SAFE public projects git trees. - safe/jmp/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 <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 DECLARE_MUTEX(mcast_mutex);
  59
  60 /* Used for all multicast joins (broadcast, IPv4 mcast and IPv6 mcast) */
  61 struct ipoib_mcast {
  62         struct ib_sa_mcmember_rec mcmember;
  63         struct ipoib_ah          *ah;
  64
  65         struct rb_node    rb_node;
  66         struct list_head  list;
  67         struct completion done;
  68
  69         int                 query_id;
  70         struct ib_sa_query *query;
  71
  72         unsigned long created;
  73         unsigned long backoff;
  74
  75         unsigned long flags;
  76         unsigned char logcount;
  77
  78         struct list_head  neigh_list;
  79
  80         struct sk_buff_head pkt_queue;
  81
  82         struct net_device *dev;
  83 };
  84
  85 struct ipoib_mcast_iter {
  86         struct net_device *dev;
  87         union ib_gid       mgid;
  88         unsigned long      created;
  89         unsigned int       queuelen;
  90         unsigned int       complete;
  91         unsigned int       send_only;
  92 };
  93
  94 static void ipoib_mcast_free(struct ipoib_mcast *mcast)
  95 {
  96         struct net_device *dev = mcast->dev;
  97         struct ipoib_dev_priv *priv = netdev_priv(dev);
  98         struct ipoib_neigh *neigh, *tmp;
  99         unsigned long flags;
 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                 /*
 109                  * It's safe to call ipoib_put_ah() inside priv->lock
 110                  * here, because we know that mcast->ah will always
 111                  * hold one more reference, so ipoib_put_ah() will
 112                  * never do more than decrement the ref count.
 113                  */
 114                 if (neigh->ah)
 115                         ipoib_put_ah(neigh->ah);
 116                 *to_ipoib_neigh(neigh->neighbour) = NULL;
 117                 neigh->neighbour->ops->destructor = NULL;
 118                 kfree(neigh);
 119         }
 120
 121         spin_unlock_irqrestore(&priv->lock, flags);
 122
 123         if (mcast->ah)
 124                 ipoib_put_ah(mcast->ah);
 125
 126         while (!skb_queue_empty(&mcast->pkt_queue))
 127                 dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
 128
 129         kfree(mcast);
 130 }
 131
 132 static struct ipoib_mcast *ipoib_mcast_alloc(struct net_device *dev,
 133                                              int can_sleep)
 134 {
 135         struct ipoib_mcast *mcast;
 136
 137         mcast = kzalloc(sizeof *mcast, can_sleep ? GFP_KERNEL : GFP_ATOMIC);
 138         if (!mcast)
 139                 return NULL;
 140
 141         mcast->dev = dev;
 142         mcast->created = jiffies;
 143         mcast->backoff = 1;
 144
 145         INIT_LIST_HEAD(&mcast->list);
 146         INIT_LIST_HEAD(&mcast->neigh_list);
 147         skb_queue_head_init(&mcast->pkt_queue);
 148
 149         return mcast;
 150 }
 151
 152 static struct ipoib_mcast *__ipoib_mcast_find(struct net_device *dev, union ib_gid *mgid)
 153 {
 154         struct ipoib_dev_priv *priv = netdev_priv(dev);
 155         struct rb_node *n = priv->multicast_tree.rb_node;
 156
 157         while (n) {
 158                 struct ipoib_mcast *mcast;
 159                 int ret;
 160
 161                 mcast = rb_entry(n, struct ipoib_mcast, rb_node);
 162
 163                 ret = memcmp(mgid->raw, mcast->mcmember.mgid.raw,
 164                              sizeof (union ib_gid));
 165                 if (ret < 0)
 166                         n = n->rb_left;
 167                 else if (ret > 0)
 168                         n = n->rb_right;
 169                 else
 170                         return mcast;
 171         }
 172
 173         return NULL;
 174 }
 175
 176 static int __ipoib_mcast_add(struct net_device *dev, struct ipoib_mcast *mcast)
 177 {
 178         struct ipoib_dev_priv *priv = netdev_priv(dev);
 179         struct rb_node **n = &priv->multicast_tree.rb_node, *pn = NULL;
 180
 181         while (*n) {
 182                 struct ipoib_mcast *tmcast;
 183                 int ret;
 184
 185                 pn = *n;
 186                 tmcast = rb_entry(pn, struct ipoib_mcast, rb_node);
 187
 188                 ret = memcmp(mcast->mcmember.mgid.raw, tmcast->mcmember.mgid.raw,
 189                              sizeof (union ib_gid));
 190                 if (ret < 0)
 191                         n = &pn->rb_left;
 192                 else if (ret > 0)
 193                         n = &pn->rb_right;
 194                 else
 195                         return -EEXIST;
 196         }
 197
 198         rb_link_node(&mcast->rb_node, pn, n);
 199         rb_insert_color(&mcast->rb_node, &priv->multicast_tree);
 200
 201         return 0;
 202 }
 203
 204 static int ipoib_mcast_join_finish(struct ipoib_mcast *mcast,
 205                                    struct ib_sa_mcmember_rec *mcmember)
 206 {
 207         struct net_device *dev = mcast->dev;
 208         struct ipoib_dev_priv *priv = netdev_priv(dev);
 209         int ret;
 210
 211         mcast->mcmember = *mcmember;
 212
 213         /* Set the cached Q_Key before we attach if it's the broadcast group */
 214         if (!memcmp(mcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
 215                     sizeof (union ib_gid))) {
 216                 priv->qkey = be32_to_cpu(priv->broadcast->mcmember.qkey);
 217                 priv->tx_wr.wr.ud.remote_qkey = priv->qkey;
 218         }
 219
 220         if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 221                 if (test_and_set_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 222                         ipoib_warn(priv, "multicast group " IPOIB_GID_FMT
 223                                    " already attached\n",
 224                                    IPOIB_GID_ARG(mcast->mcmember.mgid));
 225
 226                         return 0;
 227                 }
 228
 229                 ret = ipoib_mcast_attach(dev, be16_to_cpu(mcast->mcmember.mlid),
 230                                          &mcast->mcmember.mgid);
 231                 if (ret < 0) {
 232                         ipoib_warn(priv, "couldn't attach QP to multicast group "
 233                                    IPOIB_GID_FMT "\n",
 234                                    IPOIB_GID_ARG(mcast->mcmember.mgid));
 235
 236                         clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags);
 237                         return ret;
 238                 }
 239         }
 240
 241         {
 242                 struct ib_ah_attr av = {
 243                         .dlid          = be16_to_cpu(mcast->mcmember.mlid),
 244                         .port_num      = priv->port,
 245                         .sl            = mcast->mcmember.sl,
 246                         .ah_flags      = IB_AH_GRH,
 247                         .grh           = {
 248                                 .flow_label    = be32_to_cpu(mcast->mcmember.flow_label),
 249                                 .hop_limit     = mcast->mcmember.hop_limit,
 250                                 .sgid_index    = 0,
 251                                 .traffic_class = mcast->mcmember.traffic_class
 252                         }
 253                 };
 254                 int path_rate = ib_sa_rate_enum_to_int(mcast->mcmember.rate);
 255
 256                 av.grh.dgid = mcast->mcmember.mgid;
 257
 258                 if (path_rate > 0 && priv->local_rate > path_rate)
 259                         av.static_rate = (priv->local_rate - 1) / path_rate;
 260
 261                 ipoib_dbg_mcast(priv, "static_rate %d for local port %dX, mcmember %dX\n",
 262                                 av.static_rate, priv->local_rate,
 263                                 ib_sa_rate_enum_to_int(mcast->mcmember.rate));
 264
 265                 mcast->ah = ipoib_create_ah(dev, priv->pd, &av);
 266                 if (!mcast->ah) {
 267                         ipoib_warn(priv, "ib_address_create failed\n");
 268                 } else {
 269                         ipoib_dbg_mcast(priv, "MGID " IPOIB_GID_FMT
 270                                         " AV %p, LID 0x%04x, SL %d\n",
 271                                         IPOIB_GID_ARG(mcast->mcmember.mgid),
 272                                         mcast->ah->ah,
 273                                         be16_to_cpu(mcast->mcmember.mlid),
 274                                         mcast->mcmember.sl);
 275                 }
 276         }
 277
 278         /* actually send any queued packets */
 279         while (!skb_queue_empty(&mcast->pkt_queue)) {
 280                 struct sk_buff *skb = skb_dequeue(&mcast->pkt_queue);
 281
 282                 skb->dev = dev;
 283
 284                 if (!skb->dst || !skb->dst->neighbour) {
 285                         /* put pseudoheader back on for next time */
 286                         skb_push(skb, sizeof (struct ipoib_pseudoheader));
 287                 }
 288
 289                 if (dev_queue_xmit(skb))
 290                         ipoib_warn(priv, "dev_queue_xmit failed to requeue packet\n");
 291         }
 292
 293         return 0;
 294 }
 295
 296 static void
 297 ipoib_mcast_sendonly_join_complete(int status,
 298                                    struct ib_sa_mcmember_rec *mcmember,
 299                                    void *mcast_ptr)
 300 {
 301         struct ipoib_mcast *mcast = mcast_ptr;
 302         struct net_device *dev = mcast->dev;
 303
 304         if (!status)
 305                 ipoib_mcast_join_finish(mcast, mcmember);
 306         else {
 307                 if (mcast->logcount++ < 20)
 308                         ipoib_dbg_mcast(netdev_priv(dev), "multicast join failed for "
 309                                         IPOIB_GID_FMT ", status %d\n",
 310                                         IPOIB_GID_ARG(mcast->mcmember.mgid), status);
 311
 312                 /* Flush out any queued packets */
 313                 while (!skb_queue_empty(&mcast->pkt_queue))
 314                         dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
 315
 316                 /* Clear the busy flag so we try again */
 317                 clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 318         }
 319
 320         complete(&mcast->done);
 321 }
 322
 323 static int ipoib_mcast_sendonly_join(struct ipoib_mcast *mcast)
 324 {
 325         struct net_device *dev = mcast->dev;
 326         struct ipoib_dev_priv *priv = netdev_priv(dev);
 327         struct ib_sa_mcmember_rec rec = {
 328 #if 0                           /* Some SMs don't support send-only yet */
 329                 .join_state = 4
 330 #else
 331                 .join_state = 1
 332 #endif
 333         };
 334         int ret = 0;
 335
 336         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
 337                 ipoib_dbg_mcast(priv, "device shutting down, no multicast joins\n");
 338                 return -ENODEV;
 339         }
 340
 341         if (test_and_set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)) {
 342                 ipoib_dbg_mcast(priv, "multicast entry busy, skipping\n");
 343                 return -EBUSY;
 344         }
 345
 346         rec.mgid     = mcast->mcmember.mgid;
 347         rec.port_gid = priv->local_gid;
 348         rec.pkey     = cpu_to_be16(priv->pkey);
 349
 350         init_completion(&mcast->done);
 351
 352         ret = ib_sa_mcmember_rec_set(priv->ca, priv->port, &rec,
 353                                      IB_SA_MCMEMBER_REC_MGID            |
 354                                      IB_SA_MCMEMBER_REC_PORT_GID        |
 355                                      IB_SA_MCMEMBER_REC_PKEY            |
 356                                      IB_SA_MCMEMBER_REC_JOIN_STATE,
 357                                      1000, GFP_ATOMIC,
 358                                      ipoib_mcast_sendonly_join_complete,
 359                                      mcast, &mcast->query);
 360         if (ret < 0) {
 361                 ipoib_warn(priv, "ib_sa_mcmember_rec_set failed (ret = %d)\n",
 362                            ret);
 363         } else {
 364                 ipoib_dbg_mcast(priv, "no multicast record for " IPOIB_GID_FMT
 365                                 ", starting join\n",
 366                                 IPOIB_GID_ARG(mcast->mcmember.mgid));
 367
 368                 mcast->query_id = ret;
 369         }
 370
 371         return ret;
 372 }
 373
 374 static void ipoib_mcast_join_complete(int status,
 375                                       struct ib_sa_mcmember_rec *mcmember,
 376                                       void *mcast_ptr)
 377 {
 378         struct ipoib_mcast *mcast = mcast_ptr;
 379         struct net_device *dev = mcast->dev;
 380         struct ipoib_dev_priv *priv = netdev_priv(dev);
 381
 382         ipoib_dbg_mcast(priv, "join completion for " IPOIB_GID_FMT
 383                         " (status %d)\n",
 384                         IPOIB_GID_ARG(mcast->mcmember.mgid), status);
 385
 386         if (!status && !ipoib_mcast_join_finish(mcast, mcmember)) {
 387                 mcast->backoff = 1;
 388                 down(&mcast_mutex);
 389                 if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 390                         queue_work(ipoib_workqueue, &priv->mcast_task);
 391                 up(&mcast_mutex);
 392                 complete(&mcast->done);
 393                 return;
 394         }
 395
 396         if (status == -EINTR) {
 397                 complete(&mcast->done);
 398                 return;
 399         }
 400
 401         if (status && mcast->logcount++ < 20) {
 402                 if (status == -ETIMEDOUT || status == -EINTR) {
 403                         ipoib_dbg_mcast(priv, "multicast join failed for " IPOIB_GID_FMT
 404                                         ", status %d\n",
 405                                         IPOIB_GID_ARG(mcast->mcmember.mgid),
 406                                         status);
 407                 } else {
 408                         ipoib_warn(priv, "multicast join failed for "
 409                                    IPOIB_GID_FMT ", status %d\n",
 410                                    IPOIB_GID_ARG(mcast->mcmember.mgid),
 411                                    status);
 412                 }
 413         }
 414
 415         mcast->backoff *= 2;
 416         if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
 417                 mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
 418
 419         mcast->query = NULL;
 420
 421         down(&mcast_mutex);
 422         if (test_bit(IPOIB_MCAST_RUN, &priv->flags)) {
 423                 if (status == -ETIMEDOUT)
 424                         queue_work(ipoib_workqueue, &priv->mcast_task);
 425                 else
 426                         queue_delayed_work(ipoib_workqueue, &priv->mcast_task,
 427                                            mcast->backoff * HZ);
 428         } else
 429                 complete(&mcast->done);
 430         up(&mcast_mutex);
 431
 432         return;
 433 }
 434
 435 static void ipoib_mcast_join(struct net_device *dev, struct ipoib_mcast *mcast,
 436                              int create)
 437 {
 438         struct ipoib_dev_priv *priv = netdev_priv(dev);
 439         struct ib_sa_mcmember_rec rec = {
 440                 .join_state = 1
 441         };
 442         ib_sa_comp_mask comp_mask;
 443         int ret = 0;
 444
 445         ipoib_dbg_mcast(priv, "joining MGID " IPOIB_GID_FMT "\n",
 446                         IPOIB_GID_ARG(mcast->mcmember.mgid));
 447
 448         rec.mgid     = mcast->mcmember.mgid;
 449         rec.port_gid = priv->local_gid;
 450         rec.pkey     = cpu_to_be16(priv->pkey);
 451
 452         comp_mask =
 453                 IB_SA_MCMEMBER_REC_MGID         |
 454                 IB_SA_MCMEMBER_REC_PORT_GID     |
 455                 IB_SA_MCMEMBER_REC_PKEY         |
 456                 IB_SA_MCMEMBER_REC_JOIN_STATE;
 457
 458         if (create) {
 459                 comp_mask |=
 460                         IB_SA_MCMEMBER_REC_QKEY         |
 461                         IB_SA_MCMEMBER_REC_SL           |
 462                         IB_SA_MCMEMBER_REC_FLOW_LABEL   |
 463                         IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
 464
 465                 rec.qkey          = priv->broadcast->mcmember.qkey;
 466                 rec.sl            = priv->broadcast->mcmember.sl;
 467                 rec.flow_label    = priv->broadcast->mcmember.flow_label;
 468                 rec.traffic_class = priv->broadcast->mcmember.traffic_class;
 469         }
 470
 471         init_completion(&mcast->done);
 472
 473         ret = ib_sa_mcmember_rec_set(priv->ca, priv->port, &rec, comp_mask,
 474                                      mcast->backoff * 1000, GFP_ATOMIC,
 475                                      ipoib_mcast_join_complete,
 476                                      mcast, &mcast->query);
 477
 478         if (ret < 0) {
 479                 ipoib_warn(priv, "ib_sa_mcmember_rec_set failed, status %d\n", ret);
 480
 481                 mcast->backoff *= 2;
 482                 if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
 483                         mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
 484
 485                 down(&mcast_mutex);
 486                 if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 487                         queue_delayed_work(ipoib_workqueue,
 488                                            &priv->mcast_task,
 489                                            mcast->backoff * HZ);
 490                 up(&mcast_mutex);
 491         } else
 492                 mcast->query_id = ret;
 493 }
 494
 495 void ipoib_mcast_join_task(void *dev_ptr)
 496 {
 497         struct net_device *dev = dev_ptr;
 498         struct ipoib_dev_priv *priv = netdev_priv(dev);
 499
 500         if (!test_bit(IPOIB_MCAST_RUN, &priv->flags))
 501                 return;
 502
 503         if (ib_query_gid(priv->ca, priv->port, 0, &priv->local_gid))
 504                 ipoib_warn(priv, "ib_gid_entry_get() failed\n");
 505         else
 506                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
 507
 508         {
 509                 struct ib_port_attr attr;
 510
 511                 if (!ib_query_port(priv->ca, priv->port, &attr)) {
 512                         priv->local_lid  = attr.lid;
 513                         priv->local_rate = attr.active_speed *
 514                                 ib_width_enum_to_int(attr.active_width);
 515                 } else
 516                         ipoib_warn(priv, "ib_query_port failed\n");
 517         }
 518
 519         if (!priv->broadcast) {
 520                 priv->broadcast = ipoib_mcast_alloc(dev, 1);
 521                 if (!priv->broadcast) {
 522                         ipoib_warn(priv, "failed to allocate broadcast group\n");
 523                         down(&mcast_mutex);
 524                         if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 525                                 queue_delayed_work(ipoib_workqueue,
 526                                                    &priv->mcast_task, HZ);
 527                         up(&mcast_mutex);
 528                         return;
 529                 }
 530
 531                 memcpy(priv->broadcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
 532                        sizeof (union ib_gid));
 533
 534                 spin_lock_irq(&priv->lock);
 535                 __ipoib_mcast_add(dev, priv->broadcast);
 536                 spin_unlock_irq(&priv->lock);
 537         }
 538
 539         if (!test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
 540                 ipoib_mcast_join(dev, priv->broadcast, 0);
 541                 return;
 542         }
 543
 544         while (1) {
 545                 struct ipoib_mcast *mcast = NULL;
 546
 547                 spin_lock_irq(&priv->lock);
 548                 list_for_each_entry(mcast, &priv->multicast_list, list) {
 549                         if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)
 550                             && !test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)
 551                             && !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 552                                 /* Found the next unjoined group */
 553                                 break;
 554                         }
 555                 }
 556                 spin_unlock_irq(&priv->lock);
 557
 558                 if (&mcast->list == &priv->multicast_list) {
 559                         /* All done */
 560                         break;
 561                 }
 562
 563                 ipoib_mcast_join(dev, mcast, 1);
 564                 return;
 565         }
 566
 567         priv->mcast_mtu = ib_mtu_enum_to_int(priv->broadcast->mcmember.mtu) -
 568                 IPOIB_ENCAP_LEN;
 569         dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
 570
 571         ipoib_dbg_mcast(priv, "successfully joined all multicast groups\n");
 572
 573         clear_bit(IPOIB_MCAST_RUN, &priv->flags);
 574         netif_carrier_on(dev);
 575 }
 576
 577 int ipoib_mcast_start_thread(struct net_device *dev)
 578 {
 579         struct ipoib_dev_priv *priv = netdev_priv(dev);
 580
 581         ipoib_dbg_mcast(priv, "starting multicast thread\n");
 582
 583         down(&mcast_mutex);
 584         if (!test_and_set_bit(IPOIB_MCAST_RUN, &priv->flags))
 585                 queue_work(ipoib_workqueue, &priv->mcast_task);
 586         up(&mcast_mutex);
 587
 588         return 0;
 589 }
 590
 591 int ipoib_mcast_stop_thread(struct net_device *dev, int flush)
 592 {
 593         struct ipoib_dev_priv *priv = netdev_priv(dev);
 594         struct ipoib_mcast *mcast;
 595
 596         ipoib_dbg_mcast(priv, "stopping multicast thread\n");
 597
 598         down(&mcast_mutex);
 599         clear_bit(IPOIB_MCAST_RUN, &priv->flags);
 600         cancel_delayed_work(&priv->mcast_task);
 601         up(&mcast_mutex);
 602
 603         if (flush)
 604                 flush_workqueue(ipoib_workqueue);
 605
 606         if (priv->broadcast && priv->broadcast->query) {
 607                 ib_sa_cancel_query(priv->broadcast->query_id, priv->broadcast->query);
 608                 priv->broadcast->query = NULL;
 609                 ipoib_dbg_mcast(priv, "waiting for bcast\n");
 610                 wait_for_completion(&priv->broadcast->done);
 611         }
 612
 613         list_for_each_entry(mcast, &priv->multicast_list, list) {
 614                 if (mcast->query) {
 615                         ib_sa_cancel_query(mcast->query_id, mcast->query);
 616                         mcast->query = NULL;
 617                         ipoib_dbg_mcast(priv, "waiting for MGID " IPOIB_GID_FMT "\n",
 618                                         IPOIB_GID_ARG(mcast->mcmember.mgid));
 619                         wait_for_completion(&mcast->done);
 620                 }
 621         }
 622
 623         return 0;
 624 }
 625
 626 static int ipoib_mcast_leave(struct net_device *dev, struct ipoib_mcast *mcast)
 627 {
 628         struct ipoib_dev_priv *priv = netdev_priv(dev);
 629         struct ib_sa_mcmember_rec rec = {
 630                 .join_state = 1
 631         };
 632         int ret = 0;
 633
 634         if (!test_and_clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags))
 635                 return 0;
 636
 637         ipoib_dbg_mcast(priv, "leaving MGID " IPOIB_GID_FMT "\n",
 638                         IPOIB_GID_ARG(mcast->mcmember.mgid));
 639
 640         rec.mgid     = mcast->mcmember.mgid;
 641         rec.port_gid = priv->local_gid;
 642         rec.pkey     = cpu_to_be16(priv->pkey);
 643
 644         /* Remove ourselves from the multicast group */
 645         ret = ipoib_mcast_detach(dev, be16_to_cpu(mcast->mcmember.mlid),
 646                                  &mcast->mcmember.mgid);
 647         if (ret)
 648                 ipoib_warn(priv, "ipoib_mcast_detach failed (result = %d)\n", ret);
 649
 650         /*
 651          * Just make one shot at leaving and don't wait for a reply;
 652          * if we fail, too bad.
 653          */
 654         ret = ib_sa_mcmember_rec_delete(priv->ca, priv->port, &rec,
 655                                         IB_SA_MCMEMBER_REC_MGID         |
 656                                         IB_SA_MCMEMBER_REC_PORT_GID     |
 657                                         IB_SA_MCMEMBER_REC_PKEY         |
 658                                         IB_SA_MCMEMBER_REC_JOIN_STATE,
 659                                         0, GFP_ATOMIC, NULL,
 660                                         mcast, &mcast->query);
 661         if (ret < 0)
 662                 ipoib_warn(priv, "ib_sa_mcmember_rec_delete failed "
 663                            "for leave (result = %d)\n", ret);
 664
 665         return 0;
 666 }
 667
 668 void ipoib_mcast_send(struct net_device *dev, union ib_gid *mgid,
 669                       struct sk_buff *skb)
 670 {
 671         struct ipoib_dev_priv *priv = netdev_priv(dev);
 672         struct ipoib_mcast *mcast;
 673
 674         /*
 675          * We can only be called from ipoib_start_xmit, so we're
 676          * inside tx_lock -- no need to save/restore flags.
 677          */
 678         spin_lock(&priv->lock);
 679
 680         mcast = __ipoib_mcast_find(dev, mgid);
 681         if (!mcast) {
 682                 /* Let's create a new send only group now */
 683                 ipoib_dbg_mcast(priv, "setting up send only multicast group for "
 684                                 IPOIB_GID_FMT "\n", IPOIB_GID_ARG(*mgid));
 685
 686                 mcast = ipoib_mcast_alloc(dev, 0);
 687                 if (!mcast) {
 688                         ipoib_warn(priv, "unable to allocate memory for "
 689                                    "multicast structure\n");
 690                         dev_kfree_skb_any(skb);
 691                         goto out;
 692                 }
 693
 694                 set_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags);
 695                 mcast->mcmember.mgid = *mgid;
 696                 __ipoib_mcast_add(dev, mcast);
 697                 list_add_tail(&mcast->list, &priv->multicast_list);
 698         }
 699
 700         if (!mcast->ah) {
 701                 if (skb_queue_len(&mcast->pkt_queue) < IPOIB_MAX_MCAST_QUEUE)
 702                         skb_queue_tail(&mcast->pkt_queue, skb);
 703                 else
 704                         dev_kfree_skb_any(skb);
 705
 706                 if (mcast->query)
 707                         ipoib_dbg_mcast(priv, "no address vector, "
 708                                         "but multicast join already started\n");
 709                 else if (test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags))
 710                         ipoib_mcast_sendonly_join(mcast);
 711
 712                 /*
 713                  * If lookup completes between here and out:, don't
 714                  * want to send packet twice.
 715                  */
 716                 mcast = NULL;
 717         }
 718
 719 out:
 720         if (mcast && mcast->ah) {
 721                 if (skb->dst            &&
 722                     skb->dst->neighbour &&
 723                     !*to_ipoib_neigh(skb->dst->neighbour)) {
 724                         struct ipoib_neigh *neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
 725
 726                         if (neigh) {
 727                                 kref_get(&mcast->ah->ref);
 728                                 neigh->ah       = mcast->ah;
 729                                 neigh->neighbour = skb->dst->neighbour;
 730                                 *to_ipoib_neigh(skb->dst->neighbour) = neigh;
 731                                 list_add_tail(&neigh->list, &mcast->neigh_list);
 732                         }
 733                 }
 734
 735                 ipoib_send(dev, skb, mcast->ah, IB_MULTICAST_QPN);
 736         }
 737
 738         spin_unlock(&priv->lock);
 739 }
 740
 741 void ipoib_mcast_dev_flush(struct net_device *dev)
 742 {
 743         struct ipoib_dev_priv *priv = netdev_priv(dev);
 744         LIST_HEAD(remove_list);
 745         struct ipoib_mcast *mcast, *tmcast, *nmcast;
 746         unsigned long flags;
 747
 748         ipoib_dbg_mcast(priv, "flushing multicast list\n");
 749
 750         spin_lock_irqsave(&priv->lock, flags);
 751         list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
 752                 nmcast = ipoib_mcast_alloc(dev, 0);
 753                 if (nmcast) {
 754                         nmcast->flags =
 755                                 mcast->flags & (1 << IPOIB_MCAST_FLAG_SENDONLY);
 756
 757                         nmcast->mcmember.mgid = mcast->mcmember.mgid;
 758
 759                         /* Add the new group in before the to-be-destroyed group */
 760                         list_add_tail(&nmcast->list, &mcast->list);
 761                         list_del_init(&mcast->list);
 762
 763                         rb_replace_node(&mcast->rb_node, &nmcast->rb_node,
 764                                         &priv->multicast_tree);
 765
 766                         list_add_tail(&mcast->list, &remove_list);
 767                 } else {
 768                         ipoib_warn(priv, "could not reallocate multicast group "
 769                                    IPOIB_GID_FMT "\n",
 770                                    IPOIB_GID_ARG(mcast->mcmember.mgid));
 771                 }
 772         }
 773
 774         if (priv->broadcast) {
 775                 nmcast = ipoib_mcast_alloc(dev, 0);
 776                 if (nmcast) {
 777                         nmcast->mcmember.mgid = priv->broadcast->mcmember.mgid;
 778
 779                         rb_replace_node(&priv->broadcast->rb_node,
 780                                         &nmcast->rb_node,
 781                                         &priv->multicast_tree);
 782
 783                         list_add_tail(&priv->broadcast->list, &remove_list);
 784                         priv->broadcast = nmcast;
 785                 } else
 786                         ipoib_warn(priv, "could not reallocate broadcast group "
 787                                           IPOIB_GID_FMT "\n",
 788                                           IPOIB_GID_ARG(priv->broadcast->mcmember.mgid));
 789         }
 790
 791         spin_unlock_irqrestore(&priv->lock, flags);
 792
 793         list_for_each_entry_safe(mcast, tmcast, &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, 0);
 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_safe(mcast, tmcast, &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 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 915
 916 struct ipoib_mcast_iter *ipoib_mcast_iter_init(struct net_device *dev)
 917 {
 918         struct ipoib_mcast_iter *iter;
 919
 920         iter = kmalloc(sizeof *iter, GFP_KERNEL);
 921         if (!iter)
 922                 return NULL;
 923
 924         iter->dev = dev;
 925         memset(iter->mgid.raw, 0, 16);
 926
 927         if (ipoib_mcast_iter_next(iter)) {
 928                 kfree(iter);
 929                 return NULL;
 930         }
 931
 932         return iter;
 933 }
 934
 935 int ipoib_mcast_iter_next(struct ipoib_mcast_iter *iter)
 936 {
 937         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
 938         struct rb_node *n;
 939         struct ipoib_mcast *mcast;
 940         int ret = 1;
 941
 942         spin_lock_irq(&priv->lock);
 943
 944         n = rb_first(&priv->multicast_tree);
 945
 946         while (n) {
 947                 mcast = rb_entry(n, struct ipoib_mcast, rb_node);
 948
 949                 if (memcmp(iter->mgid.raw, mcast->mcmember.mgid.raw,
 950                            sizeof (union ib_gid)) < 0) {
 951                         iter->mgid      = mcast->mcmember.mgid;
 952                         iter->created   = mcast->created;
 953                         iter->queuelen  = skb_queue_len(&mcast->pkt_queue);
 954                         iter->complete  = !!mcast->ah;
 955                         iter->send_only = !!(mcast->flags & (1 << IPOIB_MCAST_FLAG_SENDONLY));
 956
 957                         ret = 0;
 958
 959                         break;
 960                 }
 961
 962                 n = rb_next(n);
 963         }
 964
 965         spin_unlock_irq(&priv->lock);
 966
 967         return ret;
 968 }
 969
 970 void ipoib_mcast_iter_read(struct ipoib_mcast_iter *iter,
 971                            union ib_gid *mgid,
 972                            unsigned long *created,
 973                            unsigned int *queuelen,
 974                            unsigned int *complete,
 975                            unsigned int *send_only)
 976 {
 977         *mgid      = iter->mgid;
 978         *created   = iter->created;
 979         *queuelen  = iter->queuelen;
 980         *complete  = iter->complete;
 981         *send_only = iter->send_only;
 982 }
 983
 984 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */