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/moduleparam.h>
  38 #include <linux/ip.h>
  39 #include <linux/in.h>
  40 #include <linux/igmp.h>
  41 #include <linux/inetdevice.h>
  42 #include <linux/delay.h>
  43 #include <linux/completion.h>
  44 #include <linux/slab.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 struct ipoib_mcast_iter {
  61         struct net_device *dev;
  62         union ib_gid       mgid;
  63         unsigned long      created;
  64         unsigned int       queuelen;
  65         unsigned int       complete;
  66         unsigned int       send_only;
  67 };
  68
  69 static void ipoib_mcast_free(struct ipoib_mcast *mcast)
  70 {
  71         struct net_device *dev = mcast->dev;
  72         int tx_dropped = 0;
  73
  74         ipoib_dbg_mcast(netdev_priv(dev), "deleting multicast group %pI6\n",
  75                         mcast->mcmember.mgid.raw);
  76
  77         /* remove all neigh connected to this mcast */
  78         ipoib_del_neighs_by_gid(dev, mcast->mcmember.mgid.raw);
  79
  80         if (mcast->ah)
  81                 ipoib_put_ah(mcast->ah);
  82
  83         while (!skb_queue_empty(&mcast->pkt_queue)) {
  84                 ++tx_dropped;
  85                 dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
  86         }
  87
  88         netif_tx_lock_bh(dev);
  89         dev->stats.tx_dropped += tx_dropped;
  90         netif_tx_unlock_bh(dev);
  91
  92         kfree(mcast);
  93 }
  94
  95 static struct ipoib_mcast *ipoib_mcast_alloc(struct net_device *dev,
  96                                              int can_sleep)
  97 {
  98         struct ipoib_mcast *mcast;
  99
 100         mcast = kzalloc(sizeof *mcast, can_sleep ? GFP_KERNEL : GFP_ATOMIC);
 101         if (!mcast)
 102                 return NULL;
 103
 104         mcast->dev = dev;
 105         mcast->created = jiffies;
 106         mcast->backoff = 1;
 107
 108         INIT_LIST_HEAD(&mcast->list);
 109         INIT_LIST_HEAD(&mcast->neigh_list);
 110         skb_queue_head_init(&mcast->pkt_queue);
 111
 112         return mcast;
 113 }
 114
 115 static struct ipoib_mcast *__ipoib_mcast_find(struct net_device *dev, void *mgid)
 116 {
 117         struct ipoib_dev_priv *priv = netdev_priv(dev);
 118         struct rb_node *n = priv->multicast_tree.rb_node;
 119
 120         while (n) {
 121                 struct ipoib_mcast *mcast;
 122                 int ret;
 123
 124                 mcast = rb_entry(n, struct ipoib_mcast, rb_node);
 125
 126                 ret = memcmp(mgid, mcast->mcmember.mgid.raw,
 127                              sizeof (union ib_gid));
 128                 if (ret < 0)
 129                         n = n->rb_left;
 130                 else if (ret > 0)
 131                         n = n->rb_right;
 132                 else
 133                         return mcast;
 134         }
 135
 136         return NULL;
 137 }
 138
 139 static int __ipoib_mcast_add(struct net_device *dev, struct ipoib_mcast *mcast)
 140 {
 141         struct ipoib_dev_priv *priv = netdev_priv(dev);
 142         struct rb_node **n = &priv->multicast_tree.rb_node, *pn = NULL;
 143
 144         while (*n) {
 145                 struct ipoib_mcast *tmcast;
 146                 int ret;
 147
 148                 pn = *n;
 149                 tmcast = rb_entry(pn, struct ipoib_mcast, rb_node);
 150
 151                 ret = memcmp(mcast->mcmember.mgid.raw, tmcast->mcmember.mgid.raw,
 152                              sizeof (union ib_gid));
 153                 if (ret < 0)
 154                         n = &pn->rb_left;
 155                 else if (ret > 0)
 156                         n = &pn->rb_right;
 157                 else
 158                         return -EEXIST;
 159         }
 160
 161         rb_link_node(&mcast->rb_node, pn, n);
 162         rb_insert_color(&mcast->rb_node, &priv->multicast_tree);
 163
 164         return 0;
 165 }
 166
 167 static int ipoib_mcast_join_finish(struct ipoib_mcast *mcast,
 168                                    struct ib_sa_mcmember_rec *mcmember)
 169 {
 170         struct net_device *dev = mcast->dev;
 171         struct ipoib_dev_priv *priv = netdev_priv(dev);
 172         struct ipoib_ah *ah;
 173         int ret;
 174         int set_qkey = 0;
 175
 176         mcast->mcmember = *mcmember;
 177
 178         /* Set the multicast MTU and cached Q_Key before we attach if it's
 179          * the broadcast group.
 180          */
 181         if (!memcmp(mcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
 182                     sizeof (union ib_gid))) {
 183                 spin_lock_irq(&priv->lock);
 184                 if (!priv->broadcast) {
 185                         spin_unlock_irq(&priv->lock);
 186                         return -EAGAIN;
 187                 }
 188                 priv->mcast_mtu = IPOIB_UD_MTU(ib_mtu_enum_to_int(priv->broadcast->mcmember.mtu));
 189                 priv->qkey = be32_to_cpu(priv->broadcast->mcmember.qkey);
 190                 spin_unlock_irq(&priv->lock);
 191                 priv->tx_wr.wr.ud.remote_qkey = priv->qkey;
 192                 set_qkey = 1;
 193         }
 194
 195         if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 196                 if (test_and_set_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 197                         ipoib_warn(priv, "multicast group %pI6 already attached\n",
 198                                    mcast->mcmember.mgid.raw);
 199
 200                         return 0;
 201                 }
 202
 203                 ret = ipoib_mcast_attach(dev, be16_to_cpu(mcast->mcmember.mlid),
 204                                          &mcast->mcmember.mgid, set_qkey);
 205                 if (ret < 0) {
 206                         ipoib_warn(priv, "couldn't attach QP to multicast group %pI6\n",
 207                                    mcast->mcmember.mgid.raw);
 208
 209                         clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags);
 210                         return ret;
 211                 }
 212         }
 213
 214         {
 215                 struct ib_ah_attr av = {
 216                         .dlid          = be16_to_cpu(mcast->mcmember.mlid),
 217                         .port_num      = priv->port,
 218                         .sl            = mcast->mcmember.sl,
 219                         .ah_flags      = IB_AH_GRH,
 220                         .static_rate   = mcast->mcmember.rate,
 221                         .grh           = {
 222                                 .flow_label    = be32_to_cpu(mcast->mcmember.flow_label),
 223                                 .hop_limit     = mcast->mcmember.hop_limit,
 224                                 .sgid_index    = 0,
 225                                 .traffic_class = mcast->mcmember.traffic_class
 226                         }
 227                 };
 228                 av.grh.dgid = mcast->mcmember.mgid;
 229
 230                 ah = ipoib_create_ah(dev, priv->pd, &av);
 231                 if (IS_ERR(ah)) {
 232                         ipoib_warn(priv, "ib_address_create failed %ld\n",
 233                                 -PTR_ERR(ah));
 234                         /* use original error */
 235                         return PTR_ERR(ah);
 236                 } else {
 237                         spin_lock_irq(&priv->lock);
 238                         mcast->ah = ah;
 239                         spin_unlock_irq(&priv->lock);
 240
 241                         ipoib_dbg_mcast(priv, "MGID %pI6 AV %p, LID 0x%04x, SL %d\n",
 242                                         mcast->mcmember.mgid.raw,
 243                                         mcast->ah->ah,
 244                                         be16_to_cpu(mcast->mcmember.mlid),
 245                                         mcast->mcmember.sl);
 246                 }
 247         }
 248
 249         /* actually send any queued packets */
 250         netif_tx_lock_bh(dev);
 251         while (!skb_queue_empty(&mcast->pkt_queue)) {
 252                 struct sk_buff *skb = skb_dequeue(&mcast->pkt_queue);
 253
 254                 netif_tx_unlock_bh(dev);
 255
 256                 skb->dev = dev;
 257                 if (dev_queue_xmit(skb))
 258                         ipoib_warn(priv, "dev_queue_xmit failed to requeue packet\n");
 259
 260                 netif_tx_lock_bh(dev);
 261         }
 262         netif_tx_unlock_bh(dev);
 263
 264         return 0;
 265 }
 266
 267 static int
 268 ipoib_mcast_sendonly_join_complete(int status,
 269                                    struct ib_sa_multicast *multicast)
 270 {
 271         struct ipoib_mcast *mcast = multicast->context;
 272         struct net_device *dev = mcast->dev;
 273
 274         /* We trap for port events ourselves. */
 275         if (status == -ENETRESET)
 276                 return 0;
 277
 278         if (!status)
 279                 status = ipoib_mcast_join_finish(mcast, &multicast->rec);
 280
 281         if (status) {
 282                 if (mcast->logcount++ < 20)
 283                         ipoib_dbg_mcast(netdev_priv(dev), "multicast join failed for %pI6, status %d\n",
 284                                         mcast->mcmember.mgid.raw, status);
 285
 286                 /* Flush out any queued packets */
 287                 netif_tx_lock_bh(dev);
 288                 while (!skb_queue_empty(&mcast->pkt_queue)) {
 289                         ++dev->stats.tx_dropped;
 290                         dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
 291                 }
 292                 netif_tx_unlock_bh(dev);
 293
 294                 /* Clear the busy flag so we try again */
 295                 status = test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY,
 296                                             &mcast->flags);
 297         }
 298         return status;
 299 }
 300
 301 static int ipoib_mcast_sendonly_join(struct ipoib_mcast *mcast)
 302 {
 303         struct net_device *dev = mcast->dev;
 304         struct ipoib_dev_priv *priv = netdev_priv(dev);
 305         struct ib_sa_mcmember_rec rec = {
 306 #if 0                           /* Some SMs don't support send-only yet */
 307                 .join_state = 4
 308 #else
 309                 .join_state = 1
 310 #endif
 311         };
 312         int ret = 0;
 313
 314         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
 315                 ipoib_dbg_mcast(priv, "device shutting down, no multicast joins\n");
 316                 return -ENODEV;
 317         }
 318
 319         if (test_and_set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)) {
 320                 ipoib_dbg_mcast(priv, "multicast entry busy, skipping\n");
 321                 return -EBUSY;
 322         }
 323
 324         rec.mgid     = mcast->mcmember.mgid;
 325         rec.port_gid = priv->local_gid;
 326         rec.pkey     = cpu_to_be16(priv->pkey);
 327
 328         mcast->mc = ib_sa_join_multicast(&ipoib_sa_client, priv->ca,
 329                                          priv->port, &rec,
 330                                          IB_SA_MCMEMBER_REC_MGID        |
 331                                          IB_SA_MCMEMBER_REC_PORT_GID    |
 332                                          IB_SA_MCMEMBER_REC_PKEY        |
 333                                          IB_SA_MCMEMBER_REC_JOIN_STATE,
 334                                          GFP_ATOMIC,
 335                                          ipoib_mcast_sendonly_join_complete,
 336                                          mcast);
 337         if (IS_ERR(mcast->mc)) {
 338                 ret = PTR_ERR(mcast->mc);
 339                 clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 340                 ipoib_warn(priv, "ib_sa_join_multicast failed (ret = %d)\n",
 341                            ret);
 342         } else {
 343                 ipoib_dbg_mcast(priv, "no multicast record for %pI6, starting join\n",
 344                                 mcast->mcmember.mgid.raw);
 345         }
 346
 347         return ret;
 348 }
 349
 350 void ipoib_mcast_carrier_on_task(struct work_struct *work)
 351 {
 352         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
 353                                                    carrier_on_task);
 354         struct ib_port_attr attr;
 355
 356         if (ib_query_port(priv->ca, priv->port, &attr) ||
 357             attr.state != IB_PORT_ACTIVE) {
 358                 ipoib_dbg(priv, "Keeping carrier off until IB port is active\n");
 359                 return;
 360         }
 361
 362         /*
 363          * Take rtnl_lock to avoid racing with ipoib_stop() and
 364          * turning the carrier back on while a device is being
 365          * removed.  However, ipoib_stop() will attempt to flush
 366          * the workqueue while holding the rtnl lock, so loop
 367          * on trylock until either we get the lock or we see
 368          * FLAG_OPER_UP go away as that signals that we are bailing
 369          * and can safely ignore the carrier on work.
 370          */
 371         while (!rtnl_trylock()) {
 372                 if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
 373                         return;
 374                 else
 375                         msleep(20);
 376         }
 377         if (!ipoib_cm_admin_enabled(priv->dev))
 378                 dev_set_mtu(priv->dev, min(priv->mcast_mtu, priv->admin_mtu));
 379         netif_carrier_on(priv->dev);
 380         rtnl_unlock();
 381 }
 382
 383 static int ipoib_mcast_join_complete(int status,
 384                                      struct ib_sa_multicast *multicast)
 385 {
 386         struct ipoib_mcast *mcast = multicast->context;
 387         struct net_device *dev = mcast->dev;
 388         struct ipoib_dev_priv *priv = netdev_priv(dev);
 389
 390         ipoib_dbg_mcast(priv, "join completion for %pI6 (status %d)\n",
 391                         mcast->mcmember.mgid.raw, status);
 392
 393         /* We trap for port events ourselves. */
 394         if (status == -ENETRESET) {
 395                 status = 0;
 396                 goto out;
 397         }
 398
 399         if (!status)
 400                 status = ipoib_mcast_join_finish(mcast, &multicast->rec);
 401
 402         if (!status) {
 403                 mcast->backoff = 1;
 404                 mutex_lock(&mcast_mutex);
 405                 if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 406                         queue_delayed_work(priv->wq, &priv->mcast_task, 0);
 407                 mutex_unlock(&mcast_mutex);
 408
 409                 /*
 410                  * Defer carrier on work to priv->wq to avoid a
 411                  * deadlock on rtnl_lock here.
 412                  */
 413                 if (mcast == priv->broadcast)
 414                         queue_work(priv->wq, &priv->carrier_on_task);
 415
 416                 status = 0;
 417                 goto out;
 418         }
 419
 420         if (mcast->logcount++ < 20) {
 421                 if (status == -ETIMEDOUT || status == -EAGAIN) {
 422                         ipoib_dbg_mcast(priv, "multicast join failed for %pI6, status %d\n",
 423                                         mcast->mcmember.mgid.raw, status);
 424                 } else {
 425                         ipoib_warn(priv, "multicast join failed for %pI6, status %d\n",
 426                                    mcast->mcmember.mgid.raw, status);
 427                 }
 428         }
 429
 430         mcast->backoff *= 2;
 431         if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
 432                 mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
 433
 434         /* Clear the busy flag so we try again */
 435         status = test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 436
 437         mutex_lock(&mcast_mutex);
 438         spin_lock_irq(&priv->lock);
 439         if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 440                 queue_delayed_work(priv->wq, &priv->mcast_task,
 441                                    mcast->backoff * HZ);
 442         spin_unlock_irq(&priv->lock);
 443         mutex_unlock(&mcast_mutex);
 444 out:
 445         complete(&mcast->done);
 446         return status;
 447 }
 448
 449 static void ipoib_mcast_join(struct net_device *dev, struct ipoib_mcast *mcast,
 450                              int create)
 451 {
 452         struct ipoib_dev_priv *priv = netdev_priv(dev);
 453         struct ib_sa_mcmember_rec rec = {
 454                 .join_state = 1
 455         };
 456         ib_sa_comp_mask comp_mask;
 457         int ret = 0;
 458
 459         ipoib_dbg_mcast(priv, "joining MGID %pI6\n", mcast->mcmember.mgid.raw);
 460
 461         rec.mgid     = mcast->mcmember.mgid;
 462         rec.port_gid = priv->local_gid;
 463         rec.pkey     = cpu_to_be16(priv->pkey);
 464
 465         comp_mask =
 466                 IB_SA_MCMEMBER_REC_MGID         |
 467                 IB_SA_MCMEMBER_REC_PORT_GID     |
 468                 IB_SA_MCMEMBER_REC_PKEY         |
 469                 IB_SA_MCMEMBER_REC_JOIN_STATE;
 470
 471         if (create) {
 472                 comp_mask |=
 473                         IB_SA_MCMEMBER_REC_QKEY                 |
 474                         IB_SA_MCMEMBER_REC_MTU_SELECTOR         |
 475                         IB_SA_MCMEMBER_REC_MTU                  |
 476                         IB_SA_MCMEMBER_REC_TRAFFIC_CLASS        |
 477                         IB_SA_MCMEMBER_REC_RATE_SELECTOR        |
 478                         IB_SA_MCMEMBER_REC_RATE                 |
 479                         IB_SA_MCMEMBER_REC_SL                   |
 480                         IB_SA_MCMEMBER_REC_FLOW_LABEL           |
 481                         IB_SA_MCMEMBER_REC_HOP_LIMIT;
 482
 483                 rec.qkey          = priv->broadcast->mcmember.qkey;
 484                 rec.mtu_selector  = IB_SA_EQ;
 485                 rec.mtu           = priv->broadcast->mcmember.mtu;
 486                 rec.traffic_class = priv->broadcast->mcmember.traffic_class;
 487                 rec.rate_selector = IB_SA_EQ;
 488                 rec.rate          = priv->broadcast->mcmember.rate;
 489                 rec.sl            = priv->broadcast->mcmember.sl;
 490                 rec.flow_label    = priv->broadcast->mcmember.flow_label;
 491                 rec.hop_limit     = priv->broadcast->mcmember.hop_limit;
 492         }
 493
 494         set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 495         init_completion(&mcast->done);
 496         set_bit(IPOIB_MCAST_JOIN_STARTED, &mcast->flags);
 497
 498         mcast->mc = ib_sa_join_multicast(&ipoib_sa_client, priv->ca, priv->port,
 499                                          &rec, comp_mask, GFP_KERNEL,
 500                                          ipoib_mcast_join_complete, mcast);
 501         if (IS_ERR(mcast->mc)) {
 502                 clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 503                 complete(&mcast->done);
 504                 ret = PTR_ERR(mcast->mc);
 505                 ipoib_warn(priv, "ib_sa_join_multicast failed, status %d\n", ret);
 506
 507                 mcast->backoff *= 2;
 508                 if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
 509                         mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
 510
 511                 mutex_lock(&mcast_mutex);
 512                 if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 513                         queue_delayed_work(priv->wq, &priv->mcast_task,
 514                                            mcast->backoff * HZ);
 515                 mutex_unlock(&mcast_mutex);
 516         }
 517 }
 518
 519 void ipoib_mcast_join_task(struct work_struct *work)
 520 {
 521         struct ipoib_dev_priv *priv =
 522                 container_of(work, struct ipoib_dev_priv, mcast_task.work);
 523         struct net_device *dev = priv->dev;
 524         struct ib_port_attr port_attr;
 525
 526         if (!test_bit(IPOIB_MCAST_RUN, &priv->flags))
 527                 return;
 528
 529         if (ib_query_port(priv->ca, priv->port, &port_attr) ||
 530             port_attr.state != IB_PORT_ACTIVE) {
 531                 ipoib_dbg(priv, "port state is not ACTIVE (state = %d) suspending join task\n",
 532                           port_attr.state);
 533                 return;
 534         }
 535         priv->local_lid = port_attr.lid;
 536
 537         if (ib_query_gid(priv->ca, priv->port, 0, &priv->local_gid))
 538                 ipoib_warn(priv, "ib_query_gid() failed\n");
 539         else
 540                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
 541
 542         if (!priv->broadcast) {
 543                 struct ipoib_mcast *broadcast;
 544
 545                 if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
 546                         return;
 547
 548                 broadcast = ipoib_mcast_alloc(dev, 1);
 549                 if (!broadcast) {
 550                         ipoib_warn(priv, "failed to allocate broadcast group\n");
 551                         mutex_lock(&mcast_mutex);
 552                         if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 553                                 queue_delayed_work(priv->wq, &priv->mcast_task,
 554                                                    HZ);
 555                         mutex_unlock(&mcast_mutex);
 556                         return;
 557                 }
 558
 559                 spin_lock_irq(&priv->lock);
 560                 memcpy(broadcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
 561                        sizeof (union ib_gid));
 562                 priv->broadcast = broadcast;
 563
 564                 __ipoib_mcast_add(dev, priv->broadcast);
 565                 spin_unlock_irq(&priv->lock);
 566         }
 567
 568         if (!test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
 569                 if (!test_bit(IPOIB_MCAST_FLAG_BUSY, &priv->broadcast->flags))
 570                         ipoib_mcast_join(dev, priv->broadcast, 0);
 571                 return;
 572         }
 573
 574         while (1) {
 575                 struct ipoib_mcast *mcast = NULL;
 576
 577                 spin_lock_irq(&priv->lock);
 578                 list_for_each_entry(mcast, &priv->multicast_list, list) {
 579                         if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)
 580                             && !test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)
 581                             && !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 582                                 /* Found the next unjoined group */
 583                                 break;
 584                         }
 585                 }
 586                 spin_unlock_irq(&priv->lock);
 587
 588                 if (&mcast->list == &priv->multicast_list) {
 589                         /* All done */
 590                         break;
 591                 }
 592
 593                 ipoib_mcast_join(dev, mcast, 1);
 594                 return;
 595         }
 596
 597         ipoib_dbg_mcast(priv, "successfully joined all multicast groups\n");
 598
 599         clear_bit(IPOIB_MCAST_RUN, &priv->flags);
 600 }
 601
 602 int ipoib_mcast_start_thread(struct net_device *dev)
 603 {
 604         struct ipoib_dev_priv *priv = netdev_priv(dev);
 605
 606         ipoib_dbg_mcast(priv, "starting multicast thread\n");
 607
 608         mutex_lock(&mcast_mutex);
 609         if (!test_and_set_bit(IPOIB_MCAST_RUN, &priv->flags))
 610                 queue_delayed_work(priv->wq, &priv->mcast_task, 0);
 611         mutex_unlock(&mcast_mutex);
 612
 613         return 0;
 614 }
 615
 616 int ipoib_mcast_stop_thread(struct net_device *dev, int flush)
 617 {
 618         struct ipoib_dev_priv *priv = netdev_priv(dev);
 619
 620         ipoib_dbg_mcast(priv, "stopping multicast thread\n");
 621
 622         mutex_lock(&mcast_mutex);
 623         clear_bit(IPOIB_MCAST_RUN, &priv->flags);
 624         cancel_delayed_work(&priv->mcast_task);
 625         mutex_unlock(&mcast_mutex);
 626
 627         if (flush)
 628                 flush_workqueue(priv->wq);
 629
 630         return 0;
 631 }
 632
 633 static int ipoib_mcast_leave(struct net_device *dev, struct ipoib_mcast *mcast)
 634 {
 635         struct ipoib_dev_priv *priv = netdev_priv(dev);
 636         int ret = 0;
 637
 638         if (test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
 639                 ib_sa_free_multicast(mcast->mc);
 640
 641         if (test_and_clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 642                 ipoib_dbg_mcast(priv, "leaving MGID %pI6\n",
 643                                 mcast->mcmember.mgid.raw);
 644
 645                 /* Remove ourselves from the multicast group */
 646                 ret = ib_detach_mcast(priv->qp, &mcast->mcmember.mgid,
 647                                       be16_to_cpu(mcast->mcmember.mlid));
 648                 if (ret)
 649                         ipoib_warn(priv, "ib_detach_mcast failed (result = %d)\n", ret);
 650         }
 651
 652         return 0;
 653 }
 654
 655 void ipoib_mcast_send(struct net_device *dev, u8 *daddr, struct sk_buff *skb)
 656 {
 657         struct ipoib_dev_priv *priv = netdev_priv(dev);
 658         struct ipoib_mcast *mcast;
 659         unsigned long flags;
 660         void *mgid = daddr + 4;
 661
 662         spin_lock_irqsave(&priv->lock, flags);
 663
 664         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)         ||
 665             !priv->broadcast                                    ||
 666             !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
 667                 ++dev->stats.tx_dropped;
 668                 dev_kfree_skb_any(skb);
 669                 goto unlock;
 670         }
 671
 672         mcast = __ipoib_mcast_find(dev, mgid);
 673         if (!mcast) {
 674                 /* Let's create a new send only group now */
 675                 ipoib_dbg_mcast(priv, "setting up send only multicast group for %pI6\n",
 676                                 mgid);
 677
 678                 mcast = ipoib_mcast_alloc(dev, 0);
 679                 if (!mcast) {
 680                         ipoib_warn(priv, "unable to allocate memory for "
 681                                    "multicast structure\n");
 682                         ++dev->stats.tx_dropped;
 683                         dev_kfree_skb_any(skb);
 684                         goto out;
 685                 }
 686
 687                 set_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags);
 688                 memcpy(mcast->mcmember.mgid.raw, mgid, sizeof (union ib_gid));
 689                 __ipoib_mcast_add(dev, mcast);
 690                 list_add_tail(&mcast->list, &priv->multicast_list);
 691         }
 692
 693         if (!mcast->ah) {
 694                 if (skb_queue_len(&mcast->pkt_queue) < IPOIB_MAX_MCAST_QUEUE)
 695                         skb_queue_tail(&mcast->pkt_queue, skb);
 696                 else {
 697                         ++dev->stats.tx_dropped;
 698                         dev_kfree_skb_any(skb);
 699                 }
 700
 701                 if (test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
 702                         ipoib_dbg_mcast(priv, "no address vector, "
 703                                         "but multicast join already started\n");
 704                 else if (test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags))
 705                         ipoib_mcast_sendonly_join(mcast);
 706
 707                 /*
 708                  * If lookup completes between here and out:, don't
 709                  * want to send packet twice.
 710                  */
 711                 mcast = NULL;
 712         }
 713
 714 out:
 715         if (mcast && mcast->ah) {
 716                 struct ipoib_neigh *neigh;
 717
 718                 spin_unlock_irqrestore(&priv->lock, flags);
 719                 neigh = ipoib_neigh_get(dev, daddr);
 720                 spin_lock_irqsave(&priv->lock, flags);
 721                 if (!neigh) {
 722                         neigh = ipoib_neigh_alloc(daddr, dev);
 723                         if (neigh) {
 724                                 kref_get(&mcast->ah->ref);
 725                                 neigh->ah       = mcast->ah;
 726                                 list_add_tail(&neigh->list, &mcast->neigh_list);
 727                         }
 728                 }
 729                 spin_unlock_irqrestore(&priv->lock, flags);
 730                 ipoib_send(dev, skb, mcast->ah, IB_MULTICAST_QPN);
 731                 if (neigh)
 732                         ipoib_neigh_put(neigh);
 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         /* seperate between the wait to the leave*/
 766         list_for_each_entry_safe(mcast, tmcast, &remove_list, list)
 767                 if (test_bit(IPOIB_MCAST_JOIN_STARTED, &mcast->flags))
 768                         wait_for_completion(&mcast->done);
 769
 770         list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
 771                 ipoib_mcast_leave(dev, mcast);
 772                 ipoib_mcast_free(mcast);
 773         }
 774 }
 775
 776 static int ipoib_mcast_addr_is_valid(const u8 *addr, const u8 *broadcast)
 777 {
 778         /* reserved QPN, prefix, scope */
 779         if (memcmp(addr, broadcast, 6))
 780                 return 0;
 781         /* signature lower, pkey */
 782         if (memcmp(addr + 7, broadcast + 7, 3))
 783                 return 0;
 784         return 1;
 785 }
 786
 787 void ipoib_mcast_restart_task(struct work_struct *work)
 788 {
 789         struct ipoib_dev_priv *priv =
 790                 container_of(work, struct ipoib_dev_priv, restart_task);
 791         struct net_device *dev = priv->dev;
 792         struct netdev_hw_addr *ha;
 793         struct ipoib_mcast *mcast, *tmcast;
 794         LIST_HEAD(remove_list);
 795         unsigned long flags;
 796         struct ib_sa_mcmember_rec rec;
 797
 798         ipoib_dbg_mcast(priv, "restarting multicast task\n");
 799
 800         ipoib_mcast_stop_thread(dev, 0);
 801
 802         local_irq_save(flags);
 803         netif_addr_lock(dev);
 804         spin_lock(&priv->lock);
 805
 806         /*
 807          * Unfortunately, the networking core only gives us a list of all of
 808          * the multicast hardware addresses. We need to figure out which ones
 809          * are new and which ones have been removed
 810          */
 811
 812         /* Clear out the found flag */
 813         list_for_each_entry(mcast, &priv->multicast_list, list)
 814                 clear_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
 815
 816         /* Mark all of the entries that are found or don't exist */
 817         netdev_for_each_mc_addr(ha, dev) {
 818                 union ib_gid mgid;
 819
 820                 if (!ipoib_mcast_addr_is_valid(ha->addr, dev->broadcast))
 821                         continue;
 822
 823                 memcpy(mgid.raw, ha->addr + 4, sizeof mgid);
 824
 825                 mcast = __ipoib_mcast_find(dev, &mgid);
 826                 if (!mcast || test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 827                         struct ipoib_mcast *nmcast;
 828
 829                         /* ignore group which is directly joined by userspace */
 830                         if (test_bit(IPOIB_FLAG_UMCAST, &priv->flags) &&
 831                             !ib_sa_get_mcmember_rec(priv->ca, priv->port, &mgid, &rec)) {
 832                                 ipoib_dbg_mcast(priv, "ignoring multicast entry for mgid %pI6\n",
 833                                                 mgid.raw);
 834                                 continue;
 835                         }
 836
 837                         /* Not found or send-only group, let's add a new entry */
 838                         ipoib_dbg_mcast(priv, "adding multicast entry for mgid %pI6\n",
 839                                         mgid.raw);
 840
 841                         nmcast = ipoib_mcast_alloc(dev, 0);
 842                         if (!nmcast) {
 843                                 ipoib_warn(priv, "unable to allocate memory for multicast structure\n");
 844                                 continue;
 845                         }
 846
 847                         set_bit(IPOIB_MCAST_FLAG_FOUND, &nmcast->flags);
 848
 849                         nmcast->mcmember.mgid = mgid;
 850
 851                         if (mcast) {
 852                                 /* Destroy the send only entry */
 853                                 list_move_tail(&mcast->list, &remove_list);
 854
 855                                 rb_replace_node(&mcast->rb_node,
 856                                                 &nmcast->rb_node,
 857                                                 &priv->multicast_tree);
 858                         } else
 859                                 __ipoib_mcast_add(dev, nmcast);
 860
 861                         list_add_tail(&nmcast->list, &priv->multicast_list);
 862                 }
 863
 864                 if (mcast)
 865                         set_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
 866         }
 867
 868         /* Remove all of the entries don't exist anymore */
 869         list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
 870                 if (!test_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags) &&
 871                     !test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 872                         ipoib_dbg_mcast(priv, "deleting multicast group %pI6\n",
 873                                         mcast->mcmember.mgid.raw);
 874
 875                         rb_erase(&mcast->rb_node, &priv->multicast_tree);
 876
 877                         /* Move to the remove list */
 878                         list_move_tail(&mcast->list, &remove_list);
 879                 }
 880         }
 881
 882         spin_unlock(&priv->lock);
 883         netif_addr_unlock(dev);
 884         local_irq_restore(flags);
 885
 886         /* We have to cancel outside of the spinlock */
 887         list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
 888                 ipoib_mcast_leave(mcast->dev, mcast);
 889                 ipoib_mcast_free(mcast);
 890         }
 891
 892         if (test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
 893                 ipoib_mcast_start_thread(dev);
 894 }
 895
 896 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 897
 898 struct ipoib_mcast_iter *ipoib_mcast_iter_init(struct net_device *dev)
 899 {
 900         struct ipoib_mcast_iter *iter;
 901
 902         iter = kmalloc(sizeof *iter, GFP_KERNEL);
 903         if (!iter)
 904                 return NULL;
 905
 906         iter->dev = dev;
 907         memset(iter->mgid.raw, 0, 16);
 908
 909         if (ipoib_mcast_iter_next(iter)) {
 910                 kfree(iter);
 911                 return NULL;
 912         }
 913
 914         return iter;
 915 }
 916
 917 int ipoib_mcast_iter_next(struct ipoib_mcast_iter *iter)
 918 {
 919         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
 920         struct rb_node *n;
 921         struct ipoib_mcast *mcast;
 922         int ret = 1;
 923
 924         spin_lock_irq(&priv->lock);
 925
 926         n = rb_first(&priv->multicast_tree);
 927
 928         while (n) {
 929                 mcast = rb_entry(n, struct ipoib_mcast, rb_node);
 930
 931                 if (memcmp(iter->mgid.raw, mcast->mcmember.mgid.raw,
 932                            sizeof (union ib_gid)) < 0) {
 933                         iter->mgid      = mcast->mcmember.mgid;
 934                         iter->created   = mcast->created;
 935                         iter->queuelen  = skb_queue_len(&mcast->pkt_queue);
 936                         iter->complete  = !!mcast->ah;
 937                         iter->send_only = !!(mcast->flags & (1 << IPOIB_MCAST_FLAG_SENDONLY));
 938
 939                         ret = 0;
 940
 941                         break;
 942                 }
 943
 944                 n = rb_next(n);
 945         }
 946
 947         spin_unlock_irq(&priv->lock);
 948
 949         return ret;
 950 }
 951
 952 void ipoib_mcast_iter_read(struct ipoib_mcast_iter *iter,
 953                            union ib_gid *mgid,
 954                            unsigned long *created,
 955                            unsigned int *queuelen,
 956                            unsigned int *complete,
 957                            unsigned int *send_only)
 958 {
 959         *mgid      = iter->mgid;
 960         *created   = iter->created;
 961         *queuelen  = iter->queuelen;
 962         *complete  = iter->complete;
 963         *send_only = iter->send_only;
 964 }
 965
 966 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */