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/export.h>
  38 #include <linux/slab.h>
  39 #include <linux/bitops.h>
  40 #include <linux/random.h>
  41
  42 #include <rdma/ib_cache.h>
  43 #include "sa.h"
  44
  45 static void mcast_add_one(struct ib_device *device);
  46 static void mcast_remove_one(struct ib_device *device);
  47
  48 static struct ib_client mcast_client = {
  49         .name   = "ib_multicast",
  50         .add    = mcast_add_one,
  51         .remove = mcast_remove_one
  52 };
  53
  54 static struct ib_sa_client      sa_client;
  55 static struct workqueue_struct  *mcast_wq;
  56 static union ib_gid mgid0;
  57
  58 struct mcast_device;
  59
  60 struct mcast_port {
  61         struct mcast_device     *dev;
  62         spinlock_t              lock;
  63         struct rb_root          table;
  64         atomic_t                refcount;
  65         struct completion       comp;
  66         u8                      port_num;
  67 };
  68
  69 struct mcast_device {
  70         struct ib_device        *device;
  71         struct ib_event_handler event_handler;
  72         int                     start_port;
  73         int                     end_port;
  74         struct mcast_port       port[0];
  75 };
  76
  77 enum mcast_state {
  78         MCAST_JOINING,
  79         MCAST_MEMBER,
  80         MCAST_ERROR,
  81 };
  82
  83 enum mcast_group_state {
  84         MCAST_IDLE,
  85         MCAST_BUSY,
  86         MCAST_GROUP_ERROR,
  87         MCAST_PKEY_EVENT
  88 };
  89
  90 enum {
  91         MCAST_INVALID_PKEY_INDEX = 0xFFFF
  92 };
  93
  94 struct mcast_member;
  95
  96 struct mcast_group {
  97         struct ib_sa_mcmember_rec rec;
  98         struct rb_node          node;
  99         struct mcast_port       *port;
 100         spinlock_t              lock;
 101         struct work_struct      work;
 102         struct list_head        pending_list;
 103         struct list_head        active_list;
 104         struct mcast_member     *last_join;
 105         int                     members[3];
 106         atomic_t                refcount;
 107         enum mcast_group_state  state;
 108         struct ib_sa_query      *query;
 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         return (ret > 0) ? 0 : ret;
 342 }
 343
 344 static int send_leave(struct mcast_group *group, u8 leave_state)
 345 {
 346         struct mcast_port *port = group->port;
 347         struct ib_sa_mcmember_rec rec;
 348         int ret;
 349
 350         rec = group->rec;
 351         rec.join_state = leave_state;
 352         group->leave_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         return (ret > 0) ? 0 : ret;
 362 }
 363
 364 static void join_group(struct mcast_group *group, struct mcast_member *member,
 365                        u8 join_state)
 366 {
 367         member->state = MCAST_MEMBER;
 368         adjust_membership(group, join_state, 1);
 369         group->rec.join_state |= join_state;
 370         member->multicast.rec = group->rec;
 371         member->multicast.rec.join_state = join_state;
 372         list_move(&member->list, &group->active_list);
 373 }
 374
 375 static int fail_join(struct mcast_group *group, struct mcast_member *member,
 376                      int status)
 377 {
 378         spin_lock_irq(&group->lock);
 379         list_del_init(&member->list);
 380         spin_unlock_irq(&group->lock);
 381         return member->multicast.callback(status, &member->multicast);
 382 }
 383
 384 static void process_group_error(struct mcast_group *group)
 385 {
 386         struct mcast_member *member;
 387         int ret = 0;
 388         u16 pkey_index;
 389
 390         if (group->state == MCAST_PKEY_EVENT)
 391                 ret = ib_find_pkey(group->port->dev->device,
 392                                    group->port->port_num,
 393                                    be16_to_cpu(group->rec.pkey), &pkey_index);
 394
 395         spin_lock_irq(&group->lock);
 396         if (group->state == MCAST_PKEY_EVENT && !ret &&
 397             group->pkey_index == pkey_index)
 398                 goto out;
 399
 400         while (!list_empty(&group->active_list)) {
 401                 member = list_entry(group->active_list.next,
 402                                     struct mcast_member, list);
 403                 atomic_inc(&member->refcount);
 404                 list_del_init(&member->list);
 405                 adjust_membership(group, member->multicast.rec.join_state, -1);
 406                 member->state = MCAST_ERROR;
 407                 spin_unlock_irq(&group->lock);
 408
 409                 ret = member->multicast.callback(-ENETRESET,
 410                                                  &member->multicast);
 411                 deref_member(member);
 412                 if (ret)
 413                         ib_sa_free_multicast(&member->multicast);
 414                 spin_lock_irq(&group->lock);
 415         }
 416
 417         group->rec.join_state = 0;
 418 out:
 419         group->state = MCAST_BUSY;
 420         spin_unlock_irq(&group->lock);
 421 }
 422
 423 static void mcast_work_handler(struct work_struct *work)
 424 {
 425         struct mcast_group *group;
 426         struct mcast_member *member;
 427         struct ib_sa_multicast *multicast;
 428         int status, ret;
 429         u8 join_state;
 430
 431         group = container_of(work, typeof(*group), work);
 432 retest:
 433         spin_lock_irq(&group->lock);
 434         while (!list_empty(&group->pending_list) ||
 435                (group->state != MCAST_BUSY)) {
 436
 437                 if (group->state != MCAST_BUSY) {
 438                         spin_unlock_irq(&group->lock);
 439                         process_group_error(group);
 440                         goto retest;
 441                 }
 442
 443                 member = list_entry(group->pending_list.next,
 444                                     struct mcast_member, list);
 445                 multicast = &member->multicast;
 446                 join_state = multicast->rec.join_state;
 447                 atomic_inc(&member->refcount);
 448
 449                 if (join_state == (group->rec.join_state & join_state)) {
 450                         status = cmp_rec(&group->rec, &multicast->rec,
 451                                          multicast->comp_mask);
 452                         if (!status)
 453                                 join_group(group, member, join_state);
 454                         else
 455                                 list_del_init(&member->list);
 456                         spin_unlock_irq(&group->lock);
 457                         ret = multicast->callback(status, multicast);
 458                 } else {
 459                         spin_unlock_irq(&group->lock);
 460                         status = send_join(group, member);
 461                         if (!status) {
 462                                 deref_member(member);
 463                                 return;
 464                         }
 465                         ret = fail_join(group, member, status);
 466                 }
 467
 468                 deref_member(member);
 469                 if (ret)
 470                         ib_sa_free_multicast(&member->multicast);
 471                 spin_lock_irq(&group->lock);
 472         }
 473
 474         join_state = get_leave_state(group);
 475         if (join_state) {
 476                 group->rec.join_state &= ~join_state;
 477                 spin_unlock_irq(&group->lock);
 478                 if (send_leave(group, join_state))
 479                         goto retest;
 480         } else {
 481                 group->state = MCAST_IDLE;
 482                 spin_unlock_irq(&group->lock);
 483                 release_group(group);
 484         }
 485 }
 486
 487 /*
 488  * Fail a join request if it is still active - at the head of the pending queue.
 489  */
 490 static void process_join_error(struct mcast_group *group, int status)
 491 {
 492         struct mcast_member *member;
 493         int ret;
 494
 495         spin_lock_irq(&group->lock);
 496         member = list_entry(group->pending_list.next,
 497                             struct mcast_member, list);
 498         if (group->last_join == member) {
 499                 atomic_inc(&member->refcount);
 500                 list_del_init(&member->list);
 501                 spin_unlock_irq(&group->lock);
 502                 ret = member->multicast.callback(status, &member->multicast);
 503                 deref_member(member);
 504                 if (ret)
 505                         ib_sa_free_multicast(&member->multicast);
 506         } else
 507                 spin_unlock_irq(&group->lock);
 508 }
 509
 510 static void join_handler(int status, struct ib_sa_mcmember_rec *rec,
 511                          void *context)
 512 {
 513         struct mcast_group *group = context;
 514         u16 pkey_index = MCAST_INVALID_PKEY_INDEX;
 515
 516         if (status)
 517                 process_join_error(group, status);
 518         else {
 519                 ib_find_pkey(group->port->dev->device, group->port->port_num,
 520                              be16_to_cpu(rec->pkey), &pkey_index);
 521
 522                 spin_lock_irq(&group->port->lock);
 523                 group->rec = *rec;
 524                 if (group->state == MCAST_BUSY &&
 525                     group->pkey_index == MCAST_INVALID_PKEY_INDEX)
 526                         group->pkey_index = pkey_index;
 527                 if (!memcmp(&mgid0, &group->rec.mgid, sizeof mgid0)) {
 528                         rb_erase(&group->node, &group->port->table);
 529                         mcast_insert(group->port, group, 1);
 530                 }
 531                 spin_unlock_irq(&group->port->lock);
 532         }
 533         mcast_work_handler(&group->work);
 534 }
 535
 536 static void leave_handler(int status, struct ib_sa_mcmember_rec *rec,
 537                           void *context)
 538 {
 539         struct mcast_group *group = context;
 540
 541         if (status && group->retries > 0 &&
 542             !send_leave(group, group->leave_state))
 543                 group->retries--;
 544         else
 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->retries = 3;
 569         group->port = port;
 570         group->rec.mgid = *mgid;
 571         group->pkey_index = MCAST_INVALID_PKEY_INDEX;
 572         INIT_LIST_HEAD(&group->pending_list);
 573         INIT_LIST_HEAD(&group->active_list);
 574         INIT_WORK(&group->work, mcast_work_handler);
 575         spin_lock_init(&group->lock);
 576
 577         spin_lock_irqsave(&port->lock, flags);
 578         cur_group = mcast_insert(port, group, is_mgid0);
 579         if (cur_group) {
 580                 kfree(group);
 581                 group = cur_group;
 582         } else
 583                 atomic_inc(&port->refcount);
 584 found:
 585         atomic_inc(&group->refcount);
 586         spin_unlock_irqrestore(&port->lock, flags);
 587         return group;
 588 }
 589
 590 /*
 591  * We serialize all join requests to a single group to make our lives much
 592  * easier.  Otherwise, two users could try to join the same group
 593  * simultaneously, with different configurations, one could leave while the
 594  * join is in progress, etc., which makes locking around error recovery
 595  * difficult.
 596  */
 597 struct ib_sa_multicast *
 598 ib_sa_join_multicast(struct ib_sa_client *client,
 599                      struct ib_device *device, u8 port_num,
 600                      struct ib_sa_mcmember_rec *rec,
 601                      ib_sa_comp_mask comp_mask, gfp_t gfp_mask,
 602                      int (*callback)(int status,
 603                                      struct ib_sa_multicast *multicast),
 604                      void *context)
 605 {
 606         struct mcast_device *dev;
 607         struct mcast_member *member;
 608         struct ib_sa_multicast *multicast;
 609         int ret;
 610
 611         dev = ib_get_client_data(device, &mcast_client);
 612         if (!dev)
 613                 return ERR_PTR(-ENODEV);
 614
 615         member = kmalloc(sizeof *member, gfp_mask);
 616         if (!member)
 617                 return ERR_PTR(-ENOMEM);
 618
 619         ib_sa_client_get(client);
 620         member->client = client;
 621         member->multicast.rec = *rec;
 622         member->multicast.comp_mask = comp_mask;
 623         member->multicast.callback = callback;
 624         member->multicast.context = context;
 625         init_completion(&member->comp);
 626         atomic_set(&member->refcount, 1);
 627         member->state = MCAST_JOINING;
 628
 629         member->group = acquire_group(&dev->port[port_num - dev->start_port],
 630                                       &rec->mgid, gfp_mask);
 631         if (!member->group) {
 632                 ret = -ENOMEM;
 633                 goto err;
 634         }
 635
 636         /*
 637          * The user will get the multicast structure in their callback.  They
 638          * could then free the multicast structure before we can return from
 639          * this routine.  So we save the pointer to return before queuing
 640          * any callback.
 641          */
 642         multicast = &member->multicast;
 643         queue_join(member);
 644         return multicast;
 645
 646 err:
 647         ib_sa_client_put(client);
 648         kfree(member);
 649         return ERR_PTR(ret);
 650 }
 651 EXPORT_SYMBOL(ib_sa_join_multicast);
 652
 653 void ib_sa_free_multicast(struct ib_sa_multicast *multicast)
 654 {
 655         struct mcast_member *member;
 656         struct mcast_group *group;
 657
 658         member = container_of(multicast, struct mcast_member, multicast);
 659         group = member->group;
 660
 661         spin_lock_irq(&group->lock);
 662         if (member->state == MCAST_MEMBER)
 663                 adjust_membership(group, multicast->rec.join_state, -1);
 664
 665         list_del_init(&member->list);
 666
 667         if (group->state == MCAST_IDLE) {
 668                 group->state = MCAST_BUSY;
 669                 spin_unlock_irq(&group->lock);
 670                 /* Continue to hold reference on group until callback */
 671                 queue_work(mcast_wq, &group->work);
 672         } else {
 673                 spin_unlock_irq(&group->lock);
 674                 release_group(group);
 675         }
 676
 677         deref_member(member);
 678         wait_for_completion(&member->comp);
 679         ib_sa_client_put(member->client);
 680         kfree(member);
 681 }
 682 EXPORT_SYMBOL(ib_sa_free_multicast);
 683
 684 int ib_sa_get_mcmember_rec(struct ib_device *device, u8 port_num,
 685                            union ib_gid *mgid, struct ib_sa_mcmember_rec *rec)
 686 {
 687         struct mcast_device *dev;
 688         struct mcast_port *port;
 689         struct mcast_group *group;
 690         unsigned long flags;
 691         int ret = 0;
 692
 693         dev = ib_get_client_data(device, &mcast_client);
 694         if (!dev)
 695                 return -ENODEV;
 696
 697         port = &dev->port[port_num - dev->start_port];
 698         spin_lock_irqsave(&port->lock, flags);
 699         group = mcast_find(port, mgid);
 700         if (group)
 701                 *rec = group->rec;
 702         else
 703                 ret = -EADDRNOTAVAIL;
 704         spin_unlock_irqrestore(&port->lock, flags);
 705
 706         return ret;
 707 }
 708 EXPORT_SYMBOL(ib_sa_get_mcmember_rec);
 709
 710 int ib_init_ah_from_mcmember(struct ib_device *device, u8 port_num,
 711                              struct ib_sa_mcmember_rec *rec,
 712                              struct ib_ah_attr *ah_attr)
 713 {
 714         int ret;
 715         u16 gid_index;
 716         u8 p;
 717
 718         ret = ib_find_cached_gid(device, &rec->port_gid, &p, &gid_index);
 719         if (ret)
 720                 return ret;
 721
 722         memset(ah_attr, 0, sizeof *ah_attr);
 723         ah_attr->dlid = be16_to_cpu(rec->mlid);
 724         ah_attr->sl = rec->sl;
 725         ah_attr->port_num = port_num;
 726         ah_attr->static_rate = rec->rate;
 727
 728         ah_attr->ah_flags = IB_AH_GRH;
 729         ah_attr->grh.dgid = rec->mgid;
 730
 731         ah_attr->grh.sgid_index = (u8) gid_index;
 732         ah_attr->grh.flow_label = be32_to_cpu(rec->flow_label);
 733         ah_attr->grh.hop_limit = rec->hop_limit;
 734         ah_attr->grh.traffic_class = rec->traffic_class;
 735
 736         return 0;
 737 }
 738 EXPORT_SYMBOL(ib_init_ah_from_mcmember);
 739
 740 static void mcast_groups_event(struct mcast_port *port,
 741                                enum mcast_group_state state)
 742 {
 743         struct mcast_group *group;
 744         struct rb_node *node;
 745         unsigned long flags;
 746
 747         spin_lock_irqsave(&port->lock, flags);
 748         for (node = rb_first(&port->table); node; node = rb_next(node)) {
 749                 group = rb_entry(node, struct mcast_group, node);
 750                 spin_lock(&group->lock);
 751                 if (group->state == MCAST_IDLE) {
 752                         atomic_inc(&group->refcount);
 753                         queue_work(mcast_wq, &group->work);
 754                 }
 755                 if (group->state != MCAST_GROUP_ERROR)
 756                         group->state = state;
 757                 spin_unlock(&group->lock);
 758         }
 759         spin_unlock_irqrestore(&port->lock, flags);
 760 }
 761
 762 static void mcast_event_handler(struct ib_event_handler *handler,
 763                                 struct ib_event *event)
 764 {
 765         struct mcast_device *dev;
 766         int index;
 767
 768         dev = container_of(handler, struct mcast_device, event_handler);
 769         if (rdma_port_get_link_layer(dev->device, event->element.port_num) !=
 770             IB_LINK_LAYER_INFINIBAND)
 771                 return;
 772
 773         index = event->element.port_num - dev->start_port;
 774
 775         switch (event->event) {
 776         case IB_EVENT_PORT_ERR:
 777         case IB_EVENT_LID_CHANGE:
 778         case IB_EVENT_SM_CHANGE:
 779         case IB_EVENT_CLIENT_REREGISTER:
 780                 mcast_groups_event(&dev->port[index], MCAST_GROUP_ERROR);
 781                 break;
 782         case IB_EVENT_PKEY_CHANGE:
 783                 mcast_groups_event(&dev->port[index], MCAST_PKEY_EVENT);
 784                 break;
 785         default:
 786                 break;
 787         }
 788 }
 789
 790 static void mcast_add_one(struct ib_device *device)
 791 {
 792         struct mcast_device *dev;
 793         struct mcast_port *port;
 794         int i;
 795         int count = 0;
 796
 797         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
 798                 return;
 799
 800         dev = kmalloc(sizeof *dev + device->phys_port_cnt * sizeof *port,
 801                       GFP_KERNEL);
 802         if (!dev)
 803                 return;
 804
 805         if (device->node_type == RDMA_NODE_IB_SWITCH)
 806                 dev->start_port = dev->end_port = 0;
 807         else {
 808                 dev->start_port = 1;
 809                 dev->end_port = device->phys_port_cnt;
 810         }
 811
 812         for (i = 0; i <= dev->end_port - dev->start_port; i++) {
 813                 if (rdma_port_get_link_layer(device, dev->start_port + i) !=
 814                     IB_LINK_LAYER_INFINIBAND)
 815                         continue;
 816                 port = &dev->port[i];
 817                 port->dev = dev;
 818                 port->port_num = dev->start_port + i;
 819                 spin_lock_init(&port->lock);
 820                 port->table = RB_ROOT;
 821                 init_completion(&port->comp);
 822                 atomic_set(&port->refcount, 1);
 823                 ++count;
 824         }
 825
 826         if (!count) {
 827                 kfree(dev);
 828                 return;
 829         }
 830
 831         dev->device = device;
 832         ib_set_client_data(device, &mcast_client, dev);
 833
 834         INIT_IB_EVENT_HANDLER(&dev->event_handler, device, mcast_event_handler);
 835         ib_register_event_handler(&dev->event_handler);
 836 }
 837
 838 static void mcast_remove_one(struct ib_device *device)
 839 {
 840         struct mcast_device *dev;
 841         struct mcast_port *port;
 842         int i;
 843
 844         dev = ib_get_client_data(device, &mcast_client);
 845         if (!dev)
 846                 return;
 847
 848         ib_unregister_event_handler(&dev->event_handler);
 849         flush_workqueue(mcast_wq);
 850
 851         for (i = 0; i <= dev->end_port - dev->start_port; i++) {
 852                 if (rdma_port_get_link_layer(device, dev->start_port + i) ==
 853                     IB_LINK_LAYER_INFINIBAND) {
 854                         port = &dev->port[i];
 855                         deref_port(port);
 856                         wait_for_completion(&port->comp);
 857                 }
 858         }
 859
 860         kfree(dev);
 861 }
 862
 863 int mcast_init(void)
 864 {
 865         int ret;
 866
 867         mcast_wq = create_singlethread_workqueue("ib_mcast");
 868         if (!mcast_wq)
 869                 return -ENOMEM;
 870
 871         ib_sa_register_client(&sa_client);
 872
 873         ret = ib_register_client(&mcast_client);
 874         if (ret)
 875                 goto err;
 876         return 0;
 877
 878 err:
 879         ib_sa_unregister_client(&sa_client);
 880         destroy_workqueue(mcast_wq);
 881         return ret;
 882 }
 883
 884 void mcast_cleanup(void)
 885 {
 886         ib_unregister_client(&mcast_client);
 887         ib_sa_unregister_client(&sa_client);
 888         destroy_workqueue(mcast_wq);
 889 }