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, (unsigned long)DLM_HASH_PAGES,
  72              (unsigned long)DLM_BUCKETS_PER_PAGE);
  73         return vec;
  74 out_free:
  75         dlm_free_pagevec(vec, i);
  76         return NULL;
  77 }
  78
  79 /*
  80  *
  81  * spinlock lock ordering: if multiple locks are needed, obey this ordering:
  82  *    dlm_domain_lock
  83  *    struct dlm_ctxt->spinlock
  84  *    struct dlm_lock_resource->spinlock
  85  *    struct dlm_ctxt->master_lock
  86  *    struct dlm_ctxt->ast_lock
  87  *    dlm_master_list_entry->spinlock
  88  *    dlm_lock->spinlock
  89  *
  90  */
  91
  92 DEFINE_SPINLOCK(dlm_domain_lock);
  93 LIST_HEAD(dlm_domains);
  94 static DECLARE_WAIT_QUEUE_HEAD(dlm_domain_events);
  95
  96 #define DLM_DOMAIN_BACKOFF_MS 200
  97
  98 static int dlm_query_join_handler(struct o2net_msg *msg, u32 len, void *data);
  99 static int dlm_assert_joined_handler(struct o2net_msg *msg, u32 len, void *data);
 100 static int dlm_cancel_join_handler(struct o2net_msg *msg, u32 len, void *data);
 101 static int dlm_exit_domain_handler(struct o2net_msg *msg, u32 len, void *data);
 102
 103 static void dlm_unregister_domain_handlers(struct dlm_ctxt *dlm);
 104
 105 void __dlm_unhash_lockres(struct dlm_lock_resource *lockres)
 106 {
 107         hlist_del_init(&lockres->hash_node);
 108         dlm_lockres_put(lockres);
 109 }
 110
 111 void __dlm_insert_lockres(struct dlm_ctxt *dlm,
 112                        struct dlm_lock_resource *res)
 113 {
 114         struct hlist_head *bucket;
 115         struct qstr *q;
 116
 117         assert_spin_locked(&dlm->spinlock);
 118
 119         q = &res->lockname;
 120         bucket = dlm_lockres_hash(dlm, q->hash);
 121
 122         /* get a reference for our hashtable */
 123         dlm_lockres_get(res);
 124
 125         hlist_add_head(&res->hash_node, bucket);
 126 }
 127
 128 struct dlm_lock_resource * __dlm_lookup_lockres_full(struct dlm_ctxt *dlm,
 129                                                      const char *name,
 130                                                      unsigned int len,
 131                                                      unsigned int hash)
 132 {
 133         struct hlist_head *bucket;
 134         struct hlist_node *list;
 135
 136         mlog_entry("%.*s\n", len, name);
 137
 138         assert_spin_locked(&dlm->spinlock);
 139
 140         bucket = dlm_lockres_hash(dlm, hash);
 141
 142         hlist_for_each(list, bucket) {
 143                 struct dlm_lock_resource *res = hlist_entry(list,
 144                         struct dlm_lock_resource, hash_node);
 145                 if (res->lockname.name[0] != name[0])
 146                         continue;
 147                 if (unlikely(res->lockname.len != len))
 148                         continue;
 149                 if (memcmp(res->lockname.name + 1, name + 1, len - 1))
 150                         continue;
 151                 dlm_lockres_get(res);
 152                 return res;
 153         }
 154         return NULL;
 155 }
 156
 157 /* intended to be called by functions which do not care about lock
 158  * resources which are being purged (most net _handler functions).
 159  * this will return NULL for any lock resource which is found but
 160  * currently in the process of dropping its mastery reference.
 161  * use __dlm_lookup_lockres_full when you need the lock resource
 162  * regardless (e.g. dlm_get_lock_resource) */
 163 struct dlm_lock_resource * __dlm_lookup_lockres(struct dlm_ctxt *dlm,
 164                                                 const char *name,
 165                                                 unsigned int len,
 166                                                 unsigned int hash)
 167 {
 168         struct dlm_lock_resource *res = NULL;
 169
 170         mlog_entry("%.*s\n", len, name);
 171
 172         assert_spin_locked(&dlm->spinlock);
 173
 174         res = __dlm_lookup_lockres_full(dlm, name, len, hash);
 175         if (res) {
 176                 spin_lock(&res->spinlock);
 177                 if (res->state & DLM_LOCK_RES_DROPPING_REF) {
 178                         spin_unlock(&res->spinlock);
 179                         dlm_lockres_put(res);
 180                         return NULL;
 181                 }
 182                 spin_unlock(&res->spinlock);
 183         }
 184
 185         return res;
 186 }
 187
 188 struct dlm_lock_resource * dlm_lookup_lockres(struct dlm_ctxt *dlm,
 189                                     const char *name,
 190                                     unsigned int len)
 191 {
 192         struct dlm_lock_resource *res;
 193         unsigned int hash = dlm_lockid_hash(name, len);
 194
 195         spin_lock(&dlm->spinlock);
 196         res = __dlm_lookup_lockres(dlm, name, len, hash);
 197         spin_unlock(&dlm->spinlock);
 198         return res;
 199 }
 200
 201 static struct dlm_ctxt * __dlm_lookup_domain_full(const char *domain, int len)
 202 {
 203         struct dlm_ctxt *tmp = NULL;
 204         struct list_head *iter;
 205
 206         assert_spin_locked(&dlm_domain_lock);
 207
 208         /* tmp->name here is always NULL terminated,
 209          * but domain may not be! */
 210         list_for_each(iter, &dlm_domains) {
 211                 tmp = list_entry (iter, struct dlm_ctxt, list);
 212                 if (strlen(tmp->name) == len &&
 213                     memcmp(tmp->name, domain, len)==0)
 214                         break;
 215                 tmp = NULL;
 216         }
 217
 218         return tmp;
 219 }
 220
 221 /* For null terminated domain strings ONLY */
 222 static struct dlm_ctxt * __dlm_lookup_domain(const char *domain)
 223 {
 224         assert_spin_locked(&dlm_domain_lock);
 225
 226         return __dlm_lookup_domain_full(domain, strlen(domain));
 227 }
 228
 229
 230 /* returns true on one of two conditions:
 231  * 1) the domain does not exist
 232  * 2) the domain exists and it's state is "joined" */
 233 static int dlm_wait_on_domain_helper(const char *domain)
 234 {
 235         int ret = 0;
 236         struct dlm_ctxt *tmp = NULL;
 237
 238         spin_lock(&dlm_domain_lock);
 239
 240         tmp = __dlm_lookup_domain(domain);
 241         if (!tmp)
 242                 ret = 1;
 243         else if (tmp->dlm_state == DLM_CTXT_JOINED)
 244                 ret = 1;
 245
 246         spin_unlock(&dlm_domain_lock);
 247         return ret;
 248 }
 249
 250 static void dlm_free_ctxt_mem(struct dlm_ctxt *dlm)
 251 {
 252         if (dlm->lockres_hash)
 253                 dlm_free_pagevec((void **)dlm->lockres_hash, DLM_HASH_PAGES);
 254
 255         if (dlm->name)
 256                 kfree(dlm->name);
 257
 258         kfree(dlm);
 259 }
 260
 261 /* A little strange - this function will be called while holding
 262  * dlm_domain_lock and is expected to be holding it on the way out. We
 263  * will however drop and reacquire it multiple times */
 264 static void dlm_ctxt_release(struct kref *kref)
 265 {
 266         struct dlm_ctxt *dlm;
 267
 268         dlm = container_of(kref, struct dlm_ctxt, dlm_refs);
 269
 270         BUG_ON(dlm->num_joins);
 271         BUG_ON(dlm->dlm_state == DLM_CTXT_JOINED);
 272
 273         /* we may still be in the list if we hit an error during join. */
 274         list_del_init(&dlm->list);
 275
 276         spin_unlock(&dlm_domain_lock);
 277
 278         mlog(0, "freeing memory from domain %s\n", dlm->name);
 279
 280         wake_up(&dlm_domain_events);
 281
 282         dlm_free_ctxt_mem(dlm);
 283
 284         spin_lock(&dlm_domain_lock);
 285 }
 286
 287 void dlm_put(struct dlm_ctxt *dlm)
 288 {
 289         spin_lock(&dlm_domain_lock);
 290         kref_put(&dlm->dlm_refs, dlm_ctxt_release);
 291         spin_unlock(&dlm_domain_lock);
 292 }
 293
 294 static void __dlm_get(struct dlm_ctxt *dlm)
 295 {
 296         kref_get(&dlm->dlm_refs);
 297 }
 298
 299 /* given a questionable reference to a dlm object, gets a reference if
 300  * it can find it in the list, otherwise returns NULL in which case
 301  * you shouldn't trust your pointer. */
 302 struct dlm_ctxt *dlm_grab(struct dlm_ctxt *dlm)
 303 {
 304         struct list_head *iter;
 305         struct dlm_ctxt *target = NULL;
 306
 307         spin_lock(&dlm_domain_lock);
 308
 309         list_for_each(iter, &dlm_domains) {
 310                 target = list_entry (iter, struct dlm_ctxt, list);
 311
 312                 if (target == dlm) {
 313                         __dlm_get(target);
 314                         break;
 315                 }
 316
 317                 target = NULL;
 318         }
 319
 320         spin_unlock(&dlm_domain_lock);
 321
 322         return target;
 323 }
 324
 325 int dlm_domain_fully_joined(struct dlm_ctxt *dlm)
 326 {
 327         int ret;
 328
 329         spin_lock(&dlm_domain_lock);
 330         ret = (dlm->dlm_state == DLM_CTXT_JOINED) ||
 331                 (dlm->dlm_state == DLM_CTXT_IN_SHUTDOWN);
 332         spin_unlock(&dlm_domain_lock);
 333
 334         return ret;
 335 }
 336
 337 static void dlm_destroy_dlm_worker(struct dlm_ctxt *dlm)
 338 {
 339         if (dlm->dlm_worker) {
 340                 flush_workqueue(dlm->dlm_worker);
 341                 destroy_workqueue(dlm->dlm_worker);
 342                 dlm->dlm_worker = NULL;
 343         }
 344 }
 345
 346 static void dlm_complete_dlm_shutdown(struct dlm_ctxt *dlm)
 347 {
 348         dlm_unregister_domain_handlers(dlm);
 349         dlm_complete_thread(dlm);
 350         dlm_complete_recovery_thread(dlm);
 351         dlm_destroy_dlm_worker(dlm);
 352
 353         /* We've left the domain. Now we can take ourselves out of the
 354          * list and allow the kref stuff to help us free the
 355          * memory. */
 356         spin_lock(&dlm_domain_lock);
 357         list_del_init(&dlm->list);
 358         spin_unlock(&dlm_domain_lock);
 359
 360         /* Wake up anyone waiting for us to remove this domain */
 361         wake_up(&dlm_domain_events);
 362 }
 363
 364 static int dlm_migrate_all_locks(struct dlm_ctxt *dlm)
 365 {
 366         int i, num, n, ret = 0;
 367         struct dlm_lock_resource *res;
 368         struct hlist_node *iter;
 369         struct hlist_head *bucket;
 370         int dropped;
 371
 372         mlog(0, "Migrating locks from domain %s\n", dlm->name);
 373
 374         num = 0;
 375         spin_lock(&dlm->spinlock);
 376         for (i = 0; i < DLM_HASH_BUCKETS; i++) {
 377 redo_bucket:
 378                 n = 0;
 379                 bucket = dlm_lockres_hash(dlm, i);
 380                 iter = bucket->first;
 381                 while (iter) {
 382                         n++;
 383                         res = hlist_entry(iter, struct dlm_lock_resource,
 384                                           hash_node);
 385                         dlm_lockres_get(res);
 386                         /* migrate, if necessary.  this will drop the dlm
 387                          * spinlock and retake it if it does migration. */
 388                         dropped = dlm_empty_lockres(dlm, res);
 389
 390                         spin_lock(&res->spinlock);
 391                         __dlm_lockres_calc_usage(dlm, res);
 392                         iter = res->hash_node.next;
 393                         spin_unlock(&res->spinlock);
 394
 395                         dlm_lockres_put(res);
 396
 397                         cond_resched_lock(&dlm->spinlock);
 398
 399                         if (dropped)
 400                                 goto redo_bucket;
 401                 }
 402                 num += n;
 403                 mlog(0, "%s: touched %d lockreses in bucket %d "
 404                      "(tot=%d)\n", dlm->name, n, i, num);
 405         }
 406         spin_unlock(&dlm->spinlock);
 407         wake_up(&dlm->dlm_thread_wq);
 408
 409         /* let the dlm thread take care of purging, keep scanning until
 410          * nothing remains in the hash */
 411         if (num) {
 412                 mlog(0, "%s: %d lock resources in hash last pass\n",
 413                      dlm->name, num);
 414                 ret = -EAGAIN;
 415         }
 416         mlog(0, "DONE Migrating locks from domain %s\n", dlm->name);
 417         return ret;
 418 }
 419
 420 static int dlm_no_joining_node(struct dlm_ctxt *dlm)
 421 {
 422         int ret;
 423
 424         spin_lock(&dlm->spinlock);
 425         ret = dlm->joining_node == DLM_LOCK_RES_OWNER_UNKNOWN;
 426         spin_unlock(&dlm->spinlock);
 427
 428         return ret;
 429 }
 430
 431 static void dlm_mark_domain_leaving(struct dlm_ctxt *dlm)
 432 {
 433         /* Yikes, a double spinlock! I need domain_lock for the dlm
 434          * state and the dlm spinlock for join state... Sorry! */
 435 again:
 436         spin_lock(&dlm_domain_lock);
 437         spin_lock(&dlm->spinlock);
 438
 439         if (dlm->joining_node != DLM_LOCK_RES_OWNER_UNKNOWN) {
 440                 mlog(0, "Node %d is joining, we wait on it.\n",
 441                           dlm->joining_node);
 442                 spin_unlock(&dlm->spinlock);
 443                 spin_unlock(&dlm_domain_lock);
 444
 445                 wait_event(dlm->dlm_join_events, dlm_no_joining_node(dlm));
 446                 goto again;
 447         }
 448
 449         dlm->dlm_state = DLM_CTXT_LEAVING;
 450         spin_unlock(&dlm->spinlock);
 451         spin_unlock(&dlm_domain_lock);
 452 }
 453
 454 static void __dlm_print_nodes(struct dlm_ctxt *dlm)
 455 {
 456         int node = -1;
 457
 458         assert_spin_locked(&dlm->spinlock);
 459
 460         printk(KERN_INFO "ocfs2_dlm: Nodes in domain (\"%s\"): ", dlm->name);
 461
 462         while ((node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES,
 463                                      node + 1)) < O2NM_MAX_NODES) {
 464                 printk("%d ", node);
 465         }
 466         printk("\n");
 467 }
 468
 469 static int dlm_exit_domain_handler(struct o2net_msg *msg, u32 len, void *data)
 470 {
 471         struct dlm_ctxt *dlm = data;
 472         unsigned int node;
 473         struct dlm_exit_domain *exit_msg = (struct dlm_exit_domain *) msg->buf;
 474
 475         mlog_entry("%p %u %p", msg, len, data);
 476
 477         if (!dlm_grab(dlm))
 478                 return 0;
 479
 480         node = exit_msg->node_idx;
 481
 482         printk(KERN_INFO "ocfs2_dlm: Node %u leaves domain %s\n", node, dlm->name);
 483
 484         spin_lock(&dlm->spinlock);
 485         clear_bit(node, dlm->domain_map);
 486         __dlm_print_nodes(dlm);
 487
 488         /* notify anything attached to the heartbeat events */
 489         dlm_hb_event_notify_attached(dlm, node, 0);
 490
 491         spin_unlock(&dlm->spinlock);
 492
 493         dlm_put(dlm);
 494
 495         return 0;
 496 }
 497
 498 static int dlm_send_one_domain_exit(struct dlm_ctxt *dlm,
 499                                     unsigned int node)
 500 {
 501         int status;
 502         struct dlm_exit_domain leave_msg;
 503
 504         mlog(0, "Asking node %u if we can leave the domain %s me = %u\n",
 505                   node, dlm->name, dlm->node_num);
 506
 507         memset(&leave_msg, 0, sizeof(leave_msg));
 508         leave_msg.node_idx = dlm->node_num;
 509
 510         status = o2net_send_message(DLM_EXIT_DOMAIN_MSG, dlm->key,
 511                                     &leave_msg, sizeof(leave_msg), node,
 512                                     NULL);
 513
 514         mlog(0, "status return %d from o2net_send_message\n", status);
 515
 516         return status;
 517 }
 518
 519
 520 static void dlm_leave_domain(struct dlm_ctxt *dlm)
 521 {
 522         int node, clear_node, status;
 523
 524         /* At this point we've migrated away all our locks and won't
 525          * accept mastership of new ones. The dlm is responsible for
 526          * almost nothing now. We make sure not to confuse any joining
 527          * nodes and then commence shutdown procedure. */
 528
 529         spin_lock(&dlm->spinlock);
 530         /* Clear ourselves from the domain map */
 531         clear_bit(dlm->node_num, dlm->domain_map);
 532         while ((node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES,
 533                                      0)) < O2NM_MAX_NODES) {
 534                 /* Drop the dlm spinlock. This is safe wrt the domain_map.
 535                  * -nodes cannot be added now as the
 536                  *   query_join_handlers knows to respond with OK_NO_MAP
 537                  * -we catch the right network errors if a node is
 538                  *   removed from the map while we're sending him the
 539                  *   exit message. */
 540                 spin_unlock(&dlm->spinlock);
 541
 542                 clear_node = 1;
 543
 544                 status = dlm_send_one_domain_exit(dlm, node);
 545                 if (status < 0 &&
 546                     status != -ENOPROTOOPT &&
 547                     status != -ENOTCONN) {
 548                         mlog(ML_NOTICE, "Error %d sending domain exit message "
 549                              "to node %d\n", status, node);
 550
 551                         /* Not sure what to do here but lets sleep for
 552                          * a bit in case this was a transient
 553                          * error... */
 554                         msleep(DLM_DOMAIN_BACKOFF_MS);
 555                         clear_node = 0;
 556                 }
 557
 558                 spin_lock(&dlm->spinlock);
 559                 /* If we're not clearing the node bit then we intend
 560                  * to loop back around to try again. */
 561                 if (clear_node)
 562                         clear_bit(node, dlm->domain_map);
 563         }
 564         spin_unlock(&dlm->spinlock);
 565 }
 566
 567 int dlm_joined(struct dlm_ctxt *dlm)
 568 {
 569         int ret = 0;
 570
 571         spin_lock(&dlm_domain_lock);
 572
 573         if (dlm->dlm_state == DLM_CTXT_JOINED)
 574                 ret = 1;
 575
 576         spin_unlock(&dlm_domain_lock);
 577
 578         return ret;
 579 }
 580
 581 int dlm_shutting_down(struct dlm_ctxt *dlm)
 582 {
 583         int ret = 0;
 584
 585         spin_lock(&dlm_domain_lock);
 586
 587         if (dlm->dlm_state == DLM_CTXT_IN_SHUTDOWN)
 588                 ret = 1;
 589
 590         spin_unlock(&dlm_domain_lock);
 591
 592         return ret;
 593 }
 594
 595 void dlm_unregister_domain(struct dlm_ctxt *dlm)
 596 {
 597         int leave = 0;
 598
 599         spin_lock(&dlm_domain_lock);
 600         BUG_ON(dlm->dlm_state != DLM_CTXT_JOINED);
 601         BUG_ON(!dlm->num_joins);
 602
 603         dlm->num_joins--;
 604         if (!dlm->num_joins) {
 605                 /* We mark it "in shutdown" now so new register
 606                  * requests wait until we've completely left the
 607                  * domain. Don't use DLM_CTXT_LEAVING yet as we still
 608                  * want new domain joins to communicate with us at
 609                  * least until we've completed migration of our
 610                  * resources. */
 611                 dlm->dlm_state = DLM_CTXT_IN_SHUTDOWN;
 612                 leave = 1;
 613         }
 614         spin_unlock(&dlm_domain_lock);
 615
 616         if (leave) {
 617                 mlog(0, "shutting down domain %s\n", dlm->name);
 618
 619                 /* We changed dlm state, notify the thread */
 620                 dlm_kick_thread(dlm, NULL);
 621
 622                 while (dlm_migrate_all_locks(dlm)) {
 623                         mlog(0, "%s: more migration to do\n", dlm->name);
 624                 }
 625                 dlm_mark_domain_leaving(dlm);
 626                 dlm_leave_domain(dlm);
 627                 dlm_complete_dlm_shutdown(dlm);
 628         }
 629         dlm_put(dlm);
 630 }
 631 EXPORT_SYMBOL_GPL(dlm_unregister_domain);
 632
 633 static int dlm_query_join_handler(struct o2net_msg *msg, u32 len, void *data)
 634 {
 635         struct dlm_query_join_request *query;
 636         enum dlm_query_join_response response;
 637         struct dlm_ctxt *dlm = NULL;
 638
 639         query = (struct dlm_query_join_request *) msg->buf;
 640
 641         mlog(0, "node %u wants to join domain %s\n", query->node_idx,
 642                   query->domain);
 643
 644         /*
 645          * If heartbeat doesn't consider the node live, tell it
 646          * to back off and try again.  This gives heartbeat a chance
 647          * to catch up.
 648          */
 649         if (!o2hb_check_node_heartbeating(query->node_idx)) {
 650                 mlog(0, "node %u is not in our live map yet\n",
 651                      query->node_idx);
 652
 653                 response = JOIN_DISALLOW;
 654                 goto respond;
 655         }
 656
 657         response = JOIN_OK_NO_MAP;
 658
 659         spin_lock(&dlm_domain_lock);
 660         dlm = __dlm_lookup_domain_full(query->domain, query->name_len);
 661         /* Once the dlm ctxt is marked as leaving then we don't want
 662          * to be put in someone's domain map.
 663          * Also, explicitly disallow joining at certain troublesome
 664          * times (ie. during recovery). */
 665         if (dlm && dlm->dlm_state != DLM_CTXT_LEAVING) {
 666                 int bit = query->node_idx;
 667                 spin_lock(&dlm->spinlock);
 668
 669                 if (dlm->dlm_state == DLM_CTXT_NEW &&
 670                     dlm->joining_node == DLM_LOCK_RES_OWNER_UNKNOWN) {
 671                         /*If this is a brand new context and we
 672                          * haven't started our join process yet, then
 673                          * the other node won the race. */
 674                         response = JOIN_OK_NO_MAP;
 675                 } else if (dlm->joining_node != DLM_LOCK_RES_OWNER_UNKNOWN) {
 676                         /* Disallow parallel joins. */
 677                         response = JOIN_DISALLOW;
 678                 } else if (dlm->reco.state & DLM_RECO_STATE_ACTIVE) {
 679                         mlog(ML_NOTICE, "node %u trying to join, but recovery "
 680                              "is ongoing.\n", bit);
 681                         response = JOIN_DISALLOW;
 682                 } else if (test_bit(bit, dlm->recovery_map)) {
 683                         mlog(ML_NOTICE, "node %u trying to join, but it "
 684                              "still needs recovery.\n", bit);
 685                         response = JOIN_DISALLOW;
 686                 } else if (test_bit(bit, dlm->domain_map)) {
 687                         mlog(ML_NOTICE, "node %u trying to join, but it "
 688                              "is still in the domain! needs recovery?\n",
 689                              bit);
 690                         response = JOIN_DISALLOW;
 691                 } else {
 692                         /* Alright we're fully a part of this domain
 693                          * so we keep some state as to who's joining
 694                          * and indicate to him that needs to be fixed
 695                          * up. */
 696                         response = JOIN_OK;
 697                         __dlm_set_joining_node(dlm, query->node_idx);
 698                 }
 699
 700                 spin_unlock(&dlm->spinlock);
 701         }
 702         spin_unlock(&dlm_domain_lock);
 703
 704 respond:
 705         mlog(0, "We respond with %u\n", response);
 706
 707         return response;
 708 }
 709
 710 static int dlm_assert_joined_handler(struct o2net_msg *msg, u32 len, void *data)
 711 {
 712         struct dlm_assert_joined *assert;
 713         struct dlm_ctxt *dlm = NULL;
 714
 715         assert = (struct dlm_assert_joined *) msg->buf;
 716
 717         mlog(0, "node %u asserts join on domain %s\n", assert->node_idx,
 718                   assert->domain);
 719
 720         spin_lock(&dlm_domain_lock);
 721         dlm = __dlm_lookup_domain_full(assert->domain, assert->name_len);
 722         /* XXX should we consider no dlm ctxt an error? */
 723         if (dlm) {
 724                 spin_lock(&dlm->spinlock);
 725
 726                 /* Alright, this node has officially joined our
 727                  * domain. Set him in the map and clean up our
 728                  * leftover join state. */
 729                 BUG_ON(dlm->joining_node != assert->node_idx);
 730                 set_bit(assert->node_idx, dlm->domain_map);
 731                 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
 732
 733                 printk(KERN_INFO "ocfs2_dlm: Node %u joins domain %s\n",
 734                        assert->node_idx, dlm->name);
 735                 __dlm_print_nodes(dlm);
 736
 737                 /* notify anything attached to the heartbeat events */
 738                 dlm_hb_event_notify_attached(dlm, assert->node_idx, 1);
 739
 740                 spin_unlock(&dlm->spinlock);
 741         }
 742         spin_unlock(&dlm_domain_lock);
 743
 744         return 0;
 745 }
 746
 747 static int dlm_cancel_join_handler(struct o2net_msg *msg, u32 len, void *data)
 748 {
 749         struct dlm_cancel_join *cancel;
 750         struct dlm_ctxt *dlm = NULL;
 751
 752         cancel = (struct dlm_cancel_join *) msg->buf;
 753
 754         mlog(0, "node %u cancels join on domain %s\n", cancel->node_idx,
 755                   cancel->domain);
 756
 757         spin_lock(&dlm_domain_lock);
 758         dlm = __dlm_lookup_domain_full(cancel->domain, cancel->name_len);
 759
 760         if (dlm) {
 761                 spin_lock(&dlm->spinlock);
 762
 763                 /* Yikes, this guy wants to cancel his join. No
 764                  * problem, we simply cleanup our join state. */
 765                 BUG_ON(dlm->joining_node != cancel->node_idx);
 766                 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
 767
 768                 spin_unlock(&dlm->spinlock);
 769         }
 770         spin_unlock(&dlm_domain_lock);
 771
 772         return 0;
 773 }
 774
 775 static int dlm_send_one_join_cancel(struct dlm_ctxt *dlm,
 776                                     unsigned int node)
 777 {
 778         int status;
 779         struct dlm_cancel_join cancel_msg;
 780
 781         memset(&cancel_msg, 0, sizeof(cancel_msg));
 782         cancel_msg.node_idx = dlm->node_num;
 783         cancel_msg.name_len = strlen(dlm->name);
 784         memcpy(cancel_msg.domain, dlm->name, cancel_msg.name_len);
 785
 786         status = o2net_send_message(DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY,
 787                                     &cancel_msg, sizeof(cancel_msg), node,
 788                                     NULL);
 789         if (status < 0) {
 790                 mlog_errno(status);
 791                 goto bail;
 792         }
 793
 794 bail:
 795         return status;
 796 }
 797
 798 /* map_size should be in bytes. */
 799 static int dlm_send_join_cancels(struct dlm_ctxt *dlm,
 800                                  unsigned long *node_map,
 801                                  unsigned int map_size)
 802 {
 803         int status, tmpstat;
 804         unsigned int node;
 805
 806         if (map_size != (BITS_TO_LONGS(O2NM_MAX_NODES) *
 807                          sizeof(unsigned long))) {
 808                 mlog(ML_ERROR,
 809                      "map_size %u != BITS_TO_LONGS(O2NM_MAX_NODES) %u\n",
 810                      map_size, BITS_TO_LONGS(O2NM_MAX_NODES));
 811                 return -EINVAL;
 812         }
 813
 814         status = 0;
 815         node = -1;
 816         while ((node = find_next_bit(node_map, O2NM_MAX_NODES,
 817                                      node + 1)) < O2NM_MAX_NODES) {
 818                 if (node == dlm->node_num)
 819                         continue;
 820
 821                 tmpstat = dlm_send_one_join_cancel(dlm, node);
 822                 if (tmpstat) {
 823                         mlog(ML_ERROR, "Error return %d cancelling join on "
 824                              "node %d\n", tmpstat, node);
 825                         if (!status)
 826                                 status = tmpstat;
 827                 }
 828         }
 829
 830         if (status)
 831                 mlog_errno(status);
 832         return status;
 833 }
 834
 835 static int dlm_request_join(struct dlm_ctxt *dlm,
 836                             int node,
 837                             enum dlm_query_join_response *response)
 838 {
 839         int status, retval;
 840         struct dlm_query_join_request join_msg;
 841
 842         mlog(0, "querying node %d\n", node);
 843
 844         memset(&join_msg, 0, sizeof(join_msg));
 845         join_msg.node_idx = dlm->node_num;
 846         join_msg.name_len = strlen(dlm->name);
 847         memcpy(join_msg.domain, dlm->name, join_msg.name_len);
 848
 849         status = o2net_send_message(DLM_QUERY_JOIN_MSG, DLM_MOD_KEY, &join_msg,
 850                                     sizeof(join_msg), node, &retval);
 851         if (status < 0 && status != -ENOPROTOOPT) {
 852                 mlog_errno(status);
 853                 goto bail;
 854         }
 855
 856         /* -ENOPROTOOPT from the net code means the other side isn't
 857             listening for our message type -- that's fine, it means
 858             his dlm isn't up, so we can consider him a 'yes' but not
 859             joined into the domain.  */
 860         if (status == -ENOPROTOOPT) {
 861                 status = 0;
 862                 *response = JOIN_OK_NO_MAP;
 863         } else if (retval == JOIN_DISALLOW ||
 864                    retval == JOIN_OK ||
 865                    retval == JOIN_OK_NO_MAP) {
 866                 *response = retval;
 867         } else {
 868                 status = -EINVAL;
 869                 mlog(ML_ERROR, "invalid response %d from node %u\n", retval,
 870                      node);
 871         }
 872
 873         mlog(0, "status %d, node %d response is %d\n", status, node,
 874                   *response);
 875
 876 bail:
 877         return status;
 878 }
 879
 880 static int dlm_send_one_join_assert(struct dlm_ctxt *dlm,
 881                                     unsigned int node)
 882 {
 883         int status;
 884         struct dlm_assert_joined assert_msg;
 885
 886         mlog(0, "Sending join assert to node %u\n", node);
 887
 888         memset(&assert_msg, 0, sizeof(assert_msg));
 889         assert_msg.node_idx = dlm->node_num;
 890         assert_msg.name_len = strlen(dlm->name);
 891         memcpy(assert_msg.domain, dlm->name, assert_msg.name_len);
 892
 893         status = o2net_send_message(DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY,
 894                                     &assert_msg, sizeof(assert_msg), node,
 895                                     NULL);
 896         if (status < 0)
 897                 mlog_errno(status);
 898
 899         return status;
 900 }
 901
 902 static void dlm_send_join_asserts(struct dlm_ctxt *dlm,
 903                                   unsigned long *node_map)
 904 {
 905         int status, node, live;
 906
 907         status = 0;
 908         node = -1;
 909         while ((node = find_next_bit(node_map, O2NM_MAX_NODES,
 910                                      node + 1)) < O2NM_MAX_NODES) {
 911                 if (node == dlm->node_num)
 912                         continue;
 913
 914                 do {
 915                         /* It is very important that this message be
 916                          * received so we spin until either the node
 917                          * has died or it gets the message. */
 918                         status = dlm_send_one_join_assert(dlm, node);
 919
 920                         spin_lock(&dlm->spinlock);
 921                         live = test_bit(node, dlm->live_nodes_map);
 922                         spin_unlock(&dlm->spinlock);
 923
 924                         if (status) {
 925                                 mlog(ML_ERROR, "Error return %d asserting "
 926                                      "join on node %d\n", status, node);
 927
 928                                 /* give us some time between errors... */
 929                                 if (live)
 930                                         msleep(DLM_DOMAIN_BACKOFF_MS);
 931                         }
 932                 } while (status && live);
 933         }
 934 }
 935
 936 struct domain_join_ctxt {
 937         unsigned long live_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
 938         unsigned long yes_resp_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
 939 };
 940
 941 static int dlm_should_restart_join(struct dlm_ctxt *dlm,
 942                                    struct domain_join_ctxt *ctxt,
 943                                    enum dlm_query_join_response response)
 944 {
 945         int ret;
 946
 947         if (response == JOIN_DISALLOW) {
 948                 mlog(0, "Latest response of disallow -- should restart\n");
 949                 return 1;
 950         }
 951
 952         spin_lock(&dlm->spinlock);
 953         /* For now, we restart the process if the node maps have
 954          * changed at all */
 955         ret = memcmp(ctxt->live_map, dlm->live_nodes_map,
 956                      sizeof(dlm->live_nodes_map));
 957         spin_unlock(&dlm->spinlock);
 958
 959         if (ret)
 960                 mlog(0, "Node maps changed -- should restart\n");
 961
 962         return ret;
 963 }
 964
 965 static int dlm_try_to_join_domain(struct dlm_ctxt *dlm)
 966 {
 967         int status = 0, tmpstat, node;
 968         struct domain_join_ctxt *ctxt;
 969         enum dlm_query_join_response response;
 970
 971         mlog_entry("%p", dlm);
 972
 973         ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL);
 974         if (!ctxt) {
 975                 status = -ENOMEM;
 976                 mlog_errno(status);
 977                 goto bail;
 978         }
 979
 980         /* group sem locking should work for us here -- we're already
 981          * registered for heartbeat events so filling this should be
 982          * atomic wrt getting those handlers called. */
 983         o2hb_fill_node_map(dlm->live_nodes_map, sizeof(dlm->live_nodes_map));
 984
 985         spin_lock(&dlm->spinlock);
 986         memcpy(ctxt->live_map, dlm->live_nodes_map, sizeof(ctxt->live_map));
 987
 988         __dlm_set_joining_node(dlm, dlm->node_num);
 989
 990         spin_unlock(&dlm->spinlock);
 991
 992         node = -1;
 993         while ((node = find_next_bit(ctxt->live_map, O2NM_MAX_NODES,
 994                                      node + 1)) < O2NM_MAX_NODES) {
 995                 if (node == dlm->node_num)
 996                         continue;
 997
 998                 status = dlm_request_join(dlm, node, &response);
 999                 if (status < 0) {
1000                         mlog_errno(status);
1001                         goto bail;
1002                 }
1003
1004                 /* Ok, either we got a response or the node doesn't have a
1005                  * dlm up. */
1006                 if (response == JOIN_OK)
1007                         set_bit(node, ctxt->yes_resp_map);
1008
1009                 if (dlm_should_restart_join(dlm, ctxt, response)) {
1010                         status = -EAGAIN;
1011                         goto bail;
1012                 }
1013         }
1014
1015         mlog(0, "Yay, done querying nodes!\n");
1016
1017         /* Yay, everyone agree's we can join the domain. My domain is
1018          * comprised of all nodes who were put in the
1019          * yes_resp_map. Copy that into our domain map and send a join
1020          * assert message to clean up everyone elses state. */
1021         spin_lock(&dlm->spinlock);
1022         memcpy(dlm->domain_map, ctxt->yes_resp_map,
1023                sizeof(ctxt->yes_resp_map));
1024         set_bit(dlm->node_num, dlm->domain_map);
1025         spin_unlock(&dlm->spinlock);
1026
1027         dlm_send_join_asserts(dlm, ctxt->yes_resp_map);
1028
1029         /* Joined state *must* be set before the joining node
1030          * information, otherwise the query_join handler may read no
1031          * current joiner but a state of NEW and tell joining nodes
1032          * we're not in the domain. */
1033         spin_lock(&dlm_domain_lock);
1034         dlm->dlm_state = DLM_CTXT_JOINED;
1035         dlm->num_joins++;
1036         spin_unlock(&dlm_domain_lock);
1037
1038 bail:
1039         spin_lock(&dlm->spinlock);
1040         __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
1041         if (!status)
1042                 __dlm_print_nodes(dlm);
1043         spin_unlock(&dlm->spinlock);
1044
1045         if (ctxt) {
1046                 /* Do we need to send a cancel message to any nodes? */
1047                 if (status < 0) {
1048                         tmpstat = dlm_send_join_cancels(dlm,
1049                                                         ctxt->yes_resp_map,
1050                                                         sizeof(ctxt->yes_resp_map));
1051                         if (tmpstat < 0)
1052                                 mlog_errno(tmpstat);
1053                 }
1054                 kfree(ctxt);
1055         }
1056
1057         mlog(0, "returning %d\n", status);
1058         return status;
1059 }
1060
1061 static void dlm_unregister_domain_handlers(struct dlm_ctxt *dlm)
1062 {
1063         o2hb_unregister_callback(&dlm->dlm_hb_up);
1064         o2hb_unregister_callback(&dlm->dlm_hb_down);
1065         o2net_unregister_handler_list(&dlm->dlm_domain_handlers);
1066 }
1067
1068 static int dlm_register_domain_handlers(struct dlm_ctxt *dlm)
1069 {
1070         int status;
1071
1072         mlog(0, "registering handlers.\n");
1073
1074         o2hb_setup_callback(&dlm->dlm_hb_down, O2HB_NODE_DOWN_CB,
1075                             dlm_hb_node_down_cb, dlm, DLM_HB_NODE_DOWN_PRI);
1076         status = o2hb_register_callback(&dlm->dlm_hb_down);
1077         if (status)
1078                 goto bail;
1079
1080         o2hb_setup_callback(&dlm->dlm_hb_up, O2HB_NODE_UP_CB,
1081                             dlm_hb_node_up_cb, dlm, DLM_HB_NODE_UP_PRI);
1082         status = o2hb_register_callback(&dlm->dlm_hb_up);
1083         if (status)
1084                 goto bail;
1085
1086         status = o2net_register_handler(DLM_MASTER_REQUEST_MSG, dlm->key,
1087                                         sizeof(struct dlm_master_request),
1088                                         dlm_master_request_handler,
1089                                         dlm, &dlm->dlm_domain_handlers);
1090         if (status)
1091                 goto bail;
1092
1093         status = o2net_register_handler(DLM_ASSERT_MASTER_MSG, dlm->key,
1094                                         sizeof(struct dlm_assert_master),
1095                                         dlm_assert_master_handler,
1096                                         dlm, &dlm->dlm_domain_handlers);
1097         if (status)
1098                 goto bail;
1099
1100         status = o2net_register_handler(DLM_CREATE_LOCK_MSG, dlm->key,
1101                                         sizeof(struct dlm_create_lock),
1102                                         dlm_create_lock_handler,
1103                                         dlm, &dlm->dlm_domain_handlers);
1104         if (status)
1105                 goto bail;
1106
1107         status = o2net_register_handler(DLM_CONVERT_LOCK_MSG, dlm->key,
1108                                         DLM_CONVERT_LOCK_MAX_LEN,
1109                                         dlm_convert_lock_handler,
1110                                         dlm, &dlm->dlm_domain_handlers);
1111         if (status)
1112                 goto bail;
1113
1114         status = o2net_register_handler(DLM_UNLOCK_LOCK_MSG, dlm->key,
1115                                         DLM_UNLOCK_LOCK_MAX_LEN,
1116                                         dlm_unlock_lock_handler,
1117                                         dlm, &dlm->dlm_domain_handlers);
1118         if (status)
1119                 goto bail;
1120
1121         status = o2net_register_handler(DLM_PROXY_AST_MSG, dlm->key,
1122                                         DLM_PROXY_AST_MAX_LEN,
1123                                         dlm_proxy_ast_handler,
1124                                         dlm, &dlm->dlm_domain_handlers);
1125         if (status)
1126                 goto bail;
1127
1128         status = o2net_register_handler(DLM_EXIT_DOMAIN_MSG, dlm->key,
1129                                         sizeof(struct dlm_exit_domain),
1130                                         dlm_exit_domain_handler,
1131                                         dlm, &dlm->dlm_domain_handlers);
1132         if (status)
1133                 goto bail;
1134
1135         status = o2net_register_handler(DLM_DEREF_LOCKRES_MSG, dlm->key,
1136                                         sizeof(struct dlm_deref_lockres),
1137                                         dlm_deref_lockres_handler,
1138                                         dlm, &dlm->dlm_domain_handlers);
1139         if (status)
1140                 goto bail;
1141
1142         status = o2net_register_handler(DLM_MIGRATE_REQUEST_MSG, dlm->key,
1143                                         sizeof(struct dlm_migrate_request),
1144                                         dlm_migrate_request_handler,
1145                                         dlm, &dlm->dlm_domain_handlers);
1146         if (status)
1147                 goto bail;
1148
1149         status = o2net_register_handler(DLM_MIG_LOCKRES_MSG, dlm->key,
1150                                         DLM_MIG_LOCKRES_MAX_LEN,
1151                                         dlm_mig_lockres_handler,
1152                                         dlm, &dlm->dlm_domain_handlers);
1153         if (status)
1154                 goto bail;
1155
1156         status = o2net_register_handler(DLM_MASTER_REQUERY_MSG, dlm->key,
1157                                         sizeof(struct dlm_master_requery),
1158                                         dlm_master_requery_handler,
1159                                         dlm, &dlm->dlm_domain_handlers);
1160         if (status)
1161                 goto bail;
1162
1163         status = o2net_register_handler(DLM_LOCK_REQUEST_MSG, dlm->key,
1164                                         sizeof(struct dlm_lock_request),
1165                                         dlm_request_all_locks_handler,
1166                                         dlm, &dlm->dlm_domain_handlers);
1167         if (status)
1168                 goto bail;
1169
1170         status = o2net_register_handler(DLM_RECO_DATA_DONE_MSG, dlm->key,
1171                                         sizeof(struct dlm_reco_data_done),
1172                                         dlm_reco_data_done_handler,
1173                                         dlm, &dlm->dlm_domain_handlers);
1174         if (status)
1175                 goto bail;
1176
1177         status = o2net_register_handler(DLM_BEGIN_RECO_MSG, dlm->key,
1178                                         sizeof(struct dlm_begin_reco),
1179                                         dlm_begin_reco_handler,
1180                                         dlm, &dlm->dlm_domain_handlers);
1181         if (status)
1182                 goto bail;
1183
1184         status = o2net_register_handler(DLM_FINALIZE_RECO_MSG, dlm->key,
1185                                         sizeof(struct dlm_finalize_reco),
1186                                         dlm_finalize_reco_handler,
1187                                         dlm, &dlm->dlm_domain_handlers);
1188         if (status)
1189                 goto bail;
1190
1191 bail:
1192         if (status)
1193                 dlm_unregister_domain_handlers(dlm);
1194
1195         return status;
1196 }
1197
1198 static int dlm_join_domain(struct dlm_ctxt *dlm)
1199 {
1200         int status;
1201
1202         BUG_ON(!dlm);
1203
1204         mlog(0, "Join domain %s\n", dlm->name);
1205
1206         status = dlm_register_domain_handlers(dlm);
1207         if (status) {
1208                 mlog_errno(status);
1209                 goto bail;
1210         }
1211
1212         status = dlm_launch_thread(dlm);
1213         if (status < 0) {
1214                 mlog_errno(status);
1215                 goto bail;
1216         }
1217
1218         status = dlm_launch_recovery_thread(dlm);
1219         if (status < 0) {
1220                 mlog_errno(status);
1221                 goto bail;
1222         }
1223
1224         dlm->dlm_worker = create_singlethread_workqueue("dlm_wq");
1225         if (!dlm->dlm_worker) {
1226                 status = -ENOMEM;
1227                 mlog_errno(status);
1228                 goto bail;
1229         }
1230
1231         do {
1232                 unsigned int backoff;
1233                 status = dlm_try_to_join_domain(dlm);
1234
1235                 /* If we're racing another node to the join, then we
1236                  * need to back off temporarily and let them
1237                  * complete. */
1238                 if (status == -EAGAIN) {
1239                         if (signal_pending(current)) {
1240                                 status = -ERESTARTSYS;
1241                                 goto bail;
1242                         }
1243
1244                         /*
1245                          * <chip> After you!
1246                          * <dale> No, after you!
1247                          * <chip> I insist!
1248                          * <dale> But you first!
1249                          * ...
1250                          */
1251                         backoff = (unsigned int)(jiffies & 0x3);
1252                         backoff *= DLM_DOMAIN_BACKOFF_MS;
1253                         mlog(0, "backoff %d\n", backoff);
1254                         msleep(backoff);
1255                 }
1256         } while (status == -EAGAIN);
1257
1258         if (status < 0) {
1259                 mlog_errno(status);
1260                 goto bail;
1261         }
1262
1263         status = 0;
1264 bail:
1265         wake_up(&dlm_domain_events);
1266
1267         if (status) {
1268                 dlm_unregister_domain_handlers(dlm);
1269                 dlm_complete_thread(dlm);
1270                 dlm_complete_recovery_thread(dlm);
1271                 dlm_destroy_dlm_worker(dlm);
1272         }
1273
1274         return status;
1275 }
1276
1277 static struct dlm_ctxt *dlm_alloc_ctxt(const char *domain,
1278                                 u32 key)
1279 {
1280         int i;
1281         struct dlm_ctxt *dlm = NULL;
1282
1283         dlm = kzalloc(sizeof(*dlm), GFP_KERNEL);
1284         if (!dlm) {
1285                 mlog_errno(-ENOMEM);
1286                 goto leave;
1287         }
1288
1289         dlm->name = kmalloc(strlen(domain) + 1, GFP_KERNEL);
1290         if (dlm->name == NULL) {
1291                 mlog_errno(-ENOMEM);
1292                 kfree(dlm);
1293                 dlm = NULL;
1294                 goto leave;
1295         }
1296
1297         dlm->lockres_hash = (struct hlist_head **)dlm_alloc_pagevec(DLM_HASH_PAGES);
1298         if (!dlm->lockres_hash) {
1299                 mlog_errno(-ENOMEM);
1300                 kfree(dlm->name);
1301                 kfree(dlm);
1302                 dlm = NULL;
1303                 goto leave;
1304         }
1305
1306         for (i = 0; i < DLM_HASH_BUCKETS; i++)
1307                 INIT_HLIST_HEAD(dlm_lockres_hash(dlm, i));
1308
1309         strcpy(dlm->name, domain);
1310         dlm->key = key;
1311         dlm->node_num = o2nm_this_node();
1312
1313         spin_lock_init(&dlm->spinlock);
1314         spin_lock_init(&dlm->master_lock);
1315         spin_lock_init(&dlm->ast_lock);
1316         INIT_LIST_HEAD(&dlm->list);
1317         INIT_LIST_HEAD(&dlm->dirty_list);
1318         INIT_LIST_HEAD(&dlm->reco.resources);
1319         INIT_LIST_HEAD(&dlm->reco.received);
1320         INIT_LIST_HEAD(&dlm->reco.node_data);
1321         INIT_LIST_HEAD(&dlm->purge_list);
1322         INIT_LIST_HEAD(&dlm->dlm_domain_handlers);
1323         dlm->reco.state = 0;
1324
1325         INIT_LIST_HEAD(&dlm->pending_asts);
1326         INIT_LIST_HEAD(&dlm->pending_basts);
1327
1328         mlog(0, "dlm->recovery_map=%p, &(dlm->recovery_map[0])=%p\n",
1329                   dlm->recovery_map, &(dlm->recovery_map[0]));
1330
1331         memset(dlm->recovery_map, 0, sizeof(dlm->recovery_map));
1332         memset(dlm->live_nodes_map, 0, sizeof(dlm->live_nodes_map));
1333         memset(dlm->domain_map, 0, sizeof(dlm->domain_map));
1334
1335         dlm->dlm_thread_task = NULL;
1336         dlm->dlm_reco_thread_task = NULL;
1337         dlm->dlm_worker = NULL;
1338         init_waitqueue_head(&dlm->dlm_thread_wq);
1339         init_waitqueue_head(&dlm->dlm_reco_thread_wq);
1340         init_waitqueue_head(&dlm->reco.event);
1341         init_waitqueue_head(&dlm->ast_wq);
1342         init_waitqueue_head(&dlm->migration_wq);
1343         INIT_LIST_HEAD(&dlm->master_list);
1344         INIT_LIST_HEAD(&dlm->mle_hb_events);
1345
1346         dlm->joining_node = DLM_LOCK_RES_OWNER_UNKNOWN;
1347         init_waitqueue_head(&dlm->dlm_join_events);
1348
1349         dlm->reco.new_master = O2NM_INVALID_NODE_NUM;
1350         dlm->reco.dead_node = O2NM_INVALID_NODE_NUM;
1351         atomic_set(&dlm->local_resources, 0);
1352         atomic_set(&dlm->remote_resources, 0);
1353         atomic_set(&dlm->unknown_resources, 0);
1354
1355         spin_lock_init(&dlm->work_lock);
1356         INIT_LIST_HEAD(&dlm->work_list);
1357         INIT_WORK(&dlm->dispatched_work, dlm_dispatch_work);
1358
1359         kref_init(&dlm->dlm_refs);
1360         dlm->dlm_state = DLM_CTXT_NEW;
1361
1362         INIT_LIST_HEAD(&dlm->dlm_eviction_callbacks);
1363
1364         mlog(0, "context init: refcount %u\n",
1365                   atomic_read(&dlm->dlm_refs.refcount));
1366
1367 leave:
1368         return dlm;
1369 }
1370
1371 /*
1372  * dlm_register_domain: one-time setup per "domain"
1373  */
1374 struct dlm_ctxt * dlm_register_domain(const char *domain,
1375                                u32 key)
1376 {
1377         int ret;
1378         struct dlm_ctxt *dlm = NULL;
1379         struct dlm_ctxt *new_ctxt = NULL;
1380
1381         if (strlen(domain) > O2NM_MAX_NAME_LEN) {
1382                 ret = -ENAMETOOLONG;
1383                 mlog(ML_ERROR, "domain name length too long\n");
1384                 goto leave;
1385         }
1386
1387         if (!o2hb_check_local_node_heartbeating()) {
1388                 mlog(ML_ERROR, "the local node has not been configured, or is "
1389                      "not heartbeating\n");
1390                 ret = -EPROTO;
1391                 goto leave;
1392         }
1393
1394         mlog(0, "register called for domain \"%s\"\n", domain);
1395
1396 retry:
1397         dlm = NULL;
1398         if (signal_pending(current)) {
1399                 ret = -ERESTARTSYS;
1400                 mlog_errno(ret);
1401                 goto leave;
1402         }
1403
1404         spin_lock(&dlm_domain_lock);
1405
1406         dlm = __dlm_lookup_domain(domain);
1407         if (dlm) {
1408                 if (dlm->dlm_state != DLM_CTXT_JOINED) {
1409                         spin_unlock(&dlm_domain_lock);
1410
1411                         mlog(0, "This ctxt is not joined yet!\n");
1412                         wait_event_interruptible(dlm_domain_events,
1413                                                  dlm_wait_on_domain_helper(
1414                                                          domain));
1415                         goto retry;
1416                 }
1417
1418                 __dlm_get(dlm);
1419                 dlm->num_joins++;
1420
1421                 spin_unlock(&dlm_domain_lock);
1422
1423                 ret = 0;
1424                 goto leave;
1425         }
1426
1427         /* doesn't exist */
1428         if (!new_ctxt) {
1429                 spin_unlock(&dlm_domain_lock);
1430
1431                 new_ctxt = dlm_alloc_ctxt(domain, key);
1432                 if (new_ctxt)
1433                         goto retry;
1434
1435                 ret = -ENOMEM;
1436                 mlog_errno(ret);
1437                 goto leave;
1438         }
1439
1440         /* a little variable switch-a-roo here... */
1441         dlm = new_ctxt;
1442         new_ctxt = NULL;
1443
1444         /* add the new domain */
1445         list_add_tail(&dlm->list, &dlm_domains);
1446         spin_unlock(&dlm_domain_lock);
1447
1448         ret = dlm_join_domain(dlm);
1449         if (ret) {
1450                 mlog_errno(ret);
1451                 dlm_put(dlm);
1452                 goto leave;
1453         }
1454
1455         ret = 0;
1456 leave:
1457         if (new_ctxt)
1458                 dlm_free_ctxt_mem(new_ctxt);
1459
1460         if (ret < 0)
1461                 dlm = ERR_PTR(ret);
1462
1463         return dlm;
1464 }
1465 EXPORT_SYMBOL_GPL(dlm_register_domain);
1466
1467 static LIST_HEAD(dlm_join_handlers);
1468
1469 static void dlm_unregister_net_handlers(void)
1470 {
1471         o2net_unregister_handler_list(&dlm_join_handlers);
1472 }
1473
1474 static int dlm_register_net_handlers(void)
1475 {
1476         int status = 0;
1477
1478         status = o2net_register_handler(DLM_QUERY_JOIN_MSG, DLM_MOD_KEY,
1479                                         sizeof(struct dlm_query_join_request),
1480                                         dlm_query_join_handler,
1481                                         NULL, &dlm_join_handlers);
1482         if (status)
1483                 goto bail;
1484
1485         status = o2net_register_handler(DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY,
1486                                         sizeof(struct dlm_assert_joined),
1487                                         dlm_assert_joined_handler,
1488                                         NULL, &dlm_join_handlers);
1489         if (status)
1490                 goto bail;
1491
1492         status = o2net_register_handler(DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY,
1493                                         sizeof(struct dlm_cancel_join),
1494                                         dlm_cancel_join_handler,
1495                                         NULL, &dlm_join_handlers);
1496
1497 bail:
1498         if (status < 0)
1499                 dlm_unregister_net_handlers();
1500
1501         return status;
1502 }
1503
1504 /* Domain eviction callback handling.
1505  *
1506  * The file system requires notification of node death *before* the
1507  * dlm completes it's recovery work, otherwise it may be able to
1508  * acquire locks on resources requiring recovery. Since the dlm can
1509  * evict a node from it's domain *before* heartbeat fires, a similar
1510  * mechanism is required. */
1511
1512 /* Eviction is not expected to happen often, so a per-domain lock is
1513  * not necessary. Eviction callbacks are allowed to sleep for short
1514  * periods of time. */
1515 static DECLARE_RWSEM(dlm_callback_sem);
1516
1517 void dlm_fire_domain_eviction_callbacks(struct dlm_ctxt *dlm,
1518                                         int node_num)
1519 {
1520         struct list_head *iter;
1521         struct dlm_eviction_cb *cb;
1522
1523         down_read(&dlm_callback_sem);
1524         list_for_each(iter, &dlm->dlm_eviction_callbacks) {
1525                 cb = list_entry(iter, struct dlm_eviction_cb, ec_item);
1526
1527                 cb->ec_func(node_num, cb->ec_data);
1528         }
1529         up_read(&dlm_callback_sem);
1530 }
1531
1532 void dlm_setup_eviction_cb(struct dlm_eviction_cb *cb,
1533                            dlm_eviction_func *f,
1534                            void *data)
1535 {
1536         INIT_LIST_HEAD(&cb->ec_item);
1537         cb->ec_func = f;
1538         cb->ec_data = data;
1539 }
1540 EXPORT_SYMBOL_GPL(dlm_setup_eviction_cb);
1541
1542 void dlm_register_eviction_cb(struct dlm_ctxt *dlm,
1543                               struct dlm_eviction_cb *cb)
1544 {
1545         down_write(&dlm_callback_sem);
1546         list_add_tail(&cb->ec_item, &dlm->dlm_eviction_callbacks);
1547         up_write(&dlm_callback_sem);
1548 }
1549 EXPORT_SYMBOL_GPL(dlm_register_eviction_cb);
1550
1551 void dlm_unregister_eviction_cb(struct dlm_eviction_cb *cb)
1552 {
1553         down_write(&dlm_callback_sem);
1554         list_del_init(&cb->ec_item);
1555         up_write(&dlm_callback_sem);
1556 }
1557 EXPORT_SYMBOL_GPL(dlm_unregister_eviction_cb);
1558
1559 static int __init dlm_init(void)
1560 {
1561         int status;
1562
1563         dlm_print_version();
1564
1565         status = dlm_init_mle_cache();
1566         if (status)
1567                 return -1;
1568
1569         status = dlm_register_net_handlers();
1570         if (status) {
1571                 dlm_destroy_mle_cache();
1572                 return -1;
1573         }
1574
1575         return 0;
1576 }
1577
1578 static void __exit dlm_exit (void)
1579 {
1580         dlm_unregister_net_handlers();
1581         dlm_destroy_mle_cache();
1582 }
1583
1584 MODULE_AUTHOR("Oracle");
1585 MODULE_LICENSE("GPL");
1586
1587 module_init(dlm_init);
1588 module_exit(dlm_exit);