sys/ofed/drivers/infiniband/ulp/ipoib/ipoib_multicast.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-2-Clause OR GPL-2.0
   3  *
   4  * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
   5  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
   6  * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
   7  *
   8  * This software is available to you under a choice of one of two
   9  * licenses.  You may choose to be licensed under the terms of the GNU
  10  * General Public License (GPL) Version 2, available from the file
  11  * COPYING in the main directory of this source tree, or the
  12  * OpenIB.org BSD license below:
  13  *
  14  *     Redistribution and use in source and binary forms, with or
  15  *     without modification, are permitted provided that the following
  16  *     conditions are met:
  17  *
  18  *      - Redistributions of source code must retain the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer.
  21  *
  22  *      - Redistributions in binary form must reproduce the above
  23  *        copyright notice, this list of conditions and the following
  24  *        disclaimer in the documentation and/or other materials
  25  *        provided with the distribution.
  26  *
  27  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  28  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  29  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  30  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  31  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  32  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  33  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  34  * SOFTWARE.
  35  */
  36
  37 #include "ipoib.h"
  38
  39 #include <linux/delay.h>
  40 #include <linux/completion.h>
  41
  42 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
  43 static int mcast_debug_level = 1;
  44
  45 module_param(mcast_debug_level, int, 0644);
  46 MODULE_PARM_DESC(mcast_debug_level,
  47                  "Enable multicast debug tracing if > 0");
  48 #endif
  49
  50 static DEFINE_MUTEX(mcast_mutex);
  51
  52 struct ipoib_mcast_iter {
  53         struct ipoib_dev_priv *priv;
  54         union ib_gid       mgid;
  55         unsigned long      created;
  56         unsigned int       queuelen;
  57         unsigned int       complete;
  58         unsigned int       send_only;
  59 };
  60
  61 static void ipoib_mcast_free(struct ipoib_mcast *mcast)
  62 {
  63         struct ifnet *dev = mcast->priv->dev;
  64         int tx_dropped = 0;
  65
  66         ipoib_dbg_mcast(mcast->priv, "deleting multicast group %16D\n",
  67                         mcast->mcmember.mgid.raw, ":");
  68
  69         if (mcast->ah)
  70                 ipoib_put_ah(mcast->ah);
  71
  72         tx_dropped = mcast->pkt_queue.ifq_len;
  73         _IF_DRAIN(&mcast->pkt_queue);   /* XXX Locking. */
  74
  75         if_inc_counter(dev, IFCOUNTER_OERRORS, tx_dropped);
  76
  77         kfree(mcast);
  78 }
  79
  80 static struct ipoib_mcast *ipoib_mcast_alloc(struct ipoib_dev_priv *priv,
  81                                              int can_sleep)
  82 {
  83         struct ipoib_mcast *mcast;
  84
  85         mcast = kzalloc(sizeof *mcast, can_sleep ? GFP_KERNEL : GFP_ATOMIC);
  86         if (!mcast)
  87                 return NULL;
  88
  89         mcast->priv = priv;
  90         mcast->created = jiffies;
  91         mcast->backoff = 1;
  92
  93         INIT_LIST_HEAD(&mcast->list);
  94         bzero(&mcast->pkt_queue, sizeof(mcast->pkt_queue));
  95
  96         return mcast;
  97 }
  98
  99 static struct ipoib_mcast *__ipoib_mcast_find(struct ipoib_dev_priv *priv,
 100     void *mgid)
 101 {
 102         struct rb_node *n = priv->multicast_tree.rb_node;
 103
 104         while (n) {
 105                 struct ipoib_mcast *mcast;
 106                 int ret;
 107
 108                 mcast = rb_entry(n, struct ipoib_mcast, rb_node);
 109
 110                 ret = memcmp(mgid, mcast->mcmember.mgid.raw,
 111                              sizeof (union ib_gid));
 112                 if (ret < 0)
 113                         n = n->rb_left;
 114                 else if (ret > 0)
 115                         n = n->rb_right;
 116                 else
 117                         return mcast;
 118         }
 119
 120         return NULL;
 121 }
 122
 123 static int __ipoib_mcast_add(struct ipoib_dev_priv *priv,
 124     struct ipoib_mcast *mcast)
 125 {
 126         struct rb_node **n = &priv->multicast_tree.rb_node, *pn = NULL;
 127
 128         while (*n) {
 129                 struct ipoib_mcast *tmcast;
 130                 int ret;
 131
 132                 pn = *n;
 133                 tmcast = rb_entry(pn, struct ipoib_mcast, rb_node);
 134
 135                 ret = memcmp(mcast->mcmember.mgid.raw, tmcast->mcmember.mgid.raw,
 136                              sizeof (union ib_gid));
 137                 if (ret < 0)
 138                         n = &pn->rb_left;
 139                 else if (ret > 0)
 140                         n = &pn->rb_right;
 141                 else
 142                         return -EEXIST;
 143         }
 144
 145         rb_link_node(&mcast->rb_node, pn, n);
 146         rb_insert_color(&mcast->rb_node, &priv->multicast_tree);
 147
 148         return 0;
 149 }
 150
 151 static int ipoib_mcast_join_finish(struct ipoib_mcast *mcast,
 152                                    struct ib_sa_mcmember_rec *mcmember)
 153 {
 154         struct ipoib_dev_priv *priv = mcast->priv;
 155         struct ifnet *dev = priv->dev;
 156         struct ipoib_ah *ah;
 157         int ret;
 158         int set_qkey = 0;
 159
 160         mcast->mcmember = *mcmember;
 161
 162         /* Set the cached Q_Key before we attach if it's the broadcast group */
 163         if (!memcmp(mcast->mcmember.mgid.raw, dev->if_broadcastaddr + 4,
 164                     sizeof (union ib_gid))) {
 165                 spin_lock_irq(&priv->lock);
 166                 if (!priv->broadcast) {
 167                         spin_unlock_irq(&priv->lock);
 168                         return -EAGAIN;
 169                 }
 170                 priv->qkey = be32_to_cpu(priv->broadcast->mcmember.qkey);
 171                 spin_unlock_irq(&priv->lock);
 172                 priv->tx_wr.remote_qkey = priv->qkey;
 173                 set_qkey = 1;
 174         }
 175
 176         if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 177                 if (test_and_set_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 178                         ipoib_warn(priv, "multicast group %16D already attached\n",
 179                                    mcast->mcmember.mgid.raw, ":");
 180
 181                         return 0;
 182                 }
 183
 184                 ret = ipoib_mcast_attach(priv, be16_to_cpu(mcast->mcmember.mlid),
 185                                          &mcast->mcmember.mgid, set_qkey);
 186                 if (ret < 0) {
 187                         ipoib_warn(priv, "couldn't attach QP to multicast group %16D\n",
 188                                    mcast->mcmember.mgid.raw, ":");
 189
 190                         clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags);
 191                         return ret;
 192                 }
 193         }
 194
 195         {
 196                 struct ib_ah_attr av = {
 197                         .dlid          = be16_to_cpu(mcast->mcmember.mlid),
 198                         .port_num      = priv->port,
 199                         .sl            = mcast->mcmember.sl,
 200                         .ah_flags      = IB_AH_GRH,
 201                         .static_rate   = mcast->mcmember.rate,
 202                         .grh           = {
 203                                 .flow_label    = be32_to_cpu(mcast->mcmember.flow_label),
 204                                 .hop_limit     = mcast->mcmember.hop_limit,
 205                                 .sgid_index    = 0,
 206                                 .traffic_class = mcast->mcmember.traffic_class
 207                         }
 208                 };
 209                 av.grh.dgid = mcast->mcmember.mgid;
 210
 211                 ah = ipoib_create_ah(priv, priv->pd, &av);
 212                 if (!ah) {
 213                         ipoib_warn(priv, "ib_address_create failed\n");
 214                 } else {
 215                         spin_lock_irq(&priv->lock);
 216                         mcast->ah = ah;
 217                         spin_unlock_irq(&priv->lock);
 218
 219                         ipoib_dbg_mcast(priv, "MGID %16D AV %p, LID 0x%04x, SL %d\n",
 220                                         mcast->mcmember.mgid.raw, ":",
 221                                         mcast->ah->ah,
 222                                         be16_to_cpu(mcast->mcmember.mlid),
 223                                         mcast->mcmember.sl);
 224                 }
 225         }
 226
 227         /* actually send any queued packets */
 228         while (mcast->pkt_queue.ifq_len) {
 229                 struct mbuf *mb;
 230                 _IF_DEQUEUE(&mcast->pkt_queue, mb);
 231                 mb->m_pkthdr.rcvif = dev;
 232
 233                 if (dev->if_transmit(dev, mb))
 234                         ipoib_warn(priv, "dev_queue_xmit failed to requeue packet\n");
 235         }
 236
 237         return 0;
 238 }
 239
 240 static int
 241 ipoib_mcast_sendonly_join_complete(int status,
 242                                    struct ib_sa_multicast *multicast)
 243 {
 244         struct ipoib_mcast *mcast = multicast->context;
 245         struct ipoib_dev_priv *priv = mcast->priv;
 246
 247         /* We trap for port events ourselves. */
 248         if (status == -ENETRESET)
 249                 return 0;
 250
 251         if (!status)
 252                 status = ipoib_mcast_join_finish(mcast, &multicast->rec);
 253
 254         if (status) {
 255                 if (mcast->logcount++ < 20)
 256                         ipoib_dbg_mcast(priv, "multicast join failed for %16D, status %d\n",
 257                                         mcast->mcmember.mgid.raw, ":", status);
 258
 259                 /* Flush out any queued packets */
 260                 if_inc_counter(priv->dev, IFCOUNTER_OERRORS, mcast->pkt_queue.ifq_len);
 261                 _IF_DRAIN(&mcast->pkt_queue);
 262
 263                 /* Clear the busy flag so we try again */
 264                 status = test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY,
 265                                             &mcast->flags);
 266         }
 267         return status;
 268 }
 269
 270 static int ipoib_mcast_sendonly_join(struct ipoib_mcast *mcast)
 271 {
 272         struct ipoib_dev_priv *priv = mcast->priv;
 273         struct ib_sa_mcmember_rec rec = {
 274 #if 0                           /* Some SMs don't support send-only yet */
 275                 .join_state = 4
 276 #else
 277                 .join_state = 1
 278 #endif
 279         };
 280         int ret = 0;
 281
 282         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
 283                 ipoib_dbg_mcast(priv, "device shutting down, no multicast joins\n");
 284                 return -ENODEV;
 285         }
 286
 287         if (test_and_set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)) {
 288                 ipoib_dbg_mcast(priv, "multicast entry busy, skipping\n");
 289                 return -EBUSY;
 290         }
 291
 292         rec.mgid     = mcast->mcmember.mgid;
 293         rec.port_gid = priv->local_gid;
 294         rec.pkey     = cpu_to_be16(priv->pkey);
 295
 296         mcast->mc = ib_sa_join_multicast(&ipoib_sa_client, priv->ca,
 297                                          priv->port, &rec,
 298                                          IB_SA_MCMEMBER_REC_MGID        |
 299                                          IB_SA_MCMEMBER_REC_PORT_GID    |
 300                                          IB_SA_MCMEMBER_REC_PKEY        |
 301                                          IB_SA_MCMEMBER_REC_JOIN_STATE,
 302                                          GFP_ATOMIC,
 303                                          ipoib_mcast_sendonly_join_complete,
 304                                          mcast);
 305         if (IS_ERR(mcast->mc)) {
 306                 ret = PTR_ERR(mcast->mc);
 307                 clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 308                 ipoib_warn(priv, "ib_sa_join_multicast failed (ret = %d)\n",
 309                            ret);
 310         } else {
 311                 ipoib_dbg_mcast(priv, "no multicast record for %16D, starting join\n",
 312                                 mcast->mcmember.mgid.raw, ":");
 313         }
 314
 315         return ret;
 316 }
 317
 318 void ipoib_mcast_carrier_on_task(struct work_struct *work)
 319 {
 320         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
 321                                                    carrier_on_task);
 322         struct ib_port_attr attr;
 323
 324         /*
 325          * Take rtnl_lock to avoid racing with ipoib_stop() and
 326          * turning the carrier back on while a device is being
 327          * removed.
 328          */
 329         if (ib_query_port(priv->ca, priv->port, &attr) ||
 330             attr.state != IB_PORT_ACTIVE) {
 331                 ipoib_dbg(priv, "Keeping carrier off until IB port is active\n");
 332                 return;
 333         }
 334         if_link_state_change(priv->dev, LINK_STATE_UP);
 335 }
 336
 337 static int ipoib_mcast_join_complete(int status,
 338                                      struct ib_sa_multicast *multicast)
 339 {
 340         struct ipoib_mcast *mcast = multicast->context;
 341         struct ipoib_dev_priv *priv = mcast->priv;
 342
 343         ipoib_dbg_mcast(priv, "join completion for %16D (status %d)\n",
 344                         mcast->mcmember.mgid.raw, ":", status);
 345
 346         /* We trap for port events ourselves. */
 347         if (status == -ENETRESET)
 348                 return 0;
 349
 350         if (!status)
 351                 status = ipoib_mcast_join_finish(mcast, &multicast->rec);
 352
 353         if (!status) {
 354                 mcast->backoff = 1;
 355                 mutex_lock(&mcast_mutex);
 356                 if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 357                         queue_delayed_work(ipoib_workqueue,
 358                                            &priv->mcast_task, 0);
 359                 mutex_unlock(&mcast_mutex);
 360
 361                 /*
 362                  * Defer carrier on work to ipoib_workqueue to avoid a
 363                  * deadlock on rtnl_lock here.
 364                  */
 365                 if (mcast == priv->broadcast)
 366                         queue_work(ipoib_workqueue, &priv->carrier_on_task);
 367
 368                 return 0;
 369         }
 370
 371         if (mcast->logcount++ < 20) {
 372                 if (status == -ETIMEDOUT || status == -EAGAIN) {
 373                         ipoib_dbg_mcast(priv, "multicast join failed for %16D, status %d\n",
 374                                         mcast->mcmember.mgid.raw, ":", status);
 375                 } else {
 376                         ipoib_warn(priv, "multicast join failed for %16D, status %d\n",
 377                                    mcast->mcmember.mgid.raw, ":", status);
 378                 }
 379         }
 380
 381         mcast->backoff *= 2;
 382         if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
 383                 mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
 384
 385         /* Clear the busy flag so we try again */
 386         status = test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 387
 388         mutex_lock(&mcast_mutex);
 389         spin_lock_irq(&priv->lock);
 390         if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 391                 queue_delayed_work(ipoib_workqueue, &priv->mcast_task,
 392                                    mcast->backoff * HZ);
 393         spin_unlock_irq(&priv->lock);
 394         mutex_unlock(&mcast_mutex);
 395
 396         return status;
 397 }
 398
 399 static void ipoib_mcast_join(struct ipoib_dev_priv *priv,
 400     struct ipoib_mcast *mcast, int create)
 401 {
 402         struct ib_sa_mcmember_rec rec = {
 403                 .join_state = 1
 404         };
 405         ib_sa_comp_mask comp_mask;
 406         int ret = 0;
 407
 408         ipoib_dbg_mcast(priv, "joining MGID %16D\n",
 409             mcast->mcmember.mgid.raw, ":");
 410
 411         rec.mgid     = mcast->mcmember.mgid;
 412         rec.port_gid = priv->local_gid;
 413         rec.pkey     = cpu_to_be16(priv->pkey);
 414
 415         comp_mask =
 416                 IB_SA_MCMEMBER_REC_MGID         |
 417                 IB_SA_MCMEMBER_REC_PORT_GID     |
 418                 IB_SA_MCMEMBER_REC_PKEY         |
 419                 IB_SA_MCMEMBER_REC_JOIN_STATE;
 420
 421         if (create) {
 422                 comp_mask |=
 423                         IB_SA_MCMEMBER_REC_QKEY                 |
 424                         IB_SA_MCMEMBER_REC_MTU_SELECTOR         |
 425                         IB_SA_MCMEMBER_REC_MTU                  |
 426                         IB_SA_MCMEMBER_REC_TRAFFIC_CLASS        |
 427                         IB_SA_MCMEMBER_REC_RATE_SELECTOR        |
 428                         IB_SA_MCMEMBER_REC_RATE                 |
 429                         IB_SA_MCMEMBER_REC_SL                   |
 430                         IB_SA_MCMEMBER_REC_FLOW_LABEL           |
 431                         IB_SA_MCMEMBER_REC_HOP_LIMIT;
 432
 433                 rec.qkey          = priv->broadcast->mcmember.qkey;
 434                 rec.mtu_selector  = IB_SA_EQ;
 435                 rec.mtu           = priv->broadcast->mcmember.mtu;
 436                 rec.traffic_class = priv->broadcast->mcmember.traffic_class;
 437                 rec.rate_selector = IB_SA_EQ;
 438                 rec.rate          = priv->broadcast->mcmember.rate;
 439                 rec.sl            = priv->broadcast->mcmember.sl;
 440                 rec.flow_label    = priv->broadcast->mcmember.flow_label;
 441                 rec.hop_limit     = priv->broadcast->mcmember.hop_limit;
 442         }
 443
 444         set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 445         mcast->mc = ib_sa_join_multicast(&ipoib_sa_client, priv->ca, priv->port,
 446                                          &rec, comp_mask, GFP_KERNEL,
 447                                          ipoib_mcast_join_complete, mcast);
 448         if (IS_ERR(mcast->mc)) {
 449                 clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 450                 ret = PTR_ERR(mcast->mc);
 451                 ipoib_warn(priv, "ib_sa_join_multicast failed, status %d\n", ret);
 452
 453                 mcast->backoff *= 2;
 454                 if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
 455                         mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
 456
 457                 mutex_lock(&mcast_mutex);
 458                 if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 459                         queue_delayed_work(ipoib_workqueue,
 460                                            &priv->mcast_task,
 461                                            mcast->backoff * HZ);
 462                 mutex_unlock(&mcast_mutex);
 463         }
 464 }
 465
 466 void ipoib_mcast_join_task(struct work_struct *work)
 467 {
 468         struct ipoib_dev_priv *priv =
 469                 container_of(work, struct ipoib_dev_priv, mcast_task.work);
 470         struct ifnet *dev = priv->dev;
 471         struct ib_port_attr attr;
 472
 473         ipoib_dbg_mcast(priv, "Running join task. flags 0x%lX\n", priv->flags);
 474
 475         if (!test_bit(IPOIB_MCAST_RUN, &priv->flags))
 476                 return;
 477
 478         if (ib_query_port(priv->ca, priv->port, &attr) ||
 479             attr.state != IB_PORT_ACTIVE) {
 480                 ipoib_dbg(priv, "%s: port state is not ACTIVE (state = %d) suspend task.\n",
 481                           __func__, attr.state);
 482                 return;
 483         }
 484
 485         if (ib_query_gid(priv->ca, priv->port, 0, &priv->local_gid, NULL))
 486                 ipoib_warn(priv, "ib_query_gid() failed\n");
 487         else
 488                 memcpy(IF_LLADDR(dev) + 4, priv->local_gid.raw, sizeof (union ib_gid));
 489
 490         {
 491                 struct ib_port_attr attr;
 492
 493                 if (!ib_query_port(priv->ca, priv->port, &attr))
 494                         priv->local_lid = attr.lid;
 495                 else
 496                         ipoib_warn(priv, "ib_query_port failed\n");
 497         }
 498
 499         if (!priv->broadcast) {
 500                 struct ipoib_mcast *broadcast;
 501
 502                 if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 503                         return;
 504
 505                 broadcast = ipoib_mcast_alloc(priv, 1);
 506                 if (!broadcast) {
 507                         ipoib_warn(priv, "failed to allocate broadcast group\n");
 508                         mutex_lock(&mcast_mutex);
 509                         if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
 510                                 queue_delayed_work(ipoib_workqueue,
 511                                                    &priv->mcast_task, HZ);
 512                         mutex_unlock(&mcast_mutex);
 513                         return;
 514                 }
 515
 516                 spin_lock_irq(&priv->lock);
 517                 memcpy(broadcast->mcmember.mgid.raw, dev->if_broadcastaddr + 4,
 518                        sizeof (union ib_gid));
 519                 priv->broadcast = broadcast;
 520
 521                 __ipoib_mcast_add(priv, priv->broadcast);
 522                 spin_unlock_irq(&priv->lock);
 523         }
 524
 525         if (priv->broadcast &&
 526             !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
 527                 if (priv->broadcast &&
 528                     !test_bit(IPOIB_MCAST_FLAG_BUSY, &priv->broadcast->flags))
 529                         ipoib_mcast_join(priv, priv->broadcast, 0);
 530                 return;
 531         }
 532
 533         while (1) {
 534                 struct ipoib_mcast *mcast = NULL;
 535
 536                 spin_lock_irq(&priv->lock);
 537                 list_for_each_entry(mcast, &priv->multicast_list, list) {
 538                         if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)
 539                             && !test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)
 540                             && !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 541                                 /* Found the next unjoined group */
 542                                 break;
 543                         }
 544                 }
 545                 spin_unlock_irq(&priv->lock);
 546
 547                 if (&mcast->list == &priv->multicast_list) {
 548                         /* All done */
 549                         break;
 550                 }
 551
 552                 ipoib_mcast_join(priv, mcast, 1);
 553                 return;
 554         }
 555
 556         spin_lock_irq(&priv->lock);
 557         if (priv->broadcast)
 558                 priv->mcast_mtu = IPOIB_UD_MTU(ib_mtu_enum_to_int(priv->broadcast->mcmember.mtu));
 559         else
 560                 priv->mcast_mtu = priv->admin_mtu;
 561         spin_unlock_irq(&priv->lock);
 562
 563         if (!ipoib_cm_admin_enabled(priv))
 564                 ipoib_change_mtu(priv, min(priv->mcast_mtu, priv->admin_mtu));
 565
 566         ipoib_dbg_mcast(priv, "successfully joined all multicast groups\n");
 567
 568         clear_bit(IPOIB_MCAST_RUN, &priv->flags);
 569 }
 570
 571 int ipoib_mcast_start_thread(struct ipoib_dev_priv *priv)
 572 {
 573         ipoib_dbg_mcast(priv, "starting multicast thread flags 0x%lX\n",
 574             priv->flags);
 575
 576         mutex_lock(&mcast_mutex);
 577         if (!test_and_set_bit(IPOIB_MCAST_RUN, &priv->flags))
 578                 queue_delayed_work(ipoib_workqueue, &priv->mcast_task, 0);
 579         mutex_unlock(&mcast_mutex);
 580
 581         return 0;
 582 }
 583
 584 int ipoib_mcast_stop_thread(struct ipoib_dev_priv *priv, int flush)
 585 {
 586
 587         ipoib_dbg_mcast(priv, "stopping multicast thread\n");
 588
 589         mutex_lock(&mcast_mutex);
 590         clear_bit(IPOIB_MCAST_RUN, &priv->flags);
 591         cancel_delayed_work(&priv->mcast_task);
 592         mutex_unlock(&mcast_mutex);
 593
 594         if (flush)
 595                 flush_workqueue(ipoib_workqueue);
 596
 597         return 0;
 598 }
 599
 600 static int ipoib_mcast_leave(struct ipoib_dev_priv *priv, struct ipoib_mcast *mcast)
 601 {
 602         int ret = 0;
 603
 604         if (test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
 605                 ib_sa_free_multicast(mcast->mc);
 606
 607         if (test_and_clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 608                 ipoib_dbg_mcast(priv, "leaving MGID %16D\n",
 609                                 mcast->mcmember.mgid.raw, ":");
 610
 611                 /* Remove ourselves from the multicast group */
 612                 ret = ib_detach_mcast(priv->qp, &mcast->mcmember.mgid,
 613                                       be16_to_cpu(mcast->mcmember.mlid));
 614                 if (ret)
 615                         ipoib_warn(priv, "ib_detach_mcast failed (result = %d)\n", ret);
 616         }
 617
 618         return 0;
 619 }
 620
 621 void
 622 ipoib_mcast_send(struct ipoib_dev_priv *priv, void *mgid, struct mbuf *mb)
 623 {
 624         struct ifnet *dev = priv->dev;
 625         struct ipoib_mcast *mcast;
 626
 627         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)         ||
 628             !priv->broadcast                                    ||
 629             !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
 630                 if_inc_counter(dev, IFCOUNTER_OERRORS, 1);
 631                 m_freem(mb);
 632                 return;
 633         }
 634
 635         mcast = __ipoib_mcast_find(priv, mgid);
 636         if (!mcast) {
 637                 /* Let's create a new send only group now */
 638                 ipoib_dbg_mcast(priv, "setting up send only multicast group for %16D\n",
 639                                 mgid, ":");
 640
 641                 mcast = ipoib_mcast_alloc(priv, 0);
 642                 if (!mcast) {
 643                         ipoib_warn(priv, "unable to allocate memory for "
 644                                    "multicast structure\n");
 645                         if_inc_counter(dev, IFCOUNTER_OERRORS, 1);
 646                         m_freem(mb);
 647                         goto out;
 648                 }
 649
 650                 set_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags);
 651                 memcpy(mcast->mcmember.mgid.raw, mgid, sizeof (union ib_gid));
 652                 __ipoib_mcast_add(priv, mcast);
 653                 list_add_tail(&mcast->list, &priv->multicast_list);
 654         }
 655
 656         if (!mcast->ah) {
 657                 if (mcast->pkt_queue.ifq_len < IPOIB_MAX_MCAST_QUEUE) {
 658                         _IF_ENQUEUE(&mcast->pkt_queue, mb);
 659                 } else {
 660                         if_inc_counter(dev, IFCOUNTER_OERRORS, 1);
 661                         m_freem(mb);
 662                 }
 663
 664                 if (test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
 665                         ipoib_dbg_mcast(priv, "no address vector, "
 666                                         "but multicast join already started\n");
 667                 else if (test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags))
 668                         ipoib_mcast_sendonly_join(mcast);
 669
 670                 /*
 671                  * If lookup completes between here and out:, don't
 672                  * want to send packet twice.
 673                  */
 674                 mcast = NULL;
 675         }
 676
 677 out:
 678         if (mcast && mcast->ah)
 679                 ipoib_send(priv, mb, mcast->ah, IB_MULTICAST_QPN);
 680 }
 681
 682 void ipoib_mcast_dev_flush(struct ipoib_dev_priv *priv)
 683 {
 684         LIST_HEAD(remove_list);
 685         struct ipoib_mcast *mcast, *tmcast;
 686         unsigned long flags;
 687
 688         ipoib_dbg_mcast(priv, "flushing multicast list\n");
 689
 690         spin_lock_irqsave(&priv->lock, flags);
 691
 692         list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
 693                 list_del(&mcast->list);
 694                 rb_erase(&mcast->rb_node, &priv->multicast_tree);
 695                 list_add_tail(&mcast->list, &remove_list);
 696         }
 697
 698         if (priv->broadcast) {
 699                 rb_erase(&priv->broadcast->rb_node, &priv->multicast_tree);
 700                 list_add_tail(&priv->broadcast->list, &remove_list);
 701                 priv->broadcast = NULL;
 702         }
 703
 704         spin_unlock_irqrestore(&priv->lock, flags);
 705
 706         list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
 707                 ipoib_mcast_leave(priv, mcast);
 708                 ipoib_mcast_free(mcast);
 709         }
 710 }
 711
 712 static int ipoib_mcast_addr_is_valid(const u8 *addr, unsigned int addrlen,
 713                                      const u8 *broadcast)
 714 {
 715         if (addrlen != INFINIBAND_ALEN)
 716                 return 0;
 717         /* reserved QPN, prefix, scope */
 718         if (memcmp(addr, broadcast, 6))
 719                 return 0;
 720         /* signature lower, pkey */
 721         if (memcmp(addr + 7, broadcast + 7, 3))
 722                 return 0;
 723         return 1;
 724 }
 725
 726 void ipoib_mcast_restart_task(struct work_struct *work)
 727 {
 728         struct ipoib_dev_priv *priv =
 729                 container_of(work, struct ipoib_dev_priv, restart_task);
 730         ipoib_mcast_restart(priv);
 731 }
 732
 733 void ipoib_mcast_restart(struct ipoib_dev_priv *priv)
 734 {
 735         struct ifnet *dev = priv->dev;
 736         struct ifmultiaddr *ifma;
 737         struct ipoib_mcast *mcast, *tmcast;
 738         LIST_HEAD(remove_list);
 739         struct ib_sa_mcmember_rec rec;
 740         int addrlen;
 741
 742         ipoib_dbg_mcast(priv, "restarting multicast task flags 0x%lX\n",
 743             priv->flags);
 744
 745         ipoib_mcast_stop_thread(priv, 0);
 746
 747         if_maddr_rlock(dev);
 748         spin_lock(&priv->lock);
 749
 750         /*
 751          * Unfortunately, the networking core only gives us a list of all of
 752          * the multicast hardware addresses. We need to figure out which ones
 753          * are new and which ones have been removed
 754          */
 755
 756         /* Clear out the found flag */
 757         list_for_each_entry(mcast, &priv->multicast_list, list)
 758                 clear_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
 759
 760         /* Mark all of the entries that are found or don't exist */
 761
 762
 763         TAILQ_FOREACH(ifma, &dev->if_multiaddrs, ifma_link) {
 764                 union ib_gid mgid;
 765                 uint8_t *addr;
 766
 767                 if (ifma->ifma_addr->sa_family != AF_LINK)
 768                         continue;
 769                 addr = LLADDR((struct sockaddr_dl *)ifma->ifma_addr);
 770                 addrlen = ((struct sockaddr_dl *)ifma->ifma_addr)->sdl_alen;
 771                 if (!ipoib_mcast_addr_is_valid(addr, addrlen,
 772                                                dev->if_broadcastaddr))
 773                         continue;
 774
 775                 memcpy(mgid.raw, addr + 4, sizeof mgid);
 776
 777                 mcast = __ipoib_mcast_find(priv, &mgid);
 778                 if (!mcast || test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 779                         struct ipoib_mcast *nmcast;
 780
 781                         /* ignore group which is directly joined by userspace */
 782                         if (test_bit(IPOIB_FLAG_UMCAST, &priv->flags) &&
 783                             !ib_sa_get_mcmember_rec(priv->ca, priv->port, &mgid, &rec)) {
 784                                 ipoib_dbg_mcast(priv, "ignoring multicast entry for mgid %16D\n",
 785                                                 mgid.raw, ":");
 786                                 continue;
 787                         }
 788
 789                         /* Not found or send-only group, let's add a new entry */
 790                         ipoib_dbg_mcast(priv, "adding multicast entry for mgid %16D\n",
 791                                         mgid.raw, ":");
 792
 793                         nmcast = ipoib_mcast_alloc(priv, 0);
 794                         if (!nmcast) {
 795                                 ipoib_warn(priv, "unable to allocate memory for multicast structure\n");
 796                                 continue;
 797                         }
 798
 799                         set_bit(IPOIB_MCAST_FLAG_FOUND, &nmcast->flags);
 800
 801                         nmcast->mcmember.mgid = mgid;
 802
 803                         if (mcast) {
 804                                 /* Destroy the send only entry */
 805                                 list_move_tail(&mcast->list, &remove_list);
 806
 807                                 rb_replace_node(&mcast->rb_node,
 808                                                 &nmcast->rb_node,
 809                                                 &priv->multicast_tree);
 810                         } else
 811                                 __ipoib_mcast_add(priv, nmcast);
 812
 813                         list_add_tail(&nmcast->list, &priv->multicast_list);
 814                 }
 815
 816                 if (mcast)
 817                         set_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
 818         }
 819
 820         /* Remove all of the entries don't exist anymore */
 821         list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
 822                 if (!test_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags) &&
 823                     !test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 824                         ipoib_dbg_mcast(priv, "deleting multicast group %16D\n",
 825                                         mcast->mcmember.mgid.raw, ":");
 826
 827                         rb_erase(&mcast->rb_node, &priv->multicast_tree);
 828
 829                         /* Move to the remove list */
 830                         list_move_tail(&mcast->list, &remove_list);
 831                 }
 832         }
 833
 834         spin_unlock(&priv->lock);
 835         if_maddr_runlock(dev);
 836
 837         /* We have to cancel outside of the spinlock */
 838         list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
 839                 ipoib_mcast_leave(mcast->priv, mcast);
 840                 ipoib_mcast_free(mcast);
 841         }
 842
 843         if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 844                 ipoib_mcast_start_thread(priv);
 845 }
 846
 847 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 848
 849 struct ipoib_mcast_iter *ipoib_mcast_iter_init(struct ipoib_dev_priv *priv)
 850 {
 851         struct ipoib_mcast_iter *iter;
 852
 853         iter = kmalloc(sizeof *iter, GFP_KERNEL);
 854         if (!iter)
 855                 return NULL;
 856
 857         iter->priv = priv;
 858         memset(iter->mgid.raw, 0, 16);
 859
 860         if (ipoib_mcast_iter_next(iter)) {
 861                 kfree(iter);
 862                 return NULL;
 863         }
 864
 865         return iter;
 866 }
 867
 868 int ipoib_mcast_iter_next(struct ipoib_mcast_iter *iter)
 869 {
 870         struct ipoib_dev_priv *priv = iter->priv;
 871         struct rb_node *n;
 872         struct ipoib_mcast *mcast;
 873         int ret = 1;
 874
 875         spin_lock_irq(&priv->lock);
 876
 877         n = rb_first(&priv->multicast_tree);
 878
 879         while (n) {
 880                 mcast = rb_entry(n, struct ipoib_mcast, rb_node);
 881
 882                 if (memcmp(iter->mgid.raw, mcast->mcmember.mgid.raw,
 883                            sizeof (union ib_gid)) < 0) {
 884                         iter->mgid      = mcast->mcmember.mgid;
 885                         iter->created   = mcast->created;
 886                         iter->queuelen  = mcast->pkt_queue.ifq_len;
 887                         iter->complete  = !!mcast->ah;
 888                         iter->send_only = !!(mcast->flags & (1 << IPOIB_MCAST_FLAG_SENDONLY));
 889
 890                         ret = 0;
 891
 892                         break;
 893                 }
 894
 895                 n = rb_next(n);
 896         }
 897
 898         spin_unlock_irq(&priv->lock);
 899
 900         return ret;
 901 }
 902
 903 void ipoib_mcast_iter_read(struct ipoib_mcast_iter *iter,
 904                            union ib_gid *mgid,
 905                            unsigned long *created,
 906                            unsigned int *queuelen,
 907                            unsigned int *complete,
 908                            unsigned int *send_only)
 909 {
 910         *mgid      = iter->mgid;
 911         *created   = iter->created;
 912         *queuelen  = iter->queuelen;
 913         *complete  = iter->complete;
 914         *send_only = iter->send_only;
 915 }
 916
 917 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */