struct net_device *dev)
 894 {
 895         struct ipoib_mcast_iter *iter;
 896
 897         iter = kmalloc(sizeof *iter, GFP_KERNEL);
 898         if (!iter)
 899                 return NULL;
 900
 901         iter->dev = dev;
 902         memset(iter->mgid.raw, 0, 16);
 903
 904         if (ipoib_mcast_iter_next(iter)) {
 905                 kfree(iter);
 906                 return NULL;
 907         }
 908
 909         return iter;
 910 }
 911
 912 int ipoib_mcast_iter_next(struct ipoib_mcast_iter *iter)
 913 {
 914         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
 915         struct rb_node *n;
 916         struct ipoib_mcast *mcast;
 917         int ret = 1;
 918
 919         spin_lock_irq(&priv->lock);
 920
 921         n = rb_first(&priv->multicast_tree);
 922
 923         while (n) {
 924                 mcast = rb_entry(n, struct ipoib_mcast, rb_node);
 925
 926                 if (memcmp(iter->mgid.raw, mcast->mcmember.mgid.raw,
 927                            sizeof (union ib_gid)) < 0) {
 928                         iter->mgid      = mcast->mcmember.mgid;
 929                         iter->created   = mcast->created;
 930                         iter->queuelen  = skb_queue_len(&mcast->pkt_queue);
 931                         iter->complete  = !!mcast->ah;
 932                         iter->send_only = !!(mcast->flags & (1 << IPOIB_MCAST_FLAG_SENDONLY));
 933
 934                         ret = 0;
 935
 936                         break;
 937                 }
 938
 939                 n = rb_next(n);
 940         }
 941
 942         spin_unlock_irq(&priv->lock);
 943
 944         return ret;
 945 }
 946
 947 void ipoib_mcast_iter_read(struct ipoib_mcast_iter *iter,
 948                            union ib_gid *mgid,
 949                            unsigned long *created,
 950                            unsigned int *queuelen,
 951                            unsigned int *complete,
 952                            unsigned int *send_only)
 953 {
 954         *mgid      = iter->mgid;
 955         *created   = iter->created;
 956         *queuelen  = iter->queuelen;
 957         *complete  = iter->complete;
 958         *send_only = iter->send_only;
 959 }
 960
 961 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */