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/slab.h>
  38 #include <linux/bitops.h>
  39 #include <linux/random.h>
  40
  41 #include <rdma/ib_cache.h>
  42 #include "sa.h"
  43
  44 static void mcast_add_one(struct ib_device *device);
  45 static void mcast_remove_one(struct ib_device *device);
  46
  47 static struct ib_client mcast_client = {
  48         .name   = "ib_multicast",
  49         .add    = mcast_add_one,
  50         .remove = mcast_remove_one
  51 };
  52
  53 static struct ib_sa_client      sa_client;
  54 static struct workqueue_struct  *mcast_wq;
  55 static union ib_gid mgid0;
  56
  57 struct mcast_device;
  58
  59 struct mcast_port {
  60         struct mcast_device     *dev;
  61         spinlock_t              lock;
  62         struct rb_root          table;
  63         atomic_t                refcount;
  64         struct completion       comp;
  65         u8                      port_num;
  66 };
  67
  68 struct mcast_device {
  69         struct ib_device        *device;
  70         struct ib_event_handler event_handler;
  71         int                     start_port;
  72         int                     end_port;
  73         struct mcast_port       port[0];
  74 };
  75
  76 enum mcast_state {
  77         MCAST_JOINING,
  78         MCAST_MEMBER,
  79         MCAST_ERROR,
  80 };
  81
  82 enum mcast_group_state {
  83         MCAST_IDLE,
  84         MCAST_BUSY,
  85         MCAST_GROUP_ERROR,
  86         MCAST_PKEY_EVENT
  87 };
  88
  89 enum {
  90         MCAST_INVALID_PKEY_INDEX = 0xFFFF
  91 };
  92
  93 struct mcast_member;
  94
  95 struct mcast_group {
  96         struct ib_sa_mcmember_rec rec;
  97         struct rb_node          node;
  98         struct mcast_port       *port;
  99         spinlock_t              lock;
 100         struct work_struct      work;
 101         struct list_head        pending_list;
 102         struct list_head        active_list;
 103         struct mcast_member     *last_join;
 104         int                     members[3];
 105         atomic_t                refcount;
 106         enum mcast_group_state  state;
 107         struct ib_sa_query      *query;
 108         int                     query_id;
 109         u16                     pkey_index;
 110         u8                      leave_state;
 111         int                     retries;
 112 };
 113
 114 struct mcast_member {
 115         struct ib_sa_multicast  multicast;
 116         struct ib_sa_client     *client;
 117         struct mcast_group      *group;
 118         struct list_head        list;
 119         enum mcast_state        state;
 120         atomic_t                refcount;
 121         struct completion       comp;
 122 };
 123
 124 static void join_handler(int status, struct ib_sa_mcmember_rec *rec,
 125                          void *context);
 126 static void leave_handler(int status, struct ib_sa_mcmember_rec *rec,
 127                           void *context);
 128
 129 static struct mcast_group *mcast_find(struct mcast_port *port,
 130                                       union ib_gid *mgid)
 131 {
 132         struct rb_node *node = port->table.rb_node;
 133         struct mcast_group *group;
 134         int ret;
 135
 136         while (node) {
 137                 group = rb_entry(node, struct mcast_group, node);
 138                 ret = memcmp(mgid->raw, group->rec.mgid.raw, sizeof *mgid);
 139                 if (!ret)
 140                         return group;
 141
 142                 if (ret < 0)
 143                         node = node->rb_left;
 144                 else
 145                         node = node->rb_right;
 146         }
 147         return NULL;
 148 }
 149
 150 static struct mcast_group *mcast_insert(struct mcast_port *port,
 151                                         struct mcast_group *group,
 152                                         int allow_duplicates)
 153 {
 154         struct rb_node **link = &port->table.rb_node;
 155         struct rb_node *parent = NULL;
 156         struct mcast_group *cur_group;
 157         int ret;
 158
 159         while (*link) {
 160                 parent = *link;
 161                 cur_group = rb_entry(parent, struct mcast_group, node);
 162
 163                 ret = memcmp(group->rec.mgid.raw, cur_group->rec.mgid.raw,
 164                              sizeof group->rec.mgid);
 165                 if (ret < 0)
 166                         link = &(*link)->rb_left;
 167                 else if (ret > 0)
 168                         link = &(*link)->rb_right;
 169                 else if (allow_duplicates)
 170                         link = &(*link)->rb_left;
 171                 else
 172                         return cur_group;
 173         }
 174         rb_link_node(&group->node, parent, link);
 175         rb_insert_color(&group->node, &port->table);
 176         return NULL;
 177 }
 178
 179 static void deref_port(struct mcast_port *port)
 180 {
 181         if (atomic_dec_and_test(&port->refcount))
 182                 complete(&port->comp);
 183 }
 184
 185 static void release_group(struct mcast_group *group)
 186 {
 187         struct mcast_port *port = group->port;
 188         unsigned long flags;
 189
 190         spin_lock_irqsave(&port->lock, flags);
 191         if (atomic_dec_and_test(&group->refcount)) {
 192                 rb_erase(&group->node, &port->table);
 193                 spin_unlock_irqrestore(&port->lock, flags);
 194                 kfree(group);
 195                 deref_port(port);
 196         } else
 197                 spin_unlock_irqrestore(&port->lock, flags);
 198 }
 199
 200 static void deref_member(struct mcast_member *member)
 201 {
 202         if (atomic_dec_and_test(&member->refcount))
 203                 complete(&member->comp);
 204 }
 205
 206 static void queue_join(struct mcast_member *member)
 207 {
 208         struct mcast_group *group = member->group;
 209         unsigned long flags;
 210
 211         spin_lock_irqsave(&group->lock, flags);
 212         list_add_tail(&member->list, &group->pending_list);
 213         if (group->state == MCAST_IDLE) {
 214                 group->state = MCAST_BUSY;
 215                 atomic_inc(&group->refcount);
 216                 queue_work(mcast_wq, &group->work);
 217         }
 218         spin_unlock_irqrestore(&group->lock, flags);
 219 }
 220
 221 /*
 222  * A multicast group has three types of members: full member, non member, and
 223  * send only member.  We need to keep track of the number of members of each
 224  * type based on their join state.  Adjust the number of members the belong to
 225  * the specified join states.
 226  */
 227 static void adjust_membership(struct mcast_group *group, u8 join_state, int inc)
 228 {
 229         int i;
 230
 231         for (i = 0; i < 3; i++, join_state >>= 1)
 232                 if (join_state & 0x1)
 233                         group->members[i] += inc;
 234 }
 235
 236 /*
 237  * If a multicast group has zero members left for a particular join state, but
 238  * the group is still a member with the SA, we need to leave that join state.
 239  * Determine which join states we still belong to, but that do not have any
 240  * active members.
 241  */
 242 static u8 get_leave_state(struct mcast_group *group)
 243 {
 244         u8 leave_state = 0;
 245         int i;
 246
 247         for (i = 0; i < 3; i++)
 248                 if (!group->members[i])
 249                         leave_state |= (0x1 << i);
 250
 251         return leave_state & group->rec.join_state;
 252 }
 253
 254 static int check_selector(ib_sa_comp_mask comp_mask,
 255                           ib_sa_comp_mask selector_mask,
 256                           ib_sa_comp_mask value_mask,
 257                           u8 selector, u8 src_value, u8 dst_value)
 258 {
 259         int err;
 260
 261         if (!(comp_mask & selector_mask) || !(comp_mask & value_mask))
 262                 return 0;
 263
 264         switch (selector) {
 265         case IB_SA_GT:
 266                 err = (src_value <= dst_value);
 267                 break;
 268         case IB_SA_LT:
 269                 err = (src_value >= dst_value);
 270                 break;
 271         case IB_SA_EQ:
 272                 err = (src_value != dst_value);
 273                 break;
 274         default:
 275                 err = 0;
 276                 break;
 277         }
 278
 279         return err;
 280 }
 281
 282 static int cmp_rec(struct ib_sa_mcmember_rec *src,
 283                    struct ib_sa_mcmember_rec *dst, ib_sa_comp_mask comp_mask)
 284 {
 285         /* MGID must already match */
 286
 287         if (comp_mask & IB_SA_MCMEMBER_REC_PORT_GID &&
 288             memcmp(&src->port_gid, &dst->port_gid, sizeof src->port_gid))
 289                 return -EINVAL;
 290         if (comp_mask & IB_SA_MCMEMBER_REC_QKEY && src->qkey != dst->qkey)
 291                 return -EINVAL;
 292         if (comp_mask & IB_SA_MCMEMBER_REC_MLID && src->mlid != dst->mlid)
 293                 return -EINVAL;
 294         if (check_selector(comp_mask, IB_SA_MCMEMBER_REC_MTU_SELECTOR,
 295                            IB_SA_MCMEMBER_REC_MTU, dst->mtu_selector,
 296                            src->mtu, dst->mtu))
 297                 return -EINVAL;
 298         if (comp_mask & IB_SA_MCMEMBER_REC_TRAFFIC_CLASS &&
 299             src->traffic_class != dst->traffic_class)
 300                 return -EINVAL;
 301         if (comp_mask & IB_SA_MCMEMBER_REC_PKEY && src->pkey != dst->pkey)
 302                 return -EINVAL;
 303         if (check_selector(comp_mask, IB_SA_MCMEMBER_REC_RATE_SELECTOR,
 304                            IB_SA_MCMEMBER_REC_RATE, dst->rate_selector,
 305                            src->rate, dst->rate))
 306                 return -EINVAL;
 307         if (check_selector(comp_mask,
 308                            IB_SA_MCMEMBER_REC_PACKET_LIFE_TIME_SELECTOR,
 309                            IB_SA_MCMEMBER_REC_PACKET_LIFE_TIME,
 310                            dst->packet_life_time_selector,
 311                            src->packet_life_time, dst->packet_life_time))
 312                 return -EINVAL;
 313         if (comp_mask & IB_SA_MCMEMBER_REC_SL && src->sl != dst->sl)
 314                 return -EINVAL;
 315         if (comp_mask & IB_SA_MCMEMBER_REC_FLOW_LABEL &&
 316             src->flow_label != dst->flow_label)
 317                 return -EINVAL;
 318         if (comp_mask & IB_SA_MCMEMBER_REC_HOP_LIMIT &&
 319             src->hop_limit != dst->hop_limit)
 320                 return -EINVAL;
 321         if (comp_mask & IB_SA_MCMEMBER_REC_SCOPE && src->scope != dst->scope)
 322                 return -EINVAL;
 323
 324         /* join_state checked separately, proxy_join ignored */
 325
 326         return 0;
 327 }
 328
 329 static int send_join(struct mcast_group *group, struct mcast_member *member)
 330 {
 331         struct mcast_port *port = group->port;
 332         int ret;
 333
 334         group->last_join = member;
 335         ret = ib_sa_mcmember_rec_query(&sa_client, port->dev->device,
 336                                        port->port_num, IB_MGMT_METHOD_SET,
 337                                        &member->multicast.rec,
 338                                        member->multicast.comp_mask,
 339                                        3000, GFP_KERNEL, join_handler, group,
 340                                        &group->query);
 341         if (ret >= 0) {
 342                 group->query_id = ret;
 343                 ret = 0;
 344         }
 345         return ret;
 346 }
 347
 348 static int send_leave(struct mcast_group *group, u8 leave_state)
 349 {
 350         struct mcast_port *port = group->port;
 351         struct ib_sa_mcmember_rec rec;
 352         int ret;
 353
 354         rec = group->rec;
 355         rec.join_state = leave_state;
 356         group->leave_state = leave_state;
 357
 358         ret = ib_sa_mcmember_rec_query(&sa_client, port->dev->device,
 359                                        port->port_num, IB_SA_METHOD_DELETE, &rec,
 360                                        IB_SA_MCMEMBER_REC_MGID     |
 361                                        IB_SA_MCMEMBER_REC_PORT_GID |
 362                                        IB_SA_MCMEMBER_REC_JOIN_STATE,
 363                                        3000, GFP_KERNEL, leave_handler,
 364                                        group, &group->query);
 365         if (ret >= 0) {
 366                 group->query_id = ret;
 367                 ret = 0;
 368         }
 369         return ret;
 370 }
 371
 372 static void join_group(struct mcast_group *group, struct mcast_member *member,
 373                        u8 join_state)
 374 {
 375         member->state = MCAST_MEMBER;
 376         adjust_membership(group, join_state, 1);
 377         group->rec.join_state |= join_state;
 378         member->multicast.rec = group->rec;
 379         member->multicast.rec.join_state = join_state;
 380         list_move(&member->list, &group->active_list);
 381 }
 382
 383 static int fail_join(struct mcast_group *group, struct mcast_member *member,
 384                      int status)
 385 {
 386         spin_lock_irq(&group->lock);
 387         list_del_init(&member->list);
 388         spin_unlock_irq(&group->lock);
 389         return member->multicast.callback(status, &member->multicast);
 390 }
 391
 392 static void process_group_error(struct mcast_group *group)
 393 {
 394         struct mcast_member *member;
 395         int ret = 0;
 396         u16 pkey_index;
 397
 398         if (group->state == MCAST_PKEY_EVENT)
 399                 ret = ib_find_pkey(group->port->dev->device,
 400                                    group->port->port_num,
 401                                    be16_to_cpu(group->rec.pkey), &pkey_index);
 402
 403         spin_lock_irq(&group->lock);
 404         if (group->state == MCAST_PKEY_EVENT && !ret &&
 405             group->pkey_index == pkey_index)
 406                 goto out;
 407
 408         while (!list_empty(&group->active_list)) {
 409                 member = list_entry(group->active_list.next,
 410                                     struct mcast_member, list);
 411                 atomic_inc(&member->refcount);
 412                 list_del_init(&member->list);
 413                 adjust_membership(group, member->multicast.rec.join_state, -1);
 414                 member->state = MCAST_ERROR;
 415                 spin_unlock_irq(&group->lock);
 416
 417                 ret = member->multicast.callback(-ENETRESET,
 418                                                  &member->multicast);
 419                 deref_member(member);
 420                 if (ret)
 421                         ib_sa_free_multicast(&member->multicast);
 422                 spin_lock_irq(&group->lock);
 423         }
 424
 425         group->rec.join_state = 0;
 426 out:
 427         group->state = MCAST_BUSY;
 428         spin_unlock_irq(&group->lock);
 429 }
 430
 431 static void mcast_work_handler(struct work_struct *work)
 432 {
 433         struct mcast_group *group;
 434         struct mcast_member *member;
 435         struct ib_sa_multicast *multicast;
 436         int status, ret;
 437         u8 join_state;
 438
 439         group = container_of(work, typeof(*group), work);
 440 retest:
 441         spin_lock_irq(&group->lock);
 442         while (!list_empty(&group->pending_list) ||
 443                (group->state != MCAST_BUSY)) {
 444
 445                 if (group->state != MCAST_BUSY) {
 446                         spin_unlock_irq(&group->lock);
 447                         process_group_error(group);
 448                         goto retest;
 449                 }
 450
 451                 member = list_entry(group->pending_list.next,
 452                                     struct mcast_member, list);
 453                 multicast = &member->multicast;
 454                 join_state = multicast->rec.join_state;
 455                 atomic_inc(&member->refcount);
 456
 457                 if (join_state == (group->rec.join_state & join_state)) {
 458                         status = cmp_rec(&group->rec, &multicast->rec,
 459                                          multicast->comp_mask);
 460                         if (!status)
 461                                 join_group(group, member, join_state);
 462                         else
 463                                 list_del_init(&member->list);
 464                         spin_unlock_irq(&group->lock);
 465                         ret = multicast->callback(status, multicast);
 466                 } else {
 467                         spin_unlock_irq(&group->lock);
 468                         status = send_join(group, member);
 469                         if (!status) {
 470                                 deref_member(member);
 471                                 return;
 472                         }
 473                         ret = fail_join(group, member, status);
 474                 }
 475
 476                 deref_member(member);
 477                 if (ret)
 478                         ib_sa_free_multicast(&member->multicast);
 479                 spin_lock_irq(&group->lock);
 480         }
 481
 482         join_state = get_leave_state(group);
 483         if (join_state) {
 484                 group->rec.join_state &= ~join_state;
 485                 spin_unlock_irq(&group->lock);
 486                 if (send_leave(group, join_state))
 487                         goto retest;
 488         } else {
 489                 group->state = MCAST_IDLE;
 490                 spin_unlock_irq(&group->lock);
 491                 release_group(group);
 492         }
 493 }
 494
 495 /*
 496  * Fail a join request if it is still active - at the head of the pending queue.
 497  */
 498 static void process_join_error(struct mcast_group *group, int status)
 499 {
 500         struct mcast_member *member;
 501         int ret;
 502
 503         spin_lock_irq(&group->lock);
 504         member = list_entry(group->pending_list.next,
 505                             struct mcast_member, list);
 506         if (group->last_join == member) {
 507                 atomic_inc(&member->refcount);
 508                 list_del_init(&member->list);
 509                 spin_unlock_irq(&group->lock);
 510                 ret = member->multicast.callback(status, &member->multicast);
 511                 deref_member(member);
 512                 if (ret)
 513                         ib_sa_free_multicast(&member->multicast);
 514         } else
 515                 spin_unlock_irq(&group->lock);
 516 }
 517
 518 static void join_handler(int status, struct ib_sa_mcmember_rec *rec,
 519                          void *context)
 520 {
 521         struct mcast_group *group = context;
 522         u16 pkey_index = MCAST_INVALID_PKEY_INDEX;
 523
 524         if (status)
 525                 process_join_error(group, status);
 526         else {
 527                 ib_find_pkey(group->port->dev->device, group->port->port_num,
 528                              be16_to_cpu(rec->pkey), &pkey_index);
 529
 530                 spin_lock_irq(&group->port->lock);
 531                 group->rec = *rec;
 532                 if (group->state == MCAST_BUSY &&
 533                     group->pkey_index == MCAST_INVALID_PKEY_INDEX)
 534                         group->pkey_index = pkey_index;
 535                 if (!memcmp(&mgid0, &group->rec.mgid, sizeof mgid0)) {
 536                         rb_erase(&group->node, &group->port->table);
 537                         mcast_insert(group->port, group, 1);
 538                 }
 539                 spin_unlock_irq(&group->port->lock);
 540         }
 541         mcast_work_handler(&group->work);
 542 }
 543
 544 static void leave_handler(int status, struct ib_sa_mcmember_rec *rec,
 545                           void *context)
 546 {
 547         struct mcast_group *group = context;
 548
 549         if (status && group->retries > 0 &&
 550             !send_leave(group, group->leave_state))
 551                 group->retries--;
 552         else
 553                 mcast_work_handler(&group->work);
 554 }
 555
 556 static struct mcast_group *acquire_group(struct mcast_port *port,
 557                                          union ib_gid *mgid, gfp_t gfp_mask)
 558 {
 559         struct mcast_group *group, *cur_group;
 560         unsigned long flags;
 561         int is_mgid0;
 562
 563         is_mgid0 = !memcmp(&mgid0, mgid, sizeof mgid0);
 564         if (!is_mgid0) {
 565                 spin_lock_irqsave(&port->lock, flags);
 566                 group = mcast_find(port, mgid);
 567                 if (group)
 568                         goto found;
 569                 spin_unlock_irqrestore(&port->lock, flags);
 570         }
 571
 572         group = kzalloc(sizeof *group, gfp_mask);
 573         if (!group)
 574                 return NULL;
 575
 576         group->retries = 3;
 577         group->port = port;
 578         group->rec.mgid = *mgid;
 579         group->pkey_index = MCAST_INVALID_PKEY_INDEX;
 580         INIT_LIST_HEAD(&group->pending_list);
 581         INIT_LIST_HEAD(&group->active_list);
 582         INIT_WORK(&group->work, mcast_work_handler);
 583         spin_lock_init(&group->lock);
 584
 585         spin_lock_irqsave(&port->lock, flags);
 586         cur_group = mcast_insert(port, group, is_mgid0);
 587         if (cur_group) {
 588                 kfree(group);
 589                 group = cur_group;
 590         } else
 591                 atomic_inc(&port->refcount);
 592 found:
 593         atomic_inc(&group->refcount);
 594         spin_unlock_irqrestore(&port->lock, flags);
 595         return group;
 596 }
 597
 598 /*
 599  * We serialize all join requests to a single group to make our lives much
 600  * easier.  Otherwise, two users could try to join the same group
 601  * simultaneously, with different configurations, one could leave while the
 602  * join is in progress, etc., which makes locking around error recovery
 603  * difficult.
 604  */
 605 struct ib_sa_multicast *
 606 ib_sa_join_multicast(struct ib_sa_client *client,
 607                      struct ib_device *device, u8 port_num,
 608                      struct ib_sa_mcmember_rec *rec,
 609                      ib_sa_comp_mask comp_mask, gfp_t gfp_mask,
 610                      int (*callback)(int status,
 611                                      struct ib_sa_multicast *multicast),
 612                      void *context)
 613 {
 614         struct mcast_device *dev;
 615         struct mcast_member *member;
 616         struct ib_sa_multicast *multicast;
 617         int ret;
 618
 619         dev = ib_get_client_data(device, &mcast_client);
 620         if (!dev)
 621                 return ERR_PTR(-ENODEV);
 622
 623         member = kmalloc(sizeof *member, gfp_mask);
 624         if (!member)
 625                 return ERR_PTR(-ENOMEM);
 626
 627         ib_sa_client_get(client);
 628         member->client = client;
 629         member->multicast.rec = *rec;
 630         member->multicast.comp_mask = comp_mask;
 631         member->multicast.callback = callback;
 632         member->multicast.context = context;
 633         init_completion(&member->comp);
 634         atomic_set(&member->refcount, 1);
 635         member->state = MCAST_JOINING;
 636
 637         member->group = acquire_group(&dev->port[port_num - dev->start_port],
 638                                       &rec->mgid, gfp_mask);
 639         if (!member->group) {
 640                 ret = -ENOMEM;
 641                 goto err;
 642         }
 643
 644         /*
 645          * The user will get the multicast structure in their callback.  They
 646          * could then free the multicast structure before we can return from
 647          * this routine.  So we save the pointer to return before queuing
 648          * any callback.
 649          */
 650         multicast = &member->multicast;
 651         queue_join(member);
 652         return multicast;
 653
 654 err:
 655         ib_sa_client_put(client);
 656         kfree(member);
 657         return ERR_PTR(ret);
 658 }
 659 EXPORT_SYMBOL(ib_sa_join_multicast);
 660
 661 void ib_sa_free_multicast(struct ib_sa_multicast *multicast)
 662 {
 663         struct mcast_member *member;
 664         struct mcast_group *group;
 665
 666         member = container_of(multicast, struct mcast_member, multicast);
 667         group = member->group;
 668
 669         spin_lock_irq(&group->lock);
 670         if (member->state == MCAST_MEMBER)
 671                 adjust_membership(group, multicast->rec.join_state, -1);
 672
 673         list_del_init(&member->list);
 674
 675         if (group->state == MCAST_IDLE) {
 676                 group->state = MCAST_BUSY;
 677                 spin_unlock_irq(&group->lock);
 678                 /* Continue to hold reference on group until callback */
 679                 queue_work(mcast_wq, &group->work);
 680         } else {
 681                 spin_unlock_irq(&group->lock);
 682                 release_group(group);
 683         }
 684
 685         deref_member(member);
 686         wait_for_completion(&member->comp);
 687         ib_sa_client_put(member->client);
 688         kfree(member);
 689 }
 690 EXPORT_SYMBOL(ib_sa_free_multicast);
 691
 692 int ib_sa_get_mcmember_rec(struct ib_device *device, u8 port_num,
 693                            union ib_gid *mgid, struct ib_sa_mcmember_rec *rec)
 694 {
 695         struct mcast_device *dev;
 696         struct mcast_port *port;
 697         struct mcast_group *group;
 698         unsigned long flags;
 699         int ret = 0;
 700
 701         dev = ib_get_client_data(device, &mcast_client);
 702         if (!dev)
 703                 return -ENODEV;
 704
 705         port = &dev->port[port_num - dev->start_port];
 706         spin_lock_irqsave(&port->lock, flags);
 707         group = mcast_find(port, mgid);
 708         if (group)
 709                 *rec = group->rec;
 710         else
 711                 ret = -EADDRNOTAVAIL;
 712         spin_unlock_irqrestore(&port->lock, flags);
 713
 714         return ret;
 715 }
 716 EXPORT_SYMBOL(ib_sa_get_mcmember_rec);
 717
 718 int ib_init_ah_from_mcmember(struct ib_device *device, u8 port_num,
 719                              struct ib_sa_mcmember_rec *rec,
 720                              struct ib_ah_attr *ah_attr)
 721 {
 722         int ret;
 723         u16 gid_index;
 724         u8 p;
 725
 726         ret = ib_find_cached_gid(device, &rec->port_gid, &p, &gid_index);
 727         if (ret)
 728                 return ret;
 729
 730         memset(ah_attr, 0, sizeof *ah_attr);
 731         ah_attr->dlid = be16_to_cpu(rec->mlid);
 732         ah_attr->sl = rec->sl;
 733         ah_attr->port_num = port_num;
 734         ah_attr->static_rate = rec->rate;
 735
 736         ah_attr->ah_flags = IB_AH_GRH;
 737         ah_attr->grh.dgid = rec->mgid;
 738
 739         ah_attr->grh.sgid_index = (u8) gid_index;
 740         ah_attr->grh.flow_label = be32_to_cpu(rec->flow_label);
 741         ah_attr->grh.hop_limit = rec->hop_limit;
 742         ah_attr->grh.traffic_class = rec->traffic_class;
 743
 744         return 0;
 745 }
 746 EXPORT_SYMBOL(ib_init_ah_from_mcmember);
 747
 748 static void mcast_groups_event(struct mcast_port *port,
 749                                enum mcast_group_state state)
 750 {
 751         struct mcast_group *group;
 752         struct rb_node *node;
 753         unsigned long flags;
 754
 755         spin_lock_irqsave(&port->lock, flags);
 756         for (node = rb_first(&port->table); node; node = rb_next(node)) {
 757                 group = rb_entry(node, struct mcast_group, node);
 758                 spin_lock(&group->lock);
 759                 if (group->state == MCAST_IDLE) {
 760                         atomic_inc(&group->refcount);
 761                         queue_work(mcast_wq, &group->work);
 762                 }
 763                 if (group->state != MCAST_GROUP_ERROR)
 764                         group->state = state;
 765                 spin_unlock(&group->lock);
 766         }
 767         spin_unlock_irqrestore(&port->lock, flags);
 768 }
 769
 770 static void mcast_event_handler(struct ib_event_handler *handler,
 771                                 struct ib_event *event)
 772 {
 773         struct mcast_device *dev;
 774         int index;
 775
 776         dev = container_of(handler, struct mcast_device, event_handler);
 777         if (rdma_port_get_link_layer(dev->device, event->element.port_num) !=
 778             IB_LINK_LAYER_INFINIBAND)
 779                 return;
 780
 781         index = event->element.port_num - dev->start_port;
 782
 783         switch (event->event) {
 784         case IB_EVENT_PORT_ERR:
 785         case IB_EVENT_LID_CHANGE:
 786         case IB_EVENT_SM_CHANGE:
 787         case IB_EVENT_CLIENT_REREGISTER:
 788                 mcast_groups_event(&dev->port[index], MCAST_GROUP_ERROR);
 789                 break;
 790         case IB_EVENT_PKEY_CHANGE:
 791                 mcast_groups_event(&dev->port[index], MCAST_PKEY_EVENT);
 792                 break;
 793         default:
 794                 break;
 795         }
 796 }
 797
 798 static void mcast_add_one(struct ib_device *device)
 799 {
 800         struct mcast_device *dev;
 801         struct mcast_port *port;
 802         int i;
 803         int count = 0;
 804
 805         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
 806                 return;
 807
 808         dev = kmalloc(sizeof *dev + device->phys_port_cnt * sizeof *port,
 809                       GFP_KERNEL);
 810         if (!dev)
 811                 return;
 812
 813         if (device->node_type == RDMA_NODE_IB_SWITCH)
 814                 dev->start_port = dev->end_port = 0;
 815         else {
 816                 dev->start_port = 1;
 817                 dev->end_port = device->phys_port_cnt;
 818         }
 819
 820         for (i = 0; i <= dev->end_port - dev->start_port; i++) {
 821                 if (rdma_port_get_link_layer(device, dev->start_port + i) !=
 822                     IB_LINK_LAYER_INFINIBAND)
 823                         continue;
 824                 port = &dev->port[i];
 825                 port->dev = dev;
 826                 port->port_num = dev->start_port + i;
 827                 spin_lock_init(&port->lock);
 828                 port->table = RB_ROOT;
 829                 init_completion(&port->comp);
 830                 atomic_set(&port->refcount, 1);
 831                 ++count;
 832         }
 833
 834         if (!count) {
 835                 kfree(dev);
 836                 return;
 837         }
 838
 839         dev->device = device;
 840         ib_set_client_data(device, &mcast_client, dev);
 841
 842         INIT_IB_EVENT_HANDLER(&dev->event_handler, device, mcast_event_handler);
 843         ib_register_event_handler(&dev->event_handler);
 844 }
 845
 846 static void mcast_remove_one(struct ib_device *device)
 847 {
 848         struct mcast_device *dev;
 849         struct mcast_port *port;
 850         int i;
 851
 852         dev = ib_get_client_data(device, &mcast_client);
 853         if (!dev)
 854                 return;
 855
 856         ib_unregister_event_handler(&dev->event_handler);
 857         flush_workqueue(mcast_wq);
 858
 859         for (i = 0; i <= dev->end_port - dev->start_port; i++) {
 860                 if (rdma_port_get_link_layer(device, dev->start_port + i) ==
 861                     IB_LINK_LAYER_INFINIBAND) {
 862                         port = &dev->port[i];
 863                         deref_port(port);
 864                         wait_for_completion(&port->comp);
 865                 }
 866         }
 867
 868         kfree(dev);
 869 }
 870
 871 int mcast_init(void)
 872 {
 873         int ret;
 874
 875         mcast_wq = create_singlethread_workqueue("ib_mcast");
 876         if (!mcast_wq)
 877                 return -ENOMEM;
 878
 879         ib_sa_register_client(&sa_client);
 880
 881         ret = ib_register_client(&mcast_client);
 882         if (ret)
 883                 goto err;
 884         return 0;
 885
 886 err:
 887         ib_sa_unregister_client(&sa_client);
 888         destroy_workqueue(mcast_wq);
 889         return ret;
 890 }
 891
 892 void mcast_cleanup(void)
 893 {
 894         ib_unregister_client(&mcast_client);
 895         ib_sa_unregister_client(&sa_client);
 896         destroy_workqueue(mcast_wq);
 897 }