drivers/infiniband/ulp/ipoib/ipoib_multicast.c

   1 /*
   2  * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
   3  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
   4  * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
   5  *
   6  * This software is available to you under a choice of one of two
   7  * licenses.  You may choose to be licensed under the terms of the GNU
   8  * General Public License (GPL) Version 2, available from the file
   9  * COPYING in the main directory of this source tree, or the
  10  * OpenIB.org BSD license below:
  11  *
  12  *     Redistribution and use in source and binary forms, with or
  13  *     without modification, are permitted provided that the following
  14  *     conditions are met:
  15  *
  16  *      - Redistributions of source code must retain the above
  17  *        copyright notice, this list of conditions and the following
  18  *        disclaimer.
  19  *
  20  *      - Redistributions in binary form must reproduce the above
  21  *        copyright notice, this list of conditions and the following
  22  *        disclaimer in the documentation and/or other materials
  23  *        provided with the distribution.
  24  *
  25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  32  * SOFTWARE.
  33  */
  34
  35 #include <linux/skbuff.h>
  36 #include <linux/rtnetlink.h>
  37 #include <linux/moduleparam.h>
  38 #include <linux/ip.h>
  39 #include <linux/in.h>
  40 #include <linux/igmp.h>
  41 #include <linux/inetdevice.h>
  42 #include <linux/delay.h>
  43 #include <linux/completion.h>
  44 #include <linux/slab.h>
  45
  46 #include <net/dst.h>
  47
  48 #include "ipoib.h"
  49
  50 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
  51 static int mcast_debug_level;
  52
  53 module_param(mcast_debug_level, int, 0644);
  54 MODULE_PARM_DESC(mcast_debug_level,
  55                  "Enable multicast debug tracing if > 0");
  56 #endif
  57
  58 static DEFINE_MUTEX(mcast_mutex);
  59
  60 struct ipoib_mcast_iter {
  61         struct net_device *dev;
  62         union ib_gid       mgid;
  63         unsigned long      created;
  64         unsigned int       queuelen;
  65         unsigned int       complete;
  66         unsigned int       send_only;
  67 };
  68
  69 /*
  70  * This should be called with the mcast_mutex held
  71  */
  72 static void __ipoib_mcast_schedule_join_thread(struct ipoib_dev_priv *priv,
  73                                                struct ipoib_mcast *mcast,
  74                                                bool delay)
  75 {
  76         if (!test_bit(IPOIB_MCAST_RUN, &priv->flags))
  77                 return;
  78
  79         /*
  80          * We will be scheduling *something*, so cancel whatever is
  81          * currently scheduled first
  82          */
  83         cancel_delayed_work(&priv->mcast_task);
  84         if (mcast && delay) {
  85                 /*
  86                  * We had a failure and want to schedule a retry later
  87                  */
  88                 mcast->backoff *= 2;
  89                 if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
  90                         mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
  91                 mcast->delay_until = jiffies + (mcast->backoff * HZ);
  92                 /*
  93                  * Mark this mcast for its delay, but restart the
  94                  * task immediately.  The join task will make sure to
  95                  * clear out all entries without delays, and then
  96                  * schedule itself to run again when the earliest
  97                  * delay expires
  98                  */
  99                 queue_delayed_work(priv->wq, &priv->mcast_task, 0);
 100         } else if (delay) {
 101                 /*
 102                  * Special case of retrying after a failure to
 103                  * allocate the broadcast multicast group, wait
 104                  * 1 second and try again
 105                  */
 106                 queue_delayed_work(priv->wq, &priv->mcast_task, HZ);
 107         } else
 108                 queue_delayed_work(priv->wq, &priv->mcast_task, 0);
 109 }
 110
 111 static void ipoib_mcast_free(struct ipoib_mcast *mcast)
 112 {
 113         struct net_device *dev = mcast->dev;
 114         int tx_dropped = 0;
 115
 116         ipoib_dbg_mcast(netdev_priv(dev), "deleting multicast group %pI6\n",
 117                         mcast->mcmember.mgid.raw);
 118
 119         /* remove all neigh connected to this mcast */
 120         ipoib_del_neighs_by_gid(dev, mcast->mcmember.mgid.raw);
 121
 122         if (mcast->ah)
 123                 ipoib_put_ah(mcast->ah);
 124
 125         while (!skb_queue_empty(&mcast->pkt_queue)) {
 126                 ++tx_dropped;
 127                 dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
 128         }
 129
 130         netif_tx_lock_bh(dev);
 131         dev->stats.tx_dropped += tx_dropped;
 132         netif_tx_unlock_bh(dev);
 133
 134         kfree(mcast);
 135 }
 136
 137 static struct ipoib_mcast *ipoib_mcast_alloc(struct net_device *dev,
 138                                              int can_sleep)
 139 {
 140         struct ipoib_mcast *mcast;
 141
 142         mcast = kzalloc(sizeof *mcast, can_sleep ? GFP_KERNEL : GFP_ATOMIC);
 143         if (!mcast)
 144                 return NULL;
 145
 146         mcast->dev = dev;
 147         mcast->created = jiffies;
 148         mcast->delay_until = jiffies;
 149         mcast->backoff = 1;
 150
 151         INIT_LIST_HEAD(&mcast->list);
 152         INIT_LIST_HEAD(&mcast->neigh_list);
 153         skb_queue_head_init(&mcast->pkt_queue);
 154
 155         return mcast;
 156 }
 157
 158 static struct ipoib_mcast *__ipoib_mcast_find(struct net_device *dev, void *mgid)
 159 {
 160         struct ipoib_dev_priv *priv = netdev_priv(dev);
 161         struct rb_node *n = priv->multicast_tree.rb_node;
 162
 163         while (n) {
 164                 struct ipoib_mcast *mcast;
 165                 int ret;
 166
 167                 mcast = rb_entry(n, struct ipoib_mcast, rb_node);
 168
 169                 ret = memcmp(mgid, mcast->mcmember.mgid.raw,
 170                              sizeof (union ib_gid));
 171                 if (ret < 0)
 172                         n = n->rb_left;
 173                 else if (ret > 0)
 174                         n = n->rb_right;
 175                 else
 176                         return mcast;
 177         }
 178
 179         return NULL;
 180 }
 181
 182 static int __ipoib_mcast_add(struct net_device *dev, struct ipoib_mcast *mcast)
 183 {
 184         struct ipoib_dev_priv *priv = netdev_priv(dev);
 185         struct rb_node **n = &priv->multicast_tree.rb_node, *pn = NULL;
 186
 187         while (*n) {
 188                 struct ipoib_mcast *tmcast;
 189                 int ret;
 190
 191                 pn = *n;
 192                 tmcast = rb_entry(pn, struct ipoib_mcast, rb_node);
 193
 194                 ret = memcmp(mcast->mcmember.mgid.raw, tmcast->mcmember.mgid.raw,
 195                              sizeof (union ib_gid));
 196                 if (ret < 0)
 197                         n = &pn->rb_left;
 198                 else if (ret > 0)
 199                         n = &pn->rb_right;
 200                 else
 201                         return -EEXIST;
 202         }
 203
 204         rb_link_node(&mcast->rb_node, pn, n);
 205         rb_insert_color(&mcast->rb_node, &priv->multicast_tree);
 206
 207         return 0;
 208 }
 209
 210 static int ipoib_mcast_join_finish(struct ipoib_mcast *mcast,
 211                                    struct ib_sa_mcmember_rec *mcmember)
 212 {
 213         struct net_device *dev = mcast->dev;
 214         struct ipoib_dev_priv *priv = netdev_priv(dev);
 215         struct ipoib_ah *ah;
 216         int ret;
 217         int set_qkey = 0;
 218
 219         mcast->mcmember = *mcmember;
 220
 221         /* Set the multicast MTU and cached Q_Key before we attach if it's
 222          * the broadcast group.
 223          */
 224         if (!memcmp(mcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
 225                     sizeof (union ib_gid))) {
 226                 spin_lock_irq(&priv->lock);
 227                 if (!priv->broadcast) {
 228                         spin_unlock_irq(&priv->lock);
 229                         return -EAGAIN;
 230                 }
 231                 priv->mcast_mtu = IPOIB_UD_MTU(ib_mtu_enum_to_int(priv->broadcast->mcmember.mtu));
 232                 priv->qkey = be32_to_cpu(priv->broadcast->mcmember.qkey);
 233                 spin_unlock_irq(&priv->lock);
 234                 priv->tx_wr.wr.ud.remote_qkey = priv->qkey;
 235                 set_qkey = 1;
 236         }
 237
 238         if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 239                 if (test_and_set_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 240                         ipoib_warn(priv, "multicast group %pI6 already attached\n",
 241                                    mcast->mcmember.mgid.raw);
 242
 243                         return 0;
 244                 }
 245
 246                 ret = ipoib_mcast_attach(dev, be16_to_cpu(mcast->mcmember.mlid),
 247                                          &mcast->mcmember.mgid, set_qkey);
 248                 if (ret < 0) {
 249                         ipoib_warn(priv, "couldn't attach QP to multicast group %pI6\n",
 250                                    mcast->mcmember.mgid.raw);
 251
 252                         clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags);
 253                         return ret;
 254                 }
 255         }
 256
 257         {
 258                 struct ib_ah_attr av = {
 259                         .dlid          = be16_to_cpu(mcast->mcmember.mlid),
 260                         .port_num      = priv->port,
 261                         .sl            = mcast->mcmember.sl,
 262                         .ah_flags      = IB_AH_GRH,
 263                         .static_rate   = mcast->mcmember.rate,
 264                         .grh           = {
 265                                 .flow_label    = be32_to_cpu(mcast->mcmember.flow_label),
 266                                 .hop_limit     = mcast->mcmember.hop_limit,
 267                                 .sgid_index    = 0,
 268                                 .traffic_class = mcast->mcmember.traffic_class
 269                         }
 270                 };
 271                 av.grh.dgid = mcast->mcmember.mgid;
 272
 273                 ah = ipoib_create_ah(dev, priv->pd, &av);
 274                 if (IS_ERR(ah)) {
 275                         ipoib_warn(priv, "ib_address_create failed %ld\n",
 276                                 -PTR_ERR(ah));
 277                         /* use original error */
 278                         return PTR_ERR(ah);
 279                 } else {
 280                         spin_lock_irq(&priv->lock);
 281                         mcast->ah = ah;
 282                         spin_unlock_irq(&priv->lock);
 283
 284                         ipoib_dbg_mcast(priv, "MGID %pI6 AV %p, LID 0x%04x, SL %d\n",
 285                                         mcast->mcmember.mgid.raw,
 286                                         mcast->ah->ah,
 287                                         be16_to_cpu(mcast->mcmember.mlid),
 288                                         mcast->mcmember.sl);
 289                 }
 290         }
 291
 292         /* actually send any queued packets */
 293         netif_tx_lock_bh(dev);
 294         while (!skb_queue_empty(&mcast->pkt_queue)) {
 295                 struct sk_buff *skb = skb_dequeue(&mcast->pkt_queue);
 296
 297                 netif_tx_unlock_bh(dev);
 298
 299                 skb->dev = dev;
 300                 if (dev_queue_xmit(skb))
 301                         ipoib_warn(priv, "dev_queue_xmit failed to requeue packet\n");
 302
 303                 netif_tx_lock_bh(dev);
 304         }
 305         netif_tx_unlock_bh(dev);
 306
 307         return 0;
 308 }
 309
 310 static int
 311 ipoib_mcast_sendonly_join_complete(int status,
 312                                    struct ib_sa_multicast *multicast)
 313 {
 314         struct ipoib_mcast *mcast = multicast->context;
 315         struct net_device *dev = mcast->dev;
 316         struct ipoib_dev_priv *priv = netdev_priv(dev);
 317
 318         /*
 319          * We have to take the mutex to force mcast_sendonly_join to
 320          * return from ib_sa_multicast_join and set mcast->mc to a
 321          * valid value.  Otherwise we were racing with ourselves in
 322          * that we might fail here, but get a valid return from
 323          * ib_sa_multicast_join after we had cleared mcast->mc here,
 324          * resulting in mis-matched joins and leaves and a deadlock
 325          */
 326         mutex_lock(&mcast_mutex);
 327
 328         /* We trap for port events ourselves. */
 329         if (status == -ENETRESET) {
 330                 status = 0;
 331                 goto out;
 332         }
 333
 334         if (!status)
 335                 status = ipoib_mcast_join_finish(mcast, &multicast->rec);
 336
 337         if (status) {
 338                 if (mcast->logcount++ < 20)
 339                         ipoib_dbg_mcast(netdev_priv(dev), "sendonly multicast "
 340                                         "join failed for %pI6, status %d\n",
 341                                         mcast->mcmember.mgid.raw, status);
 342
 343                 /* Flush out any queued packets */
 344                 netif_tx_lock_bh(dev);
 345                 while (!skb_queue_empty(&mcast->pkt_queue)) {
 346                         ++dev->stats.tx_dropped;
 347                         dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
 348                 }
 349                 netif_tx_unlock_bh(dev);
 350                 __ipoib_mcast_schedule_join_thread(priv, mcast, 1);
 351         } else {
 352                 mcast->backoff = 1;
 353                 mcast->delay_until = jiffies;
 354                 __ipoib_mcast_schedule_join_thread(priv, NULL, 0);
 355         }
 356 out:
 357         clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 358         if (status)
 359                 mcast->mc = NULL;
 360         complete(&mcast->done);
 361         mutex_unlock(&mcast_mutex);
 362         return status;
 363 }
 364
 365 static int ipoib_mcast_sendonly_join(struct ipoib_mcast *mcast)
 366 {
 367         struct net_device *dev = mcast->dev;
 368         struct ipoib_dev_priv *priv = netdev_priv(dev);
 369         struct ib_sa_mcmember_rec rec = {
 370 #if 0                           /* Some SMs don't support send-only yet */
 371                 .join_state = 4
 372 #else
 373                 .join_state = 1
 374 #endif
 375         };
 376         int ret = 0;
 377
 378         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
 379                 ipoib_dbg_mcast(priv, "device shutting down, no sendonly "
 380                                 "multicast joins\n");
 381                 clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 382                 complete(&mcast->done);
 383                 return -ENODEV;
 384         }
 385
 386         rec.mgid     = mcast->mcmember.mgid;
 387         rec.port_gid = priv->local_gid;
 388         rec.pkey     = cpu_to_be16(priv->pkey);
 389
 390         mutex_lock(&mcast_mutex);
 391         mcast->mc = ib_sa_join_multicast(&ipoib_sa_client, priv->ca,
 392                                          priv->port, &rec,
 393                                          IB_SA_MCMEMBER_REC_MGID        |
 394                                          IB_SA_MCMEMBER_REC_PORT_GID    |
 395                                          IB_SA_MCMEMBER_REC_PKEY        |
 396                                          IB_SA_MCMEMBER_REC_JOIN_STATE,
 397                                          GFP_ATOMIC,
 398                                          ipoib_mcast_sendonly_join_complete,
 399                                          mcast);
 400         if (IS_ERR(mcast->mc)) {
 401                 ret = PTR_ERR(mcast->mc);
 402                 clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 403                 ipoib_warn(priv, "ib_sa_join_multicast for sendonly join "
 404                            "failed (ret = %d)\n", ret);
 405                 complete(&mcast->done);
 406         } else {
 407                 ipoib_dbg_mcast(priv, "no multicast record for %pI6, starting "
 408                                 "sendonly join\n", mcast->mcmember.mgid.raw);
 409         }
 410         mutex_unlock(&mcast_mutex);
 411
 412         return ret;
 413 }
 414
 415 void ipoib_mcast_carrier_on_task(struct work_struct *work)
 416 {
 417         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
 418                                                    carrier_on_task);
 419         struct ib_port_attr attr;
 420
 421         if (ib_query_port(priv->ca, priv->port, &attr) ||
 422             attr.state != IB_PORT_ACTIVE) {
 423                 ipoib_dbg(priv, "Keeping carrier off until IB port is active\n");
 424                 return;
 425         }
 426
 427         /*
 428          * Take rtnl_lock to avoid racing with ipoib_stop() and
 429          * turning the carrier back on while a device is being
 430          * removed.  However, ipoib_stop() will attempt to flush
 431          * the workqueue while holding the rtnl lock, so loop
 432          * on trylock until either we get the lock or we see
 433          * FLAG_OPER_UP go away as that signals that we are bailing
 434          * and can safely ignore the carrier on work.
 435          */
 436         while (!rtnl_trylock()) {
 437                 if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
 438                         return;
 439                 else
 440                         msleep(20);
 441         }
 442         if (!ipoib_cm_admin_enabled(priv->dev))
 443                 dev_set_mtu(priv->dev, min(priv->mcast_mtu, priv->admin_mtu));
 444         netif_carrier_on(priv->dev);
 445         rtnl_unlock();
 446 }
 447
 448 static int ipoib_mcast_join_complete(int status,
 449                                      struct ib_sa_multicast *multicast)
 450 {
 451         struct ipoib_mcast *mcast = multicast->context;
 452         struct net_device *dev = mcast->dev;
 453         struct ipoib_dev_priv *priv = netdev_priv(dev);
 454
 455         ipoib_dbg_mcast(priv, "join completion for %pI6 (status %d)\n",
 456                         mcast->mcmember.mgid.raw, status);
 457
 458         /*
 459          * We have to take the mutex to force mcast_join to
 460          * return from ib_sa_multicast_join and set mcast->mc to a
 461          * valid value.  Otherwise we were racing with ourselves in
 462          * that we might fail here, but get a valid return from
 463          * ib_sa_multicast_join after we had cleared mcast->mc here,
 464          * resulting in mis-matched joins and leaves and a deadlock
 465          */
 466         mutex_lock(&mcast_mutex);
 467
 468         /* We trap for port events ourselves. */
 469         if (status == -ENETRESET) {
 470                 status = 0;
 471                 goto out;
 472         }
 473
 474         if (!status)
 475                 status = ipoib_mcast_join_finish(mcast, &multicast->rec);
 476
 477         if (!status) {
 478                 mcast->backoff = 1;
 479                 mcast->delay_until = jiffies;
 480                 __ipoib_mcast_schedule_join_thread(priv, NULL, 0);
 481
 482                 /*
 483                  * Defer carrier on work to priv->wq to avoid a
 484                  * deadlock on rtnl_lock here.
 485                  */
 486                 if (mcast == priv->broadcast)
 487                         queue_work(priv->wq, &priv->carrier_on_task);
 488         } else {
 489                 if (mcast->logcount++ < 20) {
 490                         if (status == -ETIMEDOUT || status == -EAGAIN) {
 491                                 ipoib_dbg_mcast(priv, "multicast join failed for %pI6, status %d\n",
 492                                                 mcast->mcmember.mgid.raw, status);
 493                         } else {
 494                                 ipoib_warn(priv, "multicast join failed for %pI6, status %d\n",
 495                                            mcast->mcmember.mgid.raw, status);
 496                         }
 497                 }
 498
 499                 /* Requeue this join task with a backoff delay */
 500                 __ipoib_mcast_schedule_join_thread(priv, mcast, 1);
 501         }
 502 out:
 503         clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 504         if (status)
 505                 mcast->mc = NULL;
 506         complete(&mcast->done);
 507         mutex_unlock(&mcast_mutex);
 508         return status;
 509 }
 510
 511 static void ipoib_mcast_join(struct net_device *dev, struct ipoib_mcast *mcast,
 512                              int create)
 513 {
 514         struct ipoib_dev_priv *priv = netdev_priv(dev);
 515         struct ib_sa_mcmember_rec rec = {
 516                 .join_state = 1
 517         };
 518         ib_sa_comp_mask comp_mask;
 519         int ret = 0;
 520
 521         ipoib_dbg_mcast(priv, "joining MGID %pI6\n", mcast->mcmember.mgid.raw);
 522
 523         rec.mgid     = mcast->mcmember.mgid;
 524         rec.port_gid = priv->local_gid;
 525         rec.pkey     = cpu_to_be16(priv->pkey);
 526
 527         comp_mask =
 528                 IB_SA_MCMEMBER_REC_MGID         |
 529                 IB_SA_MCMEMBER_REC_PORT_GID     |
 530                 IB_SA_MCMEMBER_REC_PKEY         |
 531                 IB_SA_MCMEMBER_REC_JOIN_STATE;
 532
 533         if (create) {
 534                 comp_mask |=
 535                         IB_SA_MCMEMBER_REC_QKEY                 |
 536                         IB_SA_MCMEMBER_REC_MTU_SELECTOR         |
 537                         IB_SA_MCMEMBER_REC_MTU                  |
 538                         IB_SA_MCMEMBER_REC_TRAFFIC_CLASS        |
 539                         IB_SA_MCMEMBER_REC_RATE_SELECTOR        |
 540                         IB_SA_MCMEMBER_REC_RATE                 |
 541                         IB_SA_MCMEMBER_REC_SL                   |
 542                         IB_SA_MCMEMBER_REC_FLOW_LABEL           |
 543                         IB_SA_MCMEMBER_REC_HOP_LIMIT;
 544
 545                 rec.qkey          = priv->broadcast->mcmember.qkey;
 546                 rec.mtu_selector  = IB_SA_EQ;
 547                 rec.mtu           = priv->broadcast->mcmember.mtu;
 548                 rec.traffic_class = priv->broadcast->mcmember.traffic_class;
 549                 rec.rate_selector = IB_SA_EQ;
 550                 rec.rate          = priv->broadcast->mcmember.rate;
 551                 rec.sl            = priv->broadcast->mcmember.sl;
 552                 rec.flow_label    = priv->broadcast->mcmember.flow_label;
 553                 rec.hop_limit     = priv->broadcast->mcmember.hop_limit;
 554         }
 555
 556         mutex_lock(&mcast_mutex);
 557         mcast->mc = ib_sa_join_multicast(&ipoib_sa_client, priv->ca, priv->port,
 558                                          &rec, comp_mask, GFP_KERNEL,
 559                                          ipoib_mcast_join_complete, mcast);
 560         if (IS_ERR(mcast->mc)) {
 561                 clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 562                 ret = PTR_ERR(mcast->mc);
 563                 ipoib_warn(priv, "ib_sa_join_multicast failed, status %d\n", ret);
 564                 __ipoib_mcast_schedule_join_thread(priv, mcast, 1);
 565                 complete(&mcast->done);
 566         }
 567         mutex_unlock(&mcast_mutex);
 568 }
 569
 570 void ipoib_mcast_join_task(struct work_struct *work)
 571 {
 572         struct ipoib_dev_priv *priv =
 573                 container_of(work, struct ipoib_dev_priv, mcast_task.work);
 574         struct net_device *dev = priv->dev;
 575         struct ib_port_attr port_attr;
 576         unsigned long delay_until = 0;
 577         struct ipoib_mcast *mcast = NULL;
 578         int create = 1;
 579
 580         if (!test_bit(IPOIB_MCAST_RUN, &priv->flags))
 581                 return;
 582
 583         if (ib_query_port(priv->ca, priv->port, &port_attr) ||
 584             port_attr.state != IB_PORT_ACTIVE) {
 585                 ipoib_dbg(priv, "port state is not ACTIVE (state = %d) suspending join task\n",
 586                           port_attr.state);
 587                 return;
 588         }
 589         priv->local_lid = port_attr.lid;
 590
 591         if (ib_query_gid(priv->ca, priv->port, 0, &priv->local_gid))
 592                 ipoib_warn(priv, "ib_query_gid() failed\n");
 593         else
 594                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
 595
 596         /*
 597          * We have to hold the mutex to keep from racing with the join
 598          * completion threads on setting flags on mcasts, and we have
 599          * to hold the priv->lock because dev_flush will remove entries
 600          * out from underneath us, so at a minimum we need the lock
 601          * through the time that we do the for_each loop of the mcast
 602          * list or else dev_flush can make us oops.
 603          */
 604         mutex_lock(&mcast_mutex);
 605         spin_lock_irq(&priv->lock);
 606         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
 607                 goto out;
 608
 609         if (!priv->broadcast) {
 610                 struct ipoib_mcast *broadcast;
 611
 612                 broadcast = ipoib_mcast_alloc(dev, 0);
 613                 if (!broadcast) {
 614                         ipoib_warn(priv, "failed to allocate broadcast group\n");
 615                         /*
 616                          * Restart us after a 1 second delay to retry
 617                          * creating our broadcast group and attaching to
 618                          * it.  Until this succeeds, this ipoib dev is
 619                          * completely stalled (multicast wise).
 620                          */
 621                         __ipoib_mcast_schedule_join_thread(priv, NULL, 1);
 622                         goto out;
 623                 }
 624
 625                 memcpy(broadcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
 626                        sizeof (union ib_gid));
 627                 priv->broadcast = broadcast;
 628
 629                 __ipoib_mcast_add(dev, priv->broadcast);
 630         }
 631
 632         if (!test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
 633                 if (IS_ERR_OR_NULL(priv->broadcast->mc) &&
 634                     !test_bit(IPOIB_MCAST_FLAG_BUSY, &priv->broadcast->flags)) {
 635                         mcast = priv->broadcast;
 636                         create = 0;
 637                         if (mcast->backoff > 1 &&
 638                             time_before(jiffies, mcast->delay_until)) {
 639                                 delay_until = mcast->delay_until;
 640                                 mcast = NULL;
 641                         }
 642                 }
 643                 goto out;
 644         }
 645
 646         /*
 647          * We'll never get here until the broadcast group is both allocated
 648          * and attached
 649          */
 650         list_for_each_entry(mcast, &priv->multicast_list, list) {
 651                 if (IS_ERR_OR_NULL(mcast->mc) &&
 652                     !test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags) &&
 653                     !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 654                         if (mcast->backoff == 1 ||
 655                             time_after_eq(jiffies, mcast->delay_until))
 656                                 /* Found the next unjoined group */
 657                                 break;
 658                         else if (!delay_until ||
 659                                  time_before(mcast->delay_until, delay_until))
 660                                 delay_until = mcast->delay_until;
 661                 }
 662         }
 663
 664         if (&mcast->list == &priv->multicast_list) {
 665                 /*
 666                  * All done, unless we have delayed work from
 667                  * backoff retransmissions, but we will get
 668                  * restarted when the time is right, so we are
 669                  * done for now
 670                  */
 671                 mcast = NULL;
 672                 ipoib_dbg_mcast(priv, "successfully joined all "
 673                                 "multicast groups\n");
 674         }
 675
 676 out:
 677         if (mcast) {
 678                 init_completion(&mcast->done);
 679                 set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 680         }
 681         spin_unlock_irq(&priv->lock);
 682         mutex_unlock(&mcast_mutex);
 683         if (mcast) {
 684                 if (test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags))
 685                         ipoib_mcast_sendonly_join(mcast);
 686                 else
 687                         ipoib_mcast_join(dev, mcast, create);
 688         }
 689         if (delay_until)
 690                 queue_delayed_work(priv->wq, &priv->mcast_task,
 691                                    delay_until - jiffies);
 692 }
 693
 694 int ipoib_mcast_start_thread(struct net_device *dev)
 695 {
 696         struct ipoib_dev_priv *priv = netdev_priv(dev);
 697
 698         ipoib_dbg_mcast(priv, "starting multicast thread\n");
 699
 700         mutex_lock(&mcast_mutex);
 701         set_bit(IPOIB_MCAST_RUN, &priv->flags);
 702         __ipoib_mcast_schedule_join_thread(priv, NULL, 0);
 703         mutex_unlock(&mcast_mutex);
 704
 705         return 0;
 706 }
 707
 708 int ipoib_mcast_stop_thread(struct net_device *dev)
 709 {
 710         struct ipoib_dev_priv *priv = netdev_priv(dev);
 711
 712         ipoib_dbg_mcast(priv, "stopping multicast thread\n");
 713
 714         mutex_lock(&mcast_mutex);
 715         clear_bit(IPOIB_MCAST_RUN, &priv->flags);
 716         cancel_delayed_work(&priv->mcast_task);
 717         mutex_unlock(&mcast_mutex);
 718
 719         flush_workqueue(priv->wq);
 720
 721         return 0;
 722 }
 723
 724 static int ipoib_mcast_leave(struct net_device *dev, struct ipoib_mcast *mcast)
 725 {
 726         struct ipoib_dev_priv *priv = netdev_priv(dev);
 727         int ret = 0;
 728
 729         if (test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
 730                 ipoib_warn(priv, "ipoib_mcast_leave on an in-flight join\n");
 731
 732         if (!IS_ERR_OR_NULL(mcast->mc))
 733                 ib_sa_free_multicast(mcast->mc);
 734         else
 735                 ipoib_dbg(priv, "ipoib_mcast_leave with mcast->mc invalid\n");
 736
 737         if (test_and_clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 738                 ipoib_dbg_mcast(priv, "leaving MGID %pI6\n",
 739                                 mcast->mcmember.mgid.raw);
 740
 741                 /* Remove ourselves from the multicast group */
 742                 ret = ib_detach_mcast(priv->qp, &mcast->mcmember.mgid,
 743                                       be16_to_cpu(mcast->mcmember.mlid));
 744                 if (ret)
 745                         ipoib_warn(priv, "ib_detach_mcast failed (result = %d)\n", ret);
 746         } else if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags))
 747                 ipoib_dbg(priv, "leaving with no mcmember but not a "
 748                           "SENDONLY join\n");
 749
 750         return 0;
 751 }
 752
 753 void ipoib_mcast_send(struct net_device *dev, u8 *daddr, struct sk_buff *skb)
 754 {
 755         struct ipoib_dev_priv *priv = netdev_priv(dev);
 756         struct ipoib_mcast *mcast;
 757         unsigned long flags;
 758         void *mgid = daddr + 4;
 759
 760         spin_lock_irqsave(&priv->lock, flags);
 761
 762         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)         ||
 763             !priv->broadcast                                    ||
 764             !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
 765                 ++dev->stats.tx_dropped;
 766                 dev_kfree_skb_any(skb);
 767                 goto unlock;
 768         }
 769
 770         mcast = __ipoib_mcast_find(dev, mgid);
 771         if (!mcast) {
 772                 /* Let's create a new send only group now */
 773                 ipoib_dbg_mcast(priv, "setting up send only multicast group for %pI6\n",
 774                                 mgid);
 775
 776                 mcast = ipoib_mcast_alloc(dev, 0);
 777                 if (!mcast) {
 778                         ipoib_warn(priv, "unable to allocate memory for "
 779                                    "multicast structure\n");
 780                         ++dev->stats.tx_dropped;
 781                         dev_kfree_skb_any(skb);
 782                         goto out;
 783                 }
 784
 785                 set_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags);
 786                 memcpy(mcast->mcmember.mgid.raw, mgid, sizeof (union ib_gid));
 787                 __ipoib_mcast_add(dev, mcast);
 788                 list_add_tail(&mcast->list, &priv->multicast_list);
 789                 __ipoib_mcast_schedule_join_thread(priv, NULL, 0);
 790         }
 791
 792         if (!mcast->ah) {
 793                 if (skb_queue_len(&mcast->pkt_queue) < IPOIB_MAX_MCAST_QUEUE)
 794                         skb_queue_tail(&mcast->pkt_queue, skb);
 795                 else {
 796                         ++dev->stats.tx_dropped;
 797                         dev_kfree_skb_any(skb);
 798                 }
 799                 /*
 800                  * If lookup completes between here and out:, don't
 801                  * want to send packet twice.
 802                  */
 803                 mcast = NULL;
 804         }
 805
 806 out:
 807         if (mcast && mcast->ah) {
 808                 struct ipoib_neigh *neigh;
 809
 810                 spin_unlock_irqrestore(&priv->lock, flags);
 811                 neigh = ipoib_neigh_get(dev, daddr);
 812                 spin_lock_irqsave(&priv->lock, flags);
 813                 if (!neigh) {
 814                         neigh = ipoib_neigh_alloc(daddr, dev);
 815                         if (neigh) {
 816                                 kref_get(&mcast->ah->ref);
 817                                 neigh->ah       = mcast->ah;
 818                                 list_add_tail(&neigh->list, &mcast->neigh_list);
 819                         }
 820                 }
 821                 spin_unlock_irqrestore(&priv->lock, flags);
 822                 ipoib_send(dev, skb, mcast->ah, IB_MULTICAST_QPN);
 823                 if (neigh)
 824                         ipoib_neigh_put(neigh);
 825                 return;
 826         }
 827
 828 unlock:
 829         spin_unlock_irqrestore(&priv->lock, flags);
 830 }
 831
 832 void ipoib_mcast_dev_flush(struct net_device *dev)
 833 {
 834         struct ipoib_dev_priv *priv = netdev_priv(dev);
 835         LIST_HEAD(remove_list);
 836         struct ipoib_mcast *mcast, *tmcast;
 837         unsigned long flags;
 838
 839         ipoib_dbg_mcast(priv, "flushing multicast list\n");
 840
 841         spin_lock_irqsave(&priv->lock, flags);
 842
 843         list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
 844                 list_del(&mcast->list);
 845                 rb_erase(&mcast->rb_node, &priv->multicast_tree);
 846                 list_add_tail(&mcast->list, &remove_list);
 847         }
 848
 849         if (priv->broadcast) {
 850                 rb_erase(&priv->broadcast->rb_node, &priv->multicast_tree);
 851                 list_add_tail(&priv->broadcast->list, &remove_list);
 852                 priv->broadcast = NULL;
 853         }
 854
 855         spin_unlock_irqrestore(&priv->lock, flags);
 856
 857         /*
 858          * make sure the in-flight joins have finished before we attempt
 859          * to leave
 860          */
 861         list_for_each_entry_safe(mcast, tmcast, &remove_list, list)
 862                 if (test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
 863                         wait_for_completion(&mcast->done);
 864
 865         list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
 866                 ipoib_mcast_leave(dev, mcast);
 867                 ipoib_mcast_free(mcast);
 868         }
 869 }
 870
 871 static int ipoib_mcast_addr_is_valid(const u8 *addr, const u8 *broadcast)
 872 {
 873         /* reserved QPN, prefix, scope */
 874         if (memcmp(addr, broadcast, 6))
 875                 return 0;
 876         /* signature lower, pkey */
 877         if (memcmp(addr + 7, broadcast + 7, 3))
 878                 return 0;
 879         return 1;
 880 }
 881
 882 void ipoib_mcast_restart_task(struct work_struct *work)
 883 {
 884         struct ipoib_dev_priv *priv =
 885                 container_of(work, struct ipoib_dev_priv, restart_task);
 886         struct net_device *dev = priv->dev;
 887         struct netdev_hw_addr *ha;
 888         struct ipoib_mcast *mcast, *tmcast;
 889         LIST_HEAD(remove_list);
 890         unsigned long flags;
 891         struct ib_sa_mcmember_rec rec;
 892
 893         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
 894                 /*
 895                  * shortcut...on shutdown flush is called next, just
 896                  * let it do all the work
 897                  */
 898                 return;
 899
 900         ipoib_dbg_mcast(priv, "restarting multicast task\n");
 901
 902         local_irq_save(flags);
 903         netif_addr_lock(dev);
 904         spin_lock(&priv->lock);
 905
 906         /*
 907          * Unfortunately, the networking core only gives us a list of all of
 908          * the multicast hardware addresses. We need to figure out which ones
 909          * are new and which ones have been removed
 910          */
 911
 912         /* Clear out the found flag */
 913         list_for_each_entry(mcast, &priv->multicast_list, list)
 914                 clear_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
 915
 916         /* Mark all of the entries that are found or don't exist */
 917         netdev_for_each_mc_addr(ha, dev) {
 918                 union ib_gid mgid;
 919
 920                 if (!ipoib_mcast_addr_is_valid(ha->addr, dev->broadcast))
 921                         continue;
 922
 923                 memcpy(mgid.raw, ha->addr + 4, sizeof mgid);
 924
 925                 mcast = __ipoib_mcast_find(dev, &mgid);
 926                 if (!mcast || test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 927                         struct ipoib_mcast *nmcast;
 928
 929                         /* ignore group which is directly joined by userspace */
 930                         if (test_bit(IPOIB_FLAG_UMCAST, &priv->flags) &&
 931                             !ib_sa_get_mcmember_rec(priv->ca, priv->port, &mgid, &rec)) {
 932                                 ipoib_dbg_mcast(priv, "ignoring multicast entry for mgid %pI6\n",
 933                                                 mgid.raw);
 934                                 continue;
 935                         }
 936
 937                         /* Not found or send-only group, let's add a new entry */
 938                         ipoib_dbg_mcast(priv, "adding multicast entry for mgid %pI6\n",
 939                                         mgid.raw);
 940
 941                         nmcast = ipoib_mcast_alloc(dev, 0);
 942                         if (!nmcast) {
 943                                 ipoib_warn(priv, "unable to allocate memory for multicast structure\n");
 944                                 continue;
 945                         }
 946
 947                         set_bit(IPOIB_MCAST_FLAG_FOUND, &nmcast->flags);
 948
 949                         nmcast->mcmember.mgid = mgid;
 950
 951                         if (mcast) {
 952                                 /* Destroy the send only entry */
 953                                 list_move_tail(&mcast->list, &remove_list);
 954
 955                                 rb_replace_node(&mcast->rb_node,
 956                                                 &nmcast->rb_node,
 957                                                 &priv->multicast_tree);
 958                         } else
 959                                 __ipoib_mcast_add(dev, nmcast);
 960
 961                         list_add_tail(&nmcast->list, &priv->multicast_list);
 962                 }
 963
 964                 if (mcast)
 965                         set_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
 966         }
 967
 968         /* Remove all of the entries don't exist anymore */
 969         list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
 970                 if (!test_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags) &&
 971                     !test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 972                         ipoib_dbg_mcast(priv, "deleting multicast group %pI6\n",
 973                                         mcast->mcmember.mgid.raw);
 974
 975                         rb_erase(&mcast->rb_node, &priv->multicast_tree);
 976
 977                         /* Move to the remove list */
 978                         list_move_tail(&mcast->list, &remove_list);
 979                 }
 980         }
 981
 982         spin_unlock(&priv->lock);
 983         netif_addr_unlock(dev);
 984         local_irq_restore(flags);
 985
 986         /*
 987          * make sure the in-flight joins have finished before we attempt
 988          * to leave
 989          */
 990         list_for_each_entry_safe(mcast, tmcast, &remove_list, list)
 991                 if (test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
 992                         wait_for_completion(&mcast->done);
 993
 994         list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
 995                 ipoib_mcast_leave(mcast->dev, mcast);
 996                 ipoib_mcast_free(mcast);
 997         }
 998
 999         /*
1000          * Double check that we are still up
1001          */
1002         if (test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
1003                 spin_lock_irqsave(&priv->lock, flags);
1004                 __ipoib_mcast_schedule_join_thread(priv, NULL, 0);
1005                 spin_unlock_irqrestore(&priv->lock, flags);
1006         }
1007 }
1008
1009 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
1010
1011 struct ipoib_mcast_iter *ipoib_mcast_iter_init(struct net_device *dev)
1012 {
1013         struct ipoib_mcast_iter *iter;
1014
1015         iter = kmalloc(sizeof *iter, GFP_KERNEL);
1016         if (!iter)
1017                 return NULL;
1018
1019         iter->dev = dev;
1020         memset(iter->mgid.raw, 0, 16);
1021
1022         if (ipoib_mcast_iter_next(iter)) {
1023                 kfree(iter);
1024                 return NULL;
1025         }
1026
1027         return iter;
1028 }
1029
1030 int ipoib_mcast_iter_next(struct ipoib_mcast_iter *iter)
1031 {
1032         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
1033         struct rb_node *n;
1034         struct ipoib_mcast *mcast;
1035         int ret = 1;
1036
1037         spin_lock_irq(&priv->lock);
1038
1039         n = rb_first(&priv->multicast_tree);
1040
1041         while (n) {
1042                 mcast = rb_entry(n, struct ipoib_mcast, rb_node);
1043
1044                 if (memcmp(iter->mgid.raw, mcast->mcmember.mgid.raw,
1045                            sizeof (union ib_gid)) < 0) {
1046                         iter->mgid      = mcast->mcmember.mgid;
1047                         iter->created   = mcast->created;
1048                         iter->queuelen  = skb_queue_len(&mcast->pkt_queue);
1049                         iter->complete  = !!mcast->ah;
1050                         iter->send_only = !!(mcast->flags & (1 << IPOIB_MCAST_FLAG_SENDONLY));
1051
1052                         ret = 0;
1053
1054                         break;
1055                 }
1056
1057                 n = rb_next(n);
1058         }
1059
1060         spin_unlock_irq(&priv->lock);
1061
1062         return ret;
1063 }
1064
1065 void ipoib_mcast_iter_read(struct ipoib_mcast_iter *iter,
1066                            union ib_gid *mgid,
1067                            unsigned long *created,
1068                            unsigned int *queuelen,
1069                            unsigned int *complete,
1070                            unsigned int *send_only)
1071 {
1072         *mgid      = iter->mgid;
1073         *created   = iter->created;
1074         *queuelen  = iter->queuelen;
1075         *complete  = iter->complete;
1076         *send_only = iter->send_only;
1077 }
1078
1079 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */