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         /*
 275          * We have to take the mutex to force mcast_sendonly_join to
 276          * return from ib_sa_multicast_join and set mcast->mc to a
 277          * valid value.  Otherwise we were racing with ourselves in
 278          * that we might fail here, but get a valid return from
 279          * ib_sa_multicast_join after we had cleared mcast->mc here,
 280          * resulting in mis-matched joins and leaves and a deadlock
 281          */
 282         mutex_lock(&mcast_mutex);
 283
 284         /* We trap for port events ourselves. */
 285         if (status == -ENETRESET)
 286                 goto out;
 287
 288         if (!status)
 289                 status = ipoib_mcast_join_finish(mcast, &multicast->rec);
 290
 291         if (status) {
 292                 if (mcast->logcount++ < 20)
 293                         ipoib_dbg_mcast(netdev_priv(dev), "sendonly multicast "
 294                                         "join failed for %pI6, status %d\n",
 295                                         mcast->mcmember.mgid.raw, status);
 296
 297                 /* Flush out any queued packets */
 298                 netif_tx_lock_bh(dev);
 299                 while (!skb_queue_empty(&mcast->pkt_queue)) {
 300                         ++dev->stats.tx_dropped;
 301                         dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
 302                 }
 303                 netif_tx_unlock_bh(dev);
 304         }
 305 out:
 306         clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 307         if (status)
 308                 mcast->mc = NULL;
 309         complete(&mcast->done);
 310         if (status == -ENETRESET)
 311                 status = 0;
 312         mutex_unlock(&mcast_mutex);
 313         return status;
 314 }
 315
 316 static int ipoib_mcast_sendonly_join(struct ipoib_mcast *mcast)
 317 {
 318         struct net_device *dev = mcast->dev;
 319         struct ipoib_dev_priv *priv = netdev_priv(dev);
 320         struct ib_sa_mcmember_rec rec = {
 321 #if 0                           /* Some SMs don't support send-only yet */
 322                 .join_state = 4
 323 #else
 324                 .join_state = 1
 325 #endif
 326         };
 327         int ret = 0;
 328
 329         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
 330                 ipoib_dbg_mcast(priv, "device shutting down, no sendonly "
 331                                 "multicast joins\n");
 332                 return -ENODEV;
 333         }
 334
 335         if (test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)) {
 336                 ipoib_dbg_mcast(priv, "multicast entry busy, skipping "
 337                                 "sendonly join\n");
 338                 return -EBUSY;
 339         }
 340
 341         rec.mgid     = mcast->mcmember.mgid;
 342         rec.port_gid = priv->local_gid;
 343         rec.pkey     = cpu_to_be16(priv->pkey);
 344
 345         mutex_lock(&mcast_mutex);
 346         init_completion(&mcast->done);
 347         set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 348         mcast->mc = ib_sa_join_multicast(&ipoib_sa_client, priv->ca,
 349                                          priv->port, &rec,
 350                                          IB_SA_MCMEMBER_REC_MGID        |
 351                                          IB_SA_MCMEMBER_REC_PORT_GID    |
 352                                          IB_SA_MCMEMBER_REC_PKEY        |
 353                                          IB_SA_MCMEMBER_REC_JOIN_STATE,
 354                                          GFP_ATOMIC,
 355                                          ipoib_mcast_sendonly_join_complete,
 356                                          mcast);
 357         if (IS_ERR(mcast->mc)) {
 358                 ret = PTR_ERR(mcast->mc);
 359                 clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 360                 complete(&mcast->done);
 361                 ipoib_warn(priv, "ib_sa_join_multicast for sendonly join "
 362                            "failed (ret = %d)\n", ret);
 363         } else {
 364                 ipoib_dbg_mcast(priv, "no multicast record for %pI6, starting "
 365                                 "sendonly join\n", mcast->mcmember.mgid.raw);
