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                 if (!ipoib_cm_admin_enabled(dev)) {
 195                         rtnl_lock();
 196                         dev_set_mtu(dev, min(priv->mcast_mtu, priv->admin_mtu));
 197                         rtnl_unlock();
 198                 }
 199         }
 200
 201         if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 202                 if (test_and_set_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 203                         ipoib_warn(priv, "multicast group %pI6 already attached\n",
 204                                    mcast->mcmember.mgid.raw);
 205
 206                         return 0;
 207                 }
 208
 209                 ret = ipoib_mcast_attach(dev, be16_to_cpu(mcast->mcmember.mlid),
 210                                          &mcast->mcmember.mgid, set_qkey);
 211                 if (ret < 0) {
 212                         ipoib_warn(priv, "couldn't attach QP to multicast group %pI6\n",
 213                                    mcast->mcmember.mgid.raw);
 214
 215                         clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags);
 216                         return ret;
 217                 }
 218         }
 219
 220         {
 221                 struct ib_ah_attr av = {
 222                         .dlid          = be16_to_cpu(mcast->mcmember.mlid),
 223                         .port_num      = priv->port,
 224                         .sl            = mcast->mcmember.sl,
 225                         .ah_flags      = IB_AH_GRH,
 226                         .static_rate   = mcast->mcmember.rate,
 227                         .grh           = {
 228                                 .flow_label    = be32_to_cpu(mcast->mcmember.flow_label),
 229                                 .hop_limit     = mcast->mcmember.hop_limit,
 230                                 .sgid_index    = 0,
 231                                 .traffic_class = mcast->mcmember.traffic_class
 232                         }
 233                 };
 234                 av.grh.dgid = mcast->mcmember.mgid;
 235
 236                 ah = ipoib_create_ah(dev, priv->pd, &av);
 237                 if (IS_ERR(ah)) {
 238                         ipoib_warn(priv, "ib_address_create failed %ld\n",
 239                                 -PTR_ERR(ah));
 240                         /* use original error */
 241                         return PTR_ERR(ah);
 242                 } else {
 243                         spin_lock_irq(&priv->lock);
 244                         mcast->ah = ah;
 245                         spin_unlock_irq(&priv->lock);
 246
 247                         ipoib_dbg_mcast(priv, "MGID %pI6 AV %p, LID 0x%04x, SL %d\n",
 248                                         mcast->mcmember.mgid.raw,
 249                                         mcast->ah->ah,
 250                                         be16_to_cpu(mcast->mcmember.mlid),
 251                                         mcast->mcmember.sl);
 252                 }
 253         }
 254
 255         /* actually send any queued packets */
 256         netif_tx_lock_bh(dev);
 257         while (!skb_queue_empty(&mcast->pkt_queue)) {
 258                 struct sk_buff *skb = skb_dequeue(&mcast->pkt_queue);
 259
 260                 netif_tx_unlock_bh(dev);
 261
 262                 skb->dev = dev;
 263                 if (dev_queue_xmit(skb))
 264                         ipoib_warn(priv, "dev_queue_xmit failed to requeue packet\n");
 265
 266                 netif_tx_lock_bh(dev);
 267         }
 268         netif_tx_unlock_bh(dev);
 269
 270         return 0;
 271 }
 272
 273 static int
 274 ipoib_mcast_sendonly_join_complete(int status,
 275                                    struct ib_sa_multicast *multicast)
 276 {
 277         struct ipoib_mcast *mcast = multicast->context;
 278         struct net_device *dev = mcast->dev;
 279
 280         /* We trap for port events ourselves. */
 281         if (status == -ENETRESET)
 282                 return 0;
 283
 284         if (!status)
 285                 status = ipoib_mcast_join_finish(mcast, &multicast->rec);
 286
 287         if (status) {
 288                 if (mcast->logcount++ < 20)
 289                         ipoib_dbg_mcast(netdev_priv(dev), "multicast join failed for %pI6, status %d\n",
 290                                         mcast->mcmember.mgid.raw, status);
 291
 292                 /* Flush out any queued packets */
 293                 netif_tx_lock_bh(dev);
 294                 while (!skb_queue_empty(&mcast->pkt_queue)) {
 295                         ++dev->stats.tx_dropped;
 296                         dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
 297                 }
 298                 netif_tx_unlock_bh(dev);
 299
 300                 /* Clear the busy flag so we try again */
 301                 status = test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY,
 302                                             &mcast->flags);
 303         }
 304         return status;
 305 }
 306
 307 static int ipoib_mcast_sendonly_join(struct ipoib_mcast *mcast)
 308 {
 309         struct net_device *dev = mcast->dev;
 310         struct ipoib_dev_priv *priv = netdev_priv(dev);
 311         struct ib_sa_mcmember_rec rec = {
 312 #if 0                           /* Some SMs don't support send-only yet */
 313                 .join_state = 4
 314 #else
 315                 .join_state = 1
 316 #endif
 317         };
 318         int ret = 0;
 319
 320         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
 321                 ipoib_dbg_mcast(priv, "device shutting down, no multicast joins\n");
 322                 return -ENODEV;
 323         }
 324
 325         if (test_and_set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)) {
 326                 ipoib_dbg_mcast(priv, "multicast entry busy, skipping\n");
 327                 return -EBUSY;
 328         }
 329
 330         rec.mgid     = mcast->mcmember.mgid;
 331         rec.port_gid = priv->local_gid;
 332         rec.pkey     = cpu_to_be16(priv->pkey);
 333
 334         mcast->mc = ib_sa_join_multicast(&ipoib_sa_client, priv->ca,
 335                                          priv->port, &rec,
 336                                          IB_SA_MCMEMBER_REC_MGID        |
 337                                          IB_SA_MCMEMBER_REC_PORT_GID    |
 338                                          IB_SA_MCMEMBER_REC_PKEY        |
 339                                          IB_SA_MCMEMBER_REC_JOIN_STATE,
 340                                          GFP_ATOMIC,
 341                                          ipoib_mcast_sendonly_join_complete,
 342                                          mcast);
 343         if (IS_ERR(mcast->mc)) {
 344                 ret = PTR_ERR(mcast->mc);
 345                 clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 346                 ipoib_warn(priv, "ib_sa_join_multicast failed (ret = %d)\n",
 347                            ret);
 348         } else {
 349                 ipoib_dbg_mcast(priv, "no multicast record for %pI6, starting join\n",
 350                                 mcast->mcmember.mgid.raw);
 351         }
 352
 353         return ret;
 354 }
 355
 356 void ipoib_mcast_carrier_on_task(struct work_struct *work)
 357 {
 358         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
 359                                                    carrier_on_task);
 360         struct ib_port_attr attr;
 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.
 366          */
 367         if (ib_query_port(priv->ca, priv->port, &attr) ||
 368             attr.state != IB_PORT_ACTIVE) {
 369                 ipoib_dbg(priv, "Keeping carrier off until IB port is active\n");
 370                 return;
 371         }
 372
 373         rtnl_lock();
 374         netif_carrier_on(priv->dev);
 375         rtnl_unlock();
 376 }
 377
 378 static int ipoib_mcast_join_complete(int status,
 379                                      struct ib_sa_multicast *multicast)
 380 {
 381         struct ipoib_mcast *mcast = multicast->context;
 382         struct net_device *dev = mcast->dev;
 383         struct ipoib_dev_priv *priv = netdev_priv(dev);
 384
 385         ipoib_dbg_mcast(priv, "join completion for %pI6 (status %d)\n",
 386                         mcast->mcmember.mgid.raw, status);
 387
 388         /* We trap for port events ourselves. */
 389         if (status == -ENETRESET)
 390                 return 0;
 391
 392         if (!status)
 393                 status = ipoib_mcast_join_finish(mcast, &multicast->rec);
 394
 395         if (!status) {
 396                 mcast->backoff = 1;
 397                 mutex_lock(&mcast_mutex);
 398                 if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 399                         queue_delayed_work(ipoib_workqueue,
 400                                            &priv->mcast_task, 0);
 401                 mutex_unlock(&mcast_mutex);
 402
 403                 /*
 404                  * Defer carrier on work to ipoib_workqueue to avoid a
 405                  * deadlock on rtnl_lock here.
 406                  */
 407                 if (mcast == priv->broadcast)
 408                         queue_work(ipoib_workqueue, &priv->carrier_on_task);
 409
 410                 return 0;
 411         }
 412
 413         if (mcast->logcount++ < 20) {
 414                 if (status == -ETIMEDOUT || status == -EAGAIN) {
 415                         ipoib_dbg_mcast(priv, "multicast join failed for %pI6, status %d\n",
 416                                         mcast->mcmember.mgid.raw, status);
 417                 } else {
 418                         ipoib_warn(priv, "multicast join failed for %pI6, status %d\n",
 419                                    mcast->mcmember.mgid.raw, status);
 420                 }
 421         }
 422
 423         mcast->backoff *= 2;
 424         if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
 425                 mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
 426
 427         /* Clear the busy flag so we try again */
 428         status = test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 429
 430         mutex_lock(&mcast_mutex);
 431         spin_lock_irq(&priv->lock);
 432         if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 433                 queue_delayed_work(ipoib_workqueue, &priv->mcast_task,
 434                                    mcast->backoff * HZ);
 435         spin_unlock_irq(&priv->lock);
 436         mutex_unlock(&mcast_mutex);
 437
 438         return status;
 439 }
 440
 441 static void ipoib_mcast_join(struct net_device *dev, struct ipoib_mcast *mcast,
 442                              int create)
 443 {
 444         struct ipoib_dev_priv *priv = netdev_priv(dev);
 445         struct ib_sa_mcmember_rec rec = {
 446                 .join_state = 1
 447         };
 448         ib_sa_comp_mask comp_mask;
 449         int ret = 0;
 450
 451         ipoib_dbg_mcast(priv, "joining MGID %pI6\n", mcast->mcmember.mgid.raw);
 452
 453         rec.mgid     = mcast->mcmember.mgid;
 454         rec.port_gid = priv->local_gid;
 455         rec.pkey     = cpu_to_be16(priv->pkey);
 456
 457         comp_mask =
 458                 IB_SA_MCMEMBER_REC_MGID         |
 459                 IB_SA_MCMEMBER_REC_PORT_GID     |
 460                 IB_SA_MCMEMBER_REC_PKEY         |
 461                 IB_SA_MCMEMBER_REC_JOIN_STATE;
 462
 463         if (create) {
 464                 comp_mask |=
 465                         IB_SA_MCMEMBER_REC_QKEY                 |
 466                         IB_SA_MCMEMBER_REC_MTU_SELECTOR         |
 467                         IB_SA_MCMEMBER_REC_MTU                  |
 468                         IB_SA_MCMEMBER_REC_TRAFFIC_CLASS        |
 469                         IB_SA_MCMEMBER_REC_RATE_SELECTOR        |
 470                         IB_SA_MCMEMBER_REC_RATE                 |
 471                         IB_SA_MCMEMBER_REC_SL                   |
 472                         IB_SA_MCMEMBER_REC_FLOW_LABEL           |
 473                         IB_SA_MCMEMBER_REC_HOP_LIMIT;
 474
 475                 rec.qkey          = priv->broadcast->mcmember.qkey;
 476                 rec.mtu_selector  = IB_SA_EQ;
 477                 rec.mtu           = priv->broadcast->mcmember.mtu;
 478                 rec.traffic_class = priv->broadcast->mcmember.traffic_class;
 479                 rec.rate_selector = IB_SA_EQ;
 480                 rec.rate          = priv->broadcast->mcmember.rate;
 481                 rec.sl            = priv->broadcast->mcmember.sl;
 482                 rec.flow_label    = priv->broadcast->mcmember.flow_label;
 483                 rec.hop_limit     = priv->broadcast->mcmember.hop_limit;
 484         }
 485
 486         set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 487         mcast->mc = ib_sa_join_multicast(&ipoib_sa_client, priv->ca, priv->port,
 488                                          &rec, comp_mask, GFP_KERNEL,
 489                                          ipoib_mcast_join_complete, mcast);
 490         if (IS_ERR(mcast->mc)) {
 491                 clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 492                 ret = PTR_ERR(mcast->mc);
 493                 ipoib_warn(priv, "ib_sa_join_multicast failed, status %d\n", ret);
 494
 495                 mcast->backoff *= 2;
 496                 if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
 497                         mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
 498
 499                 mutex_lock(&mcast_mutex);
 500                 if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 501                         queue_delayed_work(ipoib_workqueue,
 502                                            &priv->mcast_task,
 503                                            mcast->backoff * HZ);
 504                 mutex_unlock(&mcast_mutex);
 505         }
 506 }
 507
 508 void ipoib_mcast_join_task(struct work_struct *work)
 509 {
 510         struct ipoib_dev_priv *priv =
 511                 container_of(work, struct ipoib_dev_priv, mcast_task.work);
 512         struct net_device *dev = priv->dev;
 513
 514         if (!test_bit(IPOIB_MCAST_RUN, &priv->flags))
 515                 return;
 516
 517         if (ib_query_gid(priv->ca, priv->port, 0, &priv->local_gid))
 518                 ipoib_warn(priv, "ib_query_gid() failed\n");
 519         else
 520                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
 521
 522         {
 523                 struct ib_port_attr attr;
 524
 525                 if (!ib_query_port(priv->ca, priv->port, &attr))
 526                         priv->local_lid = attr.lid;
 527                 else
 528                         ipoib_warn(priv, "ib_query_port failed\n");
 529         }
 530
 531         if (!priv->broadcast) {
 532                 struct ipoib_mcast *broadcast;
 533
 534                 if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 535                         return;
 536
 537                 broadcast = ipoib_mcast_alloc(dev, 1);
 538                 if (!broadcast) {
 539                         ipoib_warn(priv, "failed to allocate broadcast group\n");
 540                         mutex_lock(&mcast_mutex);
 541                         if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 542                                 queue_delayed_work(ipoib_workqueue,
 543                                                    &priv->mcast_task, HZ);
 544                         mutex_unlock(&mcast_mutex);
 545                         return;
 546                 }
 547
 548                 spin_lock_irq(&priv->lock);
 549                 memcpy(broadcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
 550                        sizeof (union ib_gid));
 551                 priv->broadcast = broadcast;
 552
 553                 __ipoib_mcast_add(dev, priv->broadcast);
 554                 spin_unlock_irq(&priv->lock);
 555         }
 556
 557         if (!test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
 558                 if (!test_bit(IPOIB_MCAST_FLAG_BUSY, &priv->broadcast->flags))
 559                         ipoib_mcast_join(dev, priv->broadcast, 0);
 560                 return;
 561         }
 562
 563         while (1) {
 564                 struct ipoib_mcast *mcast = NULL;
 565
 566                 spin_lock_irq(&priv->lock);
 567                 list_for_each_entry(mcast, &priv->multicast_list, list) {
 568                         if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)
 569                             && !test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)
 570                             && !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 571                                 /* Found the next unjoined group */
 572                                 break;
 573                         }
 574                 }
 575                 spin_unlock_irq(&priv->lock);
 576
 577                 if (&mcast->list == &priv->multicast_list) {
 578                         /* All done */
 579                         break;
 580                 }
 581
 582                 ipoib_mcast_join(dev, mcast, 1);
 583                 return;
 584         }
 585
 586         ipoib_dbg_mcast(priv, "successfully joined all multicast groups\n");
 587
 588         clear_bit(IPOIB_MCAST_RUN, &priv->flags);
 589 }
 590
 591 int ipoib_mcast_start_thread(struct net_device *dev)
 592 {
 593         struct ipoib_dev_priv *priv = netdev_priv(dev);
 594
 595         ipoib_dbg_mcast(priv, "starting multicast thread\n");
 596
 597         mutex_lock(&mcast_mutex);
 598         if (!test_and_set_bit(IPOIB_MCAST_RUN, &priv->flags))
 599                 queue_delayed_work(ipoib_workqueue, &priv->mcast_task, 0);
 600         mutex_unlock(&mcast_mutex);
 601
 602         return 0;
 603 }
 604
 605 int ipoib_mcast_stop_thread(struct net_device *dev, int flush)
 606 {
 607         struct ipoib_dev_priv *priv = netdev_priv(dev);
 608
 609         ipoib_dbg_mcast(priv, "stopping multicast thread\n");
 610
 611         mutex_lock(&mcast_mutex);
 612         clear_bit(IPOIB_MCAST_RUN, &priv->flags);
 613         cancel_delayed_work(&priv->mcast_task);
 614         mutex_unlock(&mcast_mutex);
 615
 616         if (flush)
 617                 flush_workqueue(ipoib_workqueue);
 618
 619         return 0;
 620 }
 621
 622 static int ipoib_mcast_leave(struct net_device *dev, struct ipoib_mcast *mcast)
 623 {
 624         struct ipoib_dev_priv *priv = netdev_priv(dev);
 625         int ret = 0;
 626
 627         if (test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
 628                 ib_sa_free_multicast(mcast->mc);
 629
 630         if (test_and_clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 631                 ipoib_dbg_mcast(priv, "leaving MGID %pI6\n",
 632                                 mcast->mcmember.mgid.raw);
 633
 634                 /* Remove ourselves from the multicast group */
 635                 ret = ib_detach_mcast(priv->qp, &mcast->mcmember.mgid,
 636                                       be16_to_cpu(mcast->mcmember.mlid));
 637                 if (ret)
 638                         ipoib_warn(priv, "ib_detach_mcast failed (result = %d)\n", ret);
 639         }
 640
 641         return 0;
 642 }
 643
 644 void ipoib_mcast_send(struct net_device *dev, u8 *daddr, struct sk_buff *skb)
 645 {
 646         struct ipoib_dev_priv *priv = netdev_priv(dev);
 647         struct ipoib_mcast *mcast;
 648         unsigned long flags;
 649         void *mgid = daddr + 4;
 650
 651         spin_lock_irqsave(&priv->lock, flags);
 652
 653         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)         ||
 654             !priv->broadcast                                    ||
 655             !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
 656                 ++dev->stats.tx_dropped;
 657                 dev_kfree_skb_any(skb);
 658                 goto unlock;
 659         }
 660
 661         mcast = __ipoib_mcast_find(dev, mgid);
 662         if (!mcast) {
 663                 /* Let's create a new send only group now */
 664                 ipoib_dbg_mcast(priv, "setting up send only multicast group for %pI6\n",
 665                                 mgid);
 666
 667                 mcast = ipoib_mcast_alloc(dev, 0);
 668                 if (!mcast) {
 669                         ipoib_warn(priv, "unable to allocate memory for "
 670                                    "multicast structure\n");
 671                         ++dev->stats.tx_dropped;
 672                         dev_kfree_skb_any(skb);
 673                         goto out;
 674                 }
 675
 676                 set_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags);
 677                 memcpy(mcast->mcmember.mgid.raw, mgid, sizeof (union ib_gid));
 678                 __ipoib_mcast_add(dev, mcast);
 679                 list_add_tail(&mcast->list, &priv->multicast_list);
 680         }
 681
 682         if (!mcast->ah) {
 683                 if (skb_queue_len(&mcast->pkt_queue) < IPOIB_MAX_MCAST_QUEUE)
 684                         skb_queue_tail(&mcast->pkt_queue, skb);
 685                 else {
 686                         ++dev->stats.tx_dropped;
 687                         dev_kfree_skb_any(skb);
 688                 }
 689
 690                 if (test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
 691                         ipoib_dbg_mcast(priv, "no address vector, "
 692                                         "but multicast join already started\n");
 693                 else if (test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags))
 694                         ipoib_mcast_sendonly_join(mcast);
 695
 696                 /*
 697                  * If lookup completes between here and out:, don't
 698                  * want to send packet twice.
 699                  */
 700                 mcast = NULL;
 701         }
 702
 703 out:
 704         if (mcast && mcast->ah) {
 705                 struct ipoib_neigh *neigh;
 706
 707                 spin_unlock_irqrestore(&priv->lock, flags);
 708                 neigh = ipoib_neigh_get(dev, daddr);
 709                 spin_lock_irqsave(&priv->lock, flags);
 710                 if (!neigh) {
 711                         neigh = ipoib_neigh_alloc(daddr, dev);
 712                         if (neigh) {
 713                                 kref_get(&mcast->ah->ref);
 714                                 neigh->ah       = mcast->ah;
 715                                 list_add_tail(&neigh->list, &mcast->neigh_list);
 716                         }
 717                 }
 718                 spin_unlock_irqrestore(&priv->lock, flags);
 719                 ipoib_send(dev, skb, mcast->ah, IB_MULTICAST_QPN);
 720                 if (neigh)
 721                         ipoib_neigh_put(neigh);
 722                 return;
 723         }
 724
 725 unlock:
 726         spin_unlock_irqrestore(&priv->lock, flags);
 727 }
 728
 729 void ipoib_mcast_dev_flush(struct net_device *dev)
 730 {
 731         struct ipoib_dev_priv *priv = netdev_priv(dev);
 732         LIST_HEAD(remove_list);
 733         struct ipoib_mcast *mcast, *tmcast;
 734         unsigned long flags;
 735
 736         ipoib_dbg_mcast(priv, "flushing multicast list\n");
 737
 738         spin_lock_irqsave(&priv->lock, flags);
 739
 740         list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
 741                 list_del(&mcast->list);
 742                 rb_erase(&mcast->rb_node, &priv->multicast_tree);
 743                 list_add_tail(&mcast->list, &remove_list);
 744         }
 745
 746         if (priv->broadcast) {
 747                 rb_erase(&priv->broadcast->rb_node, &priv->multicast_tree);
 748                 list_add_tail(&priv->broadcast->list, &remove_list);
 749                 priv->broadcast = NULL;
 750         }
 751
 752         spin_unlock_irqrestore(&priv->lock, flags);
 753
 754         list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
 755                 ipoib_mcast_leave(dev, mcast);
 756                 ipoib_mcast_free(mcast);
 757         }
 758 }
 759
 760 static int ipoib_mcast_addr_is_valid(const u8 *addr, const u8 *broadcast)
 761 {
 762         /* reserved QPN, prefix, scope */
 763         if (memcmp(addr, broadcast, 6))
 764                 return 0;
 765         /* signature lower, pkey */
 766         if (memcmp(addr + 7, broadcast + 7, 3))
 767                 return 0;
 768         return 1;
 769 }
 770
 771 void ipoib_mcast_restart_task(struct work_struct *work)
 772 {
 773         struct ipoib_dev_priv *priv =
 774                 container_of(work, struct ipoib_dev_priv, restart_task);
 775         struct net_device *dev = priv->dev;
 776         struct netdev_hw_addr *ha;
 777         struct ipoib_mcast *mcast, *tmcast;
 778         LIST_HEAD(remove_list);
 779         unsigned long flags;
 780         struct ib_sa_mcmember_rec rec;
 781
 782         ipoib_dbg_mcast(priv, "restarting multicast task\n");
 783
 784         ipoib_mcast_stop_thread(dev, 0);
 785
 786         local_irq_save(flags);
 787         netif_addr_lock(dev);
 788         spin_lock(&priv->lock);
 789
 790         /*
 791          * Unfortunately, the networking core only gives us a list of all of
 792          * the multicast hardware addresses. We need to figure out which ones
 793          * are new and which ones have been removed
 794          */
 795
 796         /* Clear out the found flag */
 797         list_for_each_entry(mcast, &priv->multicast_list, list)
 798                 clear_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
 799
 800         /* Mark all of the entries that are found or don't exist */
 801         netdev_for_each_mc_addr(ha, dev) {
 802                 union ib_gid mgid;
 803
 804                 if (!ipoib_mcast_addr_is_valid(ha->addr, dev->broadcast))
 805                         continue;
 806
 807                 memcpy(mgid.raw, ha->addr + 4, sizeof mgid);
 808
 809                 mcast = __ipoib_mcast_find(dev, &mgid);
 810                 if (!mcast || test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 811                         struct ipoib_mcast *nmcast;
 812
 813                         /* ignore group which is directly joined by userspace */
 814                         if (test_bit(IPOIB_FLAG_UMCAST, &priv->flags) &&
 815                             !ib_sa_get_mcmember_rec(priv->ca, priv->port, &mgid, &rec)) {
 816                                 ipoib_dbg_mcast(priv, "ignoring multicast entry for mgid %pI6\n",
 817                                                 mgid.raw);
 818                                 continue;
 819                         }
 820
 821                         /* Not found or send-only group, let's add a new entry */
 822                         ipoib_dbg_mcast(priv, "adding multicast entry for mgid %pI6\n",
 823                                         mgid.raw);
 824
 825                         nmcast = ipoib_mcast_alloc(dev, 0);
 826                         if (!nmcast) {
 827                                 ipoib_warn(priv, "unable to allocate memory for multicast structure\n");
 828                                 continue;
 829                         }
 830
 831                         set_bit(IPOIB_MCAST_FLAG_FOUND, &nmcast->flags);
 832
 833                         nmcast->mcmember.mgid = mgid;
 834
 835                         if (mcast) {
 836                                 /* Destroy the send only entry */
 837                                 list_move_tail(&mcast->list, &remove_list);
 838
 839                                 rb_replace_node(&mcast->rb_node,
 840                                                 &nmcast->rb_node,
 841                                                 &priv->multicast_tree);
 842                         } else
 843                                 __ipoib_mcast_add(dev, nmcast);
 844
 845                         list_add_tail(&nmcast->list, &priv->multicast_list);
 846                 }
 847
 848                 if (mcast)
 849                         set_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
 850         }
 851
 852         /* Remove all of the entries don't exist anymore */
 853         list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
 854                 if (!test_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags) &&
 855                     !test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 856                         ipoib_dbg_mcast(priv, "deleting multicast group %pI6\n",
 857                                         mcast->mcmember.mgid.raw);
 858
 859                         rb_erase(&mcast->rb_node, &priv->multicast_tree);
 860
 861                         /* Move to the remove list */
 862                         list_move_tail(&mcast->list, &remove_list);
 863                 }
 864         }
 865
 866         spin_unlock(&priv->lock);
 867         netif_addr_unlock(dev);
 868         local_irq_restore(flags);
 869
 870         /* We have to cancel outside of the spinlock */
 871         list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
 872                 ipoib_mcast_leave(mcast->dev, mcast);
 873                 ipoib_mcast_free(mcast);
 874         }
 875
 876         if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 877                 ipoib_mcast_start_thread(dev);
 878 }
 879
 880 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 881
 882 struct ipoib_mcast_iter *ipoib_mcast_iter_init(struct net_device *dev)
 883 {
 884         struct ipoib_mcast_iter *iter;
 885
 886         iter = kmalloc(sizeof *iter, GFP_KERNEL);
 887         if (!iter)
 888                 return NULL;
 889
 890         iter->dev = dev;
 891         memset(iter->mgid.raw, 0, 16);
 892
 893         if (ipoib_mcast_iter_next(iter)) {
 894                 kfree(iter);
 895                 return NULL;
 896         }
 897
 898         return iter;
 899 }
 900
 901 int ipoib_mcast_iter_next(struct ipoib_mcast_iter *iter)
 902 {
 903         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
 904         struct rb_node *n;
 905         struct ipoib_mcast *mcast;
 906         int ret = 1;
 907
 908         spin_lock_irq(&priv->lock);
 909
 910         n = rb_first(&priv->multicast_tree);
 911
 912         while (n) {
 913                 mcast = rb_entry(n, struct ipoib_mcast, rb_node);
 914
 915                 if (memcmp(iter->mgid.raw, mcast->mcmember.mgid.raw,
 916                            sizeof (union ib_gid)) < 0) {
 917                         iter->mgid      = mcast->mcmember.mgid;
 918                         iter->created   = mcast->created;
 919                         iter->queuelen  = skb_queue_len(&mcast->pkt_queue);
 920                         iter->complete  = !!mcast->ah;
 921                         iter->send_only = !!(mcast->flags & (1 << IPOIB_MCAST_FLAG_SENDONLY));
 922
 923                         ret = 0;
 924
 925                         break;
 926                 }
 927
 928                 n = rb_next(n);
 929         }
 930
 931         spin_unlock_irq(&priv->lock);
 932
 933         return ret;
 934 }
 935
 936 void ipoib_mcast_iter_read(struct ipoib_mcast_iter *iter,
 937                            union ib_gid *mgid,
 938                            unsigned long *created,
 939                            unsigned int *queuelen,
 940                            unsigned int *complete,
 941                            unsigned int *send_only)
 942 {
 943         *mgid      = iter->mgid;
 944         *created   = iter->created;
 945         *queuelen  = iter->queuelen;
 946         *complete  = iter->complete;
 947         *send_only = iter->send_only;
 948 }
 949
 950 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */