drivers/infiniband/core/multicast.c

   1 /*
   2  * Copyright (c) 2006 Intel Corporation.  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
  33 #include <linux/completion.h>
  34 #include <linux/dma-mapping.h>
  35 #include <linux/err.h>
  36 #include <linux/interrupt.h>
  37 #include <linux/bitops.h>
  38 #include <linux/random.h>
  39
  40 #include <rdma/ib_cache.h>
  41 #include "sa.h"
  42
  43 static void mcast_add_one(struct ib_device *device);
  44 static void mcast_remove_one(struct ib_device *device);
  45
  46 static struct ib_client mcast_client = {
  47         .name   = "ib_multicast",
  48         .add    = mcast_add_one,
  49         .remove = mcast_remove_one
  50 };
  51
  52 static struct ib_sa_client      sa_client;
  53 static struct workqueue_struct  *mcast_wq;
  54 static union ib_gid mgid0;
  55
  56 struct mcast_device;
  57
  58 struct mcast_port {
  59         struct mcast_device     *dev;
  60         spinlock_t              lock;
  61         struct rb_root          table;
  62         atomic_t                refcount;
  63         struct completion       comp;
  64         u8                      port_num;
  65 };
  66
  67 struct mcast_device {
  68         struct ib_device        *device;
  69         struct ib_event_handler event_handler;
  70         int                     start_port;
  71         int                     end_port;
  72         struct mcast_port       port[0];
  73 };
  74
  75 enum mcast_state {
  76         MCAST_JOINING,
  77         MCAST_MEMBER,
  78         MCAST_ERROR,
  79 };
  80
  81 enum mcast_group_state {
  82         MCAST_IDLE,
  83         MCAST_BUSY,
  84         MCAST_GROUP_ERROR,
  85         MCAST_PKEY_EVENT
  86 };
  87
  88 enum {
  89         MCAST_INVALID_PKEY_INDEX = 0xFFFF
  90 };
  91
  92 struct mcast_member;
  93
  94 struct mcast_group {
  95         struct ib_sa_mcmember_rec rec;
  96         struct rb_node          node;
  97         struct mcast_port       *port;
  98         spinlock_t              lock;
  99         struct work_struct      work;
 100         struct list_head        pending_list;
 101         struct list_head        active_list;
 102         struct mcast_member     *last_join;
 103         int                     members[3];
 104         atomic_t                refcount;
 105         enum mcast_group_state  state;
 106         struct ib_sa_query      *query;
 107         int                     query_id;
 108         u16                     pkey_index;
 109 };
 110
 111 struct mcast_member {
 112         struct ib_sa_multicast  multicast;
 113         struct ib_sa_client     *client;
 114         struct mcast_group      *group;
 115         struct list_head        list;
 116         enum mcast_state        state;
 117         atomic_t                refcount;
 118         struct completion       comp;
 119 };
 120
 121 static void join_handler(int status, struct ib_sa_mcmember_rec *rec,
 122                          void *context);
 123 static void leave_handler(int status, struct ib_sa_mcmember_rec *rec,
 124                           void *context);
 125
 126 static struct mcast_group *mcast_find(struct mcast_port *port,
 127                                       union ib_gid *mgid)
 128 {
 129         struct rb_node *node = port->table.rb_node;
 130         struct mcast_group *group;
 131         int ret;
 132
 133         while (node) {
 134                 group = rb_entry(node, struct mcast_group, node);
 135                 ret = memcmp(mgid->raw, group->rec.mgid.raw, sizeof *mgid);
 136                 if (!ret)
 137                         return group;
 138
 139                 if (ret < 0)
 140                         node = node->rb_left;
 141                 else
 142                         node = node->rb_right;
 143         }
 144         return NULL;
 145 }
 146
 147 static struct mcast_group *mcast_insert(struct mcast_port *port,
 148                                         struct mcast_group *group,
 149                                         int allow_duplicates)
 150 {
 151         struct rb_node **link = &port->table.rb_node;
 152         struct rb_node *parent = NULL;
 153         struct mcast_group *cur_group;
 154         int ret;
 155
 156         while (*link) {
 157                 parent = *link;
 158                 cur_group = rb_entry(parent, struct mcast_group, node);
 159
 160                 ret = memcmp(group->rec.mgid.raw, cur_group->rec.mgid.raw,
 161                              sizeof group->rec.mgid);
 162                 if (ret < 0)
 163                         link = &(*link)->rb_left;
 164                 else if (ret > 0)
 165                         link = &(*link)->rb_right;
 166                 else if (allow_duplicates)
 167                         link = &(*link)->rb_left;
 168                 else
 169                         return cur_group;
 170         }
 171         rb_link_node(&group->node, parent, link);
 172         rb_insert_color(&group->node, &port->table);
 173         return NULL;
 174 }
 175
 176 static void deref_port(struct mcast_port *port)
 177 {
 178         if (atomic_dec_and_test(&port->refcount))
 179                 complete(&port->comp);
 180 }
 181
 182 static void release_group(struct mcast_group *group)
 183 {
 184         struct mcast_port *port = group->port;
 185         unsigned long flags;
 186
 187         spin_lock_irqsave(&port->lock, flags);
 188         if (atomic_dec_and_test(&group->refcount)) {
 189                 rb_erase(&group->node, &port->table);
 190                 spin_unlock_irqrestore(&port->lock, flags);
 191                 kfree(group);
 192                 deref_port(port);
 193         } else
 194                 spin_unlock_irqrestore(&port->lock, flags);
 195 }
 196
 197 static void deref_member(struct mcast_member *member)
 198 {
 199         if (atomic_dec_and_test(&member->refcount))
 200                 complete(&member->comp);
 201 }
 202
 203 static void queue_join(struct mcast_member *member)
 204 {
 205         struct mcast_group *group = member->group;
 206         unsigned long flags;
 207
 208         spin_lock_irqsave(&group->lock, flags);
 209         list_add_tail(&member->list, &group->pending_list);
 210         if (group->state == MCAST_IDLE) {
 211                 group->state = MCAST_BUSY;
 212                 atomic_inc(&group->refcount);
 213                 queue_work(mcast_wq, &group->work);
 214         }
 215         spin_unlock_irqrestore(&group->lock, flags);
 216 }
 217
 218 /*
 219  * A multicast group has three types of members: full member, non member, and
 220  * send only member.  We need to keep track of the number of members of each
 221  * type based on their join state.  Adjust the number of members the belong to
 222  * the specified join states.
 223  */
 224 static void adjust_membership(struct mcast_group *group, u8 join_state, int inc)
 225 {
 226         int i;
 227
 228         for (i = 0; i < 3; i++, join_state >>= 1)
 229                 if (join_state & 0x1)
 230                         group->members[i] += inc;
 231 }
 232
 233 /*
 234  * If a multicast group has zero members left for a particular join state, but
 235  * the group is still a member with the SA, we need to leave that join state.
 236  * Determine which join states we still belong to, but that do not have any
 237  * active members.
 238  */
 239 static u8 get_leave_state(struct mcast_group *group)
 240 {
 241         u8 leave_state = 0;
 242         int i;
 243
 244         for (i = 0; i < 3; i++)
 245                 if (!group->members[i])
 246                         leave_state |= (0x1 << i);
 247
 248         return leave_state & group->rec.join_state;
 249 }
 250
 251 static int check_selector(ib_sa_comp_mask comp_mask,
 252                           ib_sa_comp_mask selector_mask,
 253                           ib_sa_comp_mask value_mask,
 254                           u8 selector, u8 src_value, u8 dst_value)
 255 {
 256         int err;
 257
 258         if (!(comp_mask & selector_mask) || !(comp_mask & value_mask))
 259                 return 0;
 260
 261         switch (selector) {
 262         case IB_SA_GT:
 263                 err = (src_value <= dst_value);
 264                 break;
 265         case IB_SA_LT:
 266                 err = (src_value >= dst_value);
 267                 break;
 268         case IB_SA_EQ:
 269                 err = (src_value != dst_value);
 270                 break;
 271         default:
 272                 err = 0;
 273                 break;
 274         }
 275
 276         return err;
 277 }
 278
 279 static int cmp_rec(struct ib_sa_mcmember_rec *src,
 280                    struct ib_sa_mcmember_rec *dst, ib_sa_comp_mask comp_mask)
 281 {
 282         /* MGID must already match */
 283
 284         if (comp_mask & IB_SA_MCMEMBER_REC_PORT_GID &&
 285             memcmp(&src->port_gid, &dst->port_gid, sizeof src->port_gid))
 286                 return -EINVAL;
 287         if (comp_mask & IB_SA_MCMEMBER_REC_QKEY && src->qkey != dst->qkey)
 288                 return -EINVAL;
 289         if (comp_mask & IB_SA_MCMEMBER_REC_MLID && src->mlid != dst->mlid)
 290                 return -EINVAL;
 291         if (check_selector(comp_mask, IB_SA_MCMEMBER_REC_MTU_SELECTOR,
 292                            IB_SA_MCMEMBER_REC_MTU, dst->mtu_selector,
 293                            src->mtu, dst->mtu))
 294                 return -EINVAL;
 295         if (comp_mask & IB_SA_MCMEMBER_REC_TRAFFIC_CLASS &&
 296             src->traffic_class != dst->traffic_class)
 297                 return -EINVAL;
 298         if (comp_mask & IB_SA_MCMEMBER_REC_PKEY && src->pkey != dst->pkey)
 299                 return -EINVAL;
 300         if (check_selector(comp_mask, IB_SA_MCMEMBER_REC_RATE_SELECTOR,
 301                            IB_SA_MCMEMBER_REC_RATE, dst->rate_selector,
 302                            src->rate, dst->rate))
 303                 return -EINVAL;
 304         if (check_selector(comp_mask,
 305                            IB_SA_MCMEMBER_REC_PACKET_LIFE_TIME_SELECTOR,
 306                            IB_SA_MCMEMBER_REC_PACKET_LIFE_TIME,
 307                            dst->packet_life_time_selector,
 308                            src->packet_life_time, dst->packet_life_time))
 309                 return -EINVAL;
 310         if (comp_mask & IB_SA_MCMEMBER_REC_SL && src->sl != dst->sl)
 311                 return -EINVAL;
 312         if (comp_mask & IB_SA_MCMEMBER_REC_FLOW_LABEL &&
 313             src->flow_label != dst->flow_label)
 314                 return -EINVAL;
 315         if (comp_mask & IB_SA_MCMEMBER_REC_HOP_LIMIT &&
 316             src->hop_limit != dst->hop_limit)
 317                 return -EINVAL;
 318         if (comp_mask & IB_SA_MCMEMBER_REC_SCOPE && src->scope != dst->scope)
 319                 return -EINVAL;
 320
 321         /* join_state checked separately, proxy_join ignored */
 322
 323         return 0;
 324 }
 325
 326 static int send_join(struct mcast_group *group, struct mcast_member *member)
 327 {
 328         struct mcast_port *port = group->port;
 329         int ret;
 330
 331         group->last_join = member;
 332         ret = ib_sa_mcmember_rec_query(&sa_client, port->dev->device,
 333                                        port->port_num, IB_MGMT_METHOD_SET,
 334                                        &member->multicast.rec,
 335                                        member->multicast.comp_mask,
 336                                        3000, GFP_KERNEL, join_handler, group,
 337                                        &group->query);
 338         if (ret >= 0) {
 339                 group->query_id = ret;
 340                 ret = 0;
 341         }
 342         return ret;
 343 }
 344
 345 static int send_leave(struct mcast_group *group, u8 leave_state)
 346 {
 347         struct mcast_port *port = group->port;
 348         struct ib_sa_mcmember_rec rec;
 349         int ret;
 350
 351         rec = group->rec;
 352         rec.join_state = leave_state;
 353
 354         ret = ib_sa_mcmember_rec_query(&sa_client, port->dev->device,
 355                                        port->port_num, IB_SA_METHOD_DELETE, &rec,
 356                                        IB_SA_MCMEMBER_REC_MGID     |
 357                                        IB_SA_MCMEMBER_REC_PORT_GID |
 358                                        IB_SA_MCMEMBER_REC_JOIN_STATE,
 359                                        3000, GFP_KERNEL, leave_handler,
 360                                        group, &group->query);
 361         if (ret >= 0) {
 362                 group->query_id = ret;
 363                 ret = 0;
 364         }
 365         return ret;
 366 }
 367
 368 static void join_group(struct mcast_group *group, struct mcast_member *member,
 369                        u8 join_state)
 370 {
 371         member->state = MCAST_MEMBER;
 372         adjust_membership(group, join_state, 1);
 373         group->rec.join_state |= join_state;
 374         member->multicast.rec = group->rec;
 375         member->multicast.rec.join_state = join_state;
 376         list_move(&member->list, &group->active_list);
 377 }
 378
 379 static int fail_join(struct mcast_group *group, struct mcast_member *member,
 380                      int status)
 381 {
 382         spin_lock_irq(&group->lock);
 383         list_del_init(&member->list);
 384         spin_unlock_irq(&group->lock);
 385         return member->multicast.callback(status, &member->multicast);
 386 }
 387
 388 static void process_group_error(struct mcast_group *group)
 389 {
 390         struct mcast_member *member;
 391         int ret = 0;
 392         u16 pkey_index;
 393
 394         if (group->state == MCAST_PKEY_EVENT)
 395                 ret = ib_find_pkey(group->port->dev->device,
 396                                    group->port->port_num,
 397                                    be16_to_cpu(group->rec.pkey), &pkey_index);
 398
 399         spin_lock_irq(&group->lock);
 400         if (group->state == MCAST_PKEY_EVENT && !ret &&
 401             group->pkey_index == pkey_index)
 402                 goto out;
 403
 404         while (!list_empty(&group->active_list)) {
 405                 member = list_entry(group->active_list.next,
 406                                     struct mcast_member, list);
 407                 atomic_inc(&member->refcount);
 408                 list_del_init(&member->list);
 409                 adjust_membership(group, member->multicast.rec.join_state, -1);
 410                 member->state = MCAST_ERROR;
 411                 spin_unlock_irq(&group->lock);
 412
 413                 ret = member->multicast.callback(-ENETRESET,
 414                                                  &member->multicast);
 415                 deref_member(member);
 416                 if (ret)
 417                         ib_sa_free_multicast(&member->multicast);
 418                 spin_lock_irq(&group->lock);
 419         }
 420
 421         group->rec.join_state = 0;
 422 out:
 423         group->state = MCAST_BUSY;
 424         spin_unlock_irq(&group->lock);
 425 }
 426
 427 static void mcast_work_handler(struct work_struct *work)
 428 {
 429         struct mcast_group *group;
 430         struct mcast_member *member;
 431         struct ib_sa_multicast *multicast;
 432         int status, ret;
 433         u8 join_state;
 434
 435         group = container_of(work, typeof(*group), work);
 436 retest:
 437         spin_lock_irq(&group->lock);
 438         while (!list_empty(&group->pending_list) ||
 439                (group->state != MCAST_BUSY)) {
 440
 441                 if (group->state != MCAST_BUSY) {
 442                         spin_unlock_irq(&group->lock);
 443                         process_group_error(group);
 444                         goto retest;
 445                 }
 446
 447                 member = list_entry(group->pending_list.next,
 448                                     struct mcast_member, list);
 449                 multicast = &member->multicast;
 450                 join_state = multicast->rec.join_state;
 451                 atomic_inc(&member->refcount);
 452
 453                 if (join_state == (group->rec.join_state & join_state)) {
 454                         status = cmp_rec(&group->rec, &multicast->rec,
 455                                          multicast->comp_mask);
 456                         if (!status)
 457                                 join_group(group, member, join_state);
 458                         else
 459                                 list_del_init(&member->list);
 460                         spin_unlock_irq(&group->lock);
 461                         ret = multicast->callback(status, multicast);
 462                 } else {
 463                         spin_unlock_irq(&group->lock);
 464                         status = send_join(group, member);
 465                         if (!status) {
 466                                 deref_member(member);
 467                                 return;
 468                         }
 469                         ret = fail_join(group, member, status);
 470                 }
 471
 472                 deref_member(member);
 473                 if (ret)
 474                         ib_sa_free_multicast(&member->multicast);
 475                 spin_lock_irq(&group->lock);
 476         }
 477
 478         join_state = get_leave_state(group);
 479         if (join_state) {
 480                 group->rec.join_state &= ~join_state;
 481                 spin_unlock_irq(&group->lock);
 482                 if (send_leave(group, join_state))
 483                         goto retest;
 484         } else {
 485                 group->state = MCAST_IDLE;
 486                 spin_unlock_irq(&group->lock);
 487                 release_group(group);
 488         }
 489 }
 490
 491 /*
 492  * Fail a join request if it is still active - at the head of the pending queue.
 493  */
 494 static void process_join_error(struct mcast_group *group, int status)
 495 {
 496         struct mcast_member *member;
 497         int ret;
 498
 499         spin_lock_irq(&group->lock);
 500         member = list_entry(group->pending_list.next,
 501                             struct mcast_member, list);
 502         if (group->last_join == member) {
 503                 atomic_inc(&member->refcount);
 504                 list_del_init(&member->list);
 505                 spin_unlock_irq(&group->lock);
 506                 ret = member->multicast.callback(status, &member->multicast);
 507                 deref_member(member);
 508                 if (ret)
 509                         ib_sa_free_multicast(&member->multicast);
 510         } else
 511                 spin_unlock_irq(&group->lock);
 512 }
 513
 514 static void join_handler(int status, struct ib_sa_mcmember_rec *rec,
 515                          void *context)
 516 {
 517         struct mcast_group *group = context;
 518         u16 pkey_index = MCAST_INVALID_PKEY_INDEX;
 519
 520         if (status)
 521                 process_join_error(group, status);
 522         else {
 523                 ib_find_pkey(group->port->dev->device, group->port->port_num,
 524                              be16_to_cpu(rec->pkey), &pkey_index);
 525
 526                 spin_lock_irq(&group->port->lock);
 527                 group->rec = *rec;
 528                 if (group->state == MCAST_BUSY &&
 529                     group->pkey_index == MCAST_INVALID_PKEY_INDEX)
 530                         group->pkey_index = pkey_index;
 531                 if (!memcmp(&mgid0, &group->rec.mgid, sizeof mgid0)) {
 532                         rb_erase(&group->node, &group->port->table);
 533                         mcast_insert(group->port, group, 1);
 534                 }
 535                 spin_unlock_irq(&group->port->lock);
 536         }
 537         mcast_work_handler(&group->work);
 538 }
 539
 540 static void leave_handler(int status, struct ib_sa_mcmember_rec *rec,
 541                           void *context)
 542 {
 543         struct mcast_group *group = context;
 544
 545         mcast_work_handler(&group->work);
 546 }
 547
 548 static struct mcast_group *acquire_group(struct mcast_port *port,
 549                                          union ib_gid *mgid, gfp_t gfp_mask)
 550 {
 551         struct mcast_group *group, *cur_group;
 552         unsigned long flags;
 553         int is_mgid0;
 554
 555         is_mgid0 = !memcmp(&mgid0, mgid, sizeof mgid0);
 556         if (!is_mgid0) {
 557                 spin_lock_irqsave(&port->lock, flags);
 558                 group = mcast_find(port, mgid);
 559                 if (group)
 560                         goto found;
 561                 spin_unlock_irqrestore(&port->lock, flags);
 562         }
 563
 564         group = kzalloc(sizeof *group, gfp_mask);
 565         if (!group)
 566                 return NULL;
 567
 568         group->port = port;
 569         group->rec.mgid = *mgid;
 570         group->pkey_index = MCAST_INVALID_PKEY_INDEX;
 571         INIT_LIST_HEAD(&group->pending_list);
 572         INIT_LIST_HEAD(&group->active_list);
 573         INIT_WORK(&group->work, mcast_work_handler);
 574         spin_lock_init(&group->lock);
 575
 576         spin_lock_irqsave(&port->lock, flags);
 577         cur_group = mcast_insert(port, group, is_mgid0);
 578         if (cur_group) {
 579                 kfree(group);
 580                 group = cur_group;
 581         } else
 582                 atomic_inc(&port->refcount);
 583 found:
 584         atomic_inc(&group->refcount);
 585         spin_unlock_irqrestore(&port->lock, flags);
 586         return group;
 587 }
 588
 589 /*
 590  * We serialize all join requests to a single group to make our lives much
 591  * easier.  Otherwise, two users could try to join the same group
 592  * simultaneously, with different configurations, one could leave while the
 593  * join is in progress, etc., which makes locking around error recovery
 594  * difficult.
 595  */
 596 struct ib_sa_multicast *
 597 ib_sa_join_multicast(struct ib_sa_client *client,
 598                      struct ib_device *device, u8 port_num,
 599                      struct ib_sa_mcmember_rec *rec,
 600                      ib_sa_comp_mask comp_mask, gfp_t gfp_mask,
 601                      int (*callback)(int status,
 602                                      struct ib_sa_multicast *multicast),
 603                      void *context)
 604 {
 605         struct mcast_device *dev;
 606         struct mcast_member *member;
 607         struct ib_sa_multicast *multicast;
 608         int ret;
 609
 610         dev = ib_get_client_data(device, &mcast_client);
 611         if (!dev)
 612                 return ERR_PTR(-ENODEV);
 613
 614         member = kmalloc(sizeof *member, gfp_mask);
 615         if (!member)
 616                 return ERR_PTR(-ENOMEM);
 617
 618         ib_sa_client_get(client);
 619         member->client = client;
 620         member->multicast.rec = *rec;
 621         member->multicast.comp_mask = comp_mask;
 622         member->multicast.callback = callback;
 623         member->multicast.context = context;
 624         init_completion(&member->comp);
 625         atomic_set(&member->refcount, 1);
 626         member->state = MCAST_JOINING;
 627
 628         member->group = acquire_group(&dev->port[port_num - dev->start_port],
 629                                       &rec->mgid, gfp_mask);
 630         if (!member->group) {
 631                 ret = -ENOMEM;
 632                 goto err;
 633         }
 634
 635         /*
 636          * The user will get the multicast structure in their callback.  They
 637          * could then free the multicast structure before we can return from
 638          * this routine.  So we save the pointer to return before queuing
 639          * any callback.
 640          */
 641         multicast = &member->multicast;
 642         queue_join(member);
 643         return multicast;
 644
 645 err:
 646         ib_sa_client_put(client);
 647         kfree(member);
 648         return ERR_PTR(ret);
 649 }
 650 EXPORT_SYMBOL(ib_sa_join_multicast);
 651
 652 void ib_sa_free_multicast(struct ib_sa_multicast *multicast)
 653 {
 654         struct mcast_member *member;
 655         struct mcast_group *group;
 656
 657         member = container_of(multicast, struct mcast_member, multicast);
 658         group = member->group;
 659
 660         spin_lock_irq(&group->lock);
 661         if (member->state == MCAST_MEMBER)
 662                 adjust_membership(group, multicast->rec.join_state, -1);
 663
 664         list_del_init(&member->list);
 665
 666         if (group->state == MCAST_IDLE) {
 667                 group->state = MCAST_BUSY;
 668                 spin_unlock_irq(&group->lock);
 669                 /* Continue to hold reference on group until callback */
 670                 queue_work(mcast_wq, &group->work);
 671         } else {
 672                 spin_unlock_irq(&group->lock);
 673                 release_group(group);
 674         }
 675
 676         deref_member(member);
 677         wait_for_completion(&member->comp);
 678         ib_sa_client_put(member->client);
 679         kfree(member);
 680 }
 681 EXPORT_SYMBOL(ib_sa_free_multicast);
 682
 683 int ib_sa_get_mcmember_rec(struct ib_device *device, u8 port_num,
 684                            union ib_gid *mgid, struct ib_sa_mcmember_rec *rec)
 685 {
 686         struct mcast_device *dev;
 687         struct mcast_port *port;
 688         struct mcast_group *group;
 689         unsigned long flags;
 690         int ret = 0;
 691
 692         dev = ib_get_client_data(device, &mcast_client);
 693         if (!dev)
 694                 return -ENODEV;
 695
 696         port = &dev->port[port_num - dev->start_port];
 697         spin_lock_irqsave(&port->lock, flags);
 698         group = mcast_find(port, mgid);
 699         if (group)
 700                 *rec = group->rec;
 701         else
 702                 ret = -EADDRNOTAVAIL;
 703         spin_unlock_irqrestore(&port->lock, flags);
 704
 705         return ret;
 706 }
 707 EXPORT_SYMBOL(ib_sa_get_mcmember_rec);
 708
 709 int ib_init_ah_from_mcmember(struct ib_device *device, u8 port_num,
 710                              struct ib_sa_mcmember_rec *rec,
 711                              struct ib_ah_attr *ah_attr)
 712 {
 713         int ret;
 714         u16 gid_index;
 715         u8 p;
 716
 717         ret = ib_find_cached_gid(device, &rec->port_gid, &p, &gid_index);
 718         if (ret)
 719                 return ret;
 720
 721         memset(ah_attr, 0, sizeof *ah_attr);
 722         ah_attr->dlid = be16_to_cpu(rec->mlid);
 723         ah_attr->sl = rec->sl;
 724         ah_attr->port_num = port_num;
 725         ah_attr->static_rate = rec->rate;
 726
 727         ah_attr->ah_flags = IB_AH_GRH;
 728         ah_attr->grh.dgid = rec->mgid;
 729
 730         ah_attr->grh.sgid_index = (u8) gid_index;
 731         ah_attr->grh.flow_label = be32_to_cpu(rec->flow_label);
 732         ah_attr->grh.hop_limit = rec->hop_limit;
 733         ah_attr->grh.traffic_class = rec->traffic_class;
 734
 735         return 0;
 736 }
 737 EXPORT_SYMBOL(ib_init_ah_from_mcmember);
 738
 739 static void mcast_groups_event(struct mcast_port *port,
 740                                enum mcast_group_state state)
 741 {
 742         struct mcast_group *group;
 743         struct rb_node *node;
 744         unsigned long flags;
 745
 746         spin_lock_irqsave(&port->lock, flags);
 747         for (node = rb_first(&port->table); node; node = rb_next(node)) {
 748                 group = rb_entry(node, struct mcast_group, node);
 749                 spin_lock(&group->lock);
 750                 if (group->state == MCAST_IDLE) {
 751                         atomic_inc(&group->refcount);
 752                         queue_work(mcast_wq, &group->work);
 753                 }
 754                 if (group->state != MCAST_GROUP_ERROR)
 755                         group->state = state;
 756                 spin_unlock(&group->lock);
 757         }
 758         spin_unlock_irqrestore(&port->lock, flags);
 759 }
 760
 761 static void mcast_event_handler(struct ib_event_handler *handler,
 762                                 struct ib_event *event)
 763 {
 764         struct mcast_device *dev;
 765         int index;
 766
 767         dev = container_of(handler, struct mcast_device, event_handler);
 768         index = event->element.port_num - dev->start_port;
 769
 770         switch (event->event) {
 771         case IB_EVENT_PORT_ERR:
 772         case IB_EVENT_LID_CHANGE:
 773         case IB_EVENT_SM_CHANGE:
 774         case IB_EVENT_CLIENT_REREGISTER:
 775                 mcast_groups_event(&dev->port[index], MCAST_GROUP_ERROR);
 776                 break;
 777         case IB_EVENT_PKEY_CHANGE:
 778                 mcast_groups_event(&dev->port[index], MCAST_PKEY_EVENT);
 779                 break;
 780         default:
 781                 break;
 782         }
 783 }
 784
 785 static void mcast_add_one(struct ib_device *device)
 786 {
 787         struct mcast_device *dev;
 788         struct mcast_port *port;
 789         int i;
 790
 791         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
 792                 return;
 793
 794         dev = kmalloc(sizeof *dev + device->phys_port_cnt * sizeof *port,
 795                       GFP_KERNEL);
 796         if (!dev)
 797                 return;
 798
 799         if (device->node_type == RDMA_NODE_IB_SWITCH)
 800                 dev->start_port = dev->end_port = 0;
 801         else {
 802                 dev->start_port = 1;
 803                 dev->end_port = device->phys_port_cnt;
 804         }
 805
 806         for (i = 0; i <= dev->end_port - dev->start_port; i++) {
 807                 port = &dev->port[i];
 808                 port->dev = dev;
 809                 port->port_num = dev->start_port + i;
 810                 spin_lock_init(&port->lock);
 811                 port->table = RB_ROOT;
 812                 init_completion(&port->comp);
 813                 atomic_set(&port->refcount, 1);
 814         }
 815
 816         dev->device = device;
 817         ib_set_client_data(device, &mcast_client, dev);
 818
 819         INIT_IB_EVENT_HANDLER(&dev->event_handler, device, mcast_event_handler);
 820         ib_register_event_handler(&dev->event_handler);
 821 }
 822
 823 static void mcast_remove_one(struct ib_device *device)
 824 {
 825         struct mcast_device *dev;
 826         struct mcast_port *port;
 827         int i;
 828
 829         dev = ib_get_client_data(device, &mcast_client);
 830         if (!dev)
 831                 return;
 832
 833         ib_unregister_event_handler(&dev->event_handler);
 834         flush_workqueue(mcast_wq);
 835
 836         for (i = 0; i <= dev->end_port - dev->start_port; i++) {
 837                 port = &dev->port[i];
 838                 deref_port(port);
 839                 wait_for_completion(&port->comp);
 840         }
 841
 842         kfree(dev);
 843 }
 844
 845 int mcast_init(void)
 846 {
 847         int ret;
 848
 849         mcast_wq = create_singlethread_workqueue("ib_mcast");
 850         if (!mcast_wq)
 851                 return -ENOMEM;
 852
 853         ib_sa_register_client(&sa_client);
 854
 855         ret = ib_register_client(&mcast_client);
 856         if (ret)
 857                 goto err;
 858         return 0;
 859
 860 err:
 861         ib_sa_unregister_client(&sa_client);
 862         destroy_workqueue(mcast_wq);
 863         return ret;
 864 }
 865
 866 void mcast_cleanup(void)
 867 {
 868         ib_unregister_client(&mcast_client);
 869         ib_sa_unregister_client(&sa_client);
 870         destroy_workqueue(mcast_wq);
 871 }