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/init.h>
  32 #include <linux/spinlock.h>
  33 #include <linux/delay.h>
  34 #include <linux/err.h>
  35 #include <linux/debugfs.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 #include "dlmdomain.h"
  44 #include "dlmdebug.h"
  45
  46 #include "dlmver.h"
  47
  48 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_DOMAIN)
  49 #include "cluster/masklog.h"
  50
  51 /*
  52  * ocfs2 node maps are array of long int, which limits to send them freely
  53  * across the wire due to endianness issues. To workaround this, we convert
  54  * long ints to byte arrays. Following 3 routines are helper functions to
  55  * set/test/copy bits within those array of bytes
  56  */
  57 static inline void byte_set_bit(u8 nr, u8 map[])
  58 {
  59         map[nr >> 3] |= (1UL << (nr & 7));
  60 }
  61
  62 static inline int byte_test_bit(u8 nr, u8 map[])
  63 {
  64         return ((1UL << (nr & 7)) & (map[nr >> 3])) != 0;
  65 }
  66
  67 static inline void byte_copymap(u8 dmap[], unsigned long smap[],
  68                         unsigned int sz)
  69 {
  70         unsigned int nn;
  71
  72         if (!sz)
  73                 return;
  74
  75         memset(dmap, 0, ((sz + 7) >> 3));
  76         for (nn = 0 ; nn < sz; nn++)
  77                 if (test_bit(nn, smap))
  78                         byte_set_bit(nn, dmap);
  79 }
  80
  81 static void dlm_free_pagevec(void **vec, int pages)
  82 {
  83         while (pages--)
  84                 free_page((unsigned long)vec[pages]);
  85         kfree(vec);
  86 }
  87
  88 static void **dlm_alloc_pagevec(int pages)
  89 {
  90         void **vec = kmalloc(pages * sizeof(void *), GFP_KERNEL);
  91         int i;
  92
  93         if (!vec)
  94                 return NULL;
  95
  96         for (i = 0; i < pages; i++)
  97                 if (!(vec[i] = (void *)__get_free_page(GFP_KERNEL)))
  98                         goto out_free;
  99
 100         mlog(0, "Allocated DLM hash pagevec; %d pages (%lu expected), %lu buckets per page\n",
 101              pages, (unsigned long)DLM_HASH_PAGES,
 102              (unsigned long)DLM_BUCKETS_PER_PAGE);
 103         return vec;
 104 out_free:
 105         dlm_free_pagevec(vec, i);
 106         return NULL;
 107 }
 108
 109 /*
 110  *
 111  * spinlock lock ordering: if multiple locks are needed, obey this ordering:
 112  *    dlm_domain_lock
 113  *    struct dlm_ctxt->spinlock
 114  *    struct dlm_lock_resource->spinlock
 115  *    struct dlm_ctxt->master_lock
 116  *    struct dlm_ctxt->ast_lock
 117  *    dlm_master_list_entry->spinlock
 118  *    dlm_lock->spinlock
 119  *
 120  */
 121
 122 DEFINE_SPINLOCK(dlm_domain_lock);
 123 LIST_HEAD(dlm_domains);
 124 static DECLARE_WAIT_QUEUE_HEAD(dlm_domain_events);
 125
 126 /*
 127  * The supported protocol version for DLM communication.  Running domains
 128  * will have a negotiated version with the same major number and a minor
 129  * number equal or smaller.  The dlm_ctxt->dlm_locking_proto field should
 130  * be used to determine what a running domain is actually using.
 131  */
 132 static const struct dlm_protocol_version dlm_protocol = {
 133         .pv_major = 1,
 134         .pv_minor = 0,
 135 };
 136
 137 #define DLM_DOMAIN_BACKOFF_MS 200
 138
 139 static int dlm_query_join_handler(struct o2net_msg *msg, u32 len, void *data,
 140                                   void **ret_data);
 141 static int dlm_assert_joined_handler(struct o2net_msg *msg, u32 len, void *data,
 142                                      void **ret_data);
 143 static int dlm_cancel_join_handler(struct o2net_msg *msg, u32 len, void *data,
 144                                    void **ret_data);
 145 static int dlm_exit_domain_handler(struct o2net_msg *msg, u32 len, void *data,
 146                                    void **ret_data);
 147 static int dlm_protocol_compare(struct dlm_protocol_version *existing,
 148                                 struct dlm_protocol_version *request);
 149
 150 static void dlm_unregister_domain_handlers(struct dlm_ctxt *dlm);
 151
 152 void __dlm_unhash_lockres(struct dlm_lock_resource *lockres)
 153 {
 154         if (!hlist_unhashed(&lockres->hash_node)) {
 155                 hlist_del_init(&lockres->hash_node);
 156                 dlm_lockres_put(lockres);
 157         }
 158 }
 159
 160 void __dlm_insert_lockres(struct dlm_ctxt *dlm,
 161                        struct dlm_lock_resource *res)
 162 {
 163         struct hlist_head *bucket;
 164         struct qstr *q;
 165
 166         assert_spin_locked(&dlm->spinlock);
 167
 168         q = &res->lockname;
 169         bucket = dlm_lockres_hash(dlm, q->hash);
 170
 171         /* get a reference for our hashtable */
 172         dlm_lockres_get(res);
 173
 174         hlist_add_head(&res->hash_node, bucket);
 175 }
 176
 177 struct dlm_lock_resource * __dlm_lookup_lockres_full(struct dlm_ctxt *dlm,
 178                                                      const char *name,
 179                                                      unsigned int len,
 180                                                      unsigned int hash)
 181 {
 182         struct hlist_head *bucket;
 183         struct hlist_node *list;
 184
 185         mlog_entry("%.*s\n", len, name);
 186
 187         assert_spin_locked(&dlm->spinlock);
 188
 189         bucket = dlm_lockres_hash(dlm, hash);
 190
 191         hlist_for_each(list, bucket) {
 192                 struct dlm_lock_resource *res = hlist_entry(list,
 193                         struct dlm_lock_resource, hash_node);
 194                 if (res->lockname.name[0] != name[0])
 195                         continue;
 196                 if (unlikely(res->lockname.len != len))
 197                         continue;
 198                 if (memcmp(res->lockname.name + 1, name + 1, len - 1))
 199                         continue;
 200                 dlm_lockres_get(res);
 201                 return res;
 202         }
 203         return NULL;
 204 }
 205
 206 /* intended to be called by functions which do not care about lock
 207  * resources which are being purged (most net _handler functions).
 208  * this will return NULL for any lock resource which is found but
 209  * currently in the process of dropping its mastery reference.
 210  * use __dlm_lookup_lockres_full when you need the lock resource
 211  * regardless (e.g. dlm_get_lock_resource) */
 212 struct dlm_lock_resource * __dlm_lookup_lockres(struct dlm_ctxt *dlm,
 213                                                 const char *name,
 214                                                 unsigned int len,
 215                                                 unsigned int hash)
 216 {
 217         struct dlm_lock_resource *res = NULL;
 218
 219         mlog_entry("%.*s\n", len, name);
 220
 221         assert_spin_locked(&dlm->spinlock);
 222
 223         res = __dlm_lookup_lockres_full(dlm, name, len, hash);
 224         if (res) {
 225                 spin_lock(&res->spinlock);
 226                 if (res->state & DLM_LOCK_RES_DROPPING_REF) {
 227                         spin_unlock(&res->spinlock);
 228                         dlm_lockres_put(res);
 229                         return NULL;
 230                 }
 231                 spin_unlock(&res->spinlock);
 232         }
 233
 234         return res;
 235 }
 236
 237 struct dlm_lock_resource * dlm_lookup_lockres(struct dlm_ctxt *dlm,
 238                                     const char *name,
 239                                     unsigned int len)
 240 {
 241         struct dlm_lock_resource *res;
 242         unsigned int hash = dlm_lockid_hash(name, len);
 243
 244         spin_lock(&dlm->spinlock);
 245         res = __dlm_lookup_lockres(dlm, name, len, hash);
 246         spin_unlock(&dlm->spinlock);
 247         return res;
 248 }
 249
 250 static struct dlm_ctxt * __dlm_lookup_domain_full(const char *domain, int len)
 251 {
 252         struct dlm_ctxt *tmp = NULL;
 253         struct list_head *iter;
 254
 255         assert_spin_locked(&dlm_domain_lock);
 256
 257         /* tmp->name here is always NULL terminated,
 258          * but domain may not be! */
 259         list_for_each(iter, &dlm_domains) {
 260                 tmp = list_entry (iter, struct dlm_ctxt, list);
 261                 if (strlen(tmp->name) == len &&
 262                     memcmp(tmp->name, domain, len)==0)
 263                         break;
 264                 tmp = NULL;
 265         }
 266
 267         return tmp;
 268 }
 269
 270 /* For null terminated domain strings ONLY */
 271 static struct dlm_ctxt * __dlm_lookup_domain(const char *domain)
 272 {
 273         assert_spin_locked(&dlm_domain_lock);
 274
 275         return __dlm_lookup_domain_full(domain, strlen(domain));
 276 }
 277
 278
 279 /* returns true on one of two conditions:
 280  * 1) the domain does not exist
 281  * 2) the domain exists and it's state is "joined" */
 282 static int dlm_wait_on_domain_helper(const char *domain)
 283 {
 284         int ret = 0;
 285         struct dlm_ctxt *tmp = NULL;
 286
 287         spin_lock(&dlm_domain_lock);
 288
 289         tmp = __dlm_lookup_domain(domain);
 290         if (!tmp)
 291                 ret = 1;
 292         else if (tmp->dlm_state == DLM_CTXT_JOINED)
 293                 ret = 1;
 294
 295         spin_unlock(&dlm_domain_lock);
 296         return ret;
 297 }
 298
 299 static void dlm_free_ctxt_mem(struct dlm_ctxt *dlm)
 300 {
 301         dlm_destroy_debugfs_subroot(dlm);
 302
 303         if (dlm->lockres_hash)
 304                 dlm_free_pagevec((void **)dlm->lockres_hash, DLM_HASH_PAGES);
 305
 306         if (dlm->master_hash)
 307                 dlm_free_pagevec((void **)dlm->master_hash, DLM_HASH_PAGES);
 308
 309         if (dlm->name)
 310                 kfree(dlm->name);
 311
 312         kfree(dlm);
 313 }
 314
 315 /* A little strange - this function will be called while holding
 316  * dlm_domain_lock and is expected to be holding it on the way out. We
 317  * will however drop and reacquire it multiple times */
 318 static void dlm_ctxt_release(struct kref *kref)
 319 {
 320         struct dlm_ctxt *dlm;
 321
 322         dlm = container_of(kref, struct dlm_ctxt, dlm_refs);
 323
 324         BUG_ON(dlm->num_joins);
 325         BUG_ON(dlm->dlm_state == DLM_CTXT_JOINED);
 326
 327         /* we may still be in the list if we hit an error during join. */
 328         list_del_init(&dlm->list);
 329
 330         spin_unlock(&dlm_domain_lock);
 331
 332         mlog(0, "freeing memory from domain %s\n", dlm->name);
 333
 334         wake_up(&dlm_domain_events);
 335
 336         dlm_free_ctxt_mem(dlm);
 337
 338         spin_lock(&dlm_domain_lock);
 339 }
 340
 341 void dlm_put(struct dlm_ctxt *dlm)
 342 {
 343         spin_lock(&dlm_domain_lock);
 344         kref_put(&dlm->dlm_refs, dlm_ctxt_release);
 345         spin_unlock(&dlm_domain_lock);
 346 }
 347
 348 static void __dlm_get(struct dlm_ctxt *dlm)
 349 {
 350         kref_get(&dlm->dlm_refs);
 351 }
 352
 353 /* given a questionable reference to a dlm object, gets a reference if
 354  * it can find it in the list, otherwise returns NULL in which case
 355  * you shouldn't trust your pointer. */
 356 struct dlm_ctxt *dlm_grab(struct dlm_ctxt *dlm)
 357 {
 358         struct list_head *iter;
 359         struct dlm_ctxt *target = NULL;
 360
 361         spin_lock(&dlm_domain_lock);
 362
 363         list_for_each(iter, &dlm_domains) {
 364                 target = list_entry (iter, struct dlm_ctxt, list);
 365
 366                 if (target == dlm) {
 367                         __dlm_get(target);
 368                         break;
 369                 }
 370
 371                 target = NULL;
 372         }
 373
 374         spin_unlock(&dlm_domain_lock);
 375
 376         return target;
 377 }
 378
 379 int dlm_domain_fully_joined(struct dlm_ctxt *dlm)
 380 {
 381         int ret;
 382
 383         spin_lock(&dlm_domain_lock);
 384         ret = (dlm->dlm_state == DLM_CTXT_JOINED) ||
 385                 (dlm->dlm_state == DLM_CTXT_IN_SHUTDOWN);
 386         spin_unlock(&dlm_domain_lock);
 387
 388         return ret;
 389 }
 390
 391 static void dlm_destroy_dlm_worker(struct dlm_ctxt *dlm)
 392 {
 393         if (dlm->dlm_worker) {
 394                 flush_workqueue(dlm->dlm_worker);
 395                 destroy_workqueue(dlm->dlm_worker);
 396                 dlm->dlm_worker = NULL;
 397         }
 398 }
 399
 400 static void dlm_complete_dlm_shutdown(struct dlm_ctxt *dlm)
 401 {
 402         dlm_unregister_domain_handlers(dlm);
 403         dlm_debug_shutdown(dlm);
 404         dlm_complete_thread(dlm);
 405         dlm_complete_recovery_thread(dlm);
 406         dlm_destroy_dlm_worker(dlm);
 407
 408         /* We've left the domain. Now we can take ourselves out of the
 409          * list and allow the kref stuff to help us free the
 410          * memory. */
 411         spin_lock(&dlm_domain_lock);
 412         list_del_init(&dlm->list);
 413         spin_unlock(&dlm_domain_lock);
 414
 415         /* Wake up anyone waiting for us to remove this domain */
 416         wake_up(&dlm_domain_events);
 417 }
 418
 419 static int dlm_migrate_all_locks(struct dlm_ctxt *dlm)
 420 {
 421         int i, num, n, ret = 0;
 422         struct dlm_lock_resource *res;
 423         struct hlist_node *iter;
 424         struct hlist_head *bucket;
 425         int dropped;
 426
 427         mlog(0, "Migrating locks from domain %s\n", dlm->name);
 428
 429         num = 0;
 430         spin_lock(&dlm->spinlock);
 431         for (i = 0; i < DLM_HASH_BUCKETS; i++) {
 432 redo_bucket:
 433                 n = 0;
 434                 bucket = dlm_lockres_hash(dlm, i);
 435                 iter = bucket->first;
 436                 while (iter) {
 437                         n++;
 438                         res = hlist_entry(iter, struct dlm_lock_resource,
 439                                           hash_node);
 440                         dlm_lockres_get(res);
 441                         /* migrate, if necessary.  this will drop the dlm
 442                          * spinlock and retake it if it does migration. */
 443                         dropped = dlm_empty_lockres(dlm, res);
 444
 445                         spin_lock(&res->spinlock);
 446                         __dlm_lockres_calc_usage(dlm, res);
 447                         iter = res->hash_node.next;
 448                         spin_unlock(&res->spinlock);
 449
 450                         dlm_lockres_put(res);
 451
 452                         if (dropped)
 453                                 goto redo_bucket;
 454                 }
 455                 cond_resched_lock(&dlm->spinlock);
 456                 num += n;
 457                 mlog(0, "%s: touched %d lockreses in bucket %d "
 458                      "(tot=%d)\n", dlm->name, n, i, num);
 459         }
 460         spin_unlock(&dlm->spinlock);
 461         wake_up(&dlm->dlm_thread_wq);
 462
 463         /* let the dlm thread take care of purging, keep scanning until
 464          * nothing remains in the hash */
 465         if (num) {
 466                 mlog(0, "%s: %d lock resources in hash last pass\n",
 467                      dlm->name, num);
 468                 ret = -EAGAIN;
 469         }
 470         mlog(0, "DONE Migrating locks from domain %s\n", dlm->name);
 471         return ret;
 472 }
 473
 474 static int dlm_no_joining_node(struct dlm_ctxt *dlm)
 475 {
 476         int ret;
 477
 478         spin_lock(&dlm->spinlock);
 479         ret = dlm->joining_node == DLM_LOCK_RES_OWNER_UNKNOWN;
 480         spin_unlock(&dlm->spinlock);
 481
 482         return ret;
 483 }
 484
 485 static void dlm_mark_domain_leaving(struct dlm_ctxt *dlm)
 486 {
 487         /* Yikes, a double spinlock! I need domain_lock for the dlm
 488          * state and the dlm spinlock for join state... Sorry! */
 489 again:
 490         spin_lock(&dlm_domain_lock);
 491         spin_lock(&dlm->spinlock);
 492
 493         if (dlm->joining_node != DLM_LOCK_RES_OWNER_UNKNOWN) {
 494                 mlog(0, "Node %d is joining, we wait on it.\n",
 495                           dlm->joining_node);
 496                 spin_unlock(&dlm->spinlock);
 497                 spin_unlock(&dlm_domain_lock);
 498
 499                 wait_event(dlm->dlm_join_events, dlm_no_joining_node(dlm));
 500                 goto again;
 501         }
 502
 503         dlm->dlm_state = DLM_CTXT_LEAVING;
 504         spin_unlock(&dlm->spinlock);
 505         spin_unlock(&dlm_domain_lock);
 506 }
 507
 508 static void __dlm_print_nodes(struct dlm_ctxt *dlm)
 509 {
 510         int node = -1;
 511
 512         assert_spin_locked(&dlm->spinlock);
 513
 514         printk(KERN_NOTICE "o2dlm: Nodes in domain %s: ", dlm->name);
 515
 516         while ((node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES,
 517                                      node + 1)) < O2NM_MAX_NODES) {
 518                 printk("%d ", node);
 519         }
 520         printk("\n");
 521 }
 522
 523 static int dlm_exit_domain_handler(struct o2net_msg *msg, u32 len, void *data,
 524                                    void **ret_data)
 525 {
 526         struct dlm_ctxt *dlm = data;
 527         unsigned int node;
 528         struct dlm_exit_domain *exit_msg = (struct dlm_exit_domain *) msg->buf;
 529
 530         mlog_entry("%p %u %p", msg, len, data);
 531
 532         if (!dlm_grab(dlm))
 533                 return 0;
 534
 535         node = exit_msg->node_idx;
 536
 537         printk(KERN_NOTICE "o2dlm: Node %u leaves domain %s\n", node, dlm->name);
 538
 539         spin_lock(&dlm->spinlock);
 540         clear_bit(node, dlm->domain_map);
 541         __dlm_print_nodes(dlm);
 542
 543         /* notify anything attached to the heartbeat events */
 544         dlm_hb_event_notify_attached(dlm, node, 0);
 545
 546         spin_unlock(&dlm->spinlock);
 547
 548         dlm_put(dlm);
 549
 550         return 0;
 551 }
 552
 553 static int dlm_send_one_domain_exit(struct dlm_ctxt *dlm,
 554                                     unsigned int node)
 555 {
 556         int status;
 557         struct dlm_exit_domain leave_msg;
 558
 559         mlog(0, "Asking node %u if we can leave the domain %s me = %u\n",
 560                   node, dlm->name, dlm->node_num);
 561
 562         memset(&leave_msg, 0, sizeof(leave_msg));
 563         leave_msg.node_idx = dlm->node_num;
 564
 565         status = o2net_send_message(DLM_EXIT_DOMAIN_MSG, dlm->key,
 566                                     &leave_msg, sizeof(leave_msg), node,
 567                                     NULL);
 568         if (status < 0)
 569                 mlog(ML_ERROR, "Error %d when sending message %u (key 0x%x) to "
 570                      "node %u\n", status, DLM_EXIT_DOMAIN_MSG, dlm->key, node);
 571         mlog(0, "status return %d from o2net_send_message\n", status);
 572
 573         return status;
 574 }
 575
 576
 577 static void dlm_leave_domain(struct dlm_ctxt *dlm)
 578 {
 579         int node, clear_node, status;
 580
 581         /* At this point we've migrated away all our locks and won't
 582          * accept mastership of new ones. The dlm is responsible for
 583          * almost nothing now. We make sure not to confuse any joining
 584          * nodes and then commence shutdown procedure. */
 585
 586         spin_lock(&dlm->spinlock);
 587         /* Clear ourselves from the domain map */
 588         clear_bit(dlm->node_num, dlm->domain_map);
 589         while ((node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES,
 590                                      0)) < O2NM_MAX_NODES) {
 591                 /* Drop the dlm spinlock. This is safe wrt the domain_map.
 592                  * -nodes cannot be added now as the
 593                  *   query_join_handlers knows to respond with OK_NO_MAP
 594                  * -we catch the right network errors if a node is
 595                  *   removed from the map while we're sending him the
 596                  *   exit message. */
 597                 spin_unlock(&dlm->spinlock);
 598
 599                 clear_node = 1;
 600
 601                 status = dlm_send_one_domain_exit(dlm, node);
 602                 if (status < 0 &&
 603                     status != -ENOPROTOOPT &&
 604                     status != -ENOTCONN) {
 605                         mlog(ML_NOTICE, "Error %d sending domain exit message "
 606                              "to node %d\n", status, node);
 607
 608                         /* Not sure what to do here but lets sleep for
 609                          * a bit in case this was a transient
 610                          * error... */
 611                         msleep(DLM_DOMAIN_BACKOFF_MS);
 612                         clear_node = 0;
 613                 }
 614
 615                 spin_lock(&dlm->spinlock);
 616                 /* If we're not clearing the node bit then we intend
 617                  * to loop back around to try again. */
 618                 if (clear_node)
 619                         clear_bit(node, dlm->domain_map);
 620         }
 621         spin_unlock(&dlm->spinlock);
 622 }
 623
 624 int dlm_joined(struct dlm_ctxt *dlm)
 625 {
 626         int ret = 0;
 627
 628         spin_lock(&dlm_domain_lock);
 629
 630         if (dlm->dlm_state == DLM_CTXT_JOINED)
 631                 ret = 1;
 632
 633         spin_unlock(&dlm_domain_lock);
 634
 635         return ret;
 636 }
 637
 638 int dlm_shutting_down(struct dlm_ctxt *dlm)
 639 {
 640         int ret = 0;
 641
 642         spin_lock(&dlm_domain_lock);
 643
 644         if (dlm->dlm_state == DLM_CTXT_IN_SHUTDOWN)
 645                 ret = 1;
 646
 647         spin_unlock(&dlm_domain_lock);
 648
 649         return ret;
 650 }
 651
 652 void dlm_unregister_domain(struct dlm_ctxt *dlm)
 653 {
 654         int leave = 0;
 655         struct dlm_lock_resource *res;
 656
 657         spin_lock(&dlm_domain_lock);
 658         BUG_ON(dlm->dlm_state != DLM_CTXT_JOINED);
 659         BUG_ON(!dlm->num_joins);
 660
 661         dlm->num_joins--;
 662         if (!dlm->num_joins) {
 663                 /* We mark it "in shutdown" now so new register
 664                  * requests wait until we've completely left the
 665                  * domain. Don't use DLM_CTXT_LEAVING yet as we still
 666                  * want new domain joins to communicate with us at
 667                  * least until we've completed migration of our
 668                  * resources. */
 669                 dlm->dlm_state = DLM_CTXT_IN_SHUTDOWN;
 670                 leave = 1;
 671         }
 672         spin_unlock(&dlm_domain_lock);
 673
 674         if (leave) {
 675                 mlog(0, "shutting down domain %s\n", dlm->name);
 676
 677                 /* We changed dlm state, notify the thread */
 678                 dlm_kick_thread(dlm, NULL);
 679
 680                 while (dlm_migrate_all_locks(dlm)) {
 681                         /* Give dlm_thread time to purge the lockres' */
 682                         msleep(500);
 683                         mlog(0, "%s: more migration to do\n", dlm->name);
 684                 }
 685
 686                 /* This list should be empty. If not, print remaining lockres */
 687                 if (!list_empty(&dlm->tracking_list)) {
 688                         mlog(ML_ERROR, "Following lockres' are still on the "
 689                              "tracking list:\n");
 690                         list_for_each_entry(res, &dlm->tracking_list, tracking)
 691                                 dlm_print_one_lock_resource(res);
 692                 }
 693
 694                 dlm_mark_domain_leaving(dlm);
 695                 dlm_leave_domain(dlm);
 696                 dlm_force_free_mles(dlm);
 697                 dlm_complete_dlm_shutdown(dlm);
 698         }
 699         dlm_put(dlm);
 700 }
 701 EXPORT_SYMBOL_GPL(dlm_unregister_domain);
 702
 703 static int dlm_query_join_proto_check(char *proto_type, int node,
 704                                       struct dlm_protocol_version *ours,
 705                                       struct dlm_protocol_version *request)
 706 {
 707         int rc;
 708         struct dlm_protocol_version proto = *request;
 709
 710         if (!dlm_protocol_compare(ours, &proto)) {
 711                 mlog(0,
 712                      "node %u wanted to join with %s locking protocol "
 713                      "%u.%u, we respond with %u.%u\n",
 714                      node, proto_type,
 715                      request->pv_major,
 716                      request->pv_minor,
 717                      proto.pv_major, proto.pv_minor);
 718                 request->pv_minor = proto.pv_minor;
 719                 rc = 0;
 720         } else {
 721                 mlog(ML_NOTICE,
 722                      "Node %u wanted to join with %s locking "
 723                      "protocol %u.%u, but we have %u.%u, disallowing\n",
 724                      node, proto_type,
 725                      request->pv_major,
 726                      request->pv_minor,
 727                      ours->pv_major,
 728                      ours->pv_minor);
 729                 rc = 1;
 730         }
 731
 732         return rc;
 733 }
 734
 735 /*
 736  * struct dlm_query_join_packet is made up of four one-byte fields.  They
 737  * are effectively in big-endian order already.  However, little-endian
 738  * machines swap them before putting the packet on the wire (because
 739  * query_join's response is a status, and that status is treated as a u32
 740  * on the wire).  Thus, a big-endian and little-endian machines will treat
 741  * this structure differently.
 742  *
 743  * The solution is to have little-endian machines swap the structure when
 744  * converting from the structure to the u32 representation.  This will
 745  * result in the structure having the correct format on the wire no matter
 746  * the host endian format.
 747  */
 748 static void dlm_query_join_packet_to_wire(struct dlm_query_join_packet *packet,
 749                                           u32 *wire)
 750 {
 751         union dlm_query_join_response response;
 752
 753         response.packet = *packet;
 754         *wire = cpu_to_be32(response.intval);
 755 }
 756
 757 static void dlm_query_join_wire_to_packet(u32 wire,
 758                                           struct dlm_query_join_packet *packet)
 759 {
 760         union dlm_query_join_response response;
 761
 762         response.intval = cpu_to_be32(wire);
 763         *packet = response.packet;
 764 }
 765
 766 static int dlm_query_join_handler(struct o2net_msg *msg, u32 len, void *data,
 767                                   void **ret_data)
 768 {
 769         struct dlm_query_join_request *query;
 770         struct dlm_query_join_packet packet = {
 771                 .code = JOIN_DISALLOW,
 772         };
 773         struct dlm_ctxt *dlm = NULL;
 774         u32 response;
 775         u8 nodenum;
 776
 777         query = (struct dlm_query_join_request *) msg->buf;
 778
 779         mlog(0, "node %u wants to join domain %s\n", query->node_idx,
 780                   query->domain);
 781
 782         /*
 783          * If heartbeat doesn't consider the node live, tell it
 784          * to back off and try again.  This gives heartbeat a chance
 785          * to catch up.
 786          */
 787         if (!o2hb_check_node_heartbeating(query->node_idx)) {
 788                 mlog(0, "node %u is not in our live map yet\n",
 789                      query->node_idx);
 790
 791                 packet.code = JOIN_DISALLOW;
 792                 goto respond;
 793         }
 794
 795         packet.code = JOIN_OK_NO_MAP;
 796
 797         spin_lock(&dlm_domain_lock);
 798         dlm = __dlm_lookup_domain_full(query->domain, query->name_len);
 799         if (!dlm)
 800                 goto unlock_respond;
 801
 802         /*
 803          * There is a small window where the joining node may not see the
 804          * node(s) that just left but still part of the cluster. DISALLOW
 805          * join request if joining node has different node map.
 806          */
 807         nodenum=0;
 808         while (nodenum < O2NM_MAX_NODES) {
 809                 if (test_bit(nodenum, dlm->domain_map)) {
 810                         if (!byte_test_bit(nodenum, query->node_map)) {
 811                                 mlog(0, "disallow join as node %u does not "
 812                                      "have node %u in its nodemap\n",
 813                                      query->node_idx, nodenum);
 814                                 packet.code = JOIN_DISALLOW;
 815                                 goto unlock_respond;
 816                         }
 817                 }
 818                 nodenum++;
 819         }
 820
 821         /* Once the dlm ctxt is marked as leaving then we don't want
 822          * to be put in someone's domain map.
 823          * Also, explicitly disallow joining at certain troublesome
 824          * times (ie. during recovery). */
 825         if (dlm && dlm->dlm_state != DLM_CTXT_LEAVING) {
 826                 int bit = query->node_idx;
 827                 spin_lock(&dlm->spinlock);
 828
 829                 if (dlm->dlm_state == DLM_CTXT_NEW &&
 830                     dlm->joining_node == DLM_LOCK_RES_OWNER_UNKNOWN) {
 831                         /*If this is a brand new context and we
 832                          * haven't started our join process yet, then
 833                          * the other node won the race. */
 834                         packet.code = JOIN_OK_NO_MAP;
 835                 } else if (dlm->joining_node != DLM_LOCK_RES_OWNER_UNKNOWN) {
 836                         /* Disallow parallel joins. */
 837                         packet.code = JOIN_DISALLOW;
 838                 } else if (dlm->reco.state & DLM_RECO_STATE_ACTIVE) {
 839                         mlog(0, "node %u trying to join, but recovery "
 840                              "is ongoing.\n", bit);
 841                         packet.code = JOIN_DISALLOW;
 842                 } else if (test_bit(bit, dlm->recovery_map)) {
 843                         mlog(0, "node %u trying to join, but it "
 844                              "still needs recovery.\n", bit);
 845                         packet.code = JOIN_DISALLOW;
 846                 } else if (test_bit(bit, dlm->domain_map)) {
 847                         mlog(0, "node %u trying to join, but it "
 848                              "is still in the domain! needs recovery?\n",
 849                              bit);
 850                         packet.code = JOIN_DISALLOW;
 851                 } else {
 852                         /* Alright we're fully a part of this domain
 853                          * so we keep some state as to who's joining
 854                          * and indicate to him that needs to be fixed
 855                          * up. */
 856
 857                         /* Make sure we speak compatible locking protocols.  */
 858                         if (dlm_query_join_proto_check("DLM", bit,
 859                                                        &dlm->dlm_locking_proto,
 860                                                        &query->dlm_proto)) {
 861                                 packet.code = JOIN_PROTOCOL_MISMATCH;
 862                         } else if (dlm_query_join_proto_check("fs", bit,
 863                                                               &dlm->fs_locking_proto,
 864                                                               &query->fs_proto)) {
 865                                 packet.code = JOIN_PROTOCOL_MISMATCH;
 866                         } else {
 867                                 packet.dlm_minor = query->dlm_proto.pv_minor;
 868                                 packet.fs_minor = query->fs_proto.pv_minor;
 869                                 packet.code = JOIN_OK;
 870                                 __dlm_set_joining_node(dlm, query->node_idx);
 871                         }
 872                 }
 873
 874                 spin_unlock(&dlm->spinlock);
 875         }
 876 unlock_respond:
 877         spin_unlock(&dlm_domain_lock);
 878
 879 respond:
 880         mlog(0, "We respond with %u\n", packet.code);
 881
 882         dlm_query_join_packet_to_wire(&packet, &response);
 883         return response;
 884 }
 885
 886 static int dlm_assert_joined_handler(struct o2net_msg *msg, u32 len, void *data,
 887                                      void **ret_data)
 888 {
 889         struct dlm_assert_joined *assert;
 890         struct dlm_ctxt *dlm = NULL;
 891
 892         assert = (struct dlm_assert_joined *) msg->buf;
 893
 894         mlog(0, "node %u asserts join on domain %s\n", assert->node_idx,
 895                   assert->domain);
 896
 897         spin_lock(&dlm_domain_lock);
 898         dlm = __dlm_lookup_domain_full(assert->domain, assert->name_len);
 899         /* XXX should we consider no dlm ctxt an error? */
 900         if (dlm) {
 901                 spin_lock(&dlm->spinlock);
 902
 903                 /* Alright, this node has officially joined our
 904                  * domain. Set him in the map and clean up our
 905                  * leftover join state. */
 906                 BUG_ON(dlm->joining_node != assert->node_idx);
 907                 set_bit(assert->node_idx, dlm->domain_map);
 908                 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
 909
 910                 printk(KERN_NOTICE "o2dlm: Node %u joins domain %s\n",
 911                        assert->node_idx, dlm->name);
 912                 __dlm_print_nodes(dlm);
 913
 914                 /* notify anything attached to the heartbeat events */
 915                 dlm_hb_event_notify_attached(dlm, assert->node_idx, 1);
 916
 917                 spin_unlock(&dlm->spinlock);
 918         }
 919         spin_unlock(&dlm_domain_lock);
 920
 921         return 0;
 922 }
 923
 924 static int dlm_cancel_join_handler(struct o2net_msg *msg, u32 len, void *data,
 925                                    void **ret_data)
 926 {
 927         struct dlm_cancel_join *cancel;
 928         struct dlm_ctxt *dlm = NULL;
 929
 930         cancel = (struct dlm_cancel_join *) msg->buf;
 931
 932         mlog(0, "node %u cancels join on domain %s\n", cancel->node_idx,
 933                   cancel->domain);
 934
 935         spin_lock(&dlm_domain_lock);
 936         dlm = __dlm_lookup_domain_full(cancel->domain, cancel->name_len);
 937
 938         if (dlm) {
 939                 spin_lock(&dlm->spinlock);
 940
 941                 /* Yikes, this guy wants to cancel his join. No
 942                  * problem, we simply cleanup our join state. */
 943                 BUG_ON(dlm->joining_node != cancel->node_idx);
 944                 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
 945
 946                 spin_unlock(&dlm->spinlock);
 947         }
 948         spin_unlock(&dlm_domain_lock);
 949
 950         return 0;
 951 }
 952
 953 static int dlm_send_one_join_cancel(struct dlm_ctxt *dlm,
 954                                     unsigned int node)
 955 {
 956         int status;
 957         struct dlm_cancel_join cancel_msg;
 958
 959         memset(&cancel_msg, 0, sizeof(cancel_msg));
 960         cancel_msg.node_idx = dlm->node_num;
 961         cancel_msg.name_len = strlen(dlm->name);
 962         memcpy(cancel_msg.domain, dlm->name, cancel_msg.name_len);
 963
 964         status = o2net_send_message(DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY,
 965                                     &cancel_msg, sizeof(cancel_msg), node,
 966                                     NULL);
 967         if (status < 0) {
 968                 mlog(ML_ERROR, "Error %d when sending message %u (key 0x%x) to "
 969                      "node %u\n", status, DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY,
 970                      node);
 971                 goto bail;
 972         }
 973
 974 bail:
 975         return status;
 976 }
 977
 978 /* map_size should be in bytes. */
 979 static int dlm_send_join_cancels(struct dlm_ctxt *dlm,
 980                                  unsigned long *node_map,
 981                                  unsigned int map_size)
 982 {
 983         int status, tmpstat;
 984         unsigned int node;
 985
 986         if (map_size != (BITS_TO_LONGS(O2NM_MAX_NODES) *
 987                          sizeof(unsigned long))) {
 988                 mlog(ML_ERROR,
 989                      "map_size %u != BITS_TO_LONGS(O2NM_MAX_NODES) %u\n",
 990                      map_size, (unsigned)BITS_TO_LONGS(O2NM_MAX_NODES));
 991                 return -EINVAL;
 992         }
 993
 994         status = 0;
 995         node = -1;
 996         while ((node = find_next_bit(node_map, O2NM_MAX_NODES,
 997                                      node + 1)) < O2NM_MAX_NODES) {
 998                 if (node == dlm->node_num)
 999                         continue;
1000
1001                 tmpstat = dlm_send_one_join_cancel(dlm, node);
1002                 if (tmpstat) {
1003                         mlog(ML_ERROR, "Error return %d cancelling join on "
1004                              "node %d\n", tmpstat, node);
1005                         if (!status)
1006                                 status = tmpstat;
1007                 }
1008         }
1009
1010         if (status)
1011                 mlog_errno(status);
1012         return status;
1013 }
1014
1015 static int dlm_request_join(struct dlm_ctxt *dlm,
1016                             int node,
1017                             enum dlm_query_join_response_code *response)
1018 {
1019         int status;
1020         struct dlm_query_join_request join_msg;
1021         struct dlm_query_join_packet packet;
1022         u32 join_resp;
1023
1024         mlog(0, "querying node %d\n", node);
1025
1026         memset(&join_msg, 0, sizeof(join_msg));
1027         join_msg.node_idx = dlm->node_num;
1028         join_msg.name_len = strlen(dlm->name);
1029         memcpy(join_msg.domain, dlm->name, join_msg.name_len);
1030         join_msg.dlm_proto = dlm->dlm_locking_proto;
1031         join_msg.fs_proto = dlm->fs_locking_proto;
1032
1033         /* copy live node map to join message */
1034         byte_copymap(join_msg.node_map, dlm->live_nodes_map, O2NM_MAX_NODES);
1035
1036         status = o2net_send_message(DLM_QUERY_JOIN_MSG, DLM_MOD_KEY, &join_msg,
1037                                     sizeof(join_msg), node, &join_resp);
1038         if (status < 0 && status != -ENOPROTOOPT) {
1039                 mlog(ML_ERROR, "Error %d when sending message %u (key 0x%x) to "
1040                      "node %u\n", status, DLM_QUERY_JOIN_MSG, DLM_MOD_KEY,
1041                      node);
1042                 goto bail;
1043         }
1044         dlm_query_join_wire_to_packet(join_resp, &packet);
1045
1046         /* -ENOPROTOOPT from the net code means the other side isn't
1047             listening for our message type -- that's fine, it means
1048             his dlm isn't up, so we can consider him a 'yes' but not
1049             joined into the domain.  */
1050         if (status == -ENOPROTOOPT) {
1051                 status = 0;
1052                 *response = JOIN_OK_NO_MAP;
1053         } else if (packet.code == JOIN_DISALLOW ||
1054                    packet.code == JOIN_OK_NO_MAP) {
1055                 *response = packet.code;
1056         } else if (packet.code == JOIN_PROTOCOL_MISMATCH) {
1057                 mlog(ML_NOTICE,
1058                      "This node requested DLM locking protocol %u.%u and "
1059                      "filesystem locking protocol %u.%u.  At least one of "
1060                      "the protocol versions on node %d is not compatible, "
1061                      "disconnecting\n",
1062                      dlm->dlm_locking_proto.pv_major,
1063                      dlm->dlm_locking_proto.pv_minor,
1064                      dlm->fs_locking_proto.pv_major,
1065                      dlm->fs_locking_proto.pv_minor,
1066                      node);
1067                 status = -EPROTO;
1068                 *response = packet.code;
1069         } else if (packet.code == JOIN_OK) {
1070                 *response = packet.code;
1071                 /* Use the same locking protocol as the remote node */
1072                 dlm->dlm_locking_proto.pv_minor = packet.dlm_minor;
1073                 dlm->fs_locking_proto.pv_minor = packet.fs_minor;
1074                 mlog(0,
1075                      "Node %d responds JOIN_OK with DLM locking protocol "
1076                      "%u.%u and fs locking protocol %u.%u\n",
1077                      node,
1078                      dlm->dlm_locking_proto.pv_major,
1079                      dlm->dlm_locking_proto.pv_minor,
1080                      dlm->fs_locking_proto.pv_major,
1081                      dlm->fs_locking_proto.pv_minor);
1082         } else {
1083                 status = -EINVAL;
1084                 mlog(ML_ERROR, "invalid response %d from node %u\n",
1085                      packet.code, node);
1086         }
1087
1088         mlog(0, "status %d, node %d response is %d\n", status, node,
1089              *response);
1090
1091 bail:
1092         return status;
1093 }
1094
1095 static int dlm_send_one_join_assert(struct dlm_ctxt *dlm,
1096                                     unsigned int node)
1097 {
1098         int status;
1099         struct dlm_assert_joined assert_msg;
1100
1101         mlog(0, "Sending join assert to node %u\n", node);
1102
1103         memset(&assert_msg, 0, sizeof(assert_msg));
1104         assert_msg.node_idx = dlm->node_num;
1105         assert_msg.name_len = strlen(dlm->name);
1106         memcpy(assert_msg.domain, dlm->name, assert_msg.name_len);
1107
1108         status = o2net_send_message(DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY,
1109                                     &assert_msg, sizeof(assert_msg), node,
1110                                     NULL);
1111         if (status < 0)
1112                 mlog(ML_ERROR, "Error %d when sending message %u (key 0x%x) to "
1113                      "node %u\n", status, DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY,
1114                      node);
1115
1116         return status;
1117 }
1118
1119 static void dlm_send_join_asserts(struct dlm_ctxt *dlm,
1120                                   unsigned long *node_map)
1121 {
1122         int status, node, live;
1123
1124         status = 0;
1125         node = -1;
1126         while ((node = find_next_bit(node_map, O2NM_MAX_NODES,
1127                                      node + 1)) < O2NM_MAX_NODES) {
1128                 if (node == dlm->node_num)
1129                         continue;
1130
1131                 do {
1132                         /* It is very important that this message be
1133                          * received so we spin until either the node
1134                          * has died or it gets the message. */
1135                         status = dlm_send_one_join_assert(dlm, node);
1136
1137                         spin_lock(&dlm->spinlock);
1138                         live = test_bit(node, dlm->live_nodes_map);
1139                         spin_unlock(&dlm->spinlock);
1140
1141                         if (status) {
1142                                 mlog(ML_ERROR, "Error return %d asserting "
1143                                      "join on node %d\n", status, node);
1144
1145                                 /* give us some time between errors... */
1146                                 if (live)
1147                                         msleep(DLM_DOMAIN_BACKOFF_MS);
1148                         }
1149                 } while (status && live);
1150         }
1151 }
1152
1153 struct domain_join_ctxt {
1154         unsigned long live_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
1155         unsigned long yes_resp_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
1156 };
1157
1158 static int dlm_should_restart_join(struct dlm_ctxt *dlm,
1159                                    struct domain_join_ctxt *ctxt,
1160                                    enum dlm_query_join_response_code response)
1161 {
1162         int ret;
1163
1164         if (response == JOIN_DISALLOW) {
1165                 mlog(0, "Latest response of disallow -- should restart\n");
1166                 return 1;
1167         }
1168
1169         spin_lock(&dlm->spinlock);
1170         /* For now, we restart the process if the node maps have
1171          * changed at all */
1172         ret = memcmp(ctxt->live_map, dlm->live_nodes_map,
1173                      sizeof(dlm->live_nodes_map));
1174         spin_unlock(&dlm->spinlock);
1175
1176         if (ret)
1177                 mlog(0, "Node maps changed -- should restart\n");
1178
1179         return ret;
1180 }
1181
1182 static int dlm_try_to_join_domain(struct dlm_ctxt *dlm)
1183 {
1184         int status = 0, tmpstat, node;
1185         struct domain_join_ctxt *ctxt;
1186         enum dlm_query_join_response_code response = JOIN_DISALLOW;
1187
1188         mlog_entry("%p", dlm);
1189
1190         ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL);
1191         if (!ctxt) {
1192                 status = -ENOMEM;
1193                 mlog_errno(status);
1194                 goto bail;
1195         }
1196
1197         /* group sem locking should work for us here -- we're already
1198          * registered for heartbeat events so filling this should be
1199          * atomic wrt getting those handlers called. */
1200         o2hb_fill_node_map(dlm->live_nodes_map, sizeof(dlm->live_nodes_map));
1201
1202         spin_lock(&dlm->spinlock);
1203         memcpy(ctxt->live_map, dlm->live_nodes_map, sizeof(ctxt->live_map));
1204
1205         __dlm_set_joining_node(dlm, dlm->node_num);
1206
1207         spin_unlock(&dlm->spinlock);
1208
1209         node = -1;
1210         while ((node = find_next_bit(ctxt->live_map, O2NM_MAX_NODES,
1211                                      node + 1)) < O2NM_MAX_NODES) {
1212                 if (node == dlm->node_num)
1213                         continue;
1214
1215                 status = dlm_request_join(dlm, node, &response);
1216                 if (status < 0) {
1217                         mlog_errno(status);
1218                         goto bail;
1219                 }
1220
1221                 /* Ok, either we got a response or the node doesn't have a
1222                  * dlm up. */
1223                 if (response == JOIN_OK)
1224                         set_bit(node, ctxt->yes_resp_map);
1225
1226                 if (dlm_should_restart_join(dlm, ctxt, response)) {
1227                         status = -EAGAIN;
1228                         goto bail;
1229                 }
1230         }
1231
1232         mlog(0, "Yay, done querying nodes!\n");
1233
1234         /* Yay, everyone agree's we can join the domain. My domain is
1235          * comprised of all nodes who were put in the
1236          * yes_resp_map. Copy that into our domain map and send a join
1237          * assert message to clean up everyone elses state. */
1238         spin_lock(&dlm->spinlock);
1239         memcpy(dlm->domain_map, ctxt->yes_resp_map,
1240                sizeof(ctxt->yes_resp_map));
1241         set_bit(dlm->node_num, dlm->domain_map);
1242         spin_unlock(&dlm->spinlock);
1243
1244         dlm_send_join_asserts(dlm, ctxt->yes_resp_map);
1245
1246         /* Joined state *must* be set before the joining node
1247          * information, otherwise the query_join handler may read no
1248          * current joiner but a state of NEW and tell joining nodes
1249          * we're not in the domain. */
1250         spin_lock(&dlm_domain_lock);
1251         dlm->dlm_state = DLM_CTXT_JOINED;
1252         dlm->num_joins++;
1253         spin_unlock(&dlm_domain_lock);
1254
1255 bail:
1256         spin_lock(&dlm->spinlock);
1257         __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
1258         if (!status)
1259                 __dlm_print_nodes(dlm);
1260         spin_unlock(&dlm->spinlock);
1261
1262         if (ctxt) {
1263                 /* Do we need to send a cancel message to any nodes? */
1264                 if (status < 0) {
1265                         tmpstat = dlm_send_join_cancels(dlm,
1266                                                         ctxt->yes_resp_map,
1267                                                         sizeof(ctxt->yes_resp_map));
1268                         if (tmpstat < 0)
1269                                 mlog_errno(tmpstat);
1270                 }
1271                 kfree(ctxt);
1272         }
1273
1274         mlog(0, "returning %d\n", status);
1275         return status;
1276 }
1277
1278 static void dlm_unregister_domain_handlers(struct dlm_ctxt *dlm)
1279 {
1280         o2hb_unregister_callback(NULL, &dlm->dlm_hb_up);
1281         o2hb_unregister_callback(NULL, &dlm->dlm_hb_down);
1282         o2net_unregister_handler_list(&dlm->dlm_domain_handlers);
1283 }
1284
1285 static int dlm_register_domain_handlers(struct dlm_ctxt *dlm)
1286 {
1287         int status;
1288
1289         mlog(0, "registering handlers.\n");
1290
1291         o2hb_setup_callback(&dlm->dlm_hb_down, O2HB_NODE_DOWN_CB,
1292                             dlm_hb_node_down_cb, dlm, DLM_HB_NODE_DOWN_PRI);
1293         status = o2hb_register_callback(NULL, &dlm->dlm_hb_down);
1294         if (status)
1295                 goto bail;
1296
1297         o2hb_setup_callback(&dlm->dlm_hb_up, O2HB_NODE_UP_CB,
1298                             dlm_hb_node_up_cb, dlm, DLM_HB_NODE_UP_PRI);
1299         status = o2hb_register_callback(NULL, &dlm->dlm_hb_up);
1300         if (status)
1301                 goto bail;
1302
1303         status = o2net_register_handler(DLM_MASTER_REQUEST_MSG, dlm->key,
1304                                         sizeof(struct dlm_master_request),
1305                                         dlm_master_request_handler,
1306                                         dlm, NULL, &dlm->dlm_domain_handlers);
1307         if (status)
1308                 goto bail;
1309
1310         status = o2net_register_handler(DLM_ASSERT_MASTER_MSG, dlm->key,
1311                                         sizeof(struct dlm_assert_master),
1312                                         dlm_assert_master_handler,
1313                                         dlm, dlm_assert_master_post_handler,
1314                                         &dlm->dlm_domain_handlers);
1315         if (status)
1316                 goto bail;
1317
1318         status = o2net_register_handler(DLM_CREATE_LOCK_MSG, dlm->key,
1319                                         sizeof(struct dlm_create_lock),
1320                                         dlm_create_lock_handler,
1321                                         dlm, NULL, &dlm->dlm_domain_handlers);
1322         if (status)
1323                 goto bail;
1324
1325         status = o2net_register_handler(DLM_CONVERT_LOCK_MSG, dlm->key,
1326                                         DLM_CONVERT_LOCK_MAX_LEN,
1327                                         dlm_convert_lock_handler,
1328                                         dlm, NULL, &dlm->dlm_domain_handlers);
1329         if (status)
1330                 goto bail;
1331
1332         status = o2net_register_handler(DLM_UNLOCK_LOCK_MSG, dlm->key,
1333                                         DLM_UNLOCK_LOCK_MAX_LEN,
1334                                         dlm_unlock_lock_handler,
1335                                         dlm, NULL, &dlm->dlm_domain_handlers);
1336         if (status)
1337                 goto bail;
1338
1339         status = o2net_register_handler(DLM_PROXY_AST_MSG, dlm->key,
1340                                         DLM_PROXY_AST_MAX_LEN,
1341                                         dlm_proxy_ast_handler,
1342                                         dlm, NULL, &dlm->dlm_domain_handlers);
1343         if (status)
1344                 goto bail;
1345
1346         status = o2net_register_handler(DLM_EXIT_DOMAIN_MSG, dlm->key,
1347                                         sizeof(struct dlm_exit_domain),
1348                                         dlm_exit_domain_handler,
1349                                         dlm, NULL, &dlm->dlm_domain_handlers);
1350         if (status)
1351                 goto bail;
1352
1353         status = o2net_register_handler(DLM_DEREF_LOCKRES_MSG, dlm->key,
1354                                         sizeof(struct dlm_deref_lockres),
1355                                         dlm_deref_lockres_handler,
1356                                         dlm, NULL, &dlm->dlm_domain_handlers);
1357         if (status)
1358                 goto bail;
1359
1360         status = o2net_register_handler(DLM_MIGRATE_REQUEST_MSG, dlm->key,
1361                                         sizeof(struct dlm_migrate_request),
1362                                         dlm_migrate_request_handler,
1363                                         dlm, NULL, &dlm->dlm_domain_handlers);
1364         if (status)
1365                 goto bail;
1366
1367         status = o2net_register_handler(DLM_MIG_LOCKRES_MSG, dlm->key,
1368                                         DLM_MIG_LOCKRES_MAX_LEN,
1369                                         dlm_mig_lockres_handler,
1370                                         dlm, NULL, &dlm->dlm_domain_handlers);
1371         if (status)
1372                 goto bail;
1373
1374         status = o2net_register_handler(DLM_MASTER_REQUERY_MSG, dlm->key,
1375                                         sizeof(struct dlm_master_requery),
1376                                         dlm_master_requery_handler,
1377                                         dlm, NULL, &dlm->dlm_domain_handlers);
1378         if (status)
1379                 goto bail;
1380
1381         status = o2net_register_handler(DLM_LOCK_REQUEST_MSG, dlm->key,
1382                                         sizeof(struct dlm_lock_request),
1383                                         dlm_request_all_locks_handler,
1384                                         dlm, NULL, &dlm->dlm_domain_handlers);
1385         if (status)
1386                 goto bail;
1387
1388         status = o2net_register_handler(DLM_RECO_DATA_DONE_MSG, dlm->key,
1389                                         sizeof(struct dlm_reco_data_done),
1390                                         dlm_reco_data_done_handler,
1391                                         dlm, NULL, &dlm->dlm_domain_handlers);
1392         if (status)
1393                 goto bail;
1394
1395         status = o2net_register_handler(DLM_BEGIN_RECO_MSG, dlm->key,
1396                                         sizeof(struct dlm_begin_reco),
1397                                         dlm_begin_reco_handler,
1398                                         dlm, NULL, &dlm->dlm_domain_handlers);
1399         if (status)
1400                 goto bail;
1401
1402         status = o2net_register_handler(DLM_FINALIZE_RECO_MSG, dlm->key,
1403                                         sizeof(struct dlm_finalize_reco),
1404                                         dlm_finalize_reco_handler,
1405                                         dlm, NULL, &dlm->dlm_domain_handlers);
1406         if (status)
1407                 goto bail;
1408
1409 bail:
1410         if (status)
1411                 dlm_unregister_domain_handlers(dlm);
1412
1413         return status;
1414 }
1415
1416 static int dlm_join_domain(struct dlm_ctxt *dlm)
1417 {
1418         int status;
1419         unsigned int backoff;
1420         unsigned int total_backoff = 0;
1421
1422         BUG_ON(!dlm);
1423
1424         mlog(0, "Join domain %s\n", dlm->name);
1425
1426         status = dlm_register_domain_handlers(dlm);
1427         if (status) {
1428                 mlog_errno(status);
1429                 goto bail;
1430         }
1431
1432         status = dlm_debug_init(dlm);
1433         if (status < 0) {
1434                 mlog_errno(status);
1435                 goto bail;
1436         }
1437
1438         status = dlm_launch_thread(dlm);
1439         if (status < 0) {
1440                 mlog_errno(status);
1441                 goto bail;
1442         }
1443
1444         status = dlm_launch_recovery_thread(dlm);
1445         if (status < 0) {
1446                 mlog_errno(status);
1447                 goto bail;
1448         }
1449
1450         dlm->dlm_worker = create_singlethread_workqueue("dlm_wq");
1451         if (!dlm->dlm_worker) {
1452                 status = -ENOMEM;
1453                 mlog_errno(status);
1454                 goto bail;
1455         }
1456
1457         do {
1458                 status = dlm_try_to_join_domain(dlm);
1459
1460                 /* If we're racing another node to the join, then we
1461                  * need to back off temporarily and let them
1462                  * complete. */
1463 #define DLM_JOIN_TIMEOUT_MSECS  90000
1464                 if (status == -EAGAIN) {
1465                         if (signal_pending(current)) {
1466                                 status = -ERESTARTSYS;
1467                                 goto bail;
1468                         }
1469
1470                         if (total_backoff >
1471                             msecs_to_jiffies(DLM_JOIN_TIMEOUT_MSECS)) {
1472                                 status = -ERESTARTSYS;
1473                                 mlog(ML_NOTICE, "Timed out joining dlm domain "
1474                                      "%s after %u msecs\n", dlm->name,
1475                                      jiffies_to_msecs(total_backoff));
1476                                 goto bail;
1477                         }
1478
1479                         /*
1480                          * <chip> After you!
1481                          * <dale> No, after you!
1482                          * <chip> I insist!
1483                          * <dale> But you first!
1484                          * ...
1485                          */
1486                         backoff = (unsigned int)(jiffies & 0x3);
1487                         backoff *= DLM_DOMAIN_BACKOFF_MS;
1488                         total_backoff += backoff;
1489                         mlog(0, "backoff %d\n", backoff);
1490                         msleep(backoff);
1491                 }
1492         } while (status == -EAGAIN);
1493
1494         if (status < 0) {
1495                 mlog_errno(status);
1496                 goto bail;
1497         }
1498
1499         status = 0;
1500 bail:
1501         wake_up(&dlm_domain_events);
1502
1503         if (status) {
1504                 dlm_unregister_domain_handlers(dlm);
1505                 dlm_debug_shutdown(dlm);
1506                 dlm_complete_thread(dlm);
1507                 dlm_complete_recovery_thread(dlm);
1508                 dlm_destroy_dlm_worker(dlm);
1509         }
1510
1511         return status;
1512 }
1513
1514 static struct dlm_ctxt *dlm_alloc_ctxt(const char *domain,
1515                                 u32 key)
1516 {
1517         int i;
1518         int ret;
1519         struct dlm_ctxt *dlm = NULL;
1520
1521         dlm = kzalloc(sizeof(*dlm), GFP_KERNEL);
1522         if (!dlm) {
1523                 mlog_errno(-ENOMEM);
1524                 goto leave;
1525         }
1526
1527         dlm->name = kstrdup(domain, GFP_KERNEL);
1528         if (dlm->name == NULL) {
1529                 mlog_errno(-ENOMEM);
1530                 kfree(dlm);
1531                 dlm = NULL;
1532                 goto leave;
1533         }
1534
1535         dlm->lockres_hash = (struct hlist_head **)dlm_alloc_pagevec(DLM_HASH_PAGES);
1536         if (!dlm->lockres_hash) {
1537                 mlog_errno(-ENOMEM);
1538                 kfree(dlm->name);
1539                 kfree(dlm);
1540                 dlm = NULL;
1541                 goto leave;
1542         }
1543
1544         for (i = 0; i < DLM_HASH_BUCKETS; i++)
1545                 INIT_HLIST_HEAD(dlm_lockres_hash(dlm, i));
1546
1547         dlm->master_hash = (struct hlist_head **)
1548                                 dlm_alloc_pagevec(DLM_HASH_PAGES);
1549         if (!dlm->master_hash) {
1550                 mlog_errno(-ENOMEM);
1551                 dlm_free_pagevec((void **)dlm->lockres_hash, DLM_HASH_PAGES);
1552                 kfree(dlm->name);
1553                 kfree(dlm);
1554                 dlm = NULL;
1555                 goto leave;
1556         }
1557
1558         for (i = 0; i < DLM_HASH_BUCKETS; i++)
1559                 INIT_HLIST_HEAD(dlm_master_hash(dlm, i));
1560
1561         dlm->key = key;
1562         dlm->node_num = o2nm_this_node();
1563
1564         ret = dlm_create_debugfs_subroot(dlm);
1565         if (ret < 0) {
1566                 dlm_free_pagevec((void **)dlm->master_hash, DLM_HASH_PAGES);
1567                 dlm_free_pagevec((void **)dlm->lockres_hash, DLM_HASH_PAGES);
1568                 kfree(dlm->name);
1569                 kfree(dlm);
1570                 dlm = NULL;
1571                 goto leave;
1572         }
1573
1574         spin_lock_init(&dlm->spinlock);
1575         spin_lock_init(&dlm->master_lock);
1576         spin_lock_init(&dlm->ast_lock);
1577         spin_lock_init(&dlm->track_lock);
1578         INIT_LIST_HEAD(&dlm->list);
1579         INIT_LIST_HEAD(&dlm->dirty_list);
1580         INIT_LIST_HEAD(&dlm->reco.resources);
1581         INIT_LIST_HEAD(&dlm->reco.received);
1582         INIT_LIST_HEAD(&dlm->reco.node_data);
1583         INIT_LIST_HEAD(&dlm->purge_list);
1584         INIT_LIST_HEAD(&dlm->dlm_domain_handlers);
1585         INIT_LIST_HEAD(&dlm->tracking_list);
1586         dlm->reco.state = 0;
1587
1588         INIT_LIST_HEAD(&dlm->pending_asts);
1589         INIT_LIST_HEAD(&dlm->pending_basts);
1590
1591         mlog(0, "dlm->recovery_map=%p, &(dlm->recovery_map[0])=%p\n",
1592                   dlm->recovery_map, &(dlm->recovery_map[0]));
1593
1594         memset(dlm->recovery_map, 0, sizeof(dlm->recovery_map));
1595         memset(dlm->live_nodes_map, 0, sizeof(dlm->live_nodes_map));
1596         memset(dlm->domain_map, 0, sizeof(dlm->domain_map));
1597
1598         dlm->dlm_thread_task = NULL;
1599         dlm->dlm_reco_thread_task = NULL;
1600         dlm->dlm_worker = NULL;
1601         init_waitqueue_head(&dlm->dlm_thread_wq);
1602         init_waitqueue_head(&dlm->dlm_reco_thread_wq);
1603         init_waitqueue_head(&dlm->reco.event);
1604         init_waitqueue_head(&dlm->ast_wq);
1605         init_waitqueue_head(&dlm->migration_wq);
1606         INIT_LIST_HEAD(&dlm->mle_hb_events);
1607
1608         dlm->joining_node = DLM_LOCK_RES_OWNER_UNKNOWN;
1609         init_waitqueue_head(&dlm->dlm_join_events);
1610
1611         dlm->reco.new_master = O2NM_INVALID_NODE_NUM;
1612         dlm->reco.dead_node = O2NM_INVALID_NODE_NUM;
1613
1614         atomic_set(&dlm->res_tot_count, 0);
1615         atomic_set(&dlm->res_cur_count, 0);
1616         for (i = 0; i < DLM_MLE_NUM_TYPES; ++i) {
1617                 atomic_set(&dlm->mle_tot_count[i], 0);
1618                 atomic_set(&dlm->mle_cur_count[i], 0);
1619         }
1620
1621         spin_lock_init(&dlm->work_lock);
1622         INIT_LIST_HEAD(&dlm->work_list);
1623         INIT_WORK(&dlm->dispatched_work, dlm_dispatch_work);
1624
1625         kref_init(&dlm->dlm_refs);
1626         dlm->dlm_state = DLM_CTXT_NEW;
1627
1628         INIT_LIST_HEAD(&dlm->dlm_eviction_callbacks);
1629
1630         mlog(0, "context init: refcount %u\n",
1631                   atomic_read(&dlm->dlm_refs.refcount));
1632
1633 leave:
1634         return dlm;
1635 }
1636
1637 /*
1638  * Compare a requested locking protocol version against the current one.
1639  *
1640  * If the major numbers are different, they are incompatible.
1641  * If the current minor is greater than the request, they are incompatible.
1642  * If the current minor is less than or equal to the request, they are
1643  * compatible, and the requester should run at the current minor version.
1644  */
1645 static int dlm_protocol_compare(struct dlm_protocol_version *existing,
1646                                 struct dlm_protocol_version *request)
1647 {
1648         if (existing->pv_major != request->pv_major)
1649                 return 1;
1650
1651         if (existing->pv_minor > request->pv_minor)
1652                 return 1;
1653
1654         if (existing->pv_minor < request->pv_minor)
1655                 request->pv_minor = existing->pv_minor;
1656
1657         return 0;
1658 }
1659
1660 /*
1661  * dlm_register_domain: one-time setup per "domain".
1662  *
1663  * The filesystem passes in the requested locking version via proto.
1664  * If registration was successful, proto will contain the negotiated
1665  * locking protocol.
1666  */
1667 struct dlm_ctxt * dlm_register_domain(const char *domain,
1668                                u32 key,
1669                                struct dlm_protocol_version *fs_proto)
1670 {
1671         int ret;
1672         struct dlm_ctxt *dlm = NULL;
1673         struct dlm_ctxt *new_ctxt = NULL;
1674
1675         if (strlen(domain) >= O2NM_MAX_NAME_LEN) {
1676                 ret = -ENAMETOOLONG;
1677                 mlog(ML_ERROR, "domain name length too long\n");
1678                 goto leave;
1679         }
1680
1681         if (!o2hb_check_local_node_heartbeating()) {
1682                 mlog(ML_ERROR, "the local node has not been configured, or is "
1683                      "not heartbeating\n");
1684                 ret = -EPROTO;
1685                 goto leave;
1686         }
1687
1688         mlog(0, "register called for domain \"%s\"\n", domain);
1689
1690 retry:
1691         dlm = NULL;
1692         if (signal_pending(current)) {
1693                 ret = -ERESTARTSYS;
1694                 mlog_errno(ret);
1695                 goto leave;
1696         }
1697
1698         spin_lock(&dlm_domain_lock);
1699
1700         dlm = __dlm_lookup_domain(domain);
1701         if (dlm) {
1702                 if (dlm->dlm_state != DLM_CTXT_JOINED) {
1703                         spin_unlock(&dlm_domain_lock);
1704
1705                         mlog(0, "This ctxt is not joined yet!\n");
1706                         wait_event_interruptible(dlm_domain_events,
1707                                                  dlm_wait_on_domain_helper(
1708                                                          domain));
1709                         goto retry;
1710                 }
1711
1712                 if (dlm_protocol_compare(&dlm->fs_locking_proto, fs_proto)) {
1713                         spin_unlock(&dlm_domain_lock);
1714                         mlog(ML_ERROR,
1715                              "Requested locking protocol version is not "
1716                              "compatible with already registered domain "
1717                              "\"%s\"\n", domain);
1718                         ret = -EPROTO;
1719                         goto leave;
1720                 }
1721
1722                 __dlm_get(dlm);
1723                 dlm->num_joins++;
1724
1725                 spin_unlock(&dlm_domain_lock);
1726
1727                 ret = 0;
1728                 goto leave;
1729         }
1730
1731         /* doesn't exist */
1732         if (!new_ctxt) {
1733                 spin_unlock(&dlm_domain_lock);
1734
1735                 new_ctxt = dlm_alloc_ctxt(domain, key);
1736                 if (new_ctxt)
1737                         goto retry;
1738
1739                 ret = -ENOMEM;
1740                 mlog_errno(ret);
1741                 goto leave;
1742         }
1743
1744         /* a little variable switch-a-roo here... */
1745         dlm = new_ctxt;
1746         new_ctxt = NULL;
1747
1748         /* add the new domain */
1749         list_add_tail(&dlm->list, &dlm_domains);
1750         spin_unlock(&dlm_domain_lock);
1751
1752         /*
1753          * Pass the locking protocol version into the join.  If the join
1754          * succeeds, it will have the negotiated protocol set.
1755          */
1756         dlm->dlm_locking_proto = dlm_protocol;
1757         dlm->fs_locking_proto = *fs_proto;
1758
1759         ret = dlm_join_domain(dlm);
1760         if (ret) {
1761                 mlog_errno(ret);
1762                 dlm_put(dlm);
1763                 goto leave;
1764         }
1765
1766         /* Tell the caller what locking protocol we negotiated */
1767         *fs_proto = dlm->fs_locking_proto;
1768
1769         ret = 0;
1770 leave:
1771         if (new_ctxt)
1772                 dlm_free_ctxt_mem(new_ctxt);
1773
1774         if (ret < 0)
1775                 dlm = ERR_PTR(ret);
1776
1777         return dlm;
1778 }
1779 EXPORT_SYMBOL_GPL(dlm_register_domain);
1780
1781 static LIST_HEAD(dlm_join_handlers);
1782
1783 static void dlm_unregister_net_handlers(void)
1784 {
1785         o2net_unregister_handler_list(&dlm_join_handlers);
1786 }
1787
1788 static int dlm_register_net_handlers(void)
1789 {
1790         int status = 0;
1791
1792         status = o2net_register_handler(DLM_QUERY_JOIN_MSG, DLM_MOD_KEY,
1793                                         sizeof(struct dlm_query_join_request),
1794                                         dlm_query_join_handler,
1795                                         NULL, NULL, &dlm_join_handlers);
1796         if (status)
1797                 goto bail;
1798
1799         status = o2net_register_handler(DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY,
1800                                         sizeof(struct dlm_assert_joined),
1801                                         dlm_assert_joined_handler,
1802                                         NULL, NULL, &dlm_join_handlers);
1803         if (status)
1804                 goto bail;
1805
1806         status = o2net_register_handler(DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY,
1807                                         sizeof(struct dlm_cancel_join),
1808                                         dlm_cancel_join_handler,
1809                                         NULL, NULL, &dlm_join_handlers);
1810
1811 bail:
1812         if (status < 0)
1813                 dlm_unregister_net_handlers();
1814
1815         return status;
1816 }
1817
1818 /* Domain eviction callback handling.
1819  *
1820  * The file system requires notification of node death *before* the
1821  * dlm completes it's recovery work, otherwise it may be able to
1822  * acquire locks on resources requiring recovery. Since the dlm can
1823  * evict a node from it's domain *before* heartbeat fires, a similar
1824  * mechanism is required. */
1825
1826 /* Eviction is not expected to happen often, so a per-domain lock is
1827  * not necessary. Eviction callbacks are allowed to sleep for short
1828  * periods of time. */
1829 static DECLARE_RWSEM(dlm_callback_sem);
1830
1831 void dlm_fire_domain_eviction_callbacks(struct dlm_ctxt *dlm,
1832                                         int node_num)
1833 {
1834         struct list_head *iter;
1835         struct dlm_eviction_cb *cb;
1836
1837         down_read(&dlm_callback_sem);
1838         list_for_each(iter, &dlm->dlm_eviction_callbacks) {
1839                 cb = list_entry(iter, struct dlm_eviction_cb, ec_item);
1840
1841                 cb->ec_func(node_num, cb->ec_data);
1842         }
1843         up_read(&dlm_callback_sem);
1844 }
1845
1846 void dlm_setup_eviction_cb(struct dlm_eviction_cb *cb,
1847                            dlm_eviction_func *f,
1848                            void *data)
1849 {
1850         INIT_LIST_HEAD(&cb->ec_item);
1851         cb->ec_func = f;
1852         cb->ec_data = data;
1853 }
1854 EXPORT_SYMBOL_GPL(dlm_setup_eviction_cb);
1855
1856 void dlm_register_eviction_cb(struct dlm_ctxt *dlm,
1857                               struct dlm_eviction_cb *cb)
1858 {
1859         down_write(&dlm_callback_sem);
1860         list_add_tail(&cb->ec_item, &dlm->dlm_eviction_callbacks);
1861         up_write(&dlm_callback_sem);
1862 }
1863 EXPORT_SYMBOL_GPL(dlm_register_eviction_cb);
1864
1865 void dlm_unregister_eviction_cb(struct dlm_eviction_cb *cb)
1866 {
1867         down_write(&dlm_callback_sem);
1868         list_del_init(&cb->ec_item);
1869         up_write(&dlm_callback_sem);
1870 }
1871 EXPORT_SYMBOL_GPL(dlm_unregister_eviction_cb);
1872
1873 static int __init dlm_init(void)
1874 {
1875         int status;
1876
1877         dlm_print_version();
1878
1879         status = dlm_init_mle_cache();
1880         if (status) {
1881                 mlog(ML_ERROR, "Could not create o2dlm_mle slabcache\n");
1882                 goto error;
1883         }
1884
1885         status = dlm_init_master_caches();
1886         if (status) {
1887                 mlog(ML_ERROR, "Could not create o2dlm_lockres and "
1888                      "o2dlm_lockname slabcaches\n");
1889                 goto error;
1890         }
1891
1892         status = dlm_init_lock_cache();
1893         if (status) {
1894                 mlog(ML_ERROR, "Count not create o2dlm_lock slabcache\n");
1895                 goto error;
1896         }
1897
1898         status = dlm_register_net_handlers();
1899         if (status) {
1900                 mlog(ML_ERROR, "Unable to register network handlers\n");
1901                 goto error;
1902         }
1903
1904         status = dlm_create_debugfs_root();
1905         if (status)
1906                 goto error;
1907
1908         return 0;
1909 error:
1910         dlm_unregister_net_handlers();
1911         dlm_destroy_lock_cache();
1912         dlm_destroy_master_caches();
1913         dlm_destroy_mle_cache();
1914         return -1;
1915 }
1916
1917 static void __exit dlm_exit (void)
1918 {
1919         dlm_destroy_debugfs_root();
1920         dlm_unregister_net_handlers();
1921         dlm_destroy_lock_cache();
1922         dlm_destroy_master_caches();
1923         dlm_destroy_mle_cache();
1924 }
1925
1926 MODULE_AUTHOR("Oracle");
1927 MODULE_LICENSE("GPL");
1928
1929 module_init(dlm_init);
1930 module_exit(dlm_exit);