 366         }
 367         mutex_unlock(&mcast_mutex);
 368
 369         return ret;
 370 }
 371
 372 void ipoib_mcast_carrier_on_task(struct work_struct *work)
 373 {
 374         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
 375                                                    carrier_on_task);
 376         struct ib_port_attr attr;
 377
 378         if (ib_query_port(priv->ca, priv->port, &attr) ||
 379             attr.state != IB_PORT_ACTIVE) {
 380                 ipoib_dbg(priv, "Keeping carrier off until IB port is active\n");
 381                 return;
 382         }
 383
 384         /*
 385          * Take rtnl_lock to avoid racing with ipoib_stop() and
 386          * turning the carrier back on while a device is being
 387          * removed.  However, ipoib_stop() will attempt to flush
 388          * the workqueue while holding the rtnl lock, so loop
 389          * on trylock until either we get the lock or we see
 390          * FLAG_ADMIN_UP go away as that signals that we are bailing
 391          * and can safely ignore the carrier on work
 392          */
 393         while (!rtnl_trylock()) {
 394                 if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 395                         return;
 396                 else
 397                         msleep(20);
 398         }
 399         if (!ipoib_cm_admin_enabled(priv->dev))
 400                 dev_set_mtu(priv->dev, min(priv->mcast_mtu, priv->admin_mtu));
 401         netif_carrier_on(priv->dev);
 402         rtnl_unlock();
 403 }
 404
 405 static int ipoib_mcast_join_complete(int status,
 406                                      struct ib_sa_multicast *multicast)
 407 {
 408         struct ipoib_mcast *mcast = multicast->context;
 409         struct net_device *dev = mcast->dev;
 410         struct ipoib_dev_priv *priv = netdev_priv(dev);
 411
 412         ipoib_dbg_mcast(priv, "join completion for %pI6 (status %d)\n",
 413                         mcast->mcmember.mgid.raw, status);
 414
 415         /*
 416          * We have to take the mutex to force mcast_join to
 417          * return from ib_sa_multicast_join and set mcast->mc to a
 418          * valid value.  Otherwise we were racing with ourselves in
 419          * that we might fail here, but get a valid return from
 420          * ib_sa_multicast_join after we had cleared mcast->mc here,
 421          * resulting in mis-matched joins and leaves and a deadlock
 422          */
 423         mutex_lock(&mcast_mutex);
 424
 425         /* We trap for port events ourselves. */
 426         if (status == -ENETRESET)
 427                 goto out;
 428
 429         if (!status)
 430                 status = ipoib_mcast_join_finish(mcast, &multicast->rec);
 431
 432         if (!status) {
 433                 mcast->backoff = 1;
 434                 if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 435                         queue_delayed_work(ipoib_workqueue,
 436                                            &priv->mcast_task, 0);
 437
 438                 /*
 439                  * Defer carrier on work to ipoib_workqueue to avoid a
 440                  * deadlock on rtnl_lock here.
 441                  */
 442                 if (mcast == priv->broadcast)
 443                         queue_work(ipoib_workqueue, &priv->carrier_on_task);
 444         } else {
 445                 if (mcast->logcount++ < 20) {
 446                         if (status == -ETIMEDOUT || status == -EAGAIN) {
 447                                 ipoib_dbg_mcast(priv, "multicast join failed for %pI6, status %d\n",
 448                                                 mcast->mcmember.mgid.raw, status);
 449                         } else {
 450                                 ipoib_warn(priv, "multicast join failed for %pI6, status %d\n",
 451                                            mcast->mcmember.mgid.raw, status);
 452                         }
 453                 }
 454
 455                 mcast->backoff *= 2;
 456                 if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
 457                         mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
 458         }
 459 out:
 460         spin_lock_irq(&priv->lock);
 461         clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 462         if (status)
 463                 mcast->mc = NULL;
 464         complete(&mcast->done);
 465         if (status == -ENETRESET)
 466                 status = 0;
 467         if (status && test_bit(IPOIB_MCAST_RUN, &priv->flags))
 468                 queue_delayed_work(ipoib_workqueue, &priv->mcast_task,
 469                                    mcast->backoff * HZ);
 470         spin_unlock_irq(&priv->lock);
 471         mutex_unlock(&mcast_mutex);
 472
 473         return status;
 474 }
 475
 476 static void ipoib_mcast_join(struct net_device *dev, struct ipoib_mcast *mcast,
 477                              int create)
 478 {
 479         struct ipoib_dev_priv *priv = netdev_priv(dev);
 480         struct ib_sa_mcmember_rec rec = {
 481                 .join_state = 1
 482         };
 483         ib_sa_comp_mask comp_mask;
 484         int ret = 0;
 485
 486         ipoib_dbg_mcast(priv, "joining MGID %pI6\n", mcast->mcmember.mgid.raw);
 487
 488         rec.mgid     = mcast->mcmember.mgid;
 489         rec.port_gid = priv->local_gid;
 490         rec.pkey     = cpu_to_be16(priv->pkey);
 491
 492         comp_mask =
 493                 IB_SA_MCMEMBER_REC_MGID         |
 494                 IB_SA_MCMEMBER_REC_PORT_GID     |
 495                 IB_SA_MCMEMBER_REC_PKEY         |
 496                 IB_SA_MCMEMBER_REC_JOIN_STATE;
 497
 498         if (create) {
 499                 comp_mask |=
 500                         IB_SA_MCMEMBER_REC_QKEY                 |
 501                         IB_SA_MCMEMBER_REC_MTU_SELECTOR         |
 502                         IB_SA_MCMEMBER_REC_MTU                  |
 503                         IB_SA_MCMEMBER_REC_TRAFFIC_CLASS        |
 504                         IB_SA_MCMEMBER_REC_RATE_SELECTOR        |
 505                         IB_SA_MCMEMBER_REC_RATE                 |
 506                         IB_SA_MCMEMBER_REC_SL                   |
 507                         IB_SA_MCMEMBER_REC_FLOW_LABEL           |
 508                         IB_SA_MCMEMBER_REC_HOP_LIMIT;
 509
 510                 rec.qkey          = priv->broadcast->mcmember.qkey;
 511                 rec.mtu_selector  = IB_SA_EQ;
 512                 rec.mtu           = priv->broadcast->mcmember.mtu;
 513                 rec.traffic_class = priv->broadcast->mcmember.traffic_class;
 514                 rec.rate_selector = IB_SA_EQ;
 515                 rec.rate          = priv->broadcast->mcmember.rate;
 516                 rec.sl            = priv->broadcast->mcmember.sl;
 517                 rec.flow_label    = priv->broadcast->mcmember.flow_label;
 518                 rec.hop_limit     = priv->broadcast->mcmember.hop_limit;
 519         }
 520
 521         mutex_lock(&mcast_mutex);
 522         init_completion(&mcast->done);
 523         set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 524         mcast->mc = ib_sa_join_multicast(&ipoib_sa_client, priv->ca, priv->port,
 525                                          &rec, comp_mask, GFP_KERNEL,
 526                                          ipoib_mcast_join_complete, mcast);
 527         if (IS_ERR(mcast->mc)) {
 528                 clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 529                 complete(&mcast->done);
 530                 ret = PTR_ERR(mcast->mc);
 531                 ipoib_warn(priv, "ib_sa_join_multicast failed, status %d\n", ret);
 532
 533                 mcast->backoff *= 2;
 534                 if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
 535                         mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
 536
 537                 if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 538                         queue_delayed_work(ipoib_workqueue,
 539                                            &priv->mcast_task,
 540                                            mcast->backoff * HZ);
 541         }
 542         mutex_unlock(&mcast_mutex);
 543 }
 544
 545 void ipoib_mcast_join_task(struct work_struct *work)
 546 {
 547         struct ipoib_dev_priv *priv =
 548                 container_of(work, struct ipoib_dev_priv, mcast_task.work);
 549         struct net_device *dev = priv->dev;
 550         struct ib_port_attr port_attr;
 551
 552         if (!test_bit(IPOIB_MCAST_RUN, &priv->flags))
 553                 return;
 554
 555         if (ib_query_port(priv->ca, priv->port, &port_attr) ||
 556             port_attr.state != IB_PORT_ACTIVE) {
 557                 ipoib_dbg(priv, "port state is not ACTIVE (state = %d) suspending join task\n",
 558                           port_attr.state);
 559                 return;
 560         }
 561         priv->local_lid = port_attr.lid;
 562
 563         if (ib_query_gid(priv->ca, priv->port, 0, &priv->local_gid))
 564                 ipoib_warn(priv, "ib_query_gid() failed\n");
 565         else
 566                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
 567
 568         if (!priv->broadcast) {
 569                 struct ipoib_mcast *broadcast;
 570
 571                 if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 572                         return;
 573
 574                 broadcast = ipoib_mcast_alloc(dev, 1);
 575                 if (!broadcast) {
 576                         ipoib_warn(priv, "failed to allocate broadcast group\n");
 577                         mutex_lock(&mcast_mutex);
 578                         if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 579                                 queue_delayed_work(ipoib_workqueue,
 580                                                    &priv->mcast_task, HZ);
 581                         mutex_unlock(&mcast_mutex);
 582                         return;
 583                 }
 584
 585                 spin_lock_irq(&priv->lock);
 586                 memcpy(broadcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
 587                        sizeof (union ib_gid));
 588                 priv->broadcast = broadcast;
 589
 590                 __ipoib_mcast_add(dev, priv->broadcast);
 591                 spin_unlock_irq(&priv->lock);
 592         }
 593
 594         if (!test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
 595                 if (IS_ERR_OR_NULL(priv->broadcast->mc) &&
 596                     !test_bit(IPOIB_MCAST_FLAG_BUSY, &priv->broadcast->flags))
 597                         ipoib_mcast_join(dev, priv->broadcast, 0);
 598                 return;
 599         }
 600
 601         while (1) {
 602                 struct ipoib_mcast *mcast = NULL;
 603
 604                 /*
 605                  * Need the mutex so our flags are consistent, need the
 606                  * priv->lock so we don't race with list removals in either
 607                  * mcast_dev_flush or mcast_restart_task
 608                  */
 609                 mutex_lock(&mcast_mutex);
 610                 spin_lock_irq(&priv->lock);
 611                 list_for_each_entry(mcast, &priv->multicast_list, list) {
 612                         if (IS_ERR_OR_NULL(mcast->mc) &&
 613                             !test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags) &&
 614                             !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 615                                 /* Found the next unjoined group */
 616                                 break;
 617                         }
 618                 }
 619                 spin_unlock_irq(&priv->lock);
 620                 mutex_unlock(&mcast_mutex);
 621
 622                 if (&mcast->list == &priv->multicast_list) {
 623                         /* All done */
 624                         break;
 625                 }
 626
 627                 if (test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags))
 628                         ipoib_mcast_sendonly_join(mcast);
 629                 else
 630                         ipoib_mcast_join(dev, mcast, 1);
 631                 return;
 632         }
 633
 634         ipoib_dbg_mcast(priv, "successfully joined all multicast groups\n");
 635
 636         clear_bit(IPOIB_MCAST_RUN, &priv->flags);
 637 }
 638
 639 int ipoib_mcast_start_thread(struct net_device *dev)
 640 {
 641         struct ipoib_dev_priv *priv = netdev_priv(dev);
 642
 643         ipoib_dbg_mcast(priv, "starting multicast thread\n");
 644
 645         mutex_lock(&mcast_mutex);
 646         if (!test_and_set_bit(IPOIB_MCAST_RUN, &priv->flags))
 647                 queue_delayed_work(ipoib_workqueue, &priv->mcast_task, 0);
 648         mutex_unlock(&mcast_mutex);
 649
 650         return 0;
 651 }
 652
 653 int ipoib_mcast_stop_thread(struct net_device *dev, int flush)
 654 {
 655         struct ipoib_dev_priv *priv = netdev_priv(dev);
 656
 657         ipoib_dbg_mcast(priv, "stopping multicast thread\n");
 658
 659         mutex_lock(&mcast_mutex);
 660         clear_bit(IPOIB_MCAST_RUN, &priv->flags);
 661         cancel_delayed_work(&priv->mcast_task);
 662         mutex_unlock(&mcast_mutex);
 663
 664         if (flush)
 665                 flush_workqueue(ipoib_workqueue);
 666
 667         return 0;
 668 }
 669
 670 static int ipoib_mcast_leave(struct net_device *dev, struct ipoib_mcast *mcast)
 671 {
 672         struct ipoib_dev_priv *priv = netdev_priv(dev);
 673         int ret = 0;
 674
 675         if (test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
 676                 ipoib_warn(priv, "ipoib_mcast_leave on an in-flight join\n");
 677
 678         if (!IS_ERR_OR_NULL(mcast->mc))
 679                 ib_sa_free_multicast(mcast->mc);
 680
 681         if (test_and_clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 682                 ipoib_dbg_mcast(priv, "leaving MGID %pI6\n",
 683                                 mcast->mcmember.mgid.raw);
 684
 685                 /* Remove ourselves from the multicast group */
 686                 ret = ib_detach_mcast(priv->qp, &mcast->mcmember.mgid,
 687                                       be16_to_cpu(mcast->mcmember.mlid));
 688                 if (ret)
 689                         ipoib_warn(priv, "ib_detach_mcast failed (result = %d)\n", ret);
 690         }
 691
 692         return 0;
 693 }
 694
 695 void ipoib_mcast_send(struct net_device *dev, u8 *daddr, struct sk_buff *skb)
 696 {
 697         struct ipoib_dev_priv *priv = netdev_priv(dev);
 698         struct ipoib_mcast *mcast;
 699         unsigned long flags;
 700         void *mgid = daddr + 4;
 701
 702         spin_lock_irqsave(&priv->lock, flags);
 703
 704         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)         ||
 705             !priv->broadcast                                    ||
 706             !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
 707                 ++dev->stats.tx_dropped;
 708                 dev_kfree_skb_any(skb);
 709                 goto unlock;
 710         }
 711
 712         mcast = __ipoib_mcast_find(dev, mgid);
 713         if (!mcast) {
 714                 /* Let's create a new send only group now */
 715                 ipoib_dbg_mcast(priv, "setting up send only multicast group for %pI6\n",
 716                                 mgid);
 717
 718                 mcast = ipoib_mcast_alloc(dev, 0);
 719                 if (!mcast) {
 720                         ipoib_warn(priv, "unable to allocate memory for "
 721                                    "multicast structure\n");
 722                         ++dev->stats.tx_dropped;
 723                         dev_kfree_skb_any(skb);
 724                         goto out;
 725                 }
 726
 727                 set_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags);
 728                 memcpy(mcast->mcmember.mgid.raw, mgid, sizeof (union ib_gid));
 729                 __ipoib_mcast_add(dev, mcast);
 730                 list_add_tail(&mcast->list, &priv->multicast_list);
 731                 if (!test_and_set_bit(IPOIB_MCAST_RUN, &priv->flags))
 732                         queue_delayed_work(ipoib_workqueue, &priv->mcast_task, 0);
 733         }
 734
 735         if (!mcast->ah) {
 736                 if (skb_queue_len(&mcast->pkt_queue) < IPOIB_MAX_MCAST_QUEUE)
 737                         skb_queue_tail(&mcast->pkt_queue, skb);
 738                 else {
 739                         ++dev->stats.tx_dropped;
 740                         dev_kfree_skb_any(skb);
 741                 }
 742
 743                 if (test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
 744                         ipoib_dbg_mcast(priv, "no address vector, "
 745                                         "but multicast join already started\n");
 746
 747                 /*
 748                  * If lookup completes between here and out:, don't
 749                  * want to send packet twice.
 750                  */
 751                 mcast = NULL;
 752         }
 753
 754 out:
 755         if (mcast && mcast->ah) {
 756                 struct ipoib_neigh *neigh;
 757
 758                 spin_unlock_irqrestore(&priv->lock, flags);
 759                 neigh = ipoib_neigh_get(dev, daddr);
 760                 spin_lock_irqsave(&priv->lock, flags);
 761                 if (!neigh) {
 762                         neigh = ipoib_neigh_alloc(daddr, dev);
 763                         if (neigh) {
 764                                 kref_get(&mcast->ah->ref);
 765                                 neigh->ah       = mcast->ah;
 766                                 list_add_tail(&neigh->list, &mcast->neigh_list);
 767                         }
 768                 }
 769                 spin_unlock_irqrestore(&priv->lock, flags);
 770                 ipoib_send(dev, skb, mcast->ah, IB_MULTICAST_QPN);
 771                 if (neigh)
 772                         ipoib_neigh_put(neigh);
 773                 return;
 774         }
 775
 776 unlock:
 777         spin_unlock_irqrestore(&priv->lock, flags);
 778 }
 779
 780 void ipoib_mcast_dev_flush(struct net_device *dev)
 781 {
 782         struct ipoib_dev_priv *priv = netdev_priv(dev);
 783         LIST_HEAD(remove_list);
 784         struct ipoib_mcast *mcast, *tmcast;
 785         unsigned long flags;
 786
 787         ipoib_dbg_mcast(priv, "flushing multicast list\n");
 788
 789         spin_lock_irqsave(&priv->lock, flags);
 790
 791         list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
 792                 list_del(&mcast->list);
 793                 rb_erase(&mcast->rb_node, &priv->multicast_tree);
 794                 list_add_tail(&mcast->list, &remove_list);
 795         }
 796
 797         if (priv->broadcast) {
 798                 rb_erase(&priv->broadcast->rb_node, &priv->multicast_tree);
 799                 list_add_tail(&priv->broadcast->list, &remove_list);
 800                 priv->broadcast = NULL;
 801         }
 802
 803         spin_unlock_irqrestore(&priv->lock, flags);
 804
 805         /* seperate between the wait to the leave*/
 806         list_for_each_entry_safe(mcast, tmcast, &remove_list, list)
 807                 if (test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
 808                         wait_for_completion(&mcast->done);
 809
 810         list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
 811                 ipoib_mcast_leave(dev, mcast);
 812                 ipoib_mcast_free(mcast);
 813         }
 814 }
 815
 816 static int ipoib_mcast_addr_is_valid(const u8 *addr, const u8 *broadcast)
 817 {
 818         /* reserved QPN, prefix, scope */
 819         if (memcmp(addr, broadcast, 6))
 820                 return 0;
 821         /* signature lower, pkey */
 822         if (memcmp(addr + 7, broadcast + 7, 3))
 823                 return 0;
 824         return 1;
 825 }
 826
 827 void ipoib_mcast_restart_task(struct work_struct *work)
 828 {
 829         struct ipoib_dev_priv *priv =
 830                 container_of(work, struct ipoib_dev_priv, restart_task);
 831         struct net_device *dev = priv->dev;
 832         struct netdev_hw_addr *ha;
 833         struct ipoib_mcast *mcast, *tmcast;
 834         LIST_HEAD(remove_list);
 835         unsigned long flags;
 836         struct ib_sa_mcmember_rec rec;
 837
 838         ipoib_dbg_mcast(priv, "restarting multicast task\n");
 839
 840         ipoib_mcast_stop_thread(dev, 0);
 841
 842         local_irq_save(flags);
 843         netif_addr_lock(dev);
 844         spin_lock(&priv->lock);
 845
 846         /*
 847          * Unfortunately, the networking core only gives us a list of all of
 848          * the multicast hardware addresses. We need to figure out which ones
 849          * are new and which ones have been removed
 850          */
 851
 852         /* Clear out the found flag */
 853         list_for_each_entry(mcast, &priv->multicast_list, list)
 854                 clear_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
 855
 856         /* Mark all of the entries that are found or don't exist */
 857         netdev_for_each_mc_addr(ha, dev) {
 858                 union ib_gid mgid;
 859
 860                 if (!ipoib_mcast_addr_is_valid(ha->addr, dev->broadcast))
 861                         continue;
 862
 863                 memcpy(mgid.raw, ha->addr + 4, sizeof mgid);
 864
 865                 mcast = __ipoib_mcast_find(dev, &mgid);
 866                 if (!mcast || test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 867                         struct ipoib_mcast *nmcast;
 868
 869                         /* ignore group which is directly joined by userspace */
 870                         if (test_bit(IPOIB_FLAG_UMCAST, &priv->flags) &&
 871                             !ib_sa_get_mcmember_rec(priv->ca, priv->port, &mgid, &rec)) {
 872                                 ipoib_dbg_mcast(priv, "ignoring multicast entry for mgid %pI6\n",
 873                                                 mgid.raw);
 874                                 continue;
 875                         }
 876
 877                         /* Not found or send-only group, let's add a new entry */
 878                         ipoib_dbg_mcast(priv, "adding multicast entry for mgid %pI6\n",
 879                                         mgid.raw);
 880
 881                         nmcast = ipoib_mcast_alloc(dev, 0);
 882                         if (!nmcast) {
 883                                 ipoib_warn(priv, "unable to allocate memory for multicast structure\n");
 884                                 continue;
 885                         }
 886
 887                         set_bit(IPOIB_MCAST_FLAG_FOUND, &nmcast->flags);
 888
 889                         nmcast->mcmember.mgid = mgid;
 890
 891                         if (mcast) {
 892                                 /* Destroy the send only entry */
 893                                 list_move_tail(&mcast->list, &remove_list);
 894
 895                                 rb_replace_node(&mcast->rb_node,
 896                                                 &nmcast->rb_node,
 897                                                 &priv->multicast_tree);
 898                         } else
 899                                 __ipoib_mcast_add(dev, nmcast);
 900
 901                         list_add_tail(&nmcast->list, &priv->multicast_list);
 902                 }
 903
 904                 if (mcast)
 905                         set_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
 906         }
 907
 908         /* Remove all of the entries don't exist anymore */
 909         list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
 910                 if (!test_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags) &&
 911                     !test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 912                         ipoib_dbg_mcast(priv, "deleting multicast group %pI6\n",
 913                                         mcast->mcmember.mgid.raw);
 914
 915                         rb_erase(&mcast->rb_node, &priv->multicast_tree);
 916
 917                         /* Move to the remove list */
 918                         list_move_tail(&mcast->list, &remove_list);
 919                 }
 920         }
 921
 922         spin_unlock(&priv->lock);
 923         netif_addr_unlock(dev);
 924         local_irq_restore(flags);
 925
 926         /* We have to cancel outside of the spinlock */
 927         list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
 928                 ipoib_mcast_leave(mcast->dev, mcast);
 929                 ipoib_mcast_free(mcast);
 930         }
 931
 932         if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 933                 ipoib_mcast_start_thread(dev);
 934 }
 935
 936 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 937
 938 struct ipoib_mcast_iter *ipoib_mcast_iter_init(struct net_device *dev)
 939 {
 940         struct ipoib_mcast_iter *iter;
 941
 942         iter = kmalloc(sizeof *iter, GFP_KERNEL);
 943         if (!iter)
 944                 return NULL;
 945
 946         iter->dev = dev;
 947         memset(iter->mgid.raw, 0, 16);
 948
 949         if (ipoib_mcast_iter_next(iter)) {
 950                 kfree(iter);
 951                 return NULL;
 952         }
 953
 954         return iter;
 955 }
 956
 957 int ipoib_mcast_iter_next(struct ipoib_mcast_iter *iter)
 958 {
 959         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
 960         struct rb_node *n;
 961         struct ipoib_mcast *mcast;
 962         int ret = 1;
 963
 964         spin_lock_irq(&priv->lock);
 965
 966         n = rb_first(&priv->multicast_tree);
 967
 968         while (n) {
 969                 mcast = rb_entry(n, struct ipoib_mcast, rb_node);
 970
 971                 if (memcmp(iter->mgid.raw, mcast->mcmember.mgid.raw,
 972                            sizeof (union ib_gid)) < 0) {
 973                         iter->mgid      = mcast->mcmember.mgid;
 974                         iter->created   = mcast->created;
 975                         iter->queuelen  = skb_queue_len(&mcast->pkt_queue);
 976                         iter->complete  = !!mcast->ah;
 977                         iter->send_only = !!(mcast->flags & (1 << IPOIB_MCAST_FLAG_SENDONLY));
 978
 979                         ret = 0;
 980
 981                         break;
 982                 }
 983
 984                 n = rb_next(n);
 985         }
 986
 987         spin_unlock_irq(&priv->lock);
 988
 989         return ret;
 990 }
 991
 992 void ipoib_mcast_iter_read(struct ipoib_mcast_iter *iter,
 993                            union ib_gid *mgid,
 994                            unsigned long *created,
 995                            unsigned int *queuelen,
 996                            unsigned int *complete,
 997                            unsigned int *send_only)
 998 {
 999         *mgid      = iter->mgid;
1000         *created   = iter->created;
1001         *queuelen  = iter->queuelen;
1002         *complete  = iter->complete;
1003         *send_only = iter->send_only;
1004 }
1005
1006 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */