git.oblomov.eu Git - linux-2.6/blob - 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  * $Id: ipoib_multicast.c 1362 2004-12-18 15:56:29Z roland $
  35  */
  36
  37 #include <linux/skbuff.h>
  38 #include <linux/rtnetlink.h>
  39 #include <linux/ip.h>
  40 #include <linux/in.h>
  41 #include <linux/igmp.h>
  42 #include <linux/inetdevice.h>
  43 #include <linux/delay.h>
  44 #include <linux/completion.h>
  45
  46 #include "ipoib.h"
  47
  48 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
  49 static int mcast_debug_level;
  50
  51 module_param(mcast_debug_level, int, 0644);
  52 MODULE_PARM_DESC(mcast_debug_level,
  53                  "Enable multicast debug tracing if > 0");
  54 #endif
  55
  56 static DECLARE_MUTEX(mcast_mutex);
  57
  58 /* Used for all multicast joins (broadcast, IPv4 mcast and IPv6 mcast) */
  59 struct ipoib_mcast {
  60         struct ib_sa_mcmember_rec mcmember;
  61         struct ipoib_ah          *ah;
  62
  63         struct rb_node    rb_node;
  64         struct list_head  list;
  65         struct completion done;
  66
  67         int                 query_id;
  68         struct ib_sa_query *query;
  69
  70         unsigned long created;
  71         unsigned long backoff;
  72
  73         unsigned long flags;
  74         unsigned char logcount;
  75
  76         struct list_head  neigh_list;
  77
  78         struct sk_buff_head pkt_queue;
  79
  80         struct net_device *dev;
  81 };
  82
  83 struct ipoib_mcast_iter {
  84         struct net_device *dev;
  85         union ib_gid       mgid;
  86         unsigned long      created;
  87         unsigned int       queuelen;
  88         unsigned int       complete;
  89         unsigned int       send_only;
  90 };
  91
  92 static void ipoib_mcast_free(struct ipoib_mcast *mcast)
  93 {
  94         struct net_device *dev = mcast->dev;
  95         struct ipoib_dev_priv *priv = netdev_priv(dev);
  96         struct ipoib_neigh *neigh, *tmp;
  97         unsigned long flags;
  98         LIST_HEAD(ah_list);
  99         struct ipoib_ah *ah, *tah;
 100
 101         ipoib_dbg_mcast(netdev_priv(dev),
 102                         "deleting multicast group " IPOIB_GID_FMT "\n",
 103                         IPOIB_GID_ARG(mcast->mcmember.mgid));
 104
 105         spin_lock_irqsave(&priv->lock, flags);
 106
 107         list_for_each_entry_safe(neigh, tmp, &mcast->neigh_list, list) {
 108                 if (neigh->ah)
 109                         list_add_tail(&neigh->ah->list, &ah_list);
 110                 *to_ipoib_neigh(neigh->neighbour) = NULL;
 111                 neigh->neighbour->ops->destructor = NULL;
 112                 kfree(neigh);
 113         }
 114
 115         spin_unlock_irqrestore(&priv->lock, flags);
 116
 117         list_for_each_entry_safe(ah, tah, &ah_list, list)
 118                 ipoib_put_ah(ah);
 119
 120         if (mcast->ah)
 121                 ipoib_put_ah(mcast->ah);
 122
 123         while (!skb_queue_empty(&mcast->pkt_queue))
 124                 dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
 125
 126         kfree(mcast);
 127 }
 128
 129 static struct ipoib_mcast *ipoib_mcast_alloc(struct net_device *dev,
 130                                              int can_sleep)
 131 {
 132         struct ipoib_mcast *mcast;
 133
 134         mcast = kzalloc(sizeof *mcast, can_sleep ? GFP_KERNEL : GFP_ATOMIC);
 135         if (!mcast)
 136                 return NULL;
 137
 138         mcast->dev = dev;
 139         mcast->created = jiffies;
 140         mcast->backoff = 1;
 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                 int path_rate = ib_sa_rate_enum_to_int(mcast->mcmember.rate);
 256
 257                 av.grh.dgid = mcast->mcmember.mgid;
 258
 259                 if (path_rate > 0 && priv->local_rate > path_rate)
 260                         av.static_rate = (priv->local_rate - 1) / path_rate;
 261
 262                 ipoib_dbg_mcast(priv, "static_rate %d for local port %dX, mcmember %dX\n",
 263                                 av.static_rate, priv->local_rate,
 264                                 ib_sa_rate_enum_to_int(mcast->mcmember.rate));
 265
 266                 mcast->ah = ipoib_create_ah(dev, priv->pd, &av);
 267                 if (!mcast->ah) {
 268                         ipoib_warn(priv, "ib_address_create failed\n");
 269                 } else {
 270                         ipoib_dbg_mcast(priv, "MGID " IPOIB_GID_FMT
 271                                         " AV %p, LID 0x%04x, SL %d\n",
 272                                         IPOIB_GID_ARG(mcast->mcmember.mgid),
 273                                         mcast->ah->ah,
 274                                         be16_to_cpu(mcast->mcmember.mlid),
 275                                         mcast->mcmember.sl);
 276                 }
 277         }
 278
 279         /* actually send any queued packets */
 280         while (!skb_queue_empty(&mcast->pkt_queue)) {
 281                 struct sk_buff *skb = skb_dequeue(&mcast->pkt_queue);
 282
 283                 skb->dev = dev;
 284
 285                 if (!skb->dst || !skb->dst->neighbour) {
 286                         /* put pseudoheader back on for next time */
 287                         skb_push(skb, sizeof (struct ipoib_pseudoheader));
 288                 }
 289
 290                 if (dev_queue_xmit(skb))
 291                         ipoib_warn(priv, "dev_queue_xmit failed to requeue packet\n");
 292         }
 293
 294         return 0;
 295 }
 296
 297 static void
 298 ipoib_mcast_sendonly_join_complete(int status,
 299                                    struct ib_sa_mcmember_rec *mcmember,
 300                                    void *mcast_ptr)
 301 {
 302         struct ipoib_mcast *mcast = mcast_ptr;
 303         struct net_device *dev = mcast->dev;
 304
 305         if (!status)
 306                 ipoib_mcast_join_finish(mcast, mcmember);
 307         else {
 308                 if (mcast->logcount++ < 20)
 309                         ipoib_dbg_mcast(netdev_priv(dev), "multicast join failed for "
 310                                         IPOIB_GID_FMT ", status %d\n",
 311                                         IPOIB_GID_ARG(mcast->mcmember.mgid), status);
 312
 313                 /* Flush out any queued packets */
 314                 while (!skb_queue_empty(&mcast->pkt_queue))
 315                         dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
 316
 317                 /* Clear the busy flag so we try again */
 318                 clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 319         }
 320
 321         complete(&mcast->done);
 322 }
 323
 324 static int ipoib_mcast_sendonly_join(struct ipoib_mcast *mcast)
 325 {
 326         struct net_device *dev = mcast->dev;
 327         struct ipoib_dev_priv *priv = netdev_priv(dev);
 328         struct ib_sa_mcmember_rec rec = {
 329 #if 0                           /* Some SMs don't support send-only yet */
 330                 .join_state = 4
 331 #else
 332                 .join_state = 1
 333 #endif
 334         };
 335         int ret = 0;
 336
 337         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
 338                 ipoib_dbg_mcast(priv, "device shutting down, no multicast joins\n");
 339                 return -ENODEV;
 340         }
 341
 342         if (test_and_set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)) {
 343                 ipoib_dbg_mcast(priv, "multicast entry busy, skipping\n");
 344                 return -EBUSY;
 345         }
 346
 347         rec.mgid     = mcast->mcmember.mgid;
 348         rec.port_gid = priv->local_gid;
 349         rec.pkey     = cpu_to_be16(priv->pkey);
 350
 351         init_completion(&mcast->done);
 352
 353         ret = ib_sa_mcmember_rec_set(priv->ca, priv->port, &rec,
 354                                      IB_SA_MCMEMBER_REC_MGID            |
 355                                      IB_SA_MCMEMBER_REC_PORT_GID        |
 356                                      IB_SA_MCMEMBER_REC_PKEY            |
 357                                      IB_SA_MCMEMBER_REC_JOIN_STATE,
 358                                      1000, GFP_ATOMIC,
 359                                      ipoib_mcast_sendonly_join_complete,
 360                                      mcast, &mcast->query);
 361         if (ret < 0) {
 362                 ipoib_warn(priv, "ib_sa_mcmember_rec_set failed (ret = %d)\n",
 363                            ret);
 364         } else {
 365                 ipoib_dbg_mcast(priv, "no multicast record for " IPOIB_GID_FMT
 366                                 ", starting join\n",
 367                                 IPOIB_GID_ARG(mcast->mcmember.mgid));
 368
 369                 mcast->query_id = ret;
 370         }
 371
 372         return ret;
 373 }
 374
 375 static void ipoib_mcast_join_complete(int status,
 376                                       struct ib_sa_mcmember_rec *mcmember,
 377                                       void *mcast_ptr)
 378 {
 379         struct ipoib_mcast *mcast = mcast_ptr;
 380         struct net_device *dev = mcast->dev;
 381         struct ipoib_dev_priv *priv = netdev_priv(dev);
 382
 383         ipoib_dbg_mcast(priv, "join completion for " IPOIB_GID_FMT
 384                         " (status %d)\n",
 385                         IPOIB_GID_ARG(mcast->mcmember.mgid), status);
 386
 387         if (!status && !ipoib_mcast_join_finish(mcast, mcmember)) {
 388                 mcast->backoff = 1;
 389                 down(&mcast_mutex);
 390                 if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 391                         queue_work(ipoib_workqueue, &priv->mcast_task);
 392                 up(&mcast_mutex);
 393                 complete(&mcast->done);
 394                 return;
 395         }
 396
 397         if (status == -EINTR) {
 398                 complete(&mcast->done);
 399                 return;
 400         }
 401
 402         if (status && mcast->logcount++ < 20) {
 403                 if (status == -ETIMEDOUT || status == -EINTR) {
 404                         ipoib_dbg_mcast(priv, "multicast join failed for " IPOIB_GID_FMT
 405                                         ", status %d\n",
 406                                         IPOIB_GID_ARG(mcast->mcmember.mgid),
 407                                         status);
 408                 } else {
 409                         ipoib_warn(priv, "multicast join failed for "
 410                                    IPOIB_GID_FMT ", status %d\n",
 411                                    IPOIB_GID_ARG(mcast->mcmember.mgid),
 412                                    status);
 413                 }
 414         }
 415
 416         mcast->backoff *= 2;
 417         if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
 418                 mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
 419
 420         mcast->query = NULL;
 421
 422         down(&mcast_mutex);
 423         if (test_bit(IPOIB_MCAST_RUN, &priv->flags)) {
 424                 if (status == -ETIMEDOUT)
 425                         queue_work(ipoib_workqueue, &priv->mcast_task);
 426                 else
 427                         queue_delayed_work(ipoib_workqueue, &priv->mcast_task,
 428                                            mcast->backoff * HZ);
 429         } else
 430                 complete(&mcast->done);
 431         up(&mcast_mutex);
 432
 433         return;
 434 }
 435
 436 static void ipoib_mcast_join(struct net_device *dev, struct ipoib_mcast *mcast,
 437                              int create)
 438 {
 439         struct ipoib_dev_priv *priv = netdev_priv(dev);
 440         struct ib_sa_mcmember_rec rec = {
 441                 .join_state = 1
 442         };
 443         ib_sa_comp_mask comp_mask;
 444         int ret = 0;
 445
 446         ipoib_dbg_mcast(priv, "joining MGID " IPOIB_GID_FMT "\n",
 447                         IPOIB_GID_ARG(mcast->mcmember.mgid));
 448
 449         rec.mgid     = mcast->mcmember.mgid;
 450         rec.port_gid = priv->local_gid;
 451         rec.pkey     = cpu_to_be16(priv->pkey);
 452
 453         comp_mask =
 454                 IB_SA_MCMEMBER_REC_MGID         |
 455                 IB_SA_MCMEMBER_REC_PORT_GID     |
 456                 IB_SA_MCMEMBER_REC_PKEY         |
 457                 IB_SA_MCMEMBER_REC_JOIN_STATE;
 458
 459         if (create) {
 460                 comp_mask |=
 461                         IB_SA_MCMEMBER_REC_QKEY         |
 462                         IB_SA_MCMEMBER_REC_SL           |
 463                         IB_SA_MCMEMBER_REC_FLOW_LABEL   |
 464                         IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
 465
 466                 rec.qkey          = priv->broadcast->mcmember.qkey;
 467                 rec.sl            = priv->broadcast->mcmember.sl;
 468                 rec.flow_label    = priv->broadcast->mcmember.flow_label;
 469                 rec.traffic_class = priv->broadcast->mcmember.traffic_class;
 470         }
 471
 472         init_completion(&mcast->done);
 473
 474         ret = ib_sa_mcmember_rec_set(priv->ca, priv->port, &rec, comp_mask,
 475                                      mcast->backoff * 1000, GFP_ATOMIC,
 476                                      ipoib_mcast_join_complete,
 477                                      mcast, &mcast->query);
 478
 479         if (ret < 0) {
 480                 ipoib_warn(priv, "ib_sa_mcmember_rec_set failed, status %d\n", ret);
 481
 482                 mcast->backoff *= 2;
 483                 if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
 484                         mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
 485
 486                 down(&mcast_mutex);
 487                 if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 488                         queue_delayed_work(ipoib_workqueue,
 489                                            &priv->mcast_task,
 490                                            mcast->backoff * HZ);
 491                 up(&mcast_mutex);
 492         } else
 493                 mcast->query_id = ret;
 494 }
 495
 496 void ipoib_mcast_join_task(void *dev_ptr)
 497 {
 498         struct net_device *dev = dev_ptr;
 499         struct ipoib_dev_priv *priv = netdev_priv(dev);
 500
 501         if (!test_bit(IPOIB_MCAST_RUN, &priv->flags))
 502                 return;
 503
 504         if (ib_query_gid(priv->ca, priv->port, 0, &priv->local_gid))
 505                 ipoib_warn(priv, "ib_gid_entry_get() failed\n");
 506         else
 507                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
 508
 509         {
 510                 struct ib_port_attr attr;
 511
 512                 if (!ib_query_port(priv->ca, priv->port, &attr)) {
 513                         priv->local_lid  = attr.lid;
 514                         priv->local_rate = attr.active_speed *
 515                                 ib_width_enum_to_int(attr.active_width);
 516                 } else
 517                         ipoib_warn(priv, "ib_query_port failed\n");
 518         }
 519
 520         if (!priv->broadcast) {
 521                 priv->broadcast = ipoib_mcast_alloc(dev, 1);
 522                 if (!priv->broadcast) {
 523                         ipoib_warn(priv, "failed to allocate broadcast group\n");
 524                         down(&mcast_mutex);
 525                         if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 526                                 queue_delayed_work(ipoib_workqueue,
 527                                                    &priv->mcast_task, HZ);
 528                         up(&mcast_mutex);
 529                         return;
 530                 }
 531
 532                 memcpy(priv->broadcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
 533                        sizeof (union ib_gid));
 534
 535                 spin_lock_irq(&priv->lock);
 536                 __ipoib_mcast_add(dev, priv->broadcast);
 537                 spin_unlock_irq(&priv->lock);
 538         }
 539
 540         if (!test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
 541                 ipoib_mcast_join(dev, priv->broadcast, 0);
 542                 return;
 543         }
 544
 545         while (1) {
 546                 struct ipoib_mcast *mcast = NULL;
 547
 548                 spin_lock_irq(&priv->lock);
 549                 list_for_each_entry(mcast, &priv->multicast_list, list) {
 550                         if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)
 551                             && !test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)
 552                             && !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 553                                 /* Found the next unjoined group */
 554                                 break;
 555                         }
 556                 }
 557                 spin_unlock_irq(&priv->lock);
 558
 559                 if (&mcast->list == &priv->multicast_list) {
 560                         /* All done */
 561                         break;
 562                 }
 563
 564                 ipoib_mcast_join(dev, mcast, 1);
 565                 return;
 566         }
 567
 568         priv->mcast_mtu = ib_mtu_enum_to_int(priv->broadcast->mcmember.mtu) -
 569                 IPOIB_ENCAP_LEN;
 570         dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
 571
 572         ipoib_dbg_mcast(priv, "successfully joined all multicast groups\n");
 573
 574         clear_bit(IPOIB_MCAST_RUN, &priv->flags);
 575         netif_carrier_on(dev);
 576 }
 577
 578 int ipoib_mcast_start_thread(struct net_device *dev)
 579 {
 580         struct ipoib_dev_priv *priv = netdev_priv(dev);
 581
 582         ipoib_dbg_mcast(priv, "starting multicast thread\n");
 583
 584         down(&mcast_mutex);
 585         if (!test_and_set_bit(IPOIB_MCAST_RUN, &priv->flags))
 586                 queue_work(ipoib_workqueue, &priv->mcast_task);
 587         up(&mcast_mutex);
 588
 589         return 0;
 590 }
 591
 592 int ipoib_mcast_stop_thread(struct net_device *dev, int flush)
 593 {
 594         struct ipoib_dev_priv *priv = netdev_priv(dev);
 595         struct ipoib_mcast *mcast;
 596
 597         ipoib_dbg_mcast(priv, "stopping multicast thread\n");
 598
 599         down(&mcast_mutex);
 600         clear_bit(IPOIB_MCAST_RUN, &priv->flags);
 601         cancel_delayed_work(&priv->mcast_task);
 602         up(&mcast_mutex);
 603
 604         if (flush)
 605                 flush_workqueue(ipoib_workqueue);
 606
 607         if (priv->broadcast && priv->broadcast->query) {
 608                 ib_sa_cancel_query(priv->broadcast->query_id, priv->broadcast->query);
 609                 priv->broadcast->query = NULL;
 610                 ipoib_dbg_mcast(priv, "waiting for bcast\n");
 611                 wait_for_completion(&priv->broadcast->done);
 612         }
 613
 614         list_for_each_entry(mcast, &priv->multicast_list, list) {
 615                 if (mcast->query) {
 616                         ib_sa_cancel_query(mcast->query_id, mcast->query);
 617                         mcast->query = NULL;
 618                         ipoib_dbg_mcast(priv, "waiting for MGID " IPOIB_GID_FMT "\n",
 619                                         IPOIB_GID_ARG(mcast->mcmember.mgid));
 620                         wait_for_completion(&mcast->done);
 621                 }
 622         }
 623
 624         return 0;
 625 }
 626
 627 static int ipoib_mcast_leave(struct net_device *dev, struct ipoib_mcast *mcast)
 628 {
 629         struct ipoib_dev_priv *priv = netdev_priv(dev);
 630         struct ib_sa_mcmember_rec rec = {
 631                 .join_state = 1
 632         };
 633         int ret = 0;
 634
 635         if (!test_and_clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags))
 636                 return 0;
 637
 638         ipoib_dbg_mcast(priv, "leaving MGID " IPOIB_GID_FMT "\n",
 639                         IPOIB_GID_ARG(mcast->mcmember.mgid));
 640
 641         rec.mgid     = mcast->mcmember.mgid;
 642         rec.port_gid = priv->local_gid;
 643         rec.pkey     = cpu_to_be16(priv->pkey);
 644
 645         /* Remove ourselves from the multicast group */
 646         ret = ipoib_mcast_detach(dev, be16_to_cpu(mcast->mcmember.mlid),
 647                                  &mcast->mcmember.mgid);
 648         if (ret)
 649                 ipoib_warn(priv, "ipoib_mcast_detach failed (result = %d)\n", ret);
 650
 651         /*
 652          * Just make one shot at leaving and don't wait for a reply;
 653          * if we fail, too bad.
 654          */
 655         ret = ib_sa_mcmember_rec_delete(priv->ca, priv->port, &rec,
 656                                         IB_SA_MCMEMBER_REC_MGID         |
 657                                         IB_SA_MCMEMBER_REC_PORT_GID     |
 658                                         IB_SA_MCMEMBER_REC_PKEY         |
 659                                         IB_SA_MCMEMBER_REC_JOIN_STATE,
 660                                         0, GFP_ATOMIC, NULL,
 661                                         mcast, &mcast->query);
 662         if (ret < 0)
 663                 ipoib_warn(priv, "ib_sa_mcmember_rec_delete failed "
 664                            "for leave (result = %d)\n", ret);
 665
 666         return 0;
 667 }
 668
 669 void ipoib_mcast_send(struct net_device *dev, union ib_gid *mgid,
 670                       struct sk_buff *skb)
 671 {
 672         struct ipoib_dev_priv *priv = netdev_priv(dev);
 673         struct ipoib_mcast *mcast;
 674
 675         /*
 676          * We can only be called from ipoib_start_xmit, so we're
 677          * inside tx_lock -- no need to save/restore flags.
 678          */
 679         spin_lock(&priv->lock);
 680
 681         mcast = __ipoib_mcast_find(dev, mgid);
 682         if (!mcast) {
 683                 /* Let's create a new send only group now */
 684                 ipoib_dbg_mcast(priv, "setting up send only multicast group for "
 685                                 IPOIB_GID_FMT "\n", IPOIB_GID_ARG(*mgid));
 686
 687                 mcast = ipoib_mcast_alloc(dev, 0);
 688                 if (!mcast) {
 689                         ipoib_warn(priv, "unable to allocate memory for "
 690                                    "multicast structure\n");
 691                         dev_kfree_skb_any(skb);
 692                         goto out;
 693                 }
 694
 695                 set_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags);
 696                 mcast->mcmember.mgid = *mgid;
 697                 __ipoib_mcast_add(dev, mcast);
 698                 list_add_tail(&mcast->list, &priv->multicast_list);
 699         }
 700
 701         if (!mcast->ah) {
 702                 if (skb_queue_len(&mcast->pkt_queue) < IPOIB_MAX_MCAST_QUEUE)
 703                         skb_queue_tail(&mcast->pkt_queue, skb);
 704                 else
 705                         dev_kfree_skb_any(skb);
 706
 707                 if (mcast->query)
 708                         ipoib_dbg_mcast(priv, "no address vector, "
 709                                         "but multicast join already started\n");
 710                 else if (test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags))
 711                         ipoib_mcast_sendonly_join(mcast);
 712
 713                 /*
 714                  * If lookup completes between here and out:, don't
 715                  * want to send packet twice.
 716                  */
 717                 mcast = NULL;
 718         }
 719
 720 out:
 721         if (mcast && mcast->ah) {
 722                 if (skb->dst            &&
 723                     skb->dst->neighbour &&
 724                     !*to_ipoib_neigh(skb->dst->neighbour)) {
 725                         struct ipoib_neigh *neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
 726
 727                         if (neigh) {
 728                                 kref_get(&mcast->ah->ref);
 729                                 neigh->ah       = mcast->ah;
 730                                 neigh->neighbour = skb->dst->neighbour;
 731                                 *to_ipoib_neigh(skb->dst->neighbour) = neigh;
 732                                 list_add_tail(&neigh->list, &mcast->neigh_list);
 733                         }
 734                 }
 735
 736                 ipoib_send(dev, skb, mcast->ah, IB_MULTICAST_QPN);
 737         }
 738
 739         spin_unlock(&priv->lock);
 740 }
 741
 742 void ipoib_mcast_dev_flush(struct net_device *dev)
 743 {
 744         struct ipoib_dev_priv *priv = netdev_priv(dev);
 745         LIST_HEAD(remove_list);
 746         struct ipoib_mcast *mcast, *tmcast, *nmcast;
 747         unsigned long flags;
 748
 749         ipoib_dbg_mcast(priv, "flushing multicast list\n");
 750
 751         spin_lock_irqsave(&priv->lock, flags);
 752         list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
 753                 nmcast = ipoib_mcast_alloc(dev, 0);
 754                 if (nmcast) {
 755                         nmcast->flags =
 756                                 mcast->flags & (1 << IPOIB_MCAST_FLAG_SENDONLY);
 757
 758                         nmcast->mcmember.mgid = mcast->mcmember.mgid;
 759
 760                         /* Add the new group in before the to-be-destroyed group */
 761                         list_add_tail(&nmcast->list, &mcast->list);
 762                         list_del_init(&mcast->list);
 763
 764                         rb_replace_node(&mcast->rb_node, &nmcast->rb_node,
 765                                         &priv->multicast_tree);
 766
 767                         list_add_tail(&mcast->list, &remove_list);
 768                 } else {
 769                         ipoib_warn(priv, "could not reallocate multicast group "
 770                                    IPOIB_GID_FMT "\n",
 771                                    IPOIB_GID_ARG(mcast->mcmember.mgid));
 772                 }
 773         }
 774
 775         if (priv->broadcast) {
 776                 nmcast = ipoib_mcast_alloc(dev, 0);
 777                 if (nmcast) {
 778                         nmcast->mcmember.mgid = priv->broadcast->mcmember.mgid;
 779
 780                         rb_replace_node(&priv->broadcast->rb_node,
 781                                         &nmcast->rb_node,
 782                                         &priv->multicast_tree);
 783
 784                         list_add_tail(&priv->broadcast->list, &remove_list);
 785                 }
 786
 787                 priv->broadcast = nmcast;
 788         }
 789
 790         spin_unlock_irqrestore(&priv->lock, flags);
 791
 792         list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
 793                 ipoib_mcast_leave(dev, mcast);
 794                 ipoib_mcast_free(mcast);
 795         }
 796 }
 797
 798 void ipoib_mcast_dev_down(struct net_device *dev)
 799 {
 800         struct ipoib_dev_priv *priv = netdev_priv(dev);
 801         unsigned long flags;
 802
 803         /* Delete broadcast since it will be recreated */
 804         if (priv->broadcast) {
 805                 ipoib_dbg_mcast(priv, "deleting broadcast group\n");
 806
 807                 spin_lock_irqsave(&priv->lock, flags);
 808                 rb_erase(&priv->broadcast->rb_node, &priv->multicast_tree);
 809                 spin_unlock_irqrestore(&priv->lock, flags);
 810                 ipoib_mcast_leave(dev, priv->broadcast);
 811                 ipoib_mcast_free(priv->broadcast);
 812                 priv->broadcast = NULL;
 813         }
 814 }
 815
 816 void ipoib_mcast_restart_task(void *dev_ptr)
 817 {
 818         struct net_device *dev = dev_ptr;
 819         struct ipoib_dev_priv *priv = netdev_priv(dev);
 820         struct dev_mc_list *mclist;
 821         struct ipoib_mcast *mcast, *tmcast;
 822         LIST_HEAD(remove_list);
 823         unsigned long flags;
 824
 825         ipoib_dbg_mcast(priv, "restarting multicast task\n");
 826
 827         ipoib_mcast_stop_thread(dev, 0);
 828
 829         spin_lock_irqsave(&priv->lock, flags);
 830
 831         /*
 832          * Unfortunately, the networking core only gives us a list of all of
 833          * the multicast hardware addresses. We need to figure out which ones
 834          * are new and which ones have been removed
 835          */
 836
 837         /* Clear out the found flag */
 838         list_for_each_entry(mcast, &priv->multicast_list, list)
 839                 clear_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
 840
 841         /* Mark all of the entries that are found or don't exist */
 842         for (mclist = dev->mc_list; mclist; mclist = mclist->next) {
 843                 union ib_gid mgid;
 844
 845                 memcpy(mgid.raw, mclist->dmi_addr + 4, sizeof mgid);
 846
 847                 /* Add in the P_Key */
 848                 mgid.raw[4] = (priv->pkey >> 8) & 0xff;
 849                 mgid.raw[5] = priv->pkey & 0xff;
 850
 851                 mcast = __ipoib_mcast_find(dev, &mgid);
 852                 if (!mcast || test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 853                         struct ipoib_mcast *nmcast;
 854
 855                         /* Not found or send-only group, let's add a new entry */
 856                         ipoib_dbg_mcast(priv, "adding multicast entry for mgid "
 857                                         IPOIB_GID_FMT "\n", IPOIB_GID_ARG(mgid));
 858
 859                         nmcast = ipoib_mcast_alloc(dev, 0);
 860                         if (!nmcast) {
 861                                 ipoib_warn(priv, "unable to allocate memory for multicast structure\n");
 862                                 continue;
 863                         }
 864
 865                         set_bit(IPOIB_MCAST_FLAG_FOUND, &nmcast->flags);
 866
 867                         nmcast->mcmember.mgid = mgid;
 868
 869                         if (mcast) {
 870                                 /* Destroy the send only entry */
 871                                 list_del(&mcast->list);
 872                                 list_add_tail(&mcast->list, &remove_list);
 873
 874                                 rb_replace_node(&mcast->rb_node,
 875                                                 &nmcast->rb_node,
 876                                                 &priv->multicast_tree);
 877                         } else
 878                                 __ipoib_mcast_add(dev, nmcast);
 879
 880                         list_add_tail(&nmcast->list, &priv->multicast_list);
 881                 }
 882
 883                 if (mcast)
 884                         set_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
 885         }
 886
 887         /* Remove all of the entries don't exist anymore */
 888         list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
 889                 if (!test_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags) &&
 890                     !test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 891                         ipoib_dbg_mcast(priv, "deleting multicast group " IPOIB_GID_FMT "\n",
 892                                         IPOIB_GID_ARG(mcast->mcmember.mgid));
 893
 894                         rb_erase(&mcast->rb_node, &priv->multicast_tree);
 895
 896                         /* Move to the remove list */
 897                         list_del(&mcast->list);
 898                         list_add_tail(&mcast->list, &remove_list);
 899                 }
 900         }
 901         spin_unlock_irqrestore(&priv->lock, flags);
 902
 903         /* We have to cancel outside of the spinlock */
 904         list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
 905                 ipoib_mcast_leave(mcast->dev, mcast);
 906                 ipoib_mcast_free(mcast);
 907         }
 908
 909         if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 910                 ipoib_mcast_start_thread(dev);
 911 }
 912
 913 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 914
 915 struct ipoib_mcast_iter *ipoib_mcast_iter_init(struct net_device *dev)
 916 {
 917         struct ipoib_mcast_iter *iter;
 918
 919         iter = kmalloc(sizeof *iter, GFP_KERNEL);
 920         if (!iter)
 921                 return NULL;
 922
 923         iter->dev = dev;
 924         memset(iter->mgid.raw, 0, 16);
 925
 926         if (ipoib_mcast_iter_next(iter)) {
 927                 kfree(iter);
 928                 return NULL;
 929         }
 930
 931         return iter;
 932 }
 933
 934 int ipoib_mcast_iter_next(struct ipoib_mcast_iter *iter)
 935 {
 936         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
 937         struct rb_node *n;
 938         struct ipoib_mcast *mcast;
 939         int ret = 1;
 940
 941         spin_lock_irq(&priv->lock);
 942
 943         n = rb_first(&priv->multicast_tree);
 944
 945         while (n) {
 946                 mcast = rb_entry(n, struct ipoib_mcast, rb_node);
 947
 948                 if (memcmp(iter->mgid.raw, mcast->mcmember.mgid.raw,
 949                            sizeof (union ib_gid)) < 0) {
 950                         iter->mgid      = mcast->mcmember.mgid;
 951                         iter->created   = mcast->created;
 952                         iter->queuelen  = skb_queue_len(&mcast->pkt_queue);
 953                         iter->complete  = !!mcast->ah;
 954                         iter->send_only = !!(mcast->flags & (1 << IPOIB_MCAST_FLAG_SENDONLY));
 955
 956                         ret = 0;
 957
 958                         break;
 959                 }
 960
 961                 n = rb_next(n);
 962         }
 963
 964         spin_unlock_irq(&priv->lock);
 965
 966         return ret;
 967 }
 968
 969 void ipoib_mcast_iter_read(struct ipoib_mcast_iter *iter,
 970                            union ib_gid *mgid,
 971                            unsigned long *created,
 972                            unsigned int *queuelen,
 973                            unsigned int *complete,
 974                            unsigned int *send_only)
 975 {
 976         *mgid      = iter->mgid;
 977         *created   = iter->created;
 978         *queuelen  = iter->queuelen;
 979         *complete  = iter->complete;
 980         *send_only = iter->send_only;
 981 }
 982
 983 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */