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 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_DOMAIN)
  47 #include "cluster/masklog.h"
  48
  49 /*
  50  * ocfs2 node maps are array of long int, which limits to send them freely
  51  * across the wire due to endianness issues. To workaround this, we convert
  52  * long ints to byte arrays. Following 3 routines are helper functions to
  53  * set/test/copy bits within those array of bytes
  54  */
  55 static inline void byte_set_bit(u8 nr, u8 map[])
  56 {
  57         map[nr >> 3] |= (1UL << (nr & 7));
  58 }
  59
  60 static inline int byte_test_bit(u8 nr, u8 map[])
  61 {
  62         return ((1UL << (nr & 7)) & (map[nr >> 3])) != 0;
  63 }
  64
  65 static inline void byte_copymap(u8 dmap[], unsigned long smap[],
  66                         unsigned int sz)
  67 {
  68         unsigned int nn;
  69
  70         if (!sz)
  71                 return;
  72
  73         memset(dmap, 0, ((sz + 7) >> 3));
  74         for (nn = 0 ; nn < sz; nn++)
  75                 if (test_bit(nn, smap))
  76                         byte_set_bit(nn, dmap);
  77 }
  78
  79 static void dlm_free_pagevec(void **vec, int pages)
  80 {
  81         while (pages--)
  82                 free_page((unsigned long)vec[pages]);
  83         kfree(vec);
  84 }
  85
  86 static void **dlm_alloc_pagevec(int pages)
  87 {
  88         void **vec = kmalloc(pages * sizeof(void *), GFP_KERNEL);
  89         int i;
  90
  91         if (!vec)
  92                 return NULL;
  93
  94         for (i = 0; i < pages; i++)
  95                 if (!(vec[i] = (void *)__get_free_page(GFP_KERNEL)))
  96                         goto out_free;
  97
  98         mlog(0, "Allocated DLM hash pagevec; %d pages (%lu expected), %lu buckets per page\n",
  99              pages, (unsigned long)DLM_HASH_PAGES,
 100              (unsigned long)DLM_BUCKETS_PER_PAGE);
 101         return vec;
 102 out_free:
 103         dlm_free_pagevec(vec, i);
 104         return NULL;
 105 }
 106
 107 /*
 108  *
 109  * spinlock lock ordering: if multiple locks are needed, obey this ordering:
 110  *    dlm_domain_lock
 111  *    struct dlm_ctxt->spinlock
 112  *    struct dlm_lock_resource->spinlock
 113  *    struct dlm_ctxt->master_lock
 114  *    struct dlm_ctxt->ast_lock
 115  *    dlm_master_list_entry->spinlock
 116  *    dlm_lock->spinlock
 117  *
 118  */
 119
 120 DEFINE_SPINLOCK(dlm_domain_lock);
 121 LIST_HEAD(dlm_domains);
 122 static DECLARE_WAIT_QUEUE_HEAD(dlm_domain_events);
 123
 124 /*
 125  * The supported protocol version for DLM communication.  Running domains
 126  * will have a negotiated version with the same major number and a minor
 127  * number equal or smaller.  The dlm_ctxt->dlm_locking_proto field should
 128  * be used to determine what a running domain is actually using.
 129  *
 130  * New in version 1.1:
 131  *      - Message DLM_QUERY_REGION added to support global heartbeat
 132  *      - Message DLM_QUERY_NODEINFO added to allow online node removes
 133  * New in version 1.2:
 134  *      - Message DLM_BEGIN_EXIT_DOMAIN_MSG added to mark start of exit domain
 135  * New in version 1.3:
 136  *      - Message DLM_DEREF_LOCKRES_DONE added to inform non-master that the
 137  *        refmap is cleared
 138  */
 139 static const struct dlm_protocol_version dlm_protocol = {
 140         .pv_major = 1,
 141         .pv_minor = 3,
 142 };
 143
 144 #define DLM_DOMAIN_BACKOFF_MS 200
 145
 146 static int dlm_query_join_handler(struct o2net_msg *msg, u32 len, void *data,
 147                                   void **ret_data);
 148 static int dlm_assert_joined_handler(struct o2net_msg *msg, u32 len, void *data,
 149                                      void **ret_data);
 150 static int dlm_cancel_join_handler(struct o2net_msg *msg, u32 len, void *data,
 151                                    void **ret_data);
 152 static int dlm_query_region_handler(struct o2net_msg *msg, u32 len,
 153                                     void *data, void **ret_data);
 154 static int dlm_exit_domain_handler(struct o2net_msg *msg, u32 len, void *data,
 155                                    void **ret_data);
 156 static int dlm_protocol_compare(struct dlm_protocol_version *existing,
 157                                 struct dlm_protocol_version *request);
 158
