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         LIST_HEAD(ah_list);
  97         struct ipoib_ah *ah, *tah;
  98
  99         ipoib_dbg_mcast(netdev_priv(dev),
 100                         "deleting multicast group " IPOIB_GID_FMT "\n",
 101                         IPOIB_GID_ARG(mcast->mcmember.mgid));
 102
 103         spin_lock_irqsave(&priv->lock, flags);
 104
 105         list_for_each_entry_safe(neigh, tmp, &mcast->neigh_list, list) {
 106                 if (neigh->ah)
 107                         list_add_tail(&neigh->ah->list, &ah_list);
 108                 *to_ipoib_neigh(neigh->neighbour) = NULL;
 109                 neigh->neighbour->ops->destructor = NULL;
 110                 kfree(neigh);
 111         }
 112
 113         spin_unlock_irqrestore(&priv->lock, flags);
 114
 115         list_for_each_entry_safe(ah, tah, &ah_list, list)
 116                 ipoib_put_ah(ah);
 117
 118         if (mcast->ah)
 119                 ipoib_put_ah(mcast->ah);
 120
 121         while (!skb_queue_empty(&mcast->pkt_queue)) {
 122                 struct sk_buff *skb = skb_dequeue(&mcast->pkt_queue);
 123
 124                 skb->dev = dev;
 125                 dev_kfree_skb_any(skb);
 126         }
 127
 128         kfree(mcast);
 129 }
 130
 131 static struct ipoib_mcast *ipoib_mcast_alloc(struct net_device *dev,
 132                                              int can_sleep)
 133 {
 134         struct ipoib_mcast *mcast;
 135
 136         mcast = kmalloc(sizeof (*mcast), can_sleep ? GFP_KERNEL : GFP_ATOMIC);
 137         if (!mcast)
 138                 return NULL;
 139
 140         memset(mcast, 0, sizeof (*mcast));
 141
 142         init_completion(&mcast->done);
 143
 144         mcast->dev = dev;
 145         mcast->created = jiffies;
 146         mcast->backoff = HZ;
 147         mcast->logcount = 0;
 148
 149         INIT_LIST_HEAD(&mcast->list);
 150         INIT_LIST_HEAD(&mcast->neigh_list);
 151         skb_queue_head_init(&mcast->pkt_queue);
 152
 153         mcast->ah    = NULL;
 154         mcast->query = NULL;
 155
 156         return mcast;
 157 }
 158
 159 static struct ipoib_mcast *__ipoib_mcast_find(struct net_device *dev, union ib_gid *mgid)
 160 {
 161         struct ipoib_dev_priv *priv = netdev_priv(dev);
 162         struct rb_node *n = priv->multicast_tree.rb_node;
 163
 164         while (n) {
 165                 struct ipoib_mcast *mcast;
 166                 int ret;
 167
 168                 mcast = rb_entry(n, struct ipoib_mcast, rb_node);
 169
 170                 ret = memcmp(mgid->raw, mcast->mcmember.mgid.raw,
 171                              sizeof (union ib_gid));
 172                 if (ret < 0)
 173                         n = n->rb_left;
 174                 else if (ret > 0)
 175                         n = n->rb_right;
 176                 else
 177                         return mcast;
 178         }
 179
 180         return NULL;
 181 }
 182
 183 static int __ipoib_mcast_add(struct net_device *dev, struct ipoib_mcast *mcast)
 184 {
 185         struct ipoib_dev_priv *priv = netdev_priv(dev);
 186         struct rb_node **n = &priv->multicast_tree.rb_node, *pn = NULL;
 187
 188         while (*n) {
 189                 struct ipoib_mcast *tmcast;
 190                 int ret;
 191
 192                 pn = *n;
 193                 tmcast = rb_entry(pn, struct ipoib_mcast, rb_node);
 194
 195                 ret = memcmp(mcast->mcmember.mgid.raw, tmcast->mcmember.mgid.raw,
 196                              sizeof (union ib_gid));
 197                 if (ret < 0)
 198                         n = &pn->rb_left;
 199                 else if (ret > 0)
 200                         n = &pn->rb_right;
 201                 else
 202                         return -EEXIST;
 203         }
 204
 205         rb_link_node(&mcast->rb_node, pn, n);
 206         rb_insert_color(&mcast->rb_node, &priv->multicast_tree);
 207
 208         return 0;
 209 }
 210
 211 static int ipoib_mcast_join_finish(struct ipoib_mcast *mcast,
 212                                    struct ib_sa_mcmember_rec *mcmember)
 213 {
 214         struct net_device *dev = mcast->dev;
 215         struct ipoib_dev_priv *priv = netdev_priv(dev);
 216         int ret;
 217
 218         mcast->mcmember = *mcmember;
 219
 220         /* Set the cached Q_Key before we attach if it's the broadcast group */
 221         if (!memcmp(mcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
 222                     sizeof (union ib_gid))) {
 223                 priv->qkey = be32_to_cpu(priv->broadcast->mcmember.qkey);
 224                 priv->tx_wr.wr.ud.remote_qkey = priv->qkey;
 225         }
 226
 227         if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 228                 if (test_and_set_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 229                         ipoib_warn(priv, "multicast group " IPOIB_GID_FMT
 230                                    " already attached\n",
 231                                    IPOIB_GID_ARG(mcast->mcmember.mgid));
 232
 233                         return 0;
 234                 }
 235
 236                 ret = ipoib_mcast_attach(dev, be16_to_cpu(mcast->mcmember.mlid),
 237                                          &mcast->mcmember.mgid);
 238                 if (ret < 0) {
 239                         ipoib_warn(priv, "couldn't attach QP to multicast group "
 240                                    IPOIB_GID_FMT "\n",
 241                                    IPOIB_GID_ARG(mcast->mcmember.mgid));
 242
 243                         clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags);
 244                         return ret;
 245                 }
 246         }
 247
 248         {
 249                 struct ib_ah_attr av = {
 250                         .dlid          = be16_to_cpu(mcast->mcmember.mlid),
 251                         .port_num      = priv->port,
 252                         .sl            = mcast->mcmember.sl,
 253                         .ah_flags      = IB_AH_GRH,
 254                         .grh           = {
 255                                 .flow_label    = be32_to_cpu(mcast->mcmember.flow_label),
 256                                 .hop_limit     = mcast->mcmember.hop_limit,
 257                                 .sgid_index    = 0,
 258                                 .traffic_class = mcast->mcmember.traffic_class
 259                         }
 260                 };
 261                 int path_rate = ib_sa_rate_enum_to_int(mcast->mcmember.rate);
 262
 263                 av.grh.dgid = mcast->mcmember.mgid;
 264
 265                 if (path_rate > 0 && priv->local_rate > path_rate)
 266                         av.static_rate = (priv->local_rate - 1) / path_rate;
 267
 268                 ipoib_dbg_mcast(priv, "static_rate %d for local port %dX, mcmember %dX\n",
 269                                 av.static_rate, priv->local_rate,
 270                                 ib_sa_rate_enum_to_int(mcast->mcmember.rate));
 271
 272                 mcast->ah = ipoib_create_ah(dev, priv->pd, &av);
 273                 if (!mcast->ah) {
 274                         ipoib_warn(priv, "ib_address_create failed\n");
 275                 } else {
 276                         ipoib_dbg_mcast(priv, "MGID " IPOIB_GID_FMT
 277                                         " AV %p, LID 0x%04x, SL %d\n",
 278                                         IPOIB_GID_ARG(mcast->mcmember.mgid),
 279                                         mcast->ah->ah,
 280                                         be16_to_cpu(mcast->mcmember.mlid),
 281                                         mcast->mcmember.sl);
 282                 }
 283         }
 284
 285         /* actually send any queued packets */
 286         while (!skb_queue_empty(&mcast->pkt_queue)) {
 287                 struct sk_buff *skb = skb_dequeue(&mcast->pkt_queue);
 288
 289                 skb->dev = dev;
 290
 291                 if (!skb->dst || !skb->dst->neighbour) {
 292                         /* put pseudoheader back on for next time */
 293                         skb_push(skb, sizeof (struct ipoib_pseudoheader));
 294                 }
 295
 296                 if (dev_queue_xmit(skb))
 297                         ipoib_warn(priv, "dev_queue_xmit failed to requeue packet\n");
 298         }
 299
 300         return 0;
 301 }
 302
 303 static void
 304 ipoib_mcast_sendonly_join_complete(int status,
 305                                    struct ib_sa_mcmember_rec *mcmember,
 306                                    void *mcast_ptr)
 307 {
 308         struct ipoib_mcast *mcast = mcast_ptr;
 309         struct net_device *dev = mcast->dev;
 310
 311         if (!status)
 312                 ipoib_mcast_join_finish(mcast, mcmember);
 313         else {
 314                 if (mcast->logcount++ < 20)
 315                         ipoib_dbg_mcast(netdev_priv(dev), "multicast join failed for "
 316                                         IPOIB_GID_FMT ", status %d\n",
 317                                         IPOIB_GID_ARG(mcast->mcmember.mgid), status);
 318
 319                 /* Flush out any queued packets */
 320                 while (!skb_queue_empty(&mcast->pkt_queue)) {
 321                         struct sk_buff *skb = skb_dequeue(&mcast->pkt_queue);
 322
 323                         skb->dev = dev;
 324
 325                         dev_kfree_skb_any(skb);
 326                 }
 327
 328                 /* Clear the busy flag so we try again */
 329                 clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 330         }
 331
 332         complete(&mcast->done);
 333 }
 334
 335 static int ipoib_mcast_sendonly_join(struct ipoib_mcast *mcast)
 336 {
 337         struct net_device *dev = mcast->dev;
 338         struct ipoib_dev_priv *priv = netdev_priv(dev);
 339         struct ib_sa_mcmember_rec rec = {
 340 #if 0                           /* Some SMs don't support send-only yet */
 341                 .join_state = 4
 342 #else
 343                 .join_state = 1
 344 #endif
 345         };
 346         int ret = 0;
 347
 348         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
 349                 ipoib_dbg_mcast(priv, "device shutting down, no multicast joins\n");
 350                 return -ENODEV;
 351         }
 352
 353         if (test_and_set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)) {
 354                 ipoib_dbg_mcast(priv, "multicast entry busy, skipping\n");
 355                 return -EBUSY;
 356         }
 357
 358         rec.mgid     = mcast->mcmember.mgid;
 359         rec.port_gid = priv->local_gid;
 360         rec.pkey     = be16_to_cpu(priv->pkey);
 361
 362         ret = ib_sa_mcmember_rec_set(priv->ca, priv->port, &rec,
 363                                      IB_SA_MCMEMBER_REC_MGID            |
 364                                      IB_SA_MCMEMBER_REC_PORT_GID        |
 365                                      IB_SA_MCMEMBER_REC_PKEY            |
 366                                      IB_SA_MCMEMBER_REC_JOIN_STATE,
 367                                      1000, GFP_ATOMIC,
 368                                      ipoib_mcast_sendonly_join_complete,
 369                                      mcast, &mcast->query);
 370         if (ret < 0) {
 371                 ipoib_warn(priv, "ib_sa_mcmember_rec_set failed (ret = %d)\n",
 372                            ret);
 373         } else {
 374                 ipoib_dbg_mcast(priv, "no multicast record for " IPOIB_GID_FMT
 375                                 ", starting join\n",
 376                                 IPOIB_GID_ARG(mcast->mcmember.mgid));
 377
 378                 mcast->query_id = ret;
 379         }
 380
 381         return ret;
 382 }
 383
 384 static void ipoib_mcast_join_complete(int status,
 385                                       struct ib_sa_mcmember_rec *mcmember,
 386                                       void *mcast_ptr)
 387 {
 388         struct ipoib_mcast *mcast = mcast_ptr;
 389         struct net_device *dev = mcast->dev;
 390         struct ipoib_dev_priv *priv = netdev_priv(dev);
 391
 392         ipoib_dbg_mcast(priv, "join completion for " IPOIB_GID_FMT
 393                         " (status %d)\n",
 394                         IPOIB_GID_ARG(mcast->mcmember.mgid), status);
 395
 396         if (!status && !ipoib_mcast_join_finish(mcast, mcmember)) {
 397                 mcast->backoff = HZ;
 398                 down(&mcast_mutex);
 399                 if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 400                         queue_work(ipoib_workqueue, &priv->mcast_task);
 401                 up(&mcast_mutex);
 402                 complete(&mcast->done);
 403                 return;
 404         }
 405
 406         if (status == -EINTR) {
 407                 complete(&mcast->done);
 408                 return;
 409         }
 410
 411         if (status && mcast->logcount++ < 20) {
 412                 if (status == -ETIMEDOUT || status == -EINTR) {
 413                         ipoib_dbg_mcast(priv, "multicast join failed for " IPOIB_GID_FMT
 414                                         ", status %d\n",
 415                                         IPOIB_GID_ARG(mcast->mcmember.mgid),
 416                                         status);
 417                 } else {
 418                         ipoib_warn(priv, "multicast join failed for "
 419                                    IPOIB_GID_FMT ", status %d\n",
 420                                    IPOIB_GID_ARG(mcast->mcmember.mgid),
 421                                    status);
 422                 }
 423         }
 424
 425         mcast->backoff *= 2;
 426         if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
 427                 mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
 428
 429         mcast->query = NULL;
 430
 431         down(&mcast_mutex);
 432         if (test_bit(IPOIB_MCAST_RUN, &priv->flags)) {
 433                 if (status == -ETIMEDOUT)
 434                         queue_work(ipoib_workqueue, &priv->mcast_task);
 435                 else
 436                         queue_delayed_work(ipoib_workqueue, &priv->mcast_task,
 437                                            mcast->backoff * HZ);
 438         } else
 439                 complete(&mcast->done);
 440         up(&mcast_mutex);
 441
 442         return;
 443 }
 444
 445 static void ipoib_mcast_join(struct net_device *dev, struct ipoib_mcast *mcast,
 446                              int create)
 447 {
 448         struct ipoib_dev_priv *priv = netdev_priv(dev);
 449         struct ib_sa_mcmember_rec rec = {
 450                 .join_state = 1
 451         };
 452         ib_sa_comp_mask comp_mask;
 453         int ret = 0;
 454
 455         ipoib_dbg_mcast(priv, "joining MGID " IPOIB_GID_FMT "\n",
 456                         IPOIB_GID_ARG(mcast->mcmember.mgid));
 457
 458         rec.mgid     = mcast->mcmember.mgid;
 459         rec.port_gid = priv->local_gid;
 460         rec.pkey     = be16_to_cpu(priv->pkey);
 461
 462         comp_mask =
 463                 IB_SA_MCMEMBER_REC_MGID         |
 464                 IB_SA_MCMEMBER_REC_PORT_GID     |
 465                 IB_SA_MCMEMBER_REC_PKEY         |
 466                 IB_SA_MCMEMBER_REC_JOIN_STATE;
 467
 468         if (create) {
 469                 comp_mask |=
 470                         IB_SA_MCMEMBER_REC_QKEY         |
 471                         IB_SA_MCMEMBER_REC_SL           |
 472                         IB_SA_MCMEMBER_REC_FLOW_LABEL   |
 473                         IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
 474
 475                 rec.qkey          = priv->broadcast->mcmember.qkey;
 476                 rec.sl            = priv->broadcast->mcmember.sl;
 477                 rec.flow_label    = priv->broadcast->mcmember.flow_label;
 478                 rec.traffic_class = priv->broadcast->mcmember.traffic_class;
 479         }
 480
 481         ret = ib_sa_mcmember_rec_set(priv->ca, priv->port, &rec, comp_mask,
 482                                      mcast->backoff * 1000, GFP_ATOMIC,
 483                                      ipoib_mcast_join_complete,
 484                                      mcast, &mcast->query);
 485
 486         if (ret < 0) {
 487                 ipoib_warn(priv, "ib_sa_mcmember_rec_set failed, status %d\n", ret);
 488
 489                 mcast->backoff *= 2;
 490                 if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
 491                         mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
 492
 493                 down(&mcast_mutex);
 494                 if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 495                         queue_delayed_work(ipoib_workqueue,
 496                                            &priv->mcast_task,
 497                                            mcast->backoff);
 498                 up(&mcast_mutex);
 499         } else
 500                 mcast->query_id = ret;
 501 }
 502
 503 void ipoib_mcast_join_task(void *dev_ptr)
 504 {
 505         struct net_device *dev = dev_ptr;
 506         struct ipoib_dev_priv *priv = netdev_priv(dev);
 507
 508         if (!test_bit(IPOIB_MCAST_RUN, &priv->flags))
 509                 return;
 510
 511         if (ib_query_gid(priv->ca, priv->port, 0, &priv->local_gid))
 512                 ipoib_warn(priv, "ib_gid_entry_get() failed\n");
 513         else
 514                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
 515
 516         {
 517                 struct ib_port_attr attr;
 518
 519                 if (!ib_query_port(priv->ca, priv->port, &attr)) {
 520                         priv->local_lid  = attr.lid;
 521                         priv->local_rate = attr.active_speed *
 522                                 ib_width_enum_to_int(attr.active_width);
 523                 } else
 524                         ipoib_warn(priv, "ib_query_port failed\n");
 525         }
 526
 527         if (!priv->broadcast) {
 528                 priv->broadcast = ipoib_mcast_alloc(dev, 1);
 529                 if (!priv->broadcast) {
 530                         ipoib_warn(priv, "failed to allocate broadcast group\n");
 531                         down(&mcast_mutex);
 532                         if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 533                                 queue_delayed_work(ipoib_workqueue,
 534                                                    &priv->mcast_task, HZ);
 535                         up(&mcast_mutex);
 536                         return;
 537                 }
 538
 539                 memcpy(priv->broadcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
 540                        sizeof (union ib_gid));
 541
 542                 spin_lock_irq(&priv->lock);
 543                 __ipoib_mcast_add(dev, priv->broadcast);
 544                 spin_unlock_irq(&priv->lock);
 545         }
 546
 547         if (!test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
 548                 ipoib_mcast_join(dev, priv->broadcast, 0);
 549                 return;
 550         }
 551
 552         while (1) {
 553                 struct ipoib_mcast *mcast = NULL;
 554
 555                 spin_lock_irq(&priv->lock);
 556                 list_for_each_entry(mcast, &priv->multicast_list, list) {
 557                         if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)
 558                             && !test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)
 559                             && !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 560                                 /* Found the next unjoined group */
 561                                 break;
 562                         }
 563                 }
 564                 spin_unlock_irq(&priv->lock);
 565
 566                 if (&mcast->list == &priv->multicast_list) {
 567                         /* All done */
 568                         break;
 569                 }
 570
 571                 ipoib_mcast_join(dev, mcast, 1);
 572                 return;
 573         }
 574
 575         priv->mcast_mtu = ib_mtu_enum_to_int(priv->broadcast->mcmember.mtu) -
 576                 IPOIB_ENCAP_LEN;
 577         dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
 578
 579         ipoib_dbg_mcast(priv, "successfully joined all multicast groups\n");
 580
 581         clear_bit(IPOIB_MCAST_RUN, &priv->flags);
 582         netif_carrier_on(dev);
 583 }
 584
 585 int ipoib_mcast_start_thread(struct net_device *dev)
 586 {
 587         struct ipoib_dev_priv *priv = netdev_priv(dev);
 588
 589         ipoib_dbg_mcast(priv, "starting multicast thread\n");
 590
 591         down(&mcast_mutex);
 592         if (!test_and_set_bit(IPOIB_MCAST_RUN, &priv->flags))
 593                 queue_work(ipoib_workqueue, &priv->mcast_task);
 594         up(&mcast_mutex);
 595
 596         return 0;
 597 }
 598
 599 int ipoib_mcast_stop_thread(struct net_device *dev)
 600 {
 601         struct ipoib_dev_priv *priv = netdev_priv(dev);
 602         struct ipoib_mcast *mcast;
 603
 604         ipoib_dbg_mcast(priv, "stopping multicast thread\n");
 605
 606         down(&mcast_mutex);
 607         clear_bit(IPOIB_MCAST_RUN, &priv->flags);
 608         cancel_delayed_work(&priv->mcast_task);
 609         up(&mcast_mutex);
 610
 611         flush_workqueue(ipoib_workqueue);
 612
 613         if (priv->broadcast && priv->broadcast->query) {
 614                 ib_sa_cancel_query(priv->broadcast->query_id, priv->broadcast->query);
 615                 priv->broadcast->query = NULL;
 616                 ipoib_dbg_mcast(priv, "waiting for bcast\n");
 617                 wait_for_completion(&priv->broadcast->done);
 618         }
 619
 620         list_for_each_entry(mcast, &priv->multicast_list, list) {
 621                 if (mcast->query) {
 622                         ib_sa_cancel_query(mcast->query_id, mcast->query);
 623                         mcast->query = NULL;
 624                         ipoib_dbg_mcast(priv, "waiting for MGID " IPOIB_GID_FMT "\n",
 625                                         IPOIB_GID_ARG(mcast->mcmember.mgid));
 626                         wait_for_completion(&mcast->done);
 627                 }
 628         }
 629
 630         return 0;
 631 }
 632
 633 static int ipoib_mcast_leave(struct net_device *dev, struct ipoib_mcast *mcast)
 634 {
 635         struct ipoib_dev_priv *priv = netdev_priv(dev);
 636         struct ib_sa_mcmember_rec rec = {
 637                 .join_state = 1
 638         };
 639         int ret = 0;
 640
 641         if (!test_and_clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags))
 642                 return 0;
 643
 644         ipoib_dbg_mcast(priv, "leaving MGID " IPOIB_GID_FMT "\n",
 645                         IPOIB_GID_ARG(mcast->mcmember.mgid));
 646
 647         rec.mgid     = mcast->mcmember.mgid;
 648         rec.port_gid = priv->local_gid;
 649         rec.pkey     = be16_to_cpu(priv->pkey);
 650
 651         /* Remove ourselves from the multicast group */
 652         ret = ipoib_mcast_detach(dev, be16_to_cpu(mcast->mcmember.mlid),
 653                                  &mcast->mcmember.mgid);
 654         if (ret)
 655                 ipoib_warn(priv, "ipoib_mcast_detach failed (result = %d)\n", ret);
 656
 657         /*
 658          * Just make one shot at leaving and don't wait for a reply;
 659          * if we fail, too bad.
 660          */
 661         ret = ib_sa_mcmember_rec_delete(priv->ca, priv->port, &rec,
 662                                         IB_SA_MCMEMBER_REC_MGID         |
 663                                         IB_SA_MCMEMBER_REC_PORT_GID     |
 664                                         IB_SA_MCMEMBER_REC_PKEY         |
 665                                         IB_SA_MCMEMBER_REC_JOIN_STATE,
 666                                         0, GFP_ATOMIC, NULL,
 667                                         mcast, &mcast->query);
 668         if (ret < 0)
 669                 ipoib_warn(priv, "ib_sa_mcmember_rec_delete failed "
 670                            "for leave (result = %d)\n", ret);
 671
 672         return 0;
 673 }
 674
 675 void ipoib_mcast_send(struct net_device *dev, union ib_gid *mgid,
 676                       struct sk_buff *skb)
 677 {
 678         struct ipoib_dev_priv *priv = netdev_priv(dev);
 679         struct ipoib_mcast *mcast;
 680
 681         /*
 682          * We can only be called from ipoib_start_xmit, so we're
 683          * inside tx_lock -- no need to save/restore flags.
 684          */
 685         spin_lock(&priv->lock);
 686
 687         mcast = __ipoib_mcast_find(dev, mgid);
 688         if (!mcast) {
 689                 /* Let's create a new send only group now */
 690                 ipoib_dbg_mcast(priv, "setting up send only multicast group for "
 691                                 IPOIB_GID_FMT "\n", IPOIB_GID_ARG(*mgid));
 692
 693                 mcast = ipoib_mcast_alloc(dev, 0);
 694                 if (!mcast) {
 695                         ipoib_warn(priv, "unable to allocate memory for "
 696                                    "multicast structure\n");
 697                         dev_kfree_skb_any(skb);
 698                         goto out;
 699                 }
 700
 701                 set_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags);
 702                 mcast->mcmember.mgid = *mgid;
 703                 __ipoib_mcast_add(dev, mcast);
 704                 list_add_tail(&mcast->list, &priv->multicast_list);
 705         }
 706
 707         if (!mcast->ah) {
 708                 if (skb_queue_len(&mcast->pkt_queue) < IPOIB_MAX_MCAST_QUEUE)
 709                         skb_queue_tail(&mcast->pkt_queue, skb);
 710                 else
 711                         dev_kfree_skb_any(skb);
 712
 713                 if (mcast->query)
 714                         ipoib_dbg_mcast(priv, "no address vector, "
 715                                         "but multicast join already started\n");
 716                 else if (test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags))
 717                         ipoib_mcast_sendonly_join(mcast);
 718
 719                 /*
 720                  * If lookup completes between here and out:, don't
 721                  * want to send packet twice.
 722                  */
 723                 mcast = NULL;
 724         }
 725
 726 out:
 727         if (mcast && mcast->ah) {
 728                 if (skb->dst            &&
 729                     skb->dst->neighbour &&
 730                     !*to_ipoib_neigh(skb->dst->neighbour)) {
 731                         struct ipoib_neigh *neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
 732
 733                         if (neigh) {
 734                                 kref_get(&mcast->ah->ref);
 735                                 neigh->ah       = mcast->ah;
 736                                 neigh->neighbour = skb->dst->neighbour;
 737                                 *to_ipoib_neigh(skb->dst->neighbour) = neigh;
 738                                 list_add_tail(&neigh->list, &mcast->neigh_list);
 739                         }
 740                 }
 741
 742                 ipoib_send(dev, skb, mcast->ah, IB_MULTICAST_QPN);
 743         }
 744
 745         spin_unlock(&priv->lock);
 746 }
 747
 748 void ipoib_mcast_dev_flush(struct net_device *dev)
 749 {
 750         struct ipoib_dev_priv *priv = netdev_priv(dev);
 751         LIST_HEAD(remove_list);
 752         struct ipoib_mcast *mcast, *tmcast, *nmcast;
 753         unsigned long flags;
 754
 755         ipoib_dbg_mcast(priv, "flushing multicast list\n");
 756
 757         spin_lock_irqsave(&priv->lock, flags);
 758         list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
 759                 nmcast = ipoib_mcast_alloc(dev, 0);
 760                 if (nmcast) {
 761                         nmcast->flags =
 762                                 mcast->flags & (1 << IPOIB_MCAST_FLAG_SENDONLY);
 763
 764                         nmcast->mcmember.mgid = mcast->mcmember.mgid;
 765
 766                         /* Add the new group in before the to-be-destroyed group */
 767                         list_add_tail(&nmcast->list, &mcast->list);
 768                         list_del_init(&mcast->list);
 769
 770                         rb_replace_node(&mcast->rb_node, &nmcast->rb_node,
 771                                         &priv->multicast_tree);
 772
 773                         list_add_tail(&mcast->list, &remove_list);
 774                 } else {
 775                         ipoib_warn(priv, "could not reallocate multicast group "
 776                                    IPOIB_GID_FMT "\n",
 777                                    IPOIB_GID_ARG(mcast->mcmember.mgid));
 778                 }
 779         }
 780
 781         if (priv->broadcast) {
 782                 nmcast = ipoib_mcast_alloc(dev, 0);
 783                 if (nmcast) {
 784                         nmcast->mcmember.mgid = priv->broadcast->mcmember.mgid;
 785
 786                         rb_replace_node(&priv->broadcast->rb_node,
 787                                         &nmcast->rb_node,
 788                                         &priv->multicast_tree);
 789
 790                         list_add_tail(&priv->broadcast->list, &remove_list);
 791                 }
 792
 793                 priv->broadcast = nmcast;
 794         }
 795
 796         spin_unlock_irqrestore(&priv->lock, flags);
 797
 798         list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
 799                 ipoib_mcast_leave(dev, mcast);
 800                 ipoib_mcast_free(mcast);
 801         }
 802 }
 803
 804 void ipoib_mcast_dev_down(struct net_device *dev)
 805 {
 806         struct ipoib_dev_priv *priv = netdev_priv(dev);
 807         unsigned long flags;
 808
 809         /* Delete broadcast since it will be recreated */
 810         if (priv->broadcast) {
 811                 ipoib_dbg_mcast(priv, "deleting broadcast group\n");
 812
 813                 spin_lock_irqsave(&priv->lock, flags);
 814                 rb_erase(&priv->broadcast->rb_node, &priv->multicast_tree);
 815                 spin_unlock_irqrestore(&priv->lock, flags);
 816                 ipoib_mcast_leave(dev, priv->broadcast);
 817                 ipoib_mcast_free(priv->broadcast);
 818                 priv->broadcast = NULL;
 819         }
 820 }
 821
 822 void ipoib_mcast_restart_task(void *dev_ptr)
 823 {
 824         struct net_device *dev = dev_ptr;
 825         struct ipoib_dev_priv *priv = netdev_priv(dev);
 826         struct dev_mc_list *mclist;
 827         struct ipoib_mcast *mcast, *tmcast;
 828         LIST_HEAD(remove_list);
 829         unsigned long flags;
 830
 831         ipoib_dbg_mcast(priv, "restarting multicast task\n");
 832
 833         ipoib_mcast_stop_thread(dev);
 834
 835         spin_lock_irqsave(&priv->lock, flags);
 836
 837         /*
 838          * Unfortunately, the networking core only gives us a list of all of
 839          * the multicast hardware addresses. We need to figure out which ones
 840          * are new and which ones have been removed
 841          */
 842
 843         /* Clear out the found flag */
 844         list_for_each_entry(mcast, &priv->multicast_list, list)
 845                 clear_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
 846
 847         /* Mark all of the entries that are found or don't exist */
 848         for (mclist = dev->mc_list; mclist; mclist = mclist->next) {
 849                 union ib_gid mgid;
 850
 851                 memcpy(mgid.raw, mclist->dmi_addr + 4, sizeof mgid);
 852
 853                 /* Add in the P_Key */
 854                 mgid.raw[4] = (priv->pkey >> 8) & 0xff;
 855                 mgid.raw[5] = priv->pkey & 0xff;
 856
 857                 mcast = __ipoib_mcast_find(dev, &mgid);
 858                 if (!mcast || test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 859                         struct ipoib_mcast *nmcast;
 860
 861                         /* Not found or send-only group, let's add a new entry */
 862                         ipoib_dbg_mcast(priv, "adding multicast entry for mgid "
 863                                         IPOIB_GID_FMT "\n", IPOIB_GID_ARG(mgid));
 864
 865                         nmcast = ipoib_mcast_alloc(dev, 0);
 866                         if (!nmcast) {
 867                                 ipoib_warn(priv, "unable to allocate memory for multicast structure\n");
 868                                 continue;
 869                         }
 870
 871                         set_bit(IPOIB_MCAST_FLAG_FOUND, &nmcast->flags);
 872
 873                         nmcast->mcmember.mgid = mgid;
 874
 875                         if (mcast) {
 876                                 /* Destroy the send only entry */
 877                                 list_del(&mcast->list);
 878                                 list_add_tail(&mcast->list, &remove_list);
 879
 880                                 rb_replace_node(&mcast->rb_node,
 881                                                 &nmcast->rb_node,
 882                                                 &priv->multicast_tree);
 883                         } else
 884                                 __ipoib_mcast_add(dev, nmcast);
 885
 886                         list_add_tail(&nmcast->list, &priv->multicast_list);
 887                 }
 888
 889                 if (mcast)
 890                         set_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
 891         }
 892
 893         /* Remove all of the entries don't exist anymore */
 894         list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
 895                 if (!test_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags) &&
 896                     !test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 897                         ipoib_dbg_mcast(priv, "deleting multicast group " IPOIB_GID_FMT "\n",
 898                                         IPOIB_GID_ARG(mcast->mcmember.mgid));
 899
 900                         rb_erase(&mcast->rb_node, &priv->multicast_tree);
 901
 902                         /* Move to the remove list */
 903                         list_del(&mcast->list);
 904                         list_add_tail(&mcast->list, &remove_list);
 905                 }
 906         }
 907         spin_unlock_irqrestore(&priv->lock, flags);
 908
 909         /* We have to cancel outside of the spinlock */
 910         list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
 911                 ipoib_mcast_leave(mcast->dev, mcast);
 912                 ipoib_mcast_free(mcast);
 913         }
 914
 915         if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 916                 ipoib_mcast_start_thread(dev);
 917 }
 918
 919 struct ipoib_mcast_iter *ipoib_mcast_iter_init(struct net_device *dev)
 920 {
 921         struct ipoib_mcast_iter *iter;
 922
 923         iter = kmalloc(sizeof *iter, GFP_KERNEL);
 924         if (!iter)
 925                 return NULL;
 926
 927         iter->dev = dev;
 928         memset(iter->mgid.raw, 0, sizeof iter->mgid);
 929
 930         if (ipoib_mcast_iter_next(iter)) {
 931                 ipoib_mcast_iter_free(iter);
 932                 return NULL;
 933         }
 934
 935         return iter;
 936 }
 937
 938 void ipoib_mcast_iter_free(struct ipoib_mcast_iter *iter)
 939 {
 940         kfree(iter);
 941 }
 942
 943 int ipoib_mcast_iter_next(struct ipoib_mcast_iter *iter)
 944 {
 945         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
 946         struct rb_node *n;
 947         struct ipoib_mcast *mcast;
 948         int ret = 1;
 949
 950         spin_lock_irq(&priv->lock);
 951
 952         n = rb_first(&priv->multicast_tree);
 953
 954         while (n) {
 955                 mcast = rb_entry(n, struct ipoib_mcast, rb_node);
 956
 957                 if (memcmp(iter->mgid.raw, mcast->mcmember.mgid.raw,
 958                            sizeof (union ib_gid)) < 0) {
 959                         iter->mgid      = mcast->mcmember.mgid;
 960                         iter->created   = mcast->created;
 961                         iter->queuelen  = skb_queue_len(&mcast->pkt_queue);
 962                         iter->complete  = !!mcast->ah;
 963                         iter->send_only = !!(mcast->flags & (1 << IPOIB_MCAST_FLAG_SENDONLY));
 964
 965                         ret = 0;
 966
 967                         break;
 968                 }
 969
 970                 n = rb_next(n);
 971         }
 972
 973         spin_unlock_irq(&priv->lock);
 974
 975         return ret;
 976 }
 977
 978 void ipoib_mcast_iter_read(struct ipoib_mcast_iter *iter,
 979                            union ib_gid *mgid,
 980                            unsigned long *created,
 981                            unsigned int *queuelen,
 982                            unsigned int *complete,
 983                            unsigned int *send_only)
 984 {
 985         *mgid      = iter->mgid;
 986         *created   = iter->created;
 987         *queuelen  = iter->queuelen;
 988         *complete  = iter->complete;
 989         *send_only = iter->send_only;
 990 }