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, void *client_data);
  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[NUM_JOIN_MEMBERSHIP_TYPES];
 106         atomic_t                refcount;
 107         enum mcast_group_state  state;
 108         struct ib_sa_query      *query;
 109         int                     query_id;
 110         u16                     pkey_index;
 111         u8                      leave_state;
 112         int                     retries;
 113 };
 114
 115 struct mcast_member {
 116         struct ib_sa_multicast  multicast;
 117         struct ib_sa_client     *client;
 118         struct mcast_group      *group;
 119         struct list_head        list;
 120         enum mcast_state        state;
 121         atomic_t                refcount;
 122         struct completion       comp;
 123 };
 124
 125 static void join_handler(int status, struct ib_sa_mcmember_rec *rec,
 126                          void *context);
 127 static void leave_handler(int status, struct ib_sa_mcmember_rec *rec,
 128                           void *context);
 129
 130 static struct mcast_group *mcast_find(struct mcast_port *port,
 131                                       union ib_gid *mgid)
 132 {
 133         struct rb_node *node = port->table.rb_node;
 134         struct mcast_group *group;
 135         int ret;
 136
 137         while (node) {
 138                 group = rb_entry(node, struct mcast_group, node);
 139                 ret = memcmp(mgid->raw, group->rec.mgid.raw, sizeof *mgid);
 140                 if (!ret)
 141                         return group;
 142
 143                 if (ret < 0)
 144                         node = node->rb_left;
 145                 else
 146                         node = node->rb_right;
 147         }
 148         return NULL;
 149 }
 150
 151 static struct mcast_group *mcast_insert(struct mcast_port *port,
 152                                         struct mcast_group *group,
 153                                         int allow_duplicates)
 154 {
 155         struct rb_node **link = &port->table.rb_node;
 156         struct rb_node *parent = NULL;
 157         struct mcast_group *cur_group;
 158         int ret;
 159
 160         while (*link) {
 161                 parent = *link;
 162                 cur_group = rb_entry(parent, struct mcast_group, node);
 163
 164                 ret = memcmp(group->rec.mgid.raw, cur_group->rec.mgid.raw,
 165                              sizeof group->rec.mgid);
 166                 if (ret < 0)
 167                         link = &(*link)->rb_left;
 168                 else if (ret > 0)
 169                         link = &(*link)->rb_right;
 170                 else if (allow_duplicates)
 171                         link = &(*link)->rb_left;
 172                 else
 173                         return cur_group;
 174         }
 175         rb_link_node(&group->node, parent, link);
 176         rb_insert_color(&group->node, &port->table);
 177         return NULL;
 178 }
 179
 180 static void deref_port(struct mcast_port *port)
 181 {
 182         if (atomic_dec_and_test(&port->refcount))
 183                 complete(&port->comp);
 184 }
 185
 186 static void release_group(struct mcast_group *group)
 187 {
 188         struct mcast_port *port = group->port;
 189         unsigned long flags;
 190
 191         spin_lock_irqsave(&port->lock, flags);
 192         if (atomic_dec_and_test(&group->refcount)) {
 193                 rb_erase(&group->node, &port->table);
 194                 spin_unlock_irqrestore(&port->lock, flags);
 195                 kfree(group);
 196                 deref_port(port);
 197         } else
 198                 spin_unlock_irqrestore(&port->lock, flags);
 199 }
 200
 201 static void deref_member(struct mcast_member *member)
 202 {
 203         if (atomic_dec_and_test(&member->refcount))
 204                 complete(&member->comp);
 205 }
 206
 207 static void queue_join(struct mcast_member *member)
 208 {
 209         struct mcast_group *group = member->group;
 210         unsigned long flags;
 211
 212         spin_lock_irqsave(&group->lock, flags);
 213         list_add_tail(&member->list, &group->pending_list);
 214         if (group->state == MCAST_IDLE) {
 215                 group->state = MCAST_BUSY;
 216                 atomic_inc(&group->refcount);
 217                 queue_work(mcast_wq, &group->work);
 218         }
 219         spin_unlock_irqrestore(&group->lock, flags);
 220 }
 221
 222 /*
 223  * A multicast group has four types of members: full member, non member,
 224  * sendonly non member and sendonly full member.
 225  * We need to keep track of the number of members of each
 226  * type based on their join state.  Adjust the number of members the belong to
 227  * the specified join states.
 228  */
 229 static void adjust_membership(struct mcast_group *group, u8 join_state, int inc)
 230 {
 231         int i;
 232
 233         for (i = 0; i < NUM_JOIN_MEMBERSHIP_TYPES; i++, join_state >>= 1)
 234                 if (join_state & 0x1)
 235                         group->members[i] += inc;
 236 }
 237
 238 /*
 239  * If a multicast group has zero members left for a particular join state, but
 240  * the group is still a member with the SA, we need to leave that join state.
 241  * Determine which join states we still belong to, but that do not have any
 242  * active members.
 243  */
 244 static u8 get_leave_state(struct mcast_group *group)
 245 {
 246         u8 leave_state = 0;
 247         int i;
 248
 249         for (i = 0; i < NUM_JOIN_MEMBERSHIP_TYPES; i++)
 250                 if (!group->members[i])
 251                         leave_state |= (0x1 << i);
 252
 253         return leave_state & group->rec.join_state;
 254 }
 255
 256 static int check_selector(ib_sa_comp_mask comp_mask,
 257                           ib_sa_comp_mask selector_mask,
 258                           ib_sa_comp_mask value_mask,
 259                           u8 selector, u8 src_value, u8 dst_value)
 260 {
 261         int err;
 262
 263         if (!(comp_mask & selector_mask) || !(comp_mask & value_mask))
 264                 return 0;
 265
 266         switch (selector) {
 267         case IB_SA_GT:
 268                 err = (src_value <= dst_value);
 269                 break;
 270         case IB_SA_LT:
 271                 err = (src_value >= dst_value);
 272                 break;
 273         case IB_SA_EQ:
 274                 err = (src_value != dst_value);
 275                 break;
 276         default:
 277                 err = 0;
 278                 break;
 279         }
 280
 281         return err;
 282 }
 283
 284 static int cmp_rec(struct ib_sa_mcmember_rec *src,
 285                    struct ib_sa_mcmember_rec *dst, ib_sa_comp_mask comp_mask)
 286 {
 287         /* MGID must already match */
 288
 289         if (comp_mask & IB_SA_MCMEMBER_REC_PORT_GID &&
 290             memcmp(&src->port_gid, &dst->port_gid, sizeof src->port_gid))
 291                 return -EINVAL;
 292         if (comp_mask & IB_SA_MCMEMBER_REC_QKEY && src->qkey != dst->qkey)
 293                 return -EINVAL;
 294         if (comp_mask & IB_SA_MCMEMBER_REC_MLID && src->mlid != dst->mlid)
 295                 return -EINVAL;
 296         if (check_selector(comp_mask, IB_SA_MCMEMBER_REC_MTU_SELECTOR,
 297                            IB_SA_MCMEMBER_REC_MTU, dst->mtu_selector,
 298                            src->mtu, dst->mtu))
 299                 return -EINVAL;
 300         if (comp_mask & IB_SA_MCMEMBER_REC_TRAFFIC_CLASS &&
 301             src->traffic_class != dst->traffic_class)
 302                 return -EINVAL;
 303         if (comp_mask & IB_SA_MCMEMBER_REC_PKEY && src->pkey != dst->pkey)
 304                 return -EINVAL;
 305         if (check_selector(comp_mask, IB_SA_MCMEMBER_REC_RATE_SELECTOR,
 306                            IB_SA_MCMEMBER_REC_RATE, dst->rate_selector,
 307                            src->rate, dst->rate))
 308                 return -EINVAL;
 309         if (check_selector(comp_mask,
 310                            IB_SA_MCMEMBER_REC_PACKET_LIFE_TIME_SELECTOR,
 311                            IB_SA_MCMEMBER_REC_PACKET_LIFE_TIME,
 312                            dst->packet_life_time_selector,
 313                            src->packet_life_time, dst->packet_life_time))
 314                 return -EINVAL;
 315         if (comp_mask & IB_SA_MCMEMBER_REC_SL && src->sl != dst->sl)
 316                 return -EINVAL;
 317         if (comp_mask & IB_SA_MCMEMBER_REC_FLOW_LABEL &&
 318             src->flow_label != dst->flow_label)
 319                 return -EINVAL;
 320         if (comp_mask & IB_SA_MCMEMBER_REC_HOP_LIMIT &&
 321             src->hop_limit != dst->hop_limit)
 322                 return -EINVAL;
 323         if (comp_mask & IB_SA_MCMEMBER_REC_SCOPE && src->scope != dst->scope)
 324                 return -EINVAL;
 325
 326         /* join_state checked separately, proxy_join ignored */
 327
 328         return 0;
 329 }
 330
 331 static int send_join(struct mcast_group *group, struct mcast_member *member)
 332 {
 333         struct mcast_port *port = group->port;
 334         int ret;
 335
 336         group->last_join = member;
 337         ret = ib_sa_mcmember_rec_query(&sa_client, port->dev->device,
 338                                        port->port_num, IB_MGMT_METHOD_SET,
 339                                        &member->multicast.rec,
 340                                        member->multicast.comp_mask,
 341                                        3000, GFP_KERNEL, join_handler, group,
 342                                        &group->query);
 343         if (ret >= 0) {
 344                 group->query_id = ret;
 345                 ret = 0;
 346         }
 347         return ret;
 348 }
 349
 350 static int send_leave(struct mcast_group *group, u8 leave_state)
 351 {
 352         struct mcast_port *port = group->port;
 353         struct ib_sa_mcmember_rec rec;
 354         int ret;
 355
 356         rec = group->rec;
 357         rec.join_state = leave_state;
 358         group->leave_state = leave_state;
 359
 360         ret = ib_sa_mcmember_rec_query(&sa_client, port->dev->device,
 361                                        port->port_num, IB_SA_METHOD_DELETE, &rec,
 362                                        IB_SA_MCMEMBER_REC_MGID     |
 363                                        IB_SA_MCMEMBER_REC_PORT_GID |
 364                                        IB_SA_MCMEMBER_REC_JOIN_STATE,
 365                                        3000, GFP_KERNEL, leave_handler,
 366                                        group, &group->query);
 367         if (ret >= 0) {
 368                 group->query_id = ret;
 369                 ret = 0;
 370         }
 371         return ret;
 372 }
 373
 374 static void join_group(struct mcast_group *group, struct mcast_member *member,
 375                        u8 join_state)
 376 {
 377         member->state = MCAST_MEMBER;
 378         adjust_membership(group, join_state, 1);
 379         group->rec.join_state |= join_state;
 380         member->multicast.rec = group->rec;
 381         member->multicast.rec.join_state = join_state;
 382         list_move(&member->list, &group->active_list);
 383 }
 384
 385 static int fail_join(struct mcast_group *group, struct mcast_member *member,
 386                      int status)
 387 {
 388         spin_lock_irq(&group->lock);
 389         list_del_init(&member->list);
 390         spin_unlock_irq(&group->lock);
 391         return member->multicast.callback(status, &member->multicast);
 392 }
 393
 394 static void process_group_error(struct mcast_group *group)
 395 {
 396         struct mcast_member *member;
 397         int ret = 0;
 398         u16 pkey_index;
 399
 400         if (group->state == MCAST_PKEY_EVENT)
 401                 ret = ib_find_pkey(group->port->dev->device,
 402                                    group->port->port_num,
 403                                    be16_to_cpu(group->rec.pkey), &pkey_index);
 404
 405         spin_lock_irq(&group->lock);
 406         if (group->state == MCAST_PKEY_EVENT && !ret &&
 407             group->pkey_index == pkey_index)
 408                 goto out;
 409
 410         while (!list_empty(&group->active_list)) {
 411                 member = list_entry(group->active_list.next,
 412                                     struct mcast_member, list);
 413                 atomic_inc(&member->refcount);
 414                 list_del_init(&member->list);
 415                 adjust_membership(group, member->multicast.rec.join_state, -1);
 416                 member->state = MCAST_ERROR;
 417                 spin_unlock_irq(&group->lock);
 418
 419                 ret = member->multicast.callback(-ENETRESET,
 420                                                  &member->multicast);
 421                 deref_member(member);
 422                 if (ret)
 423                         ib_sa_free_multicast(&member->multicast);
 424                 spin_lock_irq(&group->lock);
 425         }
 426
 427         group->rec.join_state = 0;
 428 out:
 429         group->state = MCAST_BUSY;
 430         spin_unlock_irq(&group->lock);
 431 }
 432
 433 static void mcast_work_handler(struct work_struct *work)
 434 {
 435         struct mcast_group *group;
 436         struct mcast_member *member;
 437         struct ib_sa_multicast *multicast;
 438         int status, ret;
 439         u8 join_state;
 440
 441         group = container_of(work, typeof(*group), work);
 442 retest:
 443         spin_lock_irq(&group->lock);
 444         while (!list_empty(&group->pending_list) ||
 445                (group->state != MCAST_BUSY)) {
 446
 447                 if (group->state != MCAST_BUSY) {
 448                         spin_unlock_irq(&group->lock);
 449                         process_group_error(group);
 450                         goto retest;
 451                 }
 452
 453                 member = list_entry(group->pending_list.next,
 454                                     struct mcast_member, list);
 455                 multicast = &member->multicast;
 456                 join_state = multicast->rec.join_state;
 457                 atomic_inc(&member->refcount);
 458
 459                 if (join_state == (group->rec.join_state & join_state)) {
 460                         status = cmp_rec(&group->rec, &multicast->rec,
 461                                          multicast->comp_mask);
 462                         if (!status)
 463                                 join_group(group, member, join_state);
 464                         else
 465                                 list_del_init(&member->list);
 466                         spin_unlock_irq(&group->lock);
 467                         ret = multicast->callback(status, multicast);
 468                 } else {
 469                         spin_unlock_irq(&group->lock);
 470                         status = send_join(group, member);
 471                         if (!status) {
 472                                 deref_member(member);
 473                                 return;
 474                         }
 475                         ret = fail_join(group, member, status);
 476                 }
 477
 478                 deref_member(member);
 479                 if (ret)
 480                         ib_sa_free_multicast(&member->multicast);
 481                 spin_lock_irq(&group->lock);
 482         }
 483
 484         join_state = get_leave_state(group);
 485         if (join_state) {
 486                 group->rec.join_state &= ~join_state;
 487                 spin_unlock_irq(&group->lock);
 488                 if (send_leave(group, join_state))
 489                         goto retest;
 490         } else {
 491                 group->state = MCAST_IDLE;
 492                 spin_unlock_irq(&group->lock);
 493                 release_group(group);
 494         }
 495 }
 496
 497 /*
 498  * Fail a join request if it is still active - at the head of the pending queue.
 499  */
 500 static void process_join_error(struct mcast_group *group, int status)
 501 {
 502         struct mcast_member *member;
 503         int ret;
 504
 505         spin_lock_irq(&group->lock);
 506         member = list_entry(group->pending_list.next,
 507                             struct mcast_member, list);
 508         if (group->last_join == member) {
 509                 atomic_inc(&member->refcount);
 510                 list_del_init(&member->list);
 511                 spin_unlock_irq(&group->lock);
 512                 ret = member->multicast.callback(status, &member->multicast);
 513                 deref_member(member);
 514                 if (ret)
 515                         ib_sa_free_multicast(&member->multicast);
 516         } else
 517                 spin_unlock_irq(&group->lock);
 518 }
 519
 520 static void join_handler(int status, struct ib_sa_mcmember_rec *rec,
 521                          void *context)
 522 {
 523         struct mcast_group *group = context;
 524         u16 pkey_index = MCAST_INVALID_PKEY_INDEX;
 525
 526         if (status)
 527                 process_join_error(group, status);
 528         else {
 529                 int mgids_changed, is_mgid0;
 530                 ib_find_pkey(group->port->dev->device, group->port->port_num,
 531                              be16_to_cpu(rec->pkey), &pkey_index);
 532
 533                 spin_lock_irq(&group->port->lock);
 534                 if (group->state == MCAST_BUSY &&
 535                     group->pkey_index == MCAST_INVALID_PKEY_INDEX)
 536                         group->pkey_index = pkey_index;
 537                 mgids_changed = memcmp(&rec->mgid, &group->rec.mgid,
 538                                        sizeof(group->rec.mgid));
 539                 group->rec = *rec;
 540                 if (mgids_changed) {
 541                         rb_erase(&group->node, &group->port->table);
 542                         is_mgid0 = !memcmp(&mgid0, &group->rec.mgid,
 543                                            sizeof(mgid0));
 544                         mcast_insert(group->port, group, is_mgid0);
 545                 }
 546                 spin_unlock_irq(&group->port->lock);
 547         }
 548         mcast_work_handler(&group->work);
 549 }
 550
 551 static void leave_handler(int status, struct ib_sa_mcmember_rec *rec,
 552                           void *context)
 553 {
 554         struct mcast_group *group = context;
 555
 556         if (status && group->retries > 0 &&
 557             !send_leave(group, group->leave_state))
 558                 group->retries--;
 559         else
 560                 mcast_work_handler(&group->work);
 561 }
 562
 563 static struct mcast_group *acquire_group(struct mcast_port *port,
 564                                          union ib_gid *mgid, gfp_t gfp_mask)
 565 {
 566         struct mcast_group *group, *cur_group;
 567         unsigned long flags;
 568         int is_mgid0;
 569
 570         is_mgid0 = !memcmp(&mgid0, mgid, sizeof mgid0);
 571         if (!is_mgid0) {
 572                 spin_lock_irqsave(&port->lock, flags);
 573                 group = mcast_find(port, mgid);
 574                 if (group)
 575                         goto found;
 576                 spin_unlock_irqrestore(&port->lock, flags);
 577         }
 578
 579         group = kzalloc(sizeof *group, gfp_mask);
 580         if (!group)
 581                 return NULL;
 582
 583         group->retries = 3;
 584         group->port = port;
 585         group->rec.mgid = *mgid;
 586         group->pkey_index = MCAST_INVALID_PKEY_INDEX;
 587         INIT_LIST_HEAD(&group->pending_list);
 588         INIT_LIST_HEAD(&group->active_list);
 589         INIT_WORK(&group->work, mcast_work_handler);
 590         spin_lock_init(&group->lock);
 591
 592         spin_lock_irqsave(&port->lock, flags);
 593         cur_group = mcast_insert(port, group, is_mgid0);
 594         if (cur_group) {
 595                 kfree(group);
 596                 group = cur_group;
 597         } else
 598                 atomic_inc(&port->refcount);
 599 found:
 600         atomic_inc(&group->refcount);
 601         spin_unlock_irqrestore(&port->lock, flags);
 602         return group;
 603 }
 604
 605 /*
 606  * We serialize all join requests to a single group to make our lives much
 607  * easier.  Otherwise, two users could try to join the same group
 608  * simultaneously, with different configurations, one could leave while the
 609  * join is in progress, etc., which makes locking around error recovery
 610  * difficult.
 611  */
 612 struct ib_sa_multicast *
 613 ib_sa_join_multicast(struct ib_sa_client *client,
 614                      struct ib_device *device, u8 port_num,
 615                      struct ib_sa_mcmember_rec *rec,
 616                      ib_sa_comp_mask comp_mask, gfp_t gfp_mask,
 617                      int (*callback)(int status,
 618                                      struct ib_sa_multicast *multicast),
 619                      void *context)
 620 {
 621         struct mcast_device *dev;
 622         struct mcast_member *member;
 623         struct ib_sa_multicast *multicast;
 624         int ret;
 625
 626         dev = ib_get_client_data(device, &mcast_client);
 627         if (!dev)
 628                 return ERR_PTR(-ENODEV);
 629
 630         member = kmalloc(sizeof *member, gfp_mask);
 631         if (!member)
 632                 return ERR_PTR(-ENOMEM);
 633
 634         ib_sa_client_get(client);
 635         member->client = client;
 636         member->multicast.rec = *rec;
 637         member->multicast.comp_mask = comp_mask;
 638         member->multicast.callback = callback;
 639         member->multicast.context = context;
 640         init_completion(&member->comp);
 641         atomic_set(&member->refcount, 1);
 642         member->state = MCAST_JOINING;
 643
 644         member->group = acquire_group(&dev->port[port_num - dev->start_port],
 645                                       &rec->mgid, gfp_mask);
 646         if (!member->group) {
 647                 ret = -ENOMEM;
 648                 goto err;
 649         }
 650
 651         /*
 652          * The user will get the multicast structure in their callback.  They
 653          * could then free the multicast structure before we can return from
 654          * this routine.  So we save the pointer to return before queuing
 655          * any callback.
 656          */
 657         multicast = &member->multicast;
 658         queue_join(member);
 659         return multicast;
 660
 661 err:
 662         ib_sa_client_put(client);
 663         kfree(member);
 664         return ERR_PTR(ret);
 665 }
 666 EXPORT_SYMBOL(ib_sa_join_multicast);
 667
 668 void ib_sa_free_multicast(struct ib_sa_multicast *multicast)
 669 {
 670         struct mcast_member *member;
 671         struct mcast_group *group;
 672
 673         member = container_of(multicast, struct mcast_member, multicast);
 674         group = member->group;
 675
 676         spin_lock_irq(&group->lock);
 677         if (member->state == MCAST_MEMBER)
 678                 adjust_membership(group, multicast->rec.join_state, -1);
 679
 680         list_del_init(&member->list);
 681
 682         if (group->state == MCAST_IDLE) {
 683                 group->state = MCAST_BUSY;
 684                 spin_unlock_irq(&group->lock);
 685                 /* Continue to hold reference on group until callback */
 686                 queue_work(mcast_wq, &group->work);
 687         } else {
 688                 spin_unlock_irq(&group->lock);
 689                 release_group(group);
 690         }
 691
 692         deref_member(member);
 693         wait_for_completion(&member->comp);
 694         ib_sa_client_put(member->client);
 695         kfree(member);
 696 }
 697 EXPORT_SYMBOL(ib_sa_free_multicast);
 698
 699 int ib_sa_get_mcmember_rec(struct ib_device *device, u8 port_num,
 700                            union ib_gid *mgid, struct ib_sa_mcmember_rec *rec)
 701 {
 702         struct mcast_device *dev;
 703         struct mcast_port *port;
 704         struct mcast_group *group;
 705         unsigned long flags;
 706         int ret = 0;
 707
 708         dev = ib_get_client_data(device, &mcast_client);
 709         if (!dev)
 710                 return -ENODEV;
 711
 712         port = &dev->port[port_num - dev->start_port];
 713         spin_lock_irqsave(&port->lock, flags);
 714         group = mcast_find(port, mgid);
 715         if (group)
 716                 *rec = group->rec;
 717         else
 718                 ret = -EADDRNOTAVAIL;
 719         spin_unlock_irqrestore(&port->lock, flags);
 720
 721         return ret;
 722 }
 723 EXPORT_SYMBOL(ib_sa_get_mcmember_rec);
 724
 725 int ib_init_ah_from_mcmember(struct ib_device *device, u8 port_num,
 726                              struct ib_sa_mcmember_rec *rec,
 727                              struct net_device *ndev,
 728                              enum ib_gid_type gid_type,
 729                              struct ib_ah_attr *ah_attr)
 730 {
 731         int ret;
 732         u16 gid_index;
 733         u8 p;
 734
 735         if (rdma_protocol_roce(device, port_num)) {
 736                 ret = ib_find_cached_gid_by_port(device, &rec->port_gid,
 737                                                  gid_type, port_num,
 738                                                  ndev,
 739                                                  &gid_index);
 740         } else if (rdma_protocol_ib(device, port_num)) {
 741                 ret = ib_find_cached_gid(device, &rec->port_gid,
 742                                          IB_GID_TYPE_IB, NULL, &p,
 743                                          &gid_index);
 744         } else {
 745                 ret = -EINVAL;
 746         }
 747
 748         if (ret)
 749                 return ret;
 750
 751         memset(ah_attr, 0, sizeof *ah_attr);
 752         ah_attr->dlid = be16_to_cpu(rec->mlid);
 753         ah_attr->sl = rec->sl;
 754         ah_attr->port_num = port_num;
 755         ah_attr->static_rate = rec->rate;
 756
 757         ah_attr->ah_flags = IB_AH_GRH;
 758         ah_attr->grh.dgid = rec->mgid;
 759
 760         ah_attr->grh.sgid_index = (u8) gid_index;
 761         ah_attr->grh.flow_label = be32_to_cpu(rec->flow_label);
 762         ah_attr->grh.hop_limit = rec->hop_limit;
 763         ah_attr->grh.traffic_class = rec->traffic_class;
 764
 765         return 0;
 766 }
 767 EXPORT_SYMBOL(ib_init_ah_from_mcmember);
 768
 769 static void mcast_groups_event(struct mcast_port *port,
 770                                enum mcast_group_state state)
 771 {
 772         struct mcast_group *group;
 773         struct rb_node *node;
 774         unsigned long flags;
 775
 776         spin_lock_irqsave(&port->lock, flags);
 777         for (node = rb_first(&port->table); node; node = rb_next(node)) {
 778                 group = rb_entry(node, struct mcast_group, node);
 779                 spin_lock(&group->lock);
 780                 if (group->state == MCAST_IDLE) {
 781                         atomic_inc(&group->refcount);
 782                         queue_work(mcast_wq, &group->work);
 783                 }
 784                 if (group->state != MCAST_GROUP_ERROR)
 785                         group->state = state;
 786                 spin_unlock(&group->lock);
 787         }
 788         spin_unlock_irqrestore(&port->lock, flags);
 789 }
 790
 791 static void mcast_event_handler(struct ib_event_handler *handler,
 792                                 struct ib_event *event)
 793 {
 794         struct mcast_device *dev;
 795         int index;
 796
 797         dev = container_of(handler, struct mcast_device, event_handler);
 798         if (!rdma_cap_ib_mcast(dev->device, event->element.port_num))
 799                 return;
 800
 801         index = event->element.port_num - dev->start_port;
 802
 803         switch (event->event) {
 804         case IB_EVENT_PORT_ERR:
 805         case IB_EVENT_LID_CHANGE:
 806         case IB_EVENT_SM_CHANGE:
 807         case IB_EVENT_CLIENT_REREGISTER:
 808                 mcast_groups_event(&dev->port[index], MCAST_GROUP_ERROR);
 809                 break;
 810         case IB_EVENT_PKEY_CHANGE:
 811                 mcast_groups_event(&dev->port[index], MCAST_PKEY_EVENT);
 812                 break;
 813         default:
 814                 break;
 815         }
 816 }
 817
 818 static void mcast_add_one(struct ib_device *device)
 819 {
 820         struct mcast_device *dev;
 821         struct mcast_port *port;
 822         int i;
 823         int count = 0;
 824
 825         dev = kmalloc(sizeof *dev + device->phys_port_cnt * sizeof *port,
 826                       GFP_KERNEL);
 827         if (!dev)
 828                 return;
 829
 830         dev->start_port = rdma_start_port(device);
 831         dev->end_port = rdma_end_port(device);
 832
 833         for (i = 0; i <= dev->end_port - dev->start_port; i++) {
 834                 if (!rdma_cap_ib_mcast(device, dev->start_port + i))
 835                         continue;
 836                 port = &dev->port[i];
 837                 port->dev = dev;
 838                 port->port_num = dev->start_port + i;
 839                 spin_lock_init(&port->lock);
 840                 port->table = RB_ROOT;
 841                 init_completion(&port->comp);
 842                 atomic_set(&port->refcount, 1);
 843                 ++count;
 844         }
 845
 846         if (!count) {
 847                 kfree(dev);
 848                 return;
 849         }
 850
 851         dev->device = device;
 852         ib_set_client_data(device, &mcast_client, dev);
 853
 854         INIT_IB_EVENT_HANDLER(&dev->event_handler, device, mcast_event_handler);
 855         ib_register_event_handler(&dev->event_handler);
 856 }
 857
 858 static void mcast_remove_one(struct ib_device *device, void *client_data)
 859 {
 860         struct mcast_device *dev = client_data;
 861         struct mcast_port *port;
 862         int i;
 863
 864         if (!dev)
 865                 return;
 866
 867         ib_unregister_event_handler(&dev->event_handler);
 868         flush_workqueue(mcast_wq);
 869
 870         for (i = 0; i <= dev->end_port - dev->start_port; i++) {
 871                 if (rdma_cap_ib_mcast(device, dev->start_port + i)) {
 872                         port = &dev->port[i];
 873                         deref_port(port);
 874                         wait_for_completion(&port->comp);
 875                 }
 876         }
 877
 878         kfree(dev);
 879 }
 880
 881 int mcast_init(void)
 882 {
 883         int ret;
 884
 885         mcast_wq = alloc_ordered_workqueue("ib_mcast", WQ_MEM_RECLAIM);
 886         if (!mcast_wq)
 887                 return -ENOMEM;
 888
 889         ib_sa_register_client(&sa_client);
 890
 891         ret = ib_register_client(&mcast_client);
 892         if (ret)
 893                 goto err;
 894         return 0;
 895
 896 err:
 897         ib_sa_unregister_client(&sa_client);
 898         destroy_workqueue(mcast_wq);
 899         return ret;
 900 }
 901
 902 void mcast_cleanup(void)
 903 {
 904         ib_unregister_client(&mcast_client);
 905         ib_sa_unregister_client(&sa_client);
 906         destroy_workqueue(mcast_wq);
 907 }