 159 static void dlm_unregister_domain_handlers(struct dlm_ctxt *dlm);
 160
 161 void __dlm_unhash_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
 162 {
 163         if (hlist_unhashed(&res->hash_node))
 164                 return;
 165
 166         mlog(0, "%s: Unhash res %.*s\n", dlm->name, res->lockname.len,
 167              res->lockname.name);
 168         hlist_del_init(&res->hash_node);
 169         dlm_lockres_put(res);
 170 }
 171
 172 void __dlm_insert_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
 173 {
 174         struct hlist_head *bucket;
 175
 176         assert_spin_locked(&dlm->spinlock);
 177
 178         bucket = dlm_lockres_hash(dlm, res->lockname.hash);
 179
 180         /* get a reference for our hashtable */
 181         dlm_lockres_get(res);
 182
 183         hlist_add_head(&res->hash_node, bucket);
 184
 185         mlog(0, "%s: Hash res %.*s\n", dlm->name, res->lockname.len,
 186              res->lockname.name);
 187 }
 188
 189 struct dlm_lock_resource * __dlm_lookup_lockres_full(struct dlm_ctxt *dlm,
 190                                                      const char *name,
 191                                                      unsigned int len,
 192                                                      unsigned int hash)
 193 {
 194         struct hlist_head *bucket;
 195         struct dlm_lock_resource *res;
 196
 197         mlog(0, "%.*s\n", len, name);
 198
 199         assert_spin_locked(&dlm->spinlock);
 200
 201         bucket = dlm_lockres_hash(dlm, hash);
 202
 203         hlist_for_each_entry(res, bucket, hash_node) {
 204                 if (res->lockname.name[0] != name[0])
 205                         continue;
 206                 if (unlikely(res->lockname.len != len))
 207                         continue;
 208                 if (memcmp(res->lockname.name + 1, name + 1, len - 1))
 209                         continue;
 210                 dlm_lockres_get(res);
 211                 return res;
 212         }
 213         return NULL;
 214 }
 215
 216 /* intended to be called by functions which do not care about lock
 217  * resources which are being purged (most net _handler functions).
 218  * this will return NULL for any lock resource which is found but
 219  * currently in the process of dropping its mastery reference.
 220  * use __dlm_lookup_lockres_full when you need the lock resource
 221  * regardless (e.g. dlm_get_lock_resource) */
 222 struct dlm_lock_resource * __dlm_lookup_lockres(struct dlm_ctxt *dlm,
 223                                                 const char *name,
 224                                                 unsigned int len,
 225                                                 unsigned int hash)
 226 {
 227         struct dlm_lock_resource *res = NULL;
 228
 229         mlog(0, "%.*s\n", len, name);
 230
 231         assert_spin_locked(&dlm->spinlock);
 232
 233         res = __dlm_lookup_lockres_full(dlm, name, len, hash);
 234         if (res) {
 235                 spin_lock(&res->spinlock);
 236                 if (res->state & DLM_LOCK_RES_DROPPING_REF) {
 237                         spin_unlock(&res->spinlock);
 238                         dlm_lockres_put(res);
 239                         return NULL;
 240                 }
 241                 spin_unlock(&res->spinlock);
 242         }
 243
 244         return res;
 245 }
 246
 247 struct dlm_lock_resource * dlm_lookup_lockres(struct dlm_ctxt *dlm,
 248                                     const char *name,
 249                                     unsigned int len)
 250 {
 251         struct dlm_lock_resource *res;
 252         unsigned int hash = dlm_lockid_hash(name, len);
 253
 254         spin_lock(&dlm->spinlock);
 255         res = __dlm_lookup_lockres(dlm, name, len, hash);
 256         spin_unlock(&dlm->spinlock);
 257         return res;
 258 }
 259
 260 static struct dlm_ctxt * __dlm_lookup_domain_full(const char *domain, int len)
 261 {
 262         struct dlm_ctxt *tmp;
 263
 264         assert_spin_locked(&dlm_domain_lock);
 265
 266         /* tmp->name here is always NULL terminated,
 267          * but domain may not be! */
 268         list_for_each_entry(tmp, &dlm_domains, list) {
 269                 if (strlen(tmp->name) == len &&
 270                     memcmp(tmp->name, domain, len)==0)
 271                         return tmp;
 272         }
 273
 274         return NULL;
 275 }
 276
 277 /* For null terminated domain strings ONLY */
 278 static struct dlm_ctxt * __dlm_lookup_domain(const char *domain)
 279 {
 280         assert_spin_locked(&dlm_domain_lock);
 281
 282         return __dlm_lookup_domain_full(domain, strlen(domain));
 283 }
 284
 285
 286 /* returns true on one of two conditions:
 287  * 1) the domain does not exist
 288  * 2) the domain exists and it's state is "joined" */
 289 static int dlm_wait_on_domain_helper(const char *domain)
 290 {
 291         int ret = 0;
 292         struct dlm_ctxt *tmp = NULL;
 293
 294         spin_lock(&dlm_domain_lock);
 295
 296         tmp = __dlm_lookup_domain(domain);
 297         if (!tmp)
 298                 ret = 1;
 299         else if (tmp->dlm_state == DLM_CTXT_JOINED)
 300                 ret = 1;
 301
 302         spin_unlock(&dlm_domain_lock);
 303         return ret;
 304 }
 305
 306 static void dlm_free_ctxt_mem(struct dlm_ctxt *dlm)
 307 {
 308         dlm_destroy_debugfs_subroot(dlm);
 309
 310         if (dlm->lockres_hash)
 311                 dlm_free_pagevec((void **)dlm->lockres_hash, DLM_HASH_PAGES);
 312
 313         if (dlm->master_hash)
 314                 dlm_free_pagevec((void **)dlm->master_hash, DLM_HASH_PAGES);
 315
 316         kfree(dlm->name);
 317         kfree(dlm);
 318 }
 319
 320 /* A little strange - this function will be called while holding
 321  * dlm_domain_lock and is expected to be holding it on the way out. We
 322  * will however drop and reacquire it multiple times */
 323 static void dlm_ctxt_release(struct kref *kref)
 324 {
 325         struct dlm_ctxt *dlm;
 326
 327         dlm = container_of(kref, struct dlm_ctxt, dlm_refs);
 328
 329         BUG_ON(dlm->num_joins);
 330         BUG_ON(dlm->dlm_state == DLM_CTXT_JOINED);
 331
 332         /* we may still be in the list if we hit an error during join. */
 333         list_del_init(&dlm->list);
 334
 335         spin_unlock(&dlm_domain_lock);
 336
 337         mlog(0, "freeing memory from domain %s\n", dlm->name);
 338
 339         wake_up(&dlm_domain_events);
 340
 341         dlm_free_ctxt_mem(dlm);
 342
 343         spin_lock(&dlm_domain_lock);
 344 }
 345
 346 void dlm_put(struct dlm_ctxt *dlm)
 347 {
 348         spin_lock(&dlm_domain_lock);
 349         kref_put(&dlm->dlm_refs, dlm_ctxt_release);
 350         spin_unlock(&dlm_domain_lock);
 351 }
 352
 353 static void __dlm_get(struct dlm_ctxt *dlm)
 354 {
 355         kref_get(&dlm->dlm_refs);
 356 }
 357
 358 /* given a questionable reference to a dlm object, gets a reference if
 359  * it can find it in the list, otherwise returns NULL in which case
 360  * you shouldn't trust your pointer. */
 361 struct dlm_ctxt *dlm_grab(struct dlm_ctxt *dlm)
 362 {
 363         struct dlm_ctxt *target;
 364         struct dlm_ctxt *ret = NULL;
 365
 366         spin_lock(&dlm_domain_lock);
 367
 368         list_for_each_entry(target, &dlm_domains, list) {
 369                 if (target == dlm) {
 370                         __dlm_get(target);
 371                         ret = target;
 372                         break;
 373                 }
 374         }
 375
 376         spin_unlock(&dlm_domain_lock);
 377
 378         return ret;
 379 }
 380
 381 int dlm_domain_fully_joined(struct dlm_ctxt *dlm)
 382 {
 383         int ret;
 384
 385         spin_lock(&dlm_domain_lock);
 386         ret = (dlm->dlm_state == DLM_CTXT_JOINED) ||
 387                 (dlm->dlm_state == DLM_CTXT_IN_SHUTDOWN);
 388         spin_unlock(&dlm_domain_lock);
 389
 390         return ret;
 391 }
 392
 393 static void dlm_destroy_dlm_worker(struct dlm_ctxt *dlm)
 394 {
 395         if (dlm->dlm_worker) {
 396                 flush_workqueue(dlm->dlm_worker);
 397                 destroy_workqueue(dlm->dlm_worker);
 398                 dlm->dlm_worker = NULL;
 399         }
 400 }
 401
 402 static void dlm_complete_dlm_shutdown(struct dlm_ctxt *dlm)
 403 {
 404         dlm_unregister_domain_handlers(dlm);
 405         dlm_debug_shutdown(dlm);
 406         dlm_complete_thread(dlm);
 407         dlm_complete_recovery_thread(dlm);
 408         dlm_destroy_dlm_worker(dlm);
 409
 410         /* We've left the domain. Now we can take ourselves out of the
 411          * list and allow the kref stuff to help us free the
