SAFE public projects git trees. - safe/jmp/linux-2.6/blob - fs/ocfs2/dlm/dlmdomain.c

   1 /* -*- mode: c; c-basic-offset: 8; -*-
   2  * vim: noexpandtab sw=8 ts=8 sts=0:
   3  *
   4  * dlmdomain.c
   5  *
   6  * defines domain join / leave apis
   7  *
   8  * Copyright (C) 2004 Oracle.  All rights reserved.
   9  *
  10  * This program is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU General Public
  12  * License as published by the Free Software Foundation; either
  13  * version 2 of the License, or (at your option) any later version.
  14  *
  15  * This program is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  18  * General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU General Public
  21  * License along with this program; if not, write to the
  22  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  23  * Boston, MA 021110-1307, USA.
  24  *
  25  */
  26
  27 #include <linux/module.h>
  28 #include <linux/types.h>
  29 #include <linux/slab.h>
  30 #include <linux/highmem.h>
  31 #include <linux/utsname.h>
  32 #include <linux/init.h>
  33 #include <linux/spinlock.h>
  34 #include <linux/delay.h>
  35 #include <linux/err.h>
  36
  37 #include "cluster/heartbeat.h"
  38 #include "cluster/nodemanager.h"
  39 #include "cluster/tcp.h"
  40
  41 #include "dlmapi.h"
  42 #include "dlmcommon.h"
  43
  44 #include "dlmdomain.h"
  45
  46 #include "dlmver.h"
  47
  48 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_DOMAIN)
  49 #include "cluster/masklog.h"
  50
  51 static void dlm_free_pagevec(void **vec, int pages)
  52 {
  53         while (pages--)
  54                 free_page((unsigned long)vec[pages]);
  55         kfree(vec);
  56 }
  57
  58 static void **dlm_alloc_pagevec(int pages)
  59 {
  60         void **vec = kmalloc(pages * sizeof(void *), GFP_KERNEL);
  61         int i;
  62
  63         if (!vec)
  64                 return NULL;
  65
  66         for (i = 0; i < pages; i++)
  67                 if (!(vec[i] = (void *)__get_free_page(GFP_KERNEL)))
  68                         goto out_free;
  69
  70         mlog(0, "Allocated DLM hash pagevec; %d pages (%lu expected), %lu buckets per page\n",
  71              pages, DLM_HASH_PAGES, (unsigned long)DLM_BUCKETS_PER_PAGE);
  72         return vec;
  73 out_free:
  74         dlm_free_pagevec(vec, i);
  75         return NULL;
  76 }
  77
  78 /*
  79  *
  80  * spinlock lock ordering: if multiple locks are needed, obey this ordering:
  81  *    dlm_domain_lock
  82  *    struct dlm_ctxt->spinlock
  83  *    struct dlm_lock_resource->spinlock
  84  *    struct dlm_ctxt->master_lock
  85  *    struct dlm_ctxt->ast_lock
  86  *    dlm_master_list_entry->spinlock
  87  *    dlm_lock->spinlock
  88  *
  89  */
  90
  91 DEFINE_SPINLOCK(dlm_domain_lock);
  92 LIST_HEAD(dlm_domains);
  93 static DECLARE_WAIT_QUEUE_HEAD(dlm_domain_events);
  94
  95 #define DLM_DOMAIN_BACKOFF_MS 200
  96
  97 static int dlm_query_join_handler(struct o2net_msg *msg, u32 len, void *data);
  98 static int dlm_assert_joined_handler(struct o2net_msg *msg, u32 len, void *data);
  99 static int dlm_cancel_join_handler(struct o2net_msg *msg, u32 len, void *data);
 100 static int dlm_exit_domain_handler(struct o2net_msg *msg, u32 len, void *data);
 101
 102 static void dlm_unregister_domain_handlers(struct dlm_ctxt *dlm);
 103
 104 void __dlm_unhash_lockres(struct dlm_lock_resource *lockres)
 105 {
 106         hlist_del_init(&lockres->hash_node);
 107         dlm_lockres_put(lockres);
 108 }
 109
 110 void __dlm_insert_lockres(struct dlm_ctxt *dlm,
 111                        struct dlm_lock_resource *res)
 112 {
 113         struct hlist_head *bucket;
 114         struct qstr *q;
 115
 116         assert_spin_locked(&dlm->spinlock);
 117
 118         q = &res->lockname;
 119         bucket = dlm_lockres_hash(dlm, q->hash);
 120
 121         /* get a reference for our hashtable */
 122         dlm_lockres_get(res);
 123
 124         hlist_add_head(&res->hash_node, bucket);
 125 }
 126
 127 struct dlm_lock_resource * __dlm_lookup_lockres(struct dlm_ctxt *dlm,
 128                                                 const char *name,
 129                                                 unsigned int len,
 130                                                 unsigned int hash)
 131 {
 132         struct hlist_head *bucket;
 133         struct hlist_node *list;
 134
 135         mlog_entry("%.*s\n", len, name);
 136
 137         assert_spin_locked(&dlm->spinlock);
 138
 139         bucket = dlm_lockres_hash(dlm, hash);
 140
 141         hlist_for_each(list, bucket) {
 142                 struct dlm_lock_resource *res = hlist_entry(list,
 143                         struct dlm_lock_resource, hash_node);
 144                 if (res->lockname.name[0] != name[0])
 145                         continue;
 146                 if (unlikely(res->lockname.len != len))
 147                         continue;
 148                 if (memcmp(res->lockname.name + 1, name + 1, len - 1))
 149                         continue;
 150                 dlm_lockres_get(res);
 151                 return res;
 152         }
 153         return NULL;
 154 }
 155
 156 struct dlm_lock_resource * dlm_lookup_lockres(struct dlm_ctxt *dlm,
 157                                     const char *name,
 158                                     unsigned int len)
 159 {
 160         struct dlm_lock_resource *res;
 161         unsigned int hash = dlm_lockid_hash(name, len);
 162
 163         spin_lock(&dlm->spinlock);
 164         res = __dlm_lookup_lockres(dlm, name, len, hash);
 165         spin_unlock(&dlm->spinlock);
 166         return res;
 167 }
 168
 169 static struct dlm_ctxt * __dlm_lookup_domain_full(const char *domain, int len)
 170 {
 171         struct dlm_ctxt *tmp = NULL;
 172         struct list_head *iter;
 173
 174         assert_spin_locked(&dlm_domain_lock);
 175
 176         /* tmp->name here is always NULL terminated,
 177          * but domain may not be! */
 178         list_for_each(iter, &dlm_domains) {
 179                 tmp = list_entry (iter, struct dlm_ctxt, list);
 180                 if (strlen(tmp->name) == len &&
 181                     memcmp(tmp->name, domain, len)==0)
 182                         break;
 183                 tmp = NULL;
 184         }
 185
 186         return tmp;
 187 }
 188
 189 /* For null terminated domain strings ONLY */
 190 static struct dlm_ctxt * __dlm_lookup_domain(const char *domain)
 191 {
 192         assert_spin_locked(&dlm_domain_lock);
 193
 194         return __dlm_lookup_domain_full(domain, strlen(domain));
 195 }
 196
 197
 198 /* returns true on one of two conditions:
 199  * 1) the domain does not exist
 200  * 2) the domain exists and it's state is "joined" */
 201 static int dlm_wait_on_domain_helper(const char *domain)
 202 {
 203         int ret = 0;
 204         struct dlm_ctxt *tmp = NULL;
 205
 206         spin_lock(&dlm_domain_lock);
 207
 208         tmp = __dlm_lookup_domain(domain);
 209         if (!tmp)
 210                 ret = 1;
 211         else if (tmp->dlm_state == DLM_CTXT_JOINED)
 212                 ret = 1;
 213
 214         spin_unlock(&dlm_domain_lock);
 215         return ret;
 216 }
 217
 218 static void dlm_free_ctxt_mem(struct dlm_ctxt *dlm)
 219 {
 220         if (dlm->lockres_hash)
 221                 dlm_free_pagevec((void **)dlm->lockres_hash, DLM_HASH_PAGES);
 222
 223         if (dlm->name)
 224                 kfree(dlm->name);
 225
 226         kfree(dlm);
 227 }
 228
 229 /* A little strange - this function will be called while holding
 230  * dlm_domain_lock and is expected to be holding it on the way out. We
 231  * will however drop and reacquire it multiple times */
 232 static void dlm_ctxt_release(struct kref *kref)
 233 {
 234         struct dlm_ctxt *dlm;
 235
 236         dlm = container_of(kref, struct dlm_ctxt, dlm_refs);
 237
 238         BUG_ON(dlm->num_joins);
 239         BUG_ON(dlm->dlm_state == DLM_CTXT_JOINED);
 240
 241         /* we may still be in the list if we hit an error during join. */
 242         list_del_init(&dlm->list);
 243
 244         spin_unlock(&dlm_domain_lock);
 245
 246         mlog(0, "freeing memory from domain %s\n", dlm->name);
 247
 248         wake_up(&dlm_domain_events);
 249
 250         dlm_free_ctxt_mem(dlm);
 251
 252         spin_lock(&dlm_domain_lock);
 253 }
 254
 255 void dlm_put(struct dlm_ctxt *dlm)
 256 {
 257         spin_lock(&dlm_domain_lock);
 258         kref_put(&dlm->dlm_refs, dlm_ctxt_release);
 259         spin_unlock(&dlm_domain_lock);
 260 }
 261
 262 static void __dlm_get(struct dlm_ctxt *dlm)
 263 {
 264         kref_get(&dlm->dlm_refs);
 265 }
 266
 267 /* given a questionable reference to a dlm object, gets a reference if
 268  * it can find it in the list, otherwise returns NULL in which case
 269  * you shouldn't trust your pointer. */
 270 struct dlm_ctxt *dlm_grab(struct dlm_ctxt *dlm)
 271 {
 272         struct list_head *iter;
 273         struct dlm_ctxt *target = NULL;
 274
 275         spin_lock(&dlm_domain_lock);
 276
 277         list_for_each(iter, &dlm_domains) {
 278                 target = list_entry (iter, struct dlm_ctxt, list);
 279
 280                 if (target == dlm) {
 281                         __dlm_get(target);
 282                         break;
 283                 }
 284
 285                 target = NULL;
 286         }
 287
 288         spin_unlock(&dlm_domain_lock);
 289
 290         return target;
 291 }
 292
 293 int dlm_domain_fully_joined(struct dlm_ctxt *dlm)
 294 {
 295         int ret;
 296
 297         spin_lock(&dlm_domain_lock);
 298         ret = (dlm->dlm_state == DLM_CTXT_JOINED) ||
 299                 (dlm->dlm_state == DLM_CTXT_IN_SHUTDOWN);
 300         spin_unlock(&dlm_domain_lock);
 301
 302         return ret;
 303 }
 304
 305 static void dlm_destroy_dlm_worker(struct dlm_ctxt *dlm)
 306 {
 307         if (dlm->dlm_worker) {
 308                 flush_workqueue(dlm->dlm_worker);
 309                 destroy_workqueue(dlm->dlm_worker);
 310                 dlm->dlm_worker = NULL;
 311         }
 312 }
 313
 314 static void dlm_complete_dlm_shutdown(struct dlm_ctxt *dlm)
 315 {
 316         dlm_unregister_domain_handlers(dlm);
 317         dlm_complete_thread(dlm);
 318         dlm_complete_recovery_thread(dlm);
 319         dlm_destroy_dlm_worker(dlm);
 320
 321         /* We've left the domain. Now we can take ourselves out of the
 322          * list and allow the kref stuff to help us free the
 323          * memory. */
 324         spin_lock(&dlm_domain_lock);
 325         list_del_init(&dlm->list);
 326         spin_unlock(&dlm_domain_lock);
 327
 328         /* Wake up anyone waiting for us to remove this domain */
 329         wake_up(&dlm_domain_events);
 330 }
 331
 332 static void dlm_migrate_all_locks(struct dlm_ctxt *dlm)
 333 {
 334         int i;
 335         struct dlm_lock_resource *res;
 336
 337         mlog(0, "Migrating locks from domain %s\n", dlm->name);
 338 restart:
 339         spin_lock(&dlm->spinlock);
 340         for (i = 0; i < DLM_HASH_BUCKETS; i++) {
 341                 while (!hlist_empty(dlm_lockres_hash(dlm, i))) {
 342                         res = hlist_entry(dlm_lockres_hash(dlm, i)->first,
 343                                           struct dlm_lock_resource, hash_node);
 344                         /* need reference when manually grabbing lockres */
 345                         dlm_lockres_get(res);
 346                         /* this should unhash the lockres
 347                          * and exit with dlm->spinlock */
 348                         mlog(0, "purging res=%p\n", res);
 349                         if (dlm_lockres_is_dirty(dlm, res)) {
 350                                 /* HACK!  this should absolutely go.
 351                                  * need to figure out why some empty
 352                                  * lockreses are still marked dirty */
 353                                 mlog(ML_ERROR, "lockres %.*s dirty!\n",
 354                                      res->lockname.len, res->lockname.name);
 355
 356                                 spin_unlock(&dlm->spinlock);
 357                                 dlm_kick_thread(dlm, res);
 358                                 wait_event(dlm->ast_wq, !dlm_lockres_is_dirty(dlm, res));
 359                                 dlm_lockres_put(res);
 360                                 goto restart;
 361                         }
 362                         dlm_purge_lockres(dlm, res);
 363                         dlm_lockres_put(res);
 364                 }
 365         }
 366         spin_unlock(&dlm->spinlock);
 367
 368         mlog(0, "DONE Migrating locks from domain %s\n", dlm->name);
 369 }
 370
 371 static int dlm_no_joining_node(struct dlm_ctxt *dlm)
 372 {
 373         int ret;
 374
 375         spin_lock(&dlm->spinlock);
 376         ret = dlm->joining_node == DLM_LOCK_RES_OWNER_UNKNOWN;
 377         spin_unlock(&dlm->spinlock);
 378
 379         return ret;
 380 }
 381
 382 static void dlm_mark_domain_leaving(struct dlm_ctxt *dlm)
 383 {
 384         /* Yikes, a double spinlock! I need domain_lock for the dlm
 385          * state and the dlm spinlock for join state... Sorry! */
 386 again:
 387         spin_lock(&dlm_domain_lock);
 388         spin_lock(&dlm->spinlock);
 389
 390         if (dlm->joining_node != DLM_LOCK_RES_OWNER_UNKNOWN) {
 391                 mlog(0, "Node %d is joining, we wait on it.\n",
 392                           dlm->joining_node);
 393                 spin_unlock(&dlm->spinlock);
 394                 spin_unlock(&dlm_domain_lock);
 395
 396                 wait_event(dlm->dlm_join_events, dlm_no_joining_node(dlm));
 397                 goto again;
 398         }
 399
 400         dlm->dlm_state = DLM_CTXT_LEAVING;
 401         spin_unlock(&dlm->spinlock);
 402         spin_unlock(&dlm_domain_lock);
 403 }
 404
 405 static void __dlm_print_nodes(struct dlm_ctxt *dlm)
 406 {
 407         int node = -1;
 408
 409         assert_spin_locked(&dlm->spinlock);
 410
 411         printk(KERN_INFO "ocfs2_dlm: Nodes in domain (\"%s\"): ", dlm->name);
 412
 413         while ((node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES,
 414                                      node + 1)) < O2NM_MAX_NODES) {
 415                 printk("%d ", node);
 416         }
 417         printk("\n");
 418 }
 419
 420 static int dlm_exit_domain_handler(struct o2net_msg *msg, u32 len, void *data)
 421 {
 422         struct dlm_ctxt *dlm = data;
 423         unsigned int node;
 424         struct dlm_exit_domain *exit_msg = (struct dlm_exit_domain *) msg->buf;
 425
 426         mlog_entry("%p %u %p", msg, len, data);
 427
 428         if (!dlm_grab(dlm))
 429                 return 0;
 430
 431         node = exit_msg->node_idx;
 432
 433         printk(KERN_INFO "ocfs2_dlm: Node %u leaves domain %s\n", node, dlm->name);
 434
 435         spin_lock(&dlm->spinlock);
 436         clear_bit(node, dlm->domain_map);
 437         __dlm_print_nodes(dlm);
 438
 439         /* notify anything attached to the heartbeat events */
 440         dlm_hb_event_notify_attached(dlm, node, 0);
 441
 442         spin_unlock(&dlm->spinlock);
 443
 444         dlm_put(dlm);
 445
 446         return 0;
 447 }
 448
 449 static int dlm_send_one_domain_exit(struct dlm_ctxt *dlm,
 450                                     unsigned int node)
 451 {
 452         int status;
 453         struct dlm_exit_domain leave_msg;
 454
 455         mlog(0, "Asking node %u if we can leave the domain %s me = %u\n",
 456                   node, dlm->name, dlm->node_num);
 457
 458         memset(&leave_msg, 0, sizeof(leave_msg));
 459         leave_msg.node_idx = dlm->node_num;
 460
 461         status = o2net_send_message(DLM_EXIT_DOMAIN_MSG, dlm->key,
 462                                     &leave_msg, sizeof(leave_msg), node,
 463                                     NULL);
 464
 465         mlog(0, "status return %d from o2net_send_message\n", status);
 466
 467         return status;
 468 }
 469
 470
 471 static void dlm_leave_domain(struct dlm_ctxt *dlm)
 472 {
 473         int node, clear_node, status;
 474
 475         /* At this point we've migrated away all our locks and won't
 476          * accept mastership of new ones. The dlm is responsible for
 477          * almost nothing now. We make sure not to confuse any joining
 478          * nodes and then commence shutdown procedure. */
 479
 480         spin_lock(&dlm->spinlock);
 481         /* Clear ourselves from the domain map */
 482         clear_bit(dlm->node_num, dlm->domain_map);
 483         while ((node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES,
 484                                      0)) < O2NM_MAX_NODES) {
 485                 /* Drop the dlm spinlock. This is safe wrt the domain_map.
 486                  * -nodes cannot be added now as the
 487                  *   query_join_handlers knows to respond with OK_NO_MAP
 488                  * -we catch the right network errors if a node is
 489                  *   removed from the map while we're sending him the
 490                  *   exit message. */
 491                 spin_unlock(&dlm->spinlock);
 492
 493                 clear_node = 1;
 494
 495                 status = dlm_send_one_domain_exit(dlm, node);
 496                 if (status < 0 &&
 497                     status != -ENOPROTOOPT &&
 498                     status != -ENOTCONN) {
 499                         mlog(ML_NOTICE, "Error %d sending domain exit message "
 500                              "to node %d\n", status, node);
 501
 502                         /* Not sure what to do here but lets sleep for
 503                          * a bit in case this was a transient
 504                          * error... */
 505                         msleep(DLM_DOMAIN_BACKOFF_MS);
 506                         clear_node = 0;
 507                 }
 508
 509                 spin_lock(&dlm->spinlock);
 510                 /* If we're not clearing the node bit then we intend
 511                  * to loop back around to try again. */
 512                 if (clear_node)
 513                         clear_bit(node, dlm->domain_map);
 514         }
 515         spin_unlock(&dlm->spinlock);
 516 }
 517
 518 int dlm_joined(struct dlm_ctxt *dlm)
 519 {
 520         int ret = 0;
 521
 522         spin_lock(&dlm_domain_lock);
 523
 524         if (dlm->dlm_state == DLM_CTXT_JOINED)
 525                 ret = 1;
 526
 527         spin_unlock(&dlm_domain_lock);
 528
 529         return ret;
 530 }
 531
 532 int dlm_shutting_down(struct dlm_ctxt *dlm)
 533 {
 534         int ret = 0;
 535
 536         spin_lock(&dlm_domain_lock);
 537
 538         if (dlm->dlm_state == DLM_CTXT_IN_SHUTDOWN)
 539                 ret = 1;
 540
 541         spin_unlock(&dlm_domain_lock);
 542
 543         return ret;
 544 }
 545
 546 void dlm_unregister_domain(struct dlm_ctxt *dlm)
 547 {
 548         int leave = 0;
 549
 550         spin_lock(&dlm_domain_lock);
 551         BUG_ON(dlm->dlm_state != DLM_CTXT_JOINED);
 552         BUG_ON(!dlm->num_joins);
 553
 554         dlm->num_joins--;
 555         if (!dlm->num_joins) {
 556                 /* We mark it "in shutdown" now so new register
 557                  * requests wait until we've completely left the
 558                  * domain. Don't use DLM_CTXT_LEAVING yet as we still
 559                  * want new domain joins to communicate with us at
 560                  * least until we've completed migration of our
 561                  * resources. */
 562                 dlm->dlm_state = DLM_CTXT_IN_SHUTDOWN;
 563                 leave = 1;
 564         }
 565         spin_unlock(&dlm_domain_lock);
 566
 567         if (leave) {
 568                 mlog(0, "shutting down domain %s\n", dlm->name);
 569
 570                 /* We changed dlm state, notify the thread */
 571                 dlm_kick_thread(dlm, NULL);
 572
 573                 dlm_migrate_all_locks(dlm);
 574                 dlm_mark_domain_leaving(dlm);
 575                 dlm_leave_domain(dlm);
 576                 dlm_complete_dlm_shutdown(dlm);
 577         }
 578         dlm_put(dlm);
 579 }
 580 EXPORT_SYMBOL_GPL(dlm_unregister_domain);
 581
 582 static int dlm_query_join_handler(struct o2net_msg *msg, u32 len, void *data)
 583 {
 584         struct dlm_query_join_request *query;
 585         enum dlm_query_join_response response;
 586         struct dlm_ctxt *dlm = NULL;
 587
 588         query = (struct dlm_query_join_request *) msg->buf;
 589
 590         mlog(0, "node %u wants to join domain %s\n", query->node_idx,
 591                   query->domain);
 592
 593         /*
 594          * If heartbeat doesn't consider the node live, tell it
 595          * to back off and try again.  This gives heartbeat a chance
 596          * to catch up.
 597          */
 598         if (!o2hb_check_node_heartbeating(query->node_idx)) {
 599                 mlog(0, "node %u is not in our live map yet\n",
 600                      query->node_idx);
 601
 602                 response = JOIN_DISALLOW;
 603                 goto respond;
 604         }
 605
 606         response = JOIN_OK_NO_MAP;
 607
 608         spin_lock(&dlm_domain_lock);
 609         dlm = __dlm_lookup_domain_full(query->domain, query->name_len);
 610         /* Once the dlm ctxt is marked as leaving then we don't want
 611          * to be put in someone's domain map.
 612          * Also, explicitly disallow joining at certain troublesome
 613          * times (ie. during recovery). */
 614         if (dlm && dlm->dlm_state != DLM_CTXT_LEAVING) {
 615                 int bit = query->node_idx;
 616                 spin_lock(&dlm->spinlock);
 617
 618                 if (dlm->dlm_state == DLM_CTXT_NEW &&
 619                     dlm->joining_node == DLM_LOCK_RES_OWNER_UNKNOWN) {
 620                         /*If this is a brand new context and we
 621                          * haven't started our join process yet, then
 622                          * the other node won the race. */
 623                         response = JOIN_OK_NO_MAP;
 624                 } else if (dlm->joining_node != DLM_LOCK_RES_OWNER_UNKNOWN) {
 625                         /* Disallow parallel joins. */
 626                         response = JOIN_DISALLOW;
 627                 } else if (dlm->reco.state & DLM_RECO_STATE_ACTIVE) {
 628                         mlog(ML_NOTICE, "node %u trying to join, but recovery "
 629                              "is ongoing.\n", bit);
 630                         response = JOIN_DISALLOW;
 631                 } else if (test_bit(bit, dlm->recovery_map)) {
 632                         mlog(ML_NOTICE, "node %u trying to join, but it "
 633                              "still needs recovery.\n", bit);
 634                         response = JOIN_DISALLOW;
 635                 } else if (test_bit(bit, dlm->domain_map)) {
 636                         mlog(ML_NOTICE, "node %u trying to join, but it "
 637                              "is still in the domain! needs recovery?\n",
 638                              bit);
 639                         response = JOIN_DISALLOW;
 640                 } else {
 641                         /* Alright we're fully a part of this domain
 642                          * so we keep some state as to who's joining
 643                          * and indicate to him that needs to be fixed
 644                          * up. */
 645                         response = JOIN_OK;
 646                         __dlm_set_joining_node(dlm, query->node_idx);
 647                 }
 648
 649                 spin_unlock(&dlm->spinlock);
 650         }
 651         spin_unlock(&dlm_domain_lock);
 652
 653 respond:
 654         mlog(0, "We respond with %u\n", response);
 655
 656         return response;
 657 }
 658
 659 static int dlm_assert_joined_handler(struct o2net_msg *msg, u32 len, void *data)
 660 {
 661         struct dlm_assert_joined *assert;
 662         struct dlm_ctxt *dlm = NULL;
 663
 664         assert = (struct dlm_assert_joined *) msg->buf;
 665
 666         mlog(0, "node %u asserts join on domain %s\n", assert->node_idx,
 667                   assert->domain);
 668
 669         spin_lock(&dlm_domain_lock);
 670         dlm = __dlm_lookup_domain_full(assert->domain, assert->name_len);
 671         /* XXX should we consider no dlm ctxt an error? */
 672         if (dlm) {
 673                 spin_lock(&dlm->spinlock);
 674
 675                 /* Alright, this node has officially joined our
 676                  * domain. Set him in the map and clean up our
 677                  * leftover join state. */
 678                 BUG_ON(dlm->joining_node != assert->node_idx);
 679                 set_bit(assert->node_idx, dlm->domain_map);
 680                 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
 681
 682                 printk(KERN_INFO "ocfs2_dlm: Node %u joins domain %s\n",
 683                        assert->node_idx, dlm->name);
 684                 __dlm_print_nodes(dlm);
 685
 686                 /* notify anything attached to the heartbeat events */
 687                 dlm_hb_event_notify_attached(dlm, assert->node_idx, 1);
 688
 689                 spin_unlock(&dlm->spinlock);
 690         }
 691         spin_unlock(&dlm_domain_lock);
 692
 693         return 0;
 694 }
 695
 696 static int dlm_cancel_join_handler(struct o2net_msg *msg, u32 len, void *data)
 697 {
 698         struct dlm_cancel_join *cancel;
 699         struct dlm_ctxt *dlm = NULL;
 700
 701         cancel = (struct dlm_cancel_join *) msg->buf;
 702
 703         mlog(0, "node %u cancels join on domain %s\n", cancel->node_idx,
 704                   cancel->domain);
 705
 706         spin_lock(&dlm_domain_lock);
 707         dlm = __dlm_lookup_domain_full(cancel->domain, cancel->name_len);
 708
 709         if (dlm) {
 710                 spin_lock(&dlm->spinlock);
 711
 712                 /* Yikes, this guy wants to cancel his join. No
 713                  * problem, we simply cleanup our join state. */
 714                 BUG_ON(dlm->joining_node != cancel->node_idx);
 715                 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
 716
 717                 spin_unlock(&dlm->spinlock);
 718         }
 719         spin_unlock(&dlm_domain_lock);
 720
 721         return 0;
 722 }
 723
 724 static int dlm_send_one_join_cancel(struct dlm_ctxt *dlm,
 725                                     unsigned int node)
 726 {
 727         int status;
 728         struct dlm_cancel_join cancel_msg;
 729
 730         memset(&cancel_msg, 0, sizeof(cancel_msg));
 731         cancel_msg.node_idx = dlm->node_num;
 732         cancel_msg.name_len = strlen(dlm->name);
 733         memcpy(cancel_msg.domain, dlm->name, cancel_msg.name_len);
 734
 735         status = o2net_send_message(DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY,
 736                                     &cancel_msg, sizeof(cancel_msg), node,
 737                                     NULL);
 738         if (status < 0) {
 739                 mlog_errno(status);
 740                 goto bail;
 741         }
 742
 743 bail:
 744         return status;
 745 }
 746
 747 /* map_size should be in bytes. */
 748 static int dlm_send_join_cancels(struct dlm_ctxt *dlm,
 749                                  unsigned long *node_map,
 750                                  unsigned int map_size)
 751 {
 752         int status, tmpstat;
 753         unsigned int node;
 754
 755         if (map_size != (BITS_TO_LONGS(O2NM_MAX_NODES) *
 756                          sizeof(unsigned long))) {
 757                 mlog(ML_ERROR,
 758                      "map_size %u != BITS_TO_LONGS(O2NM_MAX_NODES) %u\n",
 759                      map_size, BITS_TO_LONGS(O2NM_MAX_NODES));
 760                 return -EINVAL;
 761         }
 762
 763         status = 0;
 764         node = -1;
 765         while ((node = find_next_bit(node_map, O2NM_MAX_NODES,
 766                                      node + 1)) < O2NM_MAX_NODES) {
 767                 if (node == dlm->node_num)
 768                         continue;
 769
 770                 tmpstat = dlm_send_one_join_cancel(dlm, node);
 771                 if (tmpstat) {
 772                         mlog(ML_ERROR, "Error return %d cancelling join on "
 773                              "node %d\n", tmpstat, node);
 774                         if (!status)
 775                                 status = tmpstat;
 776                 }
 777         }
 778
 779         if (status)
 780                 mlog_errno(status);
 781         return status;
 782 }
 783
 784 static int dlm_request_join(struct dlm_ctxt *dlm,
 785                             int node,
 786                             enum dlm_query_join_response *response)
 787 {
 788         int status, retval;
 789         struct dlm_query_join_request join_msg;
 790
 791         mlog(0, "querying node %d\n", node);
 792
 793         memset(&join_msg, 0, sizeof(join_msg));
 794         join_msg.node_idx = dlm->node_num;
 795         join_msg.name_len = strlen(dlm->name);
 796         memcpy(join_msg.domain, dlm->name, join_msg.name_len);
 797
 798         status = o2net_send_message(DLM_QUERY_JOIN_MSG, DLM_MOD_KEY, &join_msg,
 799                                     sizeof(join_msg), node, &retval);
 800         if (status < 0 && status != -ENOPROTOOPT) {
 801                 mlog_errno(status);
 802                 goto bail;
 803         }
 804
 805         /* -ENOPROTOOPT from the net code means the other side isn't
 806             listening for our message type -- that's fine, it means
 807             his dlm isn't up, so we can consider him a 'yes' but not
 808             joined into the domain.  */
 809         if (status == -ENOPROTOOPT) {
 810                 status = 0;
 811                 *response = JOIN_OK_NO_MAP;
 812         } else if (retval == JOIN_DISALLOW ||
 813                    retval == JOIN_OK ||
 814                    retval == JOIN_OK_NO_MAP) {
 815                 *response = retval;
 816         } else {
 817                 status = -EINVAL;
 818                 mlog(ML_ERROR, "invalid response %d from node %u\n", retval,
 819                      node);
 820         }
 821
 822         mlog(0, "status %d, node %d response is %d\n", status, node,
 823                   *response);
 824
 825 bail:
 826         return status;
 827 }
 828
 829 static int dlm_send_one_join_assert(struct dlm_ctxt *dlm,
 830                                     unsigned int node)
 831 {
 832         int status;
 833         struct dlm_assert_joined assert_msg;
 834
 835         mlog(0, "Sending join assert to node %u\n", node);
 836
 837         memset(&assert_msg, 0, sizeof(assert_msg));
 838         assert_msg.node_idx = dlm->node_num;
 839         assert_msg.name_len = strlen(dlm->name);
 840         memcpy(assert_msg.domain, dlm->name, assert_msg.name_len);
 841
 842         status = o2net_send_message(DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY,
 843                                     &assert_msg, sizeof(assert_msg), node,
 844                                     NULL);
 845         if (status < 0)
 846                 mlog_errno(status);
 847
 848         return status;
 849 }
 850
 851 static void dlm_send_join_asserts(struct dlm_ctxt *dlm,
 852                                   unsigned long *node_map)
 853 {
 854         int status, node, live;
 855
 856         status = 0;
 857         node = -1;
 858         while ((node = find_next_bit(node_map, O2NM_MAX_NODES,
 859                                      node + 1)) < O2NM_MAX_NODES) {
 860                 if (node == dlm->node_num)
 861                         continue;
 862
 863                 do {
 864                         /* It is very important that this message be
 865                          * received so we spin until either the node
 866                          * has died or it gets the message. */
 867                         status = dlm_send_one_join_assert(dlm, node);
 868
 869                         spin_lock(&dlm->spinlock);
 870                         live = test_bit(node, dlm->live_nodes_map);
 871                         spin_unlock(&dlm->spinlock);
 872
 873                         if (status) {
 874                                 mlog(ML_ERROR, "Error return %d asserting "
 875                                      "join on node %d\n", status, node);
 876
 877                                 /* give us some time between errors... */
 878                                 if (live)
 879                                         msleep(DLM_DOMAIN_BACKOFF_MS);
 880                         }
 881                 } while (status && live);
 882         }
 883 }
 884
 885 struct domain_join_ctxt {
 886         unsigned long live_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
 887         unsigned long yes_resp_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
 888 };
 889
 890 static int dlm_should_restart_join(struct dlm_ctxt *dlm,
 891                                    struct domain_join_ctxt *ctxt,
 892                                    enum dlm_query_join_response response)
 893 {
 894         int ret;
 895
 896         if (response == JOIN_DISALLOW) {
 897                 mlog(0, "Latest response of disallow -- should restart\n");
 898                 return 1;
 899         }
 900
 901         spin_lock(&dlm->spinlock);
 902         /* For now, we restart the process if the node maps have
 903          * changed at all */
 904         ret = memcmp(ctxt->live_map, dlm->live_nodes_map,
 905                      sizeof(dlm->live_nodes_map));
 906         spin_unlock(&dlm->spinlock);
 907
 908         if (ret)
 909                 mlog(0, "Node maps changed -- should restart\n");
 910
 911         return ret;
 912 }
 913
 914 static int dlm_try_to_join_domain(struct dlm_ctxt *dlm)
 915 {
 916         int status = 0, tmpstat, node;
 917         struct domain_join_ctxt *ctxt;
 918         enum dlm_query_join_response response;
 919
 920         mlog_entry("%p", dlm);
 921
 922         ctxt = kcalloc(1, sizeof(*ctxt), GFP_KERNEL);
 923         if (!ctxt) {
 924                 status = -ENOMEM;
 925                 mlog_errno(status);
 926                 goto bail;
 927         }
 928
 929         /* group sem locking should work for us here -- we're already
 930          * registered for heartbeat events so filling this should be
 931          * atomic wrt getting those handlers called. */
 932         o2hb_fill_node_map(dlm->live_nodes_map, sizeof(dlm->live_nodes_map));
 933
 934         spin_lock(&dlm->spinlock);
 935         memcpy(ctxt->live_map, dlm->live_nodes_map, sizeof(ctxt->live_map));
 936
 937         __dlm_set_joining_node(dlm, dlm->node_num);
 938
 939         spin_unlock(&dlm->spinlock);
 940
 941         node = -1;
 942         while ((node = find_next_bit(ctxt->live_map, O2NM_MAX_NODES,
 943                                      node + 1)) < O2NM_MAX_NODES) {
 944                 if (node == dlm->node_num)
 945                         continue;
 946
 947                 status = dlm_request_join(dlm, node, &response);
 948                 if (status < 0) {
 949                         mlog_errno(status);
 950                         goto bail;
 951                 }
 952
 953                 /* Ok, either we got a response or the node doesn't have a
 954                  * dlm up. */
 955                 if (response == JOIN_OK)
 956                         set_bit(node, ctxt->yes_resp_map);
 957
 958                 if (dlm_should_restart_join(dlm, ctxt, response)) {
 959                         status = -EAGAIN;
 960                         goto bail;
 961                 }
 962         }
 963
 964         mlog(0, "Yay, done querying nodes!\n");
 965
 966         /* Yay, everyone agree's we can join the domain. My domain is
 967          * comprised of all nodes who were put in the
 968          * yes_resp_map. Copy that into our domain map and send a join
 969          * assert message to clean up everyone elses state. */
 970         spin_lock(&dlm->spinlock);
 971         memcpy(dlm->domain_map, ctxt->yes_resp_map,
 972                sizeof(ctxt->yes_resp_map));
 973         set_bit(dlm->node_num, dlm->domain_map);
 974         spin_unlock(&dlm->spinlock);
 975
 976         dlm_send_join_asserts(dlm, ctxt->yes_resp_map);
 977
 978         /* Joined state *must* be set before the joining node
 979          * information, otherwise the query_join handler may read no
 980          * current joiner but a state of NEW and tell joining nodes
 981          * we're not in the domain. */
 982         spin_lock(&dlm_domain_lock);
 983         dlm->dlm_state = DLM_CTXT_JOINED;
 984         dlm->num_joins++;
 985         spin_unlock(&dlm_domain_lock);
 986
 987 bail:
 988         spin_lock(&dlm->spinlock);
 989         __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
 990         if (!status)
 991                 __dlm_print_nodes(dlm);
 992         spin_unlock(&dlm->spinlock);
 993
 994         if (ctxt) {
 995                 /* Do we need to send a cancel message to any nodes? */
 996                 if (status < 0) {
 997                         tmpstat = dlm_send_join_cancels(dlm,
 998                                                         ctxt->yes_resp_map,
 999                                                         sizeof(ctxt->yes_resp_map));
1000                         if (tmpstat < 0)
1001                                 mlog_errno(tmpstat);
1002                 }
1003                 kfree(ctxt);
1004         }
1005
1006         mlog(0, "returning %d\n", status);
1007         return status;
1008 }
1009
1010 static void dlm_unregister_domain_handlers(struct dlm_ctxt *dlm)
1011 {
1012         o2hb_unregister_callback(&dlm->dlm_hb_up);
1013         o2hb_unregister_callback(&dlm->dlm_hb_down);
1014         o2net_unregister_handler_list(&dlm->dlm_domain_handlers);
1015 }
1016
1017 static int dlm_register_domain_handlers(struct dlm_ctxt *dlm)
1018 {
1019         int status;
1020
1021         mlog(0, "registering handlers.\n");
1022
1023         o2hb_setup_callback(&dlm->dlm_hb_down, O2HB_NODE_DOWN_CB,
1024                             dlm_hb_node_down_cb, dlm, DLM_HB_NODE_DOWN_PRI);
1025         status = o2hb_register_callback(&dlm->dlm_hb_down);
1026         if (status)
1027                 goto bail;
1028
1029         o2hb_setup_callback(&dlm->dlm_hb_up, O2HB_NODE_UP_CB,
1030                             dlm_hb_node_up_cb, dlm, DLM_HB_NODE_UP_PRI);
1031         status = o2hb_register_callback(&dlm->dlm_hb_up);
1032         if (status)
1033                 goto bail;
1034
1035         status = o2net_register_handler(DLM_MASTER_REQUEST_MSG, dlm->key,
1036                                         sizeof(struct dlm_master_request),
1037                                         dlm_master_request_handler,
1038                                         dlm, &dlm->dlm_domain_handlers);
1039         if (status)
1040                 goto bail;
1041
1042         status = o2net_register_handler(DLM_ASSERT_MASTER_MSG, dlm->key,
1043                                         sizeof(struct dlm_assert_master),
1044                                         dlm_assert_master_handler,
1045                                         dlm, &dlm->dlm_domain_handlers);
1046         if (status)
1047                 goto bail;
1048
1049         status = o2net_register_handler(DLM_CREATE_LOCK_MSG, dlm->key,
1050                                         sizeof(struct dlm_create_lock),
1051                                         dlm_create_lock_handler,
1052                                         dlm, &dlm->dlm_domain_handlers);
1053         if (status)
1054                 goto bail;
1055
1056         status = o2net_register_handler(DLM_CONVERT_LOCK_MSG, dlm->key,
1057                                         DLM_CONVERT_LOCK_MAX_LEN,
1058                                         dlm_convert_lock_handler,
1059                                         dlm, &dlm->dlm_domain_handlers);
1060         if (status)
1061                 goto bail;
1062
1063         status = o2net_register_handler(DLM_UNLOCK_LOCK_MSG, dlm->key,
1064                                         DLM_UNLOCK_LOCK_MAX_LEN,
1065                                         dlm_unlock_lock_handler,
1066                                         dlm, &dlm->dlm_domain_handlers);
1067         if (status)
1068                 goto bail;
1069
1070         status = o2net_register_handler(DLM_PROXY_AST_MSG, dlm->key,
1071                                         DLM_PROXY_AST_MAX_LEN,
1072                                         dlm_proxy_ast_handler,
1073                                         dlm, &dlm->dlm_domain_handlers);
1074         if (status)
1075                 goto bail;
1076
1077         status = o2net_register_handler(DLM_EXIT_DOMAIN_MSG, dlm->key,
1078                                         sizeof(struct dlm_exit_domain),
1079                                         dlm_exit_domain_handler,
1080                                         dlm, &dlm->dlm_domain_handlers);
1081         if (status)
1082                 goto bail;
1083
1084         status = o2net_register_handler(DLM_MIGRATE_REQUEST_MSG, dlm->key,
1085                                         sizeof(struct dlm_migrate_request),
1086                                         dlm_migrate_request_handler,
1087                                         dlm, &dlm->dlm_domain_handlers);
1088         if (status)
1089                 goto bail;
1090
1091         status = o2net_register_handler(DLM_MIG_LOCKRES_MSG, dlm->key,
1092                                         DLM_MIG_LOCKRES_MAX_LEN,
1093                                         dlm_mig_lockres_handler,
1094                                         dlm, &dlm->dlm_domain_handlers);
1095         if (status)
1096                 goto bail;
1097
1098         status = o2net_register_handler(DLM_MASTER_REQUERY_MSG, dlm->key,
1099                                         sizeof(struct dlm_master_requery),
1100                                         dlm_master_requery_handler,
1101                                         dlm, &dlm->dlm_domain_handlers);
1102         if (status)
1103                 goto bail;
1104
1105         status = o2net_register_handler(DLM_LOCK_REQUEST_MSG, dlm->key,
1106                                         sizeof(struct dlm_lock_request),
1107                                         dlm_request_all_locks_handler,
1108                                         dlm, &dlm->dlm_domain_handlers);
1109         if (status)
1110                 goto bail;
1111
1112         status = o2net_register_handler(DLM_RECO_DATA_DONE_MSG, dlm->key,
1113                                         sizeof(struct dlm_reco_data_done),
1114                                         dlm_reco_data_done_handler,
1115                                         dlm, &dlm->dlm_domain_handlers);
1116         if (status)
1117                 goto bail;
1118
1119         status = o2net_register_handler(DLM_BEGIN_RECO_MSG, dlm->key,
1120                                         sizeof(struct dlm_begin_reco),
1121                                         dlm_begin_reco_handler,
1122                                         dlm, &dlm->dlm_domain_handlers);
1123         if (status)
1124                 goto bail;
1125
1126         status = o2net_register_handler(DLM_FINALIZE_RECO_MSG, dlm->key,
1127                                         sizeof(struct dlm_finalize_reco),
1128                                         dlm_finalize_reco_handler,
1129                                         dlm, &dlm->dlm_domain_handlers);
1130         if (status)
1131                 goto bail;
1132
1133 bail:
1134         if (status)
1135                 dlm_unregister_domain_handlers(dlm);
1136
1137         return status;
1138 }
1139
1140 static int dlm_join_domain(struct dlm_ctxt *dlm)
1141 {
1142         int status;
1143
1144         BUG_ON(!dlm);
1145
1146         mlog(0, "Join domain %s\n", dlm->name);
1147
1148         status = dlm_register_domain_handlers(dlm);
1149         if (status) {
1150                 mlog_errno(status);
1151                 goto bail;
1152         }
1153
1154         status = dlm_launch_thread(dlm);
1155         if (status < 0) {
1156                 mlog_errno(status);
1157                 goto bail;
1158         }
1159
1160         status = dlm_launch_recovery_thread(dlm);
1161         if (status < 0) {
1162                 mlog_errno(status);
1163                 goto bail;
1164         }
1165
1166         dlm->dlm_worker = create_singlethread_workqueue("dlm_wq");
1167         if (!dlm->dlm_worker) {
1168                 status = -ENOMEM;
1169                 mlog_errno(status);
1170                 goto bail;
1171         }
1172
1173         do {
1174                 unsigned int backoff;
1175                 status = dlm_try_to_join_domain(dlm);
1176
1177                 /* If we're racing another node to the join, then we
1178                  * need to back off temporarily and let them
1179                  * complete. */
1180                 if (status == -EAGAIN) {
1181                         if (signal_pending(current)) {
1182                                 status = -ERESTARTSYS;
1183                                 goto bail;
1184                         }
1185
1186                         /*
1187                          * <chip> After you!
1188                          * <dale> No, after you!
1189                          * <chip> I insist!
1190                          * <dale> But you first!
1191                          * ...
1192                          */
1193                         backoff = (unsigned int)(jiffies & 0x3);
1194                         backoff *= DLM_DOMAIN_BACKOFF_MS;
1195                         mlog(0, "backoff %d\n", backoff);
1196                         msleep(backoff);
1197                 }
1198         } while (status == -EAGAIN);
1199
1200         if (status < 0) {
1201                 mlog_errno(status);
1202                 goto bail;
1203         }
1204
1205         status = 0;
1206 bail:
1207         wake_up(&dlm_domain_events);
1208
1209         if (status) {
1210                 dlm_unregister_domain_handlers(dlm);
1211                 dlm_complete_thread(dlm);
1212                 dlm_complete_recovery_thread(dlm);
1213                 dlm_destroy_dlm_worker(dlm);
1214         }
1215
1216         return status;
1217 }
1218
1219 static struct dlm_ctxt *dlm_alloc_ctxt(const char *domain,
1220                                 u32 key)
1221 {
1222         int i;
1223         struct dlm_ctxt *dlm = NULL;
1224
1225         dlm = kcalloc(1, sizeof(*dlm), GFP_KERNEL);
1226         if (!dlm) {
1227                 mlog_errno(-ENOMEM);
1228                 goto leave;
1229         }
1230
1231         dlm->name = kmalloc(strlen(domain) + 1, GFP_KERNEL);
1232         if (dlm->name == NULL) {
1233                 mlog_errno(-ENOMEM);
1234                 kfree(dlm);
1235                 dlm = NULL;
1236                 goto leave;
1237         }
1238
1239         dlm->lockres_hash = (struct hlist_head **)dlm_alloc_pagevec(DLM_HASH_PAGES);
1240         if (!dlm->lockres_hash) {
1241                 mlog_errno(-ENOMEM);
1242                 kfree(dlm->name);
1243                 kfree(dlm);
1244                 dlm = NULL;
1245                 goto leave;
1246         }
1247
1248         for (i = 0; i < DLM_HASH_BUCKETS; i++)
1249                 INIT_HLIST_HEAD(dlm_lockres_hash(dlm, i));
1250
1251         strcpy(dlm->name, domain);
1252         dlm->key = key;
1253         dlm->node_num = o2nm_this_node();
1254
1255         spin_lock_init(&dlm->spinlock);
1256         spin_lock_init(&dlm->master_lock);
1257         spin_lock_init(&dlm->ast_lock);
1258         INIT_LIST_HEAD(&dlm->list);
1259         INIT_LIST_HEAD(&dlm->dirty_list);
1260         INIT_LIST_HEAD(&dlm->reco.resources);
1261         INIT_LIST_HEAD(&dlm->reco.received);
1262         INIT_LIST_HEAD(&dlm->reco.node_data);
1263         INIT_LIST_HEAD(&dlm->purge_list);
1264         INIT_LIST_HEAD(&dlm->dlm_domain_handlers);
1265         dlm->reco.state = 0;
1266
1267         INIT_LIST_HEAD(&dlm->pending_asts);
1268         INIT_LIST_HEAD(&dlm->pending_basts);
1269
1270         mlog(0, "dlm->recovery_map=%p, &(dlm->recovery_map[0])=%p\n",
1271                   dlm->recovery_map, &(dlm->recovery_map[0]));
1272
1273         memset(dlm->recovery_map, 0, sizeof(dlm->recovery_map));
1274         memset(dlm->live_nodes_map, 0, sizeof(dlm->live_nodes_map));
1275         memset(dlm->domain_map, 0, sizeof(dlm->domain_map));
1276
1277         dlm->dlm_thread_task = NULL;
1278         dlm->dlm_reco_thread_task = NULL;
1279         dlm->dlm_worker = NULL;
1280         init_waitqueue_head(&dlm->dlm_thread_wq);
1281         init_waitqueue_head(&dlm->dlm_reco_thread_wq);
1282         init_waitqueue_head(&dlm->reco.event);
1283         init_waitqueue_head(&dlm->ast_wq);
1284         init_waitqueue_head(&dlm->migration_wq);
1285         INIT_LIST_HEAD(&dlm->master_list);
1286         INIT_LIST_HEAD(&dlm->mle_hb_events);
1287
1288         dlm->joining_node = DLM_LOCK_RES_OWNER_UNKNOWN;
1289         init_waitqueue_head(&dlm->dlm_join_events);
1290
1291         dlm->reco.new_master = O2NM_INVALID_NODE_NUM;
1292         dlm->reco.dead_node = O2NM_INVALID_NODE_NUM;
1293         atomic_set(&dlm->local_resources, 0);
1294         atomic_set(&dlm->remote_resources, 0);
1295         atomic_set(&dlm->unknown_resources, 0);
1296
1297         spin_lock_init(&dlm->work_lock);
1298         INIT_LIST_HEAD(&dlm->work_list);
1299         INIT_WORK(&dlm->dispatched_work, dlm_dispatch_work);
1300
1301         kref_init(&dlm->dlm_refs);
1302         dlm->dlm_state = DLM_CTXT_NEW;
1303
1304         INIT_LIST_HEAD(&dlm->dlm_eviction_callbacks);
1305
1306         mlog(0, "context init: refcount %u\n",
1307                   atomic_read(&dlm->dlm_refs.refcount));
1308
1309 leave:
1310         return dlm;
1311 }
1312
1313 /*
1314  * dlm_register_domain: one-time setup per "domain"
1315  */
1316 struct dlm_ctxt * dlm_register_domain(const char *domain,
1317                                u32 key)
1318 {
1319         int ret;
1320         struct dlm_ctxt *dlm = NULL;
1321         struct dlm_ctxt *new_ctxt = NULL;
1322
1323         if (strlen(domain) > O2NM_MAX_NAME_LEN) {
1324                 ret = -ENAMETOOLONG;
1325                 mlog(ML_ERROR, "domain name length too long\n");
1326                 goto leave;
1327         }
1328
1329         if (!o2hb_check_local_node_heartbeating()) {
1330                 mlog(ML_ERROR, "the local node has not been configured, or is "
1331                      "not heartbeating\n");
1332                 ret = -EPROTO;
1333                 goto leave;
1334         }
1335
1336         mlog(0, "register called for domain \"%s\"\n", domain);
1337
1338 retry:
1339         dlm = NULL;
1340         if (signal_pending(current)) {
1341                 ret = -ERESTARTSYS;
1342                 mlog_errno(ret);
1343                 goto leave;
1344         }
1345
1346         spin_lock(&dlm_domain_lock);
1347
1348         dlm = __dlm_lookup_domain(domain);
1349         if (dlm) {
1350                 if (dlm->dlm_state != DLM_CTXT_JOINED) {
1351                         spin_unlock(&dlm_domain_lock);
1352
1353                         mlog(0, "This ctxt is not joined yet!\n");
1354                         wait_event_interruptible(dlm_domain_events,
1355                                                  dlm_wait_on_domain_helper(
1356                                                          domain));
1357                         goto retry;
1358                 }
1359
1360                 __dlm_get(dlm);
1361                 dlm->num_joins++;
1362
1363                 spin_unlock(&dlm_domain_lock);
1364
1365                 ret = 0;
1366                 goto leave;
1367         }
1368
1369         /* doesn't exist */
1370         if (!new_ctxt) {
1371                 spin_unlock(&dlm_domain_lock);
1372
1373                 new_ctxt = dlm_alloc_ctxt(domain, key);
1374                 if (new_ctxt)
1375                         goto retry;
1376
1377                 ret = -ENOMEM;
1378                 mlog_errno(ret);
1379                 goto leave;
1380         }
1381
1382         /* a little variable switch-a-roo here... */
1383         dlm = new_ctxt;
1384         new_ctxt = NULL;
1385
1386         /* add the new domain */
1387         list_add_tail(&dlm->list, &dlm_domains);
1388         spin_unlock(&dlm_domain_lock);
1389
1390         ret = dlm_join_domain(dlm);
1391         if (ret) {
1392                 mlog_errno(ret);
1393                 dlm_put(dlm);
1394                 goto leave;
1395         }
1396
1397         ret = 0;
1398 leave:
1399         if (new_ctxt)
1400                 dlm_free_ctxt_mem(new_ctxt);
1401
1402         if (ret < 0)
1403                 dlm = ERR_PTR(ret);
1404
1405         return dlm;
1406 }
1407 EXPORT_SYMBOL_GPL(dlm_register_domain);
1408
1409 static LIST_HEAD(dlm_join_handlers);
1410
1411 static void dlm_unregister_net_handlers(void)
1412 {
1413         o2net_unregister_handler_list(&dlm_join_handlers);
1414 }
1415
1416 static int dlm_register_net_handlers(void)
1417 {
1418         int status = 0;
1419
1420         status = o2net_register_handler(DLM_QUERY_JOIN_MSG, DLM_MOD_KEY,
1421                                         sizeof(struct dlm_query_join_request),
1422                                         dlm_query_join_handler,
1423                                         NULL, &dlm_join_handlers);
1424         if (status)
1425                 goto bail;
1426
1427         status = o2net_register_handler(DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY,
1428                                         sizeof(struct dlm_assert_joined),
1429                                         dlm_assert_joined_handler,
1430                                         NULL, &dlm_join_handlers);
1431         if (status)
1432                 goto bail;
1433
1434         status = o2net_register_handler(DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY,
1435                                         sizeof(struct dlm_cancel_join),
1436                                         dlm_cancel_join_handler,
1437                                         NULL, &dlm_join_handlers);
1438
1439 bail:
1440         if (status < 0)
1441                 dlm_unregister_net_handlers();
1442
1443         return status;
1444 }
1445
1446 /* Domain eviction callback handling.
1447  *
1448  * The file system requires notification of node death *before* the
1449  * dlm completes it's recovery work, otherwise it may be able to
1450  * acquire locks on resources requiring recovery. Since the dlm can
1451  * evict a node from it's domain *before* heartbeat fires, a similar
1452  * mechanism is required. */
1453
1454 /* Eviction is not expected to happen often, so a per-domain lock is
1455  * not necessary. Eviction callbacks are allowed to sleep for short
1456  * periods of time. */
1457 static DECLARE_RWSEM(dlm_callback_sem);
1458
1459 void dlm_fire_domain_eviction_callbacks(struct dlm_ctxt *dlm,
1460                                         int node_num)
1461 {
1462         struct list_head *iter;
1463         struct dlm_eviction_cb *cb;
1464
1465         down_read(&dlm_callback_sem);
1466         list_for_each(iter, &dlm->dlm_eviction_callbacks) {
1467                 cb = list_entry(iter, struct dlm_eviction_cb, ec_item);
1468
1469                 cb->ec_func(node_num, cb->ec_data);
1470         }
1471         up_read(&dlm_callback_sem);
1472 }
1473
1474 void dlm_setup_eviction_cb(struct dlm_eviction_cb *cb,
1475                            dlm_eviction_func *f,
1476                            void *data)
1477 {
1478         INIT_LIST_HEAD(&cb->ec_item);
1479         cb->ec_func = f;
1480         cb->ec_data = data;
1481 }
1482 EXPORT_SYMBOL_GPL(dlm_setup_eviction_cb);
1483
1484 void dlm_register_eviction_cb(struct dlm_ctxt *dlm,
1485                               struct dlm_eviction_cb *cb)
1486 {
1487         down_write(&dlm_callback_sem);
1488         list_add_tail(&cb->ec_item, &dlm->dlm_eviction_callbacks);
1489         up_write(&dlm_callback_sem);
1490 }
1491 EXPORT_SYMBOL_GPL(dlm_register_eviction_cb);
1492
1493 void dlm_unregister_eviction_cb(struct dlm_eviction_cb *cb)
1494 {
1495         down_write(&dlm_callback_sem);
1496         list_del_init(&cb->ec_item);
1497         up_write(&dlm_callback_sem);
1498 }
1499 EXPORT_SYMBOL_GPL(dlm_unregister_eviction_cb);
1500
1501 static int __init dlm_init(void)
1502 {
1503         int status;
1504
1505         dlm_print_version();
1506
1507         status = dlm_init_mle_cache();
1508         if (status)
1509                 return -1;
1510
1511         status = dlm_register_net_handlers();
1512         if (status) {
1513                 dlm_destroy_mle_cache();
1514                 return -1;
1515         }
1516
1517         return 0;
1518 }
1519
1520 static void __exit dlm_exit (void)
1521 {
1522         dlm_unregister_net_handlers();
1523         dlm_destroy_mle_cache();
1524 }
1525
1526 MODULE_AUTHOR("Oracle");
1527 MODULE_LICENSE("GPL");
1528
1529 module_init(dlm_init);
1530 module_exit(dlm_exit);