drivers/infiniband/ulp/ipoib/ipoib_multicast.c

   1 /*
   2  * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
   3  *
   4  * This software is available to you under a choice of one of two
   5  * licenses.  You may choose to be licensed under the terms of the GNU
   6  * General Public License (GPL) Version 2, available from the file
   7  * COPYING in the main directory of this source tree, or the
   8  * OpenIB.org BSD license below:
   9  *
  10  *     Redistribution and use in source and binary forms, with or
  11  *     without modification, are permitted provided that the following
  12  *     conditions are met:
  13  *
  14  *      - Redistributions of source code must retain the above
  15  *        copyright notice, this list of conditions and the following
  16  *        disclaimer.
  17  *
  18  *      - Redistributions in binary form must reproduce the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer in the documentation and/or other materials
  21  *        provided with the distribution.
  22  *
  23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30  * SOFTWARE.
  31  *
  32  * $Id: ipoib_multicast.c 1362 2004-12-18 15:56:29Z roland $
  33  */
  34
  35 #include <linux/skbuff.h>
  36 #include <linux/rtnetlink.h>
  37 #include <linux/ip.h>
  38 #include <linux/in.h>
  39 #include <linux/igmp.h>
  40 #include <linux/inetdevice.h>
  41 #include <linux/delay.h>
  42 #include <linux/completion.h>
  43
  44 #include "ipoib.h"
  45
  46 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
  47 static int mcast_debug_level;
  48
  49 module_param(mcast_debug_level, int, 0644);
  50 MODULE_PARM_DESC(mcast_debug_level,
  51                  "Enable multicast debug tracing if > 0");
  52 #endif
  53
  54 static DECLARE_MUTEX(mcast_mutex);
  55
  56 /* Used for all multicast joins (broadcast, IPv4 mcast and IPv6 mcast) */
  57 struct ipoib_mcast {
  58         struct ib_sa_mcmember_rec mcmember;
  59         struct ipoib_ah          *ah;
  60
  61         struct rb_node    rb_node;
  62         struct list_head  list;
  63         struct completion done;
  64
  65         int                 query_id;
  66         struct ib_sa_query *query;
  67
  68         unsigned long created;
  69         unsigned long backoff;
  70
  71         unsigned long flags;
  72         unsigned char logcount;
  73
  74         struct list_head  neigh_list;
  75
  76         struct sk_buff_head pkt_queue;
  77
  78         struct net_device *dev;
  79 };
  80
  81 struct ipoib_mcast_iter {
  82         struct net_device *dev;
  83         union ib_gid       mgid;
  84         unsigned long      created;
  85         unsigned int       queuelen;
  86         unsigned int       complete;
  87         unsigned int       send_only;
  88 };
  89
  90 static void ipoib_mcast_free(struct ipoib_mcast *mcast)
  91 {
  92         struct net_device *dev = mcast->dev;
  93         struct ipoib_dev_priv *priv = netdev_priv(dev);
  94         struct ipoib_neigh *neigh, *tmp;
  95         unsigned long flags;
  96
  97         ipoib_dbg_mcast(netdev_priv(dev),
  98                         "deleting multicast group " IPOIB_GID_FMT "\n",
  99                         IPOIB_GID_ARG(mcast->mcmember.mgid));
 100
 101         spin_lock_irqsave(&priv->lock, flags);
 102
 103         list_for_each_entry_safe(neigh, tmp, &mcast->neigh_list, list) {
 104                 ipoib_put_ah(neigh->ah);
 105                 *to_ipoib_neigh(neigh->neighbour) = NULL;
 106                 neigh->neighbour->ops->destructor = NULL;
 107                 kfree(neigh);
 108         }
 109
 110         spin_unlock_irqrestore(&priv->lock, flags);
 111
 112         if (mcast->ah)
 113                 ipoib_put_ah(mcast->ah);
 114
 115         while (!skb_queue_empty(&mcast->pkt_queue)) {
 116                 struct sk_buff *skb = skb_dequeue(&mcast->pkt_queue);
 117
 118                 skb->dev = dev;
 119                 dev_kfree_skb_any(skb);
 120         }
 121
 122         kfree(mcast);
 123 }
 124
 125 static struct ipoib_mcast *ipoib_mcast_alloc(struct net_device *dev,
 126                                              int can_sleep)
 127 {
 128         struct ipoib_mcast *mcast;
 129
 130         mcast = kmalloc(sizeof (*mcast), can_sleep ? GFP_KERNEL : GFP_ATOMIC);
 131         if (!mcast)
 132                 return NULL;
 133
 134         memset(mcast, 0, sizeof (*mcast));
 135
 136         init_completion(&mcast->done);
 137
 138         mcast->dev = dev;
 139         mcast->created = jiffies;
 140         mcast->backoff = HZ;
 141         mcast->logcount = 0;
 142
 143         INIT_LIST_HEAD(&mcast->list);
 144         INIT_LIST_HEAD(&mcast->neigh_list);
 145         skb_queue_head_init(&mcast->pkt_queue);
 146
 147         mcast->ah    = NULL;
 148         mcast->query = NULL;
 149
 150         return mcast;
 151 }
 152
 153 static struct ipoib_mcast *__ipoib_mcast_find(struct net_device *dev, union ib_gid *mgid)
 154 {
 155         struct ipoib_dev_priv *priv = netdev_priv(dev);
 156         struct rb_node *n = priv->multicast_tree.rb_node;
 157
 158         while (n) {
 159                 struct ipoib_mcast *mcast;
 160                 int ret;
 161
 162                 mcast = rb_entry(n, struct ipoib_mcast, rb_node);
 163
 164                 ret = memcmp(mgid->raw, mcast->mcmember.mgid.raw,
 165                              sizeof (union ib_gid));
 166                 if (ret < 0)
 167                         n = n->rb_left;
 168                 else if (ret > 0)
 169                         n = n->rb_right;
 170                 else
 171                         return mcast;
 172         }
 173
 174         return NULL;
 175 }
 176
 177 static int __ipoib_mcast_add(struct net_device *dev, struct ipoib_mcast *mcast)
 178 {
 179         struct ipoib_dev_priv *priv = netdev_priv(dev);
 180         struct rb_node **n = &priv->multicast_tree.rb_node, *pn = NULL;
 181
 182         while (*n) {
 183                 struct ipoib_mcast *tmcast;
 184                 int ret;
 185
 186                 pn = *n;
 187                 tmcast = rb_entry(pn, struct ipoib_mcast, rb_node);
 188
 189                 ret = memcmp(mcast->mcmember.mgid.raw, tmcast->mcmember.mgid.raw,
 190                              sizeof (union ib_gid));
 191                 if (ret < 0)
 192                         n = &pn->rb_left;
 193                 else if (ret > 0)
 194                         n = &pn->rb_right;
 195                 else
 196                         return -EEXIST;
 197         }
 198
 199         rb_link_node(&mcast->rb_node, pn, n);
 200         rb_insert_color(&mcast->rb_node, &priv->multicast_tree);
 201
 202         return 0;
 203 }
 204
 205 static int ipoib_mcast_join_finish(struct ipoib_mcast *mcast,
 206                                    struct ib_sa_mcmember_rec *mcmember)
 207 {
 208         struct net_device *dev = mcast->dev;
 209         struct ipoib_dev_priv *priv = netdev_priv(dev);
 210         int ret;
 211
 212         mcast->mcmember = *mcmember;
 213
 214         /* Set the cached Q_Key before we attach if it's the broadcast group */
 215         if (!memcmp(mcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
 216                     sizeof (union ib_gid))) {
 217                 priv->qkey = be32_to_cpu(priv->broadcast->mcmember.qkey);
 218                 priv->tx_wr.wr.ud.remote_qkey = priv->qkey;
 219         }
 220
 221         if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 222                 if (test_and_set_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 223                         ipoib_warn(priv, "multicast group " IPOIB_GID_FMT
 224                                    " already attached\n",
 225                                    IPOIB_GID_ARG(mcast->mcmember.mgid));
 226
 227                         return 0;
 228                 }
 229
 230                 ret = ipoib_mcast_attach(dev, be16_to_cpu(mcast->mcmember.mlid),
 231                                          &mcast->mcmember.mgid);
 232                 if (ret < 0) {
 233                         ipoib_warn(priv, "couldn't attach QP to multicast group "
 234                                    IPOIB_GID_FMT "\n",
 235                                    IPOIB_GID_ARG(mcast->mcmember.mgid));
 236
 237                         clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags);
 238                         return ret;
 239                 }
 240         }
 241
 242         {
 243                 struct ib_ah_attr av = {
 244                         .dlid          = be16_to_cpu(mcast->mcmember.mlid),
 245                         .port_num      = priv->port,
 246                         .sl            = mcast->mcmember.sl,
 247                         .ah_flags      = IB_AH_GRH,
 248                         .grh           = {
 249                                 .flow_label    = be32_to_cpu(mcast->mcmember.flow_label),
 250                                 .hop_limit     = mcast->mcmember.hop_limit,
 251                                 .sgid_index    = 0,
 252                                 .traffic_class = mcast->mcmember.traffic_class
 253                         }
 254                 };
 255
 256                 av.grh.dgid = mcast->mcmember.mgid;
 257
 258                 if (ib_sa_rate_enum_to_int(mcast->mcmember.rate) > 0)
 259                         av.static_rate = (2 * priv->local_rate -
 260                                           ib_sa_rate_enum_to_int(mcast->mcmember.rate) - 1) /
 261                                 (priv->local_rate ? priv->local_rate : 1);
 262
 263                 ipoib_dbg_mcast(priv, "static_rate %d for local port %dX, mcmember %dX\n",
 264                                 av.static_rate, priv->local_rate,
 265                                 ib_sa_rate_enum_to_int(mcast->mcmember.rate));
 266
 267                 mcast->ah = ipoib_create_ah(dev, priv->pd, &av);
 268                 if (!mcast->ah) {
 269                         ipoib_warn(priv, "ib_address_create failed\n");
 270                 } else {
 271                         ipoib_dbg_mcast(priv, "MGID " IPOIB_GID_FMT
 272                                         " AV %p, LID 0x%04x, SL %d\n",
 273                                         IPOIB_GID_ARG(mcast->mcmember.mgid),
 274                                         mcast->ah->ah,
 275                                         be16_to_cpu(mcast->mcmember.mlid),
 276                                         mcast->mcmember.sl);
 277                 }
 278         }
 279
 280         /* actually send any queued packets */
 281         while (!skb_queue_empty(&mcast->pkt_queue)) {
 282                 struct sk_buff *skb = skb_dequeue(&mcast->pkt_queue);
 283
 284                 skb->dev = dev;
 285
 286                 if (!skb->dst || !skb->dst->neighbour) {
 287                         /* put pseudoheader back on for next time */
 288                         skb_push(skb, sizeof (struct ipoib_pseudoheader));
 289                 }
 290
 291                 if (dev_queue_xmit(skb))
 292                         ipoib_warn(priv, "dev_queue_xmit failed to requeue packet\n");
 293         }
 294
 295         return 0;
 296 }
 297
 298 static void
 299 ipoib_mcast_sendonly_join_complete(int status,
 300                                    struct ib_sa_mcmember_rec *mcmember,
 301                                    void *mcast_ptr)
 302 {
 303         struct ipoib_mcast *mcast = mcast_ptr;
 304         struct net_device *dev = mcast->dev;
 305
 306         if (!status)
 307                 ipoib_mcast_join_finish(mcast, mcmember);
 308         else {
 309                 if (mcast->logcount++ < 20)
 310                         ipoib_dbg_mcast(netdev_priv(dev), "multicast join failed for "
 311                                         IPOIB_GID_FMT ", status %d\n",
 312                                         IPOIB_GID_ARG(mcast->mcmember.mgid), status);
 313
 314                 /* Flush out any queued packets */
 315                 while (!skb_queue_empty(&mcast->pkt_queue)) {
 316                         struct sk_buff *skb = skb_dequeue(&mcast->pkt_queue);
 317
 318                         skb->dev = dev;
 319
 320                         dev_kfree_skb_any(skb);
 321                 }
 322
 323                 /* Clear the busy flag so we try again */
 324                 clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 325         }
 326
 327         complete(&mcast->done);
 328 }
 329
 330 static int ipoib_mcast_sendonly_join(struct ipoib_mcast *mcast)
 331 {
 332         struct net_device *dev = mcast->dev;
 333         struct ipoib_dev_priv *priv = netdev_priv(dev);
 334         struct ib_sa_mcmember_rec rec = {
 335 #if 0                           /* Some SMs don't support send-only yet */
 336                 .join_state = 4
 337 #else
 338                 .join_state = 1
 339 #endif
 340         };
 341         int ret = 0;
 342
 343         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
 344                 ipoib_dbg_mcast(priv, "device shutting down, no multicast joins\n");
 345                 return -ENODEV;
 346         }
 347
 348         if (test_and_set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)) {
 349                 ipoib_dbg_mcast(priv, "multicast entry busy, skipping\n");
 350                 return -EBUSY;
 351         }
 352
 353         rec.mgid     = mcast->mcmember.mgid;
 354         rec.port_gid = priv->local_gid;
 355         rec.pkey     = be16_to_cpu(priv->pkey);
 356
 357         ret = ib_sa_mcmember_rec_set(priv->ca, priv->port, &rec,
 358                                      IB_SA_MCMEMBER_REC_MGID            |
 359                                      IB_SA_MCMEMBER_REC_PORT_GID        |
 360                                      IB_SA_MCMEMBER_REC_PKEY            |
 361                                      IB_SA_MCMEMBER_REC_JOIN_STATE,
 362                                      1000, GFP_ATOMIC,
 363                                      ipoib_mcast_sendonly_join_complete,
 364                                      mcast, &mcast->query);
 365         if (ret < 0) {
 366                 ipoib_warn(priv, "ib_sa_mcmember_rec_set failed (ret = %d)\n",
 367                            ret);
 368         } else {
 369                 ipoib_dbg_mcast(priv, "no multicast record for " IPOIB_GID_FMT
 370                                 ", starting join\n",
 371                                 IPOIB_GID_ARG(mcast->mcmember.mgid));
 372
 373                 mcast->query_id = ret;
 374         }
 375
 376         return ret;
 377 }
 378
 379 static void ipoib_mcast_join_complete(int status,
 380                                       struct ib_sa_mcmember_rec *mcmember,
 381                                       void *mcast_ptr)
 382 {
 383         struct ipoib_mcast *mcast = mcast_ptr;
 384         struct net_device *dev = mcast->dev;
 385         struct ipoib_dev_priv *priv = netdev_priv(dev);
 386
 387         ipoib_dbg_mcast(priv, "join completion for " IPOIB_GID_FMT
 388                         " (status %d)\n",
 389                         IPOIB_GID_ARG(mcast->mcmember.mgid), status);
 390
 391         if (!status && !ipoib_mcast_join_finish(mcast, mcmember)) {
 392                 mcast->backoff = HZ;
 393                 down(&mcast_mutex);
 394                 if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 395                         queue_work(ipoib_workqueue, &priv->mcast_task);
 396                 up(&mcast_mutex);
 397                 complete(&mcast->done);
 398                 return;
 399         }
 400
 401         if (status == -EINTR) {
 402                 complete(&mcast->done);
 403                 return;
 404         }
 405
 406         if (status && mcast->logcount++ < 20) {
 407                 if (status == -ETIMEDOUT || status == -EINTR) {
 408                         ipoib_dbg_mcast(priv, "multicast join failed for " IPOIB_GID_FMT
 409                                         ", status %d\n",
 410                                         IPOIB_GID_ARG(mcast->mcmember.mgid),
 411                                         status);
 412                 } else {
 413                         ipoib_warn(priv, "multicast join failed for "
 414                                    IPOIB_GID_FMT ", status %d\n",
 415                                    IPOIB_GID_ARG(mcast->mcmember.mgid),
 416                                    status);
 417                 }
 418         }
 419
 420         mcast->backoff *= 2;
 421         if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
 422                 mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
 423
 424         mcast->query = NULL;
 425
 426         down(&mcast_mutex);
 427         if (test_bit(IPOIB_MCAST_RUN, &priv->flags)) {
 428                 if (status == -ETIMEDOUT)
 429                         queue_work(ipoib_workqueue, &priv->mcast_task);
 430                 else
 431                         queue_delayed_work(ipoib_workqueue, &priv->mcast_task,
 432                                            mcast->backoff * HZ);
 433         } else
 434                 complete(&mcast->done);
 435         up(&mcast_mutex);
 436
 437         return;
 438 }
 439
 440 static void ipoib_mcast_join(struct net_device *dev, struct ipoib_mcast *mcast,
 441                              int create)
 442 {
 443         struct ipoib_dev_priv *priv = netdev_priv(dev);
 444         struct ib_sa_mcmember_rec rec = {
 445                 .join_state = 1
 446         };
 447         ib_sa_comp_mask comp_mask;
 448         int ret = 0;
 449
 450         ipoib_dbg_mcast(priv, "joining MGID " IPOIB_GID_FMT "\n",
 451                         IPOIB_GID_ARG(mcast->mcmember.mgid));
 452
 453         rec.mgid     = mcast->mcmember.mgid;
 454         rec.port_gid = priv->local_gid;
 455         rec.pkey     = be16_to_cpu(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_SL           |
 467                         IB_SA_MCMEMBER_REC_FLOW_LABEL   |
 468                         IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
 469
 470                 rec.qkey          = priv->broadcast->mcmember.qkey;
 471                 rec.sl            = priv->broadcast->mcmember.sl;
 472                 rec.flow_label    = priv->broadcast->mcmember.flow_label;
 473                 rec.traffic_class = priv->broadcast->mcmember.traffic_class;
 474         }
 475
 476         ret = ib_sa_mcmember_rec_set(priv->ca, priv->port, &rec, comp_mask,
 477                                      mcast->backoff * 1000, GFP_ATOMIC,
 478                                      ipoib_mcast_join_complete,
 479                                      mcast, &mcast->query);
 480
 481         if (ret < 0) {
 482                 ipoib_warn(priv, "ib_sa_mcmember_rec_set failed, status %d\n", ret);
 483
 484                 mcast->backoff *= 2;
 485                 if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
 486                         mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
 487
 488                 down(&mcast_mutex);
 489                 if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 490                         queue_delayed_work(ipoib_workqueue,
 491                                            &priv->mcast_task,
 492                                            mcast->backoff);
 493                 up(&mcast_mutex);
 494         } else
 495                 mcast->query_id = ret;
 496 }
 497
 498 void ipoib_mcast_join_task(void *dev_ptr)
 499 {
 500         struct net_device *dev = dev_ptr;
 501         struct ipoib_dev_priv *priv = netdev_priv(dev);
 502
 503         if (!test_bit(IPOIB_MCAST_RUN, &priv->flags))
 504                 return;
 505
 506         if (ib_query_gid(priv->ca, priv->port, 0, &priv->local_gid))
 507                 ipoib_warn(priv, "ib_gid_entry_get() failed\n");
 508         else
 509                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
 510
 511         {
 512                 struct ib_port_attr attr;
 513
 514                 if (!ib_query_port(priv->ca, priv->port, &attr)) {
 515                         priv->local_lid  = attr.lid;
 516                         priv->local_rate = attr.active_speed *
 517                                 ib_width_enum_to_int(attr.active_width);
 518                 } else
 519                         ipoib_warn(priv, "ib_query_port failed\n");
 520         }
 521
 522         if (!priv->broadcast) {
 523                 priv->broadcast = ipoib_mcast_alloc(dev, 1);
 524                 if (!priv->broadcast) {
 525                         ipoib_warn(priv, "failed to allocate broadcast group\n");
 526                         down(&mcast_mutex);
 527                         if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 528                                 queue_delayed_work(ipoib_workqueue,
 529                                                    &priv->mcast_task, HZ);
 530                         up(&mcast_mutex);
 531                         return;
 532                 }
 533
 534                 memcpy(priv->broadcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
 535                        sizeof (union ib_gid));
 536
 537                 spin_lock_irq(&priv->lock);
 538                 __ipoib_mcast_add(dev, priv->broadcast);
 539                 spin_unlock_irq(&priv->lock);
 540         }
 541
 542         if (!test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
 543                 ipoib_mcast_join(dev, priv->broadcast, 0);
 544                 return;
 545         }
 546
 547         while (1) {
 548                 struct ipoib_mcast *mcast = NULL;
 549
 550                 spin_lock_irq(&priv->lock);
 551                 list_for_each_entry(mcast, &priv->multicast_list, list) {
 552                         if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)
 553                             && !test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)
 554                             && !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 555                                 /* Found the next unjoined group */
 556                                 break;
 557                         }
 558                 }
 559                 spin_unlock_irq(&priv->lock);
 560
 561                 if (&mcast->list == &priv->multicast_list) {
 562                         /* All done */
 563                         break;
 564                 }
 565
 566                 ipoib_mcast_join(dev, mcast, 1);
 567                 return;
 568         }
 569
 570         priv->mcast_mtu = ib_mtu_enum_to_int(priv->broadcast->mcmember.mtu) -
 571                 IPOIB_ENCAP_LEN;
 572         dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
 573
 574         ipoib_dbg_mcast(priv, "successfully joined all multicast groups\n");
 575
 576         clear_bit(IPOIB_MCAST_RUN, &priv->flags);
 577         netif_carrier_on(dev);
 578 }
 579
 580 int ipoib_mcast_start_thread(struct net_device *dev)
 581 {
 582         struct ipoib_dev_priv *priv = netdev_priv(dev);
 583
 584         ipoib_dbg_mcast(priv, "starting multicast thread\n");
 585
 586         down(&mcast_mutex);
 587         if (!test_and_set_bit(IPOIB_MCAST_RUN, &priv->flags))
 588                 queue_work(ipoib_workqueue, &priv->mcast_task);
 589         up(&mcast_mutex);
 590
 591         return 0;
 592 }
 593
 594 int ipoib_mcast_stop_thread(struct net_device *dev)
 595 {
 596         struct ipoib_dev_priv *priv = netdev_priv(dev);
 597         struct ipoib_mcast *mcast;
 598
 599         ipoib_dbg_mcast(priv, "stopping multicast thread\n");
 600
 601         down(&mcast_mutex);
 602         clear_bit(IPOIB_MCAST_RUN, &priv->flags);
 603         cancel_delayed_work(&priv->mcast_task);
 604         up(&mcast_mutex);
 605
 606         flush_workqueue(ipoib_workqueue);
 607
 608         if (priv->broadcast && priv->broadcast->query) {
 609                 ib_sa_cancel_query(priv->broadcast->query_id, priv->broadcast->query);
 610                 priv->broadcast->query = NULL;
 611                 ipoib_dbg_mcast(priv, "waiting for bcast\n");
 612                 wait_for_completion(&priv->broadcast->done);
 613         }
 614
 615         list_for_each_entry(mcast, &priv->multicast_list, list) {
 616                 if (mcast->query) {
 617                         ib_sa_cancel_query(mcast->query_id, mcast->query);
 618                         mcast->query = NULL;
 619                         ipoib_dbg_mcast(priv, "waiting for MGID " IPOIB_GID_FMT "\n",
 620                                         IPOIB_GID_ARG(mcast->mcmember.mgid));
 621                         wait_for_completion(&mcast->done);
 622                 }
 623         }
 624
 625         return 0;
 626 }
 627
 628 static int ipoib_mcast_leave(struct net_device *dev, struct ipoib_mcast *mcast)
 629 {
 630         struct ipoib_dev_priv *priv = netdev_priv(dev);
 631         struct ib_sa_mcmember_rec rec = {
 632                 .join_state = 1
 633         };
 634         int ret = 0;
 635
 636         if (!test_and_clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags))
 637                 return 0;
 638
 639         ipoib_dbg_mcast(priv, "leaving MGID " IPOIB_GID_FMT "\n",
 640                         IPOIB_GID_ARG(mcast->mcmember.mgid));
 641
 642         rec.mgid     = mcast->mcmember.mgid;
 643         rec.port_gid = priv->local_gid;
 644         rec.pkey     = be16_to_cpu(priv->pkey);
 645
 646         /* Remove ourselves from the multicast group */
 647         ret = ipoib_mcast_detach(dev, be16_to_cpu(mcast->mcmember.mlid),
 648                                  &mcast->mcmember.mgid);
 649         if (ret)
 650                 ipoib_warn(priv, "ipoib_mcast_detach failed (result = %d)\n", ret);
 651
 652         /*
 653          * Just make one shot at leaving and don't wait for a reply;
 654          * if we fail, too bad.
 655          */
 656         ret = ib_sa_mcmember_rec_delete(priv->ca, priv->port, &rec,
 657                                         IB_SA_MCMEMBER_REC_MGID         |
 658                                         IB_SA_MCMEMBER_REC_PORT_GID     |
 659                                         IB_SA_MCMEMBER_REC_PKEY         |
 660                                         IB_SA_MCMEMBER_REC_JOIN_STATE,
 661                                         0, GFP_ATOMIC, NULL,
 662                                         mcast, &mcast->query);
 663         if (ret < 0)
 664                 ipoib_warn(priv, "ib_sa_mcmember_rec_delete failed "
 665                            "for leave (result = %d)\n", ret);
 666
 667         return 0;
 668 }
 669
 670 void ipoib_mcast_send(struct net_device *dev, union ib_gid *mgid,
 671                       struct sk_buff *skb)
 672 {
 673         struct ipoib_dev_priv *priv = netdev_priv(dev);
 674         struct ipoib_mcast *mcast;
 675
 676         /*
 677          * We can only be called from ipoib_start_xmit, so we're
 678          * inside tx_lock -- no need to save/restore flags.
 679          */
 680         spin_lock(&priv->lock);
 681
 682         mcast = __ipoib_mcast_find(dev, mgid);
 683         if (!mcast) {
 684                 /* Let's create a new send only group now */
 685                 ipoib_dbg_mcast(priv, "setting up send only multicast group for "
 686                                 IPOIB_GID_FMT "\n", IPOIB_GID_ARG(*mgid));
 687
 688                 mcast = ipoib_mcast_alloc(dev, 0);
 689                 if (!mcast) {
 690                         ipoib_warn(priv, "unable to allocate memory for "
 691                                    "multicast structure\n");
 692                         dev_kfree_skb_any(skb);
 693                         goto out;
 694                 }
 695
 696                 set_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags);
 697                 mcast->mcmember.mgid = *mgid;
 698                 __ipoib_mcast_add(dev, mcast);
 699                 list_add_tail(&mcast->list, &priv->multicast_list);
 700         }
 701
 702         if (!mcast->ah) {
 703                 if (skb_queue_len(&mcast->pkt_queue) < IPOIB_MAX_MCAST_QUEUE)
 704                         skb_queue_tail(&mcast->pkt_queue, skb);
 705                 else
 706                         dev_kfree_skb_any(skb);
 707
 708                 if (mcast->query)
 709                         ipoib_dbg_mcast(priv, "no address vector, "
 710                                         "but multicast join already started\n");
 711                 else if (test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags))
 712                         ipoib_mcast_sendonly_join(mcast);
 713
 714                 /*
 715                  * If lookup completes between here and out:, don't
 716                  * want to send packet twice.
 717                  */
 718                 mcast = NULL;
 719         }
 720
 721 out:
 722         if (mcast && mcast->ah) {
 723                 if (skb->dst            &&
 724                     skb->dst->neighbour &&
 725                     !*to_ipoib_neigh(skb->dst->neighbour)) {
 726                         struct ipoib_neigh *neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
 727
 728                         if (neigh) {
 729                                 kref_get(&mcast->ah->ref);
 730                                 neigh->ah       = mcast->ah;
 731                                 neigh->neighbour = skb->dst->neighbour;
 732                                 *to_ipoib_neigh(skb->dst->neighbour) = neigh;
 733                                 list_add_tail(&neigh->list, &mcast->neigh_list);
 734                         }
 735                 }
 736
 737                 ipoib_send(dev, skb, mcast->ah, IB_MULTICAST_QPN);
 738         }
 739
 740         spin_unlock(&priv->lock);
 741 }
 742
 743 void ipoib_mcast_dev_flush(struct net_device *dev)
 744 {
 745         struct ipoib_dev_priv *priv = netdev_priv(dev);
 746         LIST_HEAD(remove_list);
 747         struct ipoib_mcast *mcast, *tmcast, *nmcast;
 748         unsigned long flags;
 749
 750         ipoib_dbg_mcast(priv, "flushing multicast list\n");
 751
 752         spin_lock_irqsave(&priv->lock, flags);
 753         list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
 754                 nmcast = ipoib_mcast_alloc(dev, 0);
 755                 if (nmcast) {
 756                         nmcast->flags =
 757                                 mcast->flags & (1 << IPOIB_MCAST_FLAG_SENDONLY);
 758
 759                         nmcast->mcmember.mgid = mcast->mcmember.mgid;
 760
 761                         /* Add the new group in before the to-be-destroyed group */
 762                         list_add_tail(&nmcast->list, &mcast->list);
 763                         list_del_init(&mcast->list);
 764
 765                         rb_replace_node(&mcast->rb_node, &nmcast->rb_node,
 766                                         &priv->multicast_tree);
 767
 768                         list_add_tail(&mcast->list, &remove_list);
 769                 } else {
 770                         ipoib_warn(priv, "could not reallocate multicast group "
 771                                    IPOIB_GID_FMT "\n",
 772                                    IPOIB_GID_ARG(mcast->mcmember.mgid));
 773                 }
 774         }
 775
 776         if (priv->broadcast) {
 777                 nmcast = ipoib_mcast_alloc(dev, 0);
 778                 if (nmcast) {
 779                         nmcast->mcmember.mgid = priv->broadcast->mcmember.mgid;
 780
 781                         rb_replace_node(&priv->broadcast->rb_node,
 782                                         &nmcast->rb_node,
 783                                         &priv->multicast_tree);
 784
 785                         list_add_tail(&priv->broadcast->list, &remove_list);
 786                 }
 787
 788                 priv->broadcast = nmcast;
 789         }
 790
 791         spin_unlock_irqrestore(&priv->lock, flags);
 792
 793         list_for_each_entry(mcast, &remove_list, list) {
 794                 ipoib_mcast_leave(dev, mcast);
 795                 ipoib_mcast_free(mcast);
 796         }
 797 }
 798
 799 void ipoib_mcast_dev_down(struct net_device *dev)
 800 {
 801         struct ipoib_dev_priv *priv = netdev_priv(dev);
 802         unsigned long flags;
 803
 804         /* Delete broadcast since it will be recreated */
 805         if (priv->broadcast) {
 806                 ipoib_dbg_mcast(priv, "deleting broadcast group\n");
 807
 808                 spin_lock_irqsave(&priv->lock, flags);
 809                 rb_erase(&priv->broadcast->rb_node, &priv->multicast_tree);
 810                 spin_unlock_irqrestore(&priv->lock, flags);
 811                 ipoib_mcast_leave(dev, priv->broadcast);
 812                 ipoib_mcast_free(priv->broadcast);
 813                 priv->broadcast = NULL;
 814         }
 815 }
 816
 817 void ipoib_mcast_restart_task(void *dev_ptr)
 818 {
 819         struct net_device *dev = dev_ptr;
 820         struct ipoib_dev_priv *priv = netdev_priv(dev);
 821         struct dev_mc_list *mclist;
 822         struct ipoib_mcast *mcast, *tmcast;
 823         LIST_HEAD(remove_list);
 824         unsigned long flags;
 825
 826         ipoib_dbg_mcast(priv, "restarting multicast task\n");
 827
 828         ipoib_mcast_stop_thread(dev);
 829
 830         spin_lock_irqsave(&priv->lock, flags);
 831
 832         /*
 833          * Unfortunately, the networking core only gives us a list of all of
 834          * the multicast hardware addresses. We need to figure out which ones
 835          * are new and which ones have been removed
 836          */
 837
 838         /* Clear out the found flag */
 839         list_for_each_entry(mcast, &priv->multicast_list, list)
 840                 clear_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
 841
 842         /* Mark all of the entries that are found or don't exist */
 843         for (mclist = dev->mc_list; mclist; mclist = mclist->next) {
 844                 union ib_gid mgid;
 845
 846                 memcpy(mgid.raw, mclist->dmi_addr + 4, sizeof mgid);
 847
 848                 /* Add in the P_Key */
 849                 mgid.raw[4] = (priv->pkey >> 8) & 0xff;
 850                 mgid.raw[5] = priv->pkey & 0xff;
 851
 852                 mcast = __ipoib_mcast_find(dev, &mgid);
 853                 if (!mcast || test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 854                         struct ipoib_mcast *nmcast;
 855
 856                         /* Not found or send-only group, let's add a new entry */
 857                         ipoib_dbg_mcast(priv, "adding multicast entry for mgid "
 858                                         IPOIB_GID_FMT "\n", IPOIB_GID_ARG(mgid));
 859
 860                         nmcast = ipoib_mcast_alloc(dev, 0);
 861                         if (!nmcast) {
 862                                 ipoib_warn(priv, "unable to allocate memory for multicast structure\n");
 863                                 continue;
 864                         }
 865
 866                         set_bit(IPOIB_MCAST_FLAG_FOUND, &nmcast->flags);
 867
 868                         nmcast->mcmember.mgid = mgid;
 869
 870                         if (mcast) {
 871                                 /* Destroy the send only entry */
 872                                 list_del(&mcast->list);
 873                                 list_add_tail(&mcast->list, &remove_list);
 874
 875                                 rb_replace_node(&mcast->rb_node,
 876                                                 &nmcast->rb_node,
 877                                                 &priv->multicast_tree);
 878                         } else
 879                                 __ipoib_mcast_add(dev, nmcast);
 880
 881                         list_add_tail(&nmcast->list, &priv->multicast_list);
 882                 }
 883
 884                 if (mcast)
 885                         set_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
 886         }
 887
 888         /* Remove all of the entries don't exist anymore */
 889         list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
 890                 if (!test_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags) &&
 891                     !test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 892                         ipoib_dbg_mcast(priv, "deleting multicast group " IPOIB_GID_FMT "\n",
 893                                         IPOIB_GID_ARG(mcast->mcmember.mgid));
 894
 895                         rb_erase(&mcast->rb_node, &priv->multicast_tree);
 896
 897                         /* Move to the remove list */
 898                         list_del(&mcast->list);
 899                         list_add_tail(&mcast->list, &remove_list);
 900                 }
 901         }
 902         spin_unlock_irqrestore(&priv->lock, flags);
 903
 904         /* We have to cancel outside of the spinlock */
 905         list_for_each_entry(mcast, &remove_list, list) {
 906                 ipoib_mcast_leave(mcast->dev, mcast);
 907                 ipoib_mcast_free(mcast);
 908         }
 909
 910         if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 911                 ipoib_mcast_start_thread(dev);
 912 }
 913
 914 struct ipoib_mcast_iter *ipoib_mcast_iter_init(struct net_device *dev)
 915 {
 916         struct ipoib_mcast_iter *iter;
 917
 918         iter = kmalloc(sizeof *iter, GFP_KERNEL);
 919         if (!iter)
 920                 return NULL;
 921
 922         iter->dev = dev;
 923         memset(iter->mgid.raw, 0, sizeof iter->mgid);
 924
 925         if (ipoib_mcast_iter_next(iter)) {
 926                 ipoib_mcast_iter_free(iter);
 927                 return NULL;
 928         }
 929
 930         return iter;
 931 }
 932
 933 void ipoib_mcast_iter_free(struct ipoib_mcast_iter *iter)
 934 {
 935         kfree(iter);
 936 }
 937
 938 int ipoib_mcast_iter_next(struct ipoib_mcast_iter *iter)
 939 {
 940         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
 941         struct rb_node *n;
 942         struct ipoib_mcast *mcast;
 943         int ret = 1;
 944
 945         spin_lock_irq(&priv->lock);
 946
 947         n = rb_first(&priv->multicast_tree);
 948
 949         while (n) {
 950                 mcast = rb_entry(n, struct ipoib_mcast, rb_node);
 951
 952                 if (memcmp(iter->mgid.raw, mcast->mcmember.mgid.raw,
 953                            sizeof (union ib_gid)) < 0) {
 954                         iter->mgid      = mcast->mcmember.mgid;
 955                         iter->created   = mcast->created;
 956                         iter->queuelen  = skb_queue_len(&mcast->pkt_queue);
 957                         iter->complete  = !!mcast->ah;
 958                         iter->send_only = !!(mcast->flags & (1 << IPOIB_MCAST_FLAG_SENDONLY));
 959
 960                         ret = 0;
 961
 962                         break;
 963                 }
 964
 965                 n = rb_next(n);
 966         }
 967
 968         spin_unlock_irq(&priv->lock);
 969
 970         return ret;
 971 }
 972
 973 void ipoib_mcast_iter_read(struct ipoib_mcast_iter *iter,
 974                            union ib_gid *mgid,
 975                            unsigned long *created,
 976                            unsigned int *queuelen,
 977                            unsigned int *complete,
 978                            unsigned int *send_only)
 979 {
 980         *mgid      = iter->mgid;
 981         *created   = iter->created;
 982         *queuelen  = iter->queuelen;
 983         *complete  = iter->complete;
 984         *send_only = iter->send_only;
 985 }