 412          * memory. */
 413         spin_lock(&dlm_domain_lock);
 414         list_del_init(&dlm->list);
 415         spin_unlock(&dlm_domain_lock);
 416
 417         /* Wake up anyone waiting for us to remove this domain */
 418         wake_up(&dlm_domain_events);
 419 }
 420
 421 static int dlm_migrate_all_locks(struct dlm_ctxt *dlm)
 422 {
 423         int i, num, n, ret = 0;
 424         struct dlm_lock_resource *res;
 425         struct hlist_node *iter;
 426         struct hlist_head *bucket;
 427         int dropped;
 428
 429         mlog(0, "Migrating locks from domain %s\n", dlm->name);
 430
 431         num = 0;
 432         spin_lock(&dlm->spinlock);
 433         for (i = 0; i < DLM_HASH_BUCKETS; i++) {
 434 redo_bucket:
 435                 n = 0;
 436                 bucket = dlm_lockres_hash(dlm, i);
 437                 iter = bucket->first;
 438                 while (iter) {
 439                         n++;
 440                         res = hlist_entry(iter, struct dlm_lock_resource,
 441                                           hash_node);
 442                         dlm_lockres_get(res);
 443                         /* migrate, if necessary.  this will drop the dlm
 444                          * spinlock and retake it if it does migration. */
 445                         dropped = dlm_empty_lockres(dlm, res);
 446
 447                         spin_lock(&res->spinlock);
 448                         if (dropped)
 449                                 __dlm_lockres_calc_usage(dlm, res);
 450                         else
 451                                 iter = res->hash_node.next;
 452                         spin_unlock(&res->spinlock);
 453
 454                         dlm_lockres_put(res);
 455
 456                         if (dropped) {
 457                                 cond_resched_lock(&dlm->spinlock);
 458                                 goto redo_bucket;
 459                         }
 460                 }
 461                 cond_resched_lock(&dlm->spinlock);
 462                 num += n;
 463         }
 464         spin_unlock(&dlm->spinlock);
 465         wake_up(&dlm->dlm_thread_wq);
 466
 467         /* let the dlm thread take care of purging, keep scanning until
 468          * nothing remains in the hash */
 469         if (num) {
 470                 mlog(0, "%s: %d lock resources in hash last pass\n",
 471                      dlm->name, num);
 472                 ret = -EAGAIN;
 473         }
 474         mlog(0, "DONE Migrating locks from domain %s\n", dlm->name);
 475         return ret;
 476 }
 477
 478 static int dlm_no_joining_node(struct dlm_ctxt *dlm)
 479 {
 480         int ret;
 481
 482         spin_lock(&dlm->spinlock);
 483         ret = dlm->joining_node == DLM_LOCK_RES_OWNER_UNKNOWN;
 484         spin_unlock(&dlm->spinlock);
 485
 486         return ret;
 487 }
 488
 489 static int dlm_begin_exit_domain_handler(struct o2net_msg *msg, u32 len,
 490                                          void *data, void **ret_data)
 491 {
 492         struct dlm_ctxt *dlm = data;
 493         unsigned int node;
 494         struct dlm_exit_domain *exit_msg = (struct dlm_exit_domain *) msg->buf;
 495
 496         if (!dlm_grab(dlm))
 497                 return 0;
 498
 499         node = exit_msg->node_idx;
 500         mlog(0, "%s: Node %u sent a begin exit domain message\n", dlm->name, node);
 501
 502         spin_lock(&dlm->spinlock);
 503         set_bit(node, dlm->exit_domain_map);
 504         spin_unlock(&dlm->spinlock);
 505
 506         dlm_put(dlm);
 507
 508         return 0;
 509 }
 510
 511 static void dlm_mark_domain_leaving(struct dlm_ctxt *dlm)
 512 {
 513         /* Yikes, a double spinlock! I need domain_lock for the dlm
 514          * state and the dlm spinlock for join state... Sorry! */
 515 again:
 516         spin_lock(&dlm_domain_lock);
 517         spin_lock(&dlm->spinlock);
 518
 519         if (dlm->joining_node != DLM_LOCK_RES_OWNER_UNKNOWN) {
 520                 mlog(0, "Node %d is joining, we wait on it.\n",
 521                           dlm->joining_node);
 522                 spin_unlock(&dlm->spinlock);
 523                 spin_unlock(&dlm_domain_lock);
 524
 525                 wait_event(dlm->dlm_join_events, dlm_no_joining_node(dlm));
 526                 goto again;
 527         }
 528
 529         dlm->dlm_state = DLM_CTXT_LEAVING;
 530         spin_unlock(&dlm->spinlock);
 531         spin_unlock(&dlm_domain_lock);
 532 }
 533
 534 static void __dlm_print_nodes(struct dlm_ctxt *dlm)
 535 {
 536         int node = -1, num = 0;
 537
 538         assert_spin_locked(&dlm->spinlock);
 539
 540         printk("( ");
 541         while ((node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES,
 542                                      node + 1)) < O2NM_MAX_NODES) {
 543                 printk("%d ", node);
 544                 ++num;
 545         }
 546         printk(") %u nodes\n", num);
 547 }
 548
 549 static int dlm_exit_domain_handler(struct o2net_msg *msg, u32 len, void *data,
 550                                    void **ret_data)
 551 {
 552         struct dlm_ctxt *dlm = data;
 553         unsigned int node;
 554         struct dlm_exit_domain *exit_msg = (struct dlm_exit_domain *) msg->buf;
 555
 556         mlog(0, "%p %u %p", msg, len, data);
 557
 558         if (!dlm_grab(dlm))
 559                 return 0;
 560
 561         node = exit_msg->node_idx;
 562
 563         spin_lock(&dlm->spinlock);
 564         clear_bit(node, dlm->domain_map);
 565         clear_bit(node, dlm->exit_domain_map);
 566         printk(KERN_NOTICE "o2dlm: Node %u leaves domain %s ", node, dlm->name);
 567         __dlm_print_nodes(dlm);
 568
 569         /* notify anything attached to the heartbeat events */
 570         dlm_hb_event_notify_attached(dlm, node, 0);
 571
 572         spin_unlock(&dlm->spinlock);
 573
 574         dlm_put(dlm);
 575
 576         return 0;
 577 }
 578
 579 static int dlm_send_one_domain_exit(struct dlm_ctxt *dlm, u32 msg_type,
 580                                     unsigned int node)
 581 {
 582         int status;
 583         struct dlm_exit_domain leave_msg;
 584
 585         mlog(0, "%s: Sending domain exit message %u to node %u\n", dlm->name,
 586              msg_type, node);
 587
 588         memset(&leave_msg, 0, sizeof(leave_msg));
 589         leave_msg.node_idx = dlm->node_num;
 590
 591         status = o2net_send_message(msg_type, dlm->key, &leave_msg,
 592                                     sizeof(leave_msg), node, NULL);
 593         if (status < 0)
 594                 mlog(ML_ERROR, "Error %d sending domain exit message %u "
 595                      "to node %u on domain %s\n", status, msg_type, node,
 596                      dlm->name);
 597
 598         return status;
 599 }
 600
 601 static void dlm_begin_exit_domain(struct dlm_ctxt *dlm)
 602 {
 603         int node = -1;
 604
 605         /* Support for begin exit domain was added in 1.2 */
 606         if (dlm->dlm_locking_proto.pv_major == 1 &&
 607             dlm->dlm_locking_proto.pv_minor < 2)
 608                 return;
 609
 610         /*
 611          * Unlike DLM_EXIT_DOMAIN_MSG, DLM_BEGIN_EXIT_DOMAIN_MSG is purely
 612          * informational. Meaning if a node does not receive the message,
 613          * so be it.
 614          */
 615         spin_lock(&dlm->spinlock);
 616         while (1) {
 617                 node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES, node + 1);
 618                 if (node >= O2NM_MAX_NODES)
 619                         break;
 620                 if (node == dlm->node_num)
 621                         continue;
 622
 623                 spin_unlock(&dlm->spinlock);
 624                 dlm_send_one_domain_exit(dlm, DLM_BEGIN_EXIT_DOMAIN_MSG, node);
 625                 spin_lock(&dlm->spinlock);
 626         }
 627         spin_unlock(&dlm->spinlock);
 628 }
 629
 630 static void dlm_leave_domain(struct dlm_ctxt *dlm)
 631 {
 632         int node, clear_node, status;
 633
 634         /* At this point we've migrated away all our locks and won't
 635          * accept mastership of new ones. The dlm is responsible for
 636          * almost nothing now. We make sure not to confuse any joining
 637          * nodes and then commence shutdown procedure. */
 638
 639         spin_lock(&dlm->spinlock);
 640         /* Clear ourselves from the domain map */
 641         clear_bit(dlm->node_num, dlm->domain_map);
 642         while ((node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES,
 643                                      0)) < O2NM_MAX_NODES) {
 644                 /* Drop the dlm spinlock. This is safe wrt the domain_map.
 645                  * -nodes cannot be added now as the
 646                  *   query_join_handlers knows to respond with OK_NO_MAP
 647                  * -we catch the right network errors if a node is
 648                  *   removed from the map while we're sending him the
 649                  *   exit message. */
 650                 spin_unlock(&dlm->spinlock);
 651
 652                 clear_node = 1;
 653
 654                 status = dlm_send_one_domain_exit(dlm, DLM_EXIT_DOMAIN_MSG,
 655                                                   node);
 656                 if (status < 0 &&
 657                     status != -ENOPROTOOPT &&
 658                     status != -ENOTCONN) {
 659                         mlog(ML_NOTICE, "Error %d sending domain exit message "
 660                              "to node %d\n", status, node);
 661
 662                         /* Not sure what to do here but lets sleep for
 663                          * a bit in case this was a transient
 664                          * error... */
 665                         msleep(DLM_DOMAIN_BACKOFF_MS);
 666                         clear_node = 0;
 667                 }
 668
 669                 spin_lock(&dlm->spinlock);
 670                 /* If we're not clearing the node bit then we intend
 671                  * to loop back around to try again. */
 672                 if (clear_node)
 673                         clear_bit(node, dlm->domain_map);
 674         }
 675         spin_unlock(&dlm->spinlock);
 676 }
 677
 678 int dlm_shutting_down(struct dlm_ctxt *dlm)
 679 {
 680         int ret = 0;
 681
 682         spin_lock(&dlm_domain_lock);
 683
 684         if (dlm->dlm_state == DLM_CTXT_IN_SHUTDOWN)
 685                 ret = 1;
 686
 687         spin_unlock(&dlm_domain_lock);
 688
 689         return ret;
 690 }
 691
 692 void dlm_unregister_domain(struct dlm_ctxt *dlm)
 693 {
 694         int leave = 0;
 695         struct dlm_lock_resource *res;
 696
 697         spin_lock(&dlm_domain_lock);
 698         BUG_ON(dlm->dlm_state != DLM_CTXT_JOINED);
 699         BUG_ON(!dlm->num_joins);
 700
 701         dlm->num_joins--;
 702         if (!dlm->num_joins) {
 703                 /* We mark it "in shutdown" now so new register
 704                  * requests wait until we've completely left the
 705                  * domain. Don't use DLM_CTXT_LEAVING yet as we still
 706                  * want new domain joins to communicate with us at
 707                  * least until we've completed migration of our
 708                  * resources. */
 709                 dlm->dlm_state = DLM_CTXT_IN_SHUTDOWN;
 710                 leave = 1;
 711         }
 712         spin_unlock(&dlm_domain_lock);
 713
 714         if (leave) {
 715                 mlog(0, "shutting down domain %s\n", dlm->name);
 716                 dlm_begin_exit_domain(dlm);
 717
 718                 /* We changed dlm state, notify the thread */
 719                 dlm_kick_thread(dlm, NULL);
 720
 721                 while (dlm_migrate_all_locks(dlm)) {
 722                         /* Give dlm_thread time to purge the lockres' */
 723                         msleep(500);
 724                         mlog(0, "%s: more migration to do\n", dlm->name);
 725                 }
 726
 727                 /* This list should be empty. If not, print remaining lockres */
 728                 if (!list_empty(&dlm->tracking_list)) {
 729                         mlog(ML_ERROR, "Following lockres' are still on the "
 730                              "tracking list:\n");
 731                         list_for_each_entry(res, &dlm->tracking_list, tracking)
 732                                 dlm_print_one_lock_resource(res);
 733                 }
 734
 735                 dlm_mark_domain_leaving(dlm);
 736                 dlm_leave_domain(dlm);
 737                 printk(KERN_NOTICE "o2dlm: Leaving domain %s\n", dlm->name);
 738                 dlm_force_free_mles(dlm);
 739                 dlm_complete_dlm_shutdown(dlm);
 740         }
 741         dlm_put(dlm);
 742 }
 743 EXPORT_SYMBOL_GPL(dlm_unregister_domain);
 744
 745 static int dlm_query_join_proto_check(char *proto_type, int node,
 746                                       struct dlm_protocol_version *ours,
 747                                       struct dlm_protocol_version *request)
 748 {
 749         int rc;
 750         struct dlm_protocol_version proto = *request;
 751
 752         if (!dlm_protocol_compare(ours, &proto)) {
 753                 mlog(0,
 754                      "node %u wanted to join with %s locking protocol "
 755                      "%u.%u, we respond with %u.%u\n",
 756                      node, proto_type,
 757                      request->pv_major,
 758                      request->pv_minor,
 759                      proto.pv_major, proto.pv_minor);
 760                 request->pv_minor = proto.pv_minor;
 761                 rc = 0;
 762         } else {
 763                 mlog(ML_NOTICE,
 764                      "Node %u wanted to join with %s locking "
 765                      "protocol %u.%u, but we have %u.%u, disallowing\n",
 766                      node, proto_type,
 767                      request->pv_major,
 768                      request->pv_minor,
 769                      ours->pv_major,
 770                      ours->pv_minor);
 771                 rc = 1;
 772         }
 773
 774         return rc;
 775 }
 776
 777 /*
 778  * struct dlm_query_join_packet is made up of four one-byte fields.  They
 779  * are effectively in big-endian order already.  However, little-endian
 780  * machines swap them before putting the packet on the wire (because
 781  * query_join's response is a status, and that status is treated as a u32
 782  * on the wire).  Thus, a big-endian and little-endian machines will treat
 783  * this structure differently.
 784  *
 785  * The solution is to have little-endian machines swap the structure when
 786  * converting from the structure to the u32 representation.  This will
 787  * result in the structure having the correct format on the wire no matter
 788  * the host endian format.
 789  */
 790 static void dlm_query_join_packet_to_wire(struct dlm_query_join_packet *packet,
 791                                           u32 *wire)
 792 {
 793         union dlm_query_join_response response;
 794
 795         response.packet = *packet;
 796         *wire = be32_to_cpu(response.intval);
 797 }
 798
 799 static void dlm_query_join_wire_to_packet(u32 wire,
 800                                           struct dlm_query_join_packet *packet)
 801 {
 802         union dlm_query_join_response response;
 803
 804         response.intval = cpu_to_be32(wire);
 805         *packet = response.packet;
 806 }
 807
 808 static int dlm_query_join_handler(struct o2net_msg *msg, u32 len, void *data,
 809                                   void **ret_data)
 810 {
 811         struct dlm_query_join_request *query;
 812         struct dlm_query_join_packet packet = {
 813                 .code = JOIN_DISALLOW,
 814         };
 815         struct dlm_ctxt *dlm = NULL;
 816         u32 response;
 817         u8 nodenum;
 818
 819         query = (struct dlm_query_join_request *) msg->buf;
 820
 821         mlog(0, "node %u wants to join domain %s\n", query->node_idx,
 822                   query->domain);
 823
 824         /*
 825          * If heartbeat doesn't consider the node live, tell it
 826          * to back off and try again.  This gives heartbeat a chance
 827          * to catch up.
 828          */
 829         if (!o2hb_check_node_heartbeating_no_sem(query->node_idx)) {
 830                 mlog(0, "node %u is not in our live map yet\n",
 831                      query->node_idx);
 832
 833                 packet.code = JOIN_DISALLOW;
 834                 goto respond;
 835         }
 836
 837         packet.code = JOIN_OK_NO_MAP;
 838
 839         spin_lock(&dlm_domain_lock);
 840         dlm = __dlm_lookup_domain_full(query->domain, query->name_len);
 841         if (!dlm)
 842                 goto unlock_respond;
 843
 844         /*
 845          * There is a small window where the joining node may not see the
 846          * node(s) that just left but still part of the cluster. DISALLOW
 847          * join request if joining node has different node map.
 848          */
 849         nodenum=0;
 850         while (nodenum < O2NM_MAX_NODES) {
 851                 if (test_bit(nodenum, dlm->domain_map)) {
 852                         if (!byte_test_bit(nodenum, query->node_map)) {
 853                                 mlog(0, "disallow join as node %u does not "
 854                                      "have node %u in its nodemap\n",
 855                                      query->node_idx, nodenum);
 856                                 packet.code = JOIN_DISALLOW;
 857                                 goto unlock_respond;
 858                         }
 859                 }
 860                 nodenum++;
 861         }
 862
 863         /* Once the dlm ctxt is marked as leaving then we don't want
 864          * to be put in someone's domain map.
 865          * Also, explicitly disallow joining at certain troublesome
 866          * times (ie. during recovery). */
 867         if (dlm->dlm_state != DLM_CTXT_LEAVING) {
 868                 int bit = query->node_idx;
 869                 spin_lock(&dlm->spinlock);
 870
 871                 if (dlm->dlm_state == DLM_CTXT_NEW &&
 872                     dlm->joining_node == DLM_LOCK_RES_OWNER_UNKNOWN) {
 873                         /*If this is a brand new context and we
 874                          * haven't started our join process yet, then
 875                          * the other node won the race. */
 876                         packet.code = JOIN_OK_NO_MAP;
 877                 } else if (dlm->joining_node != DLM_LOCK_RES_OWNER_UNKNOWN) {
 878                         /* Disallow parallel joins. */
 879                         packet.code = JOIN_DISALLOW;
 880                 } else if (dlm->reco.state & DLM_RECO_STATE_ACTIVE) {
 881                         mlog(0, "node %u trying to join, but recovery "
 882                              "is ongoing.\n", bit);
 883                         packet.code = JOIN_DISALLOW;
 884                 } else if (test_bit(bit, dlm->recovery_map)) {
 885                         mlog(0, "node %u trying to join, but it "
 886                              "still needs recovery.\n", bit);
 887                         packet.code = JOIN_DISALLOW;
 888                 } else if (test_bit(bit, dlm->domain_map)) {
 889                         mlog(0, "node %u trying to join, but it "
 890                              "is still in the domain! needs recovery?\n",
 891                              bit);
 892                         packet.code = JOIN_DISALLOW;
 893                 } else {
 894                         /* Alright we're fully a part of this domain
 895                          * so we keep some state as to who's joining
 896                          * and indicate to him that needs to be fixed
 897                          * up. */
 898
 899                         /* Make sure we speak compatible locking protocols.  */
 900                         if (dlm_query_join_proto_check("DLM", bit,
 901                                                        &dlm->dlm_locking_proto,
 902                                                        &query->dlm_proto)) {
 903                                 packet.code = JOIN_PROTOCOL_MISMATCH;
 904                         } else if (dlm_query_join_proto_check("fs", bit,
 905                                                               &dlm->fs_locking_proto,
 906                                                               &query->fs_proto)) {
 907                                 packet.code = JOIN_PROTOCOL_MISMATCH;
 908                         } else {
 909                                 packet.dlm_minor = query->dlm_proto.pv_minor;
 910                                 packet.fs_minor = query->fs_proto.pv_minor;
 911                                 packet.code = JOIN_OK;
 912                                 __dlm_set_joining_node(dlm, query->node_idx);
 913                         }
 914                 }
 915
 916                 spin_unlock(&dlm->spinlock);
 917         }
 918 unlock_respond:
 919         spin_unlock(&dlm_domain_lock);
 920
 921 respond:
 922         mlog(0, "We respond with %u\n", packet.code);
 923
 924         dlm_query_join_packet_to_wire(&packet, &response);
 925         return response;
 926 }
 927
 928 static int dlm_assert_joined_handler(struct o2net_msg *msg, u32 len, void *data,
 929                                      void **ret_data)
 930 {
 931         struct dlm_assert_joined *assert;
 932         struct dlm_ctxt *dlm = NULL;
 933
 934         assert = (struct dlm_assert_joined *) msg->buf;
 935
 936         mlog(0, "node %u asserts join on domain %s\n", assert->node_idx,
 937                   assert->domain);
 938
 939         spin_lock(&dlm_domain_lock);
 940         dlm = __dlm_lookup_domain_full(assert->domain, assert->name_len);
 941         /* XXX should we consider no dlm ctxt an error? */
 942         if (dlm) {
 943                 spin_lock(&dlm->spinlock);
 944
 945                 /* Alright, this node has officially joined our
 946                  * domain. Set him in the map and clean up our
 947                  * leftover join state. */
 948                 BUG_ON(dlm->joining_node != assert->node_idx);
 949
 950                 if (dlm->reco.state & DLM_RECO_STATE_ACTIVE) {
 951                         mlog(0, "dlm recovery is ongoing, disallow join\n");
 952                         spin_unlock(&dlm->spinlock);
 953                         spin_unlock(&dlm_domain_lock);
 954                         return -EAGAIN;
 955                 }
 956
 957                 set_bit(assert->node_idx, dlm->domain_map);
 958                 clear_bit(assert->node_idx, dlm->exit_domain_map);
 959                 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
 960
 961                 printk(KERN_NOTICE "o2dlm: Node %u joins domain %s ",
 962                        assert->node_idx, dlm->name);
 963                 __dlm_print_nodes(dlm);
 964
 965                 /* notify anything attached to the heartbeat events */
 966                 dlm_hb_event_notify_attached(dlm, assert->node_idx, 1);
 967
 968                 spin_unlock(&dlm->spinlock);
 969         }
 970         spin_unlock(&dlm_domain_lock);
 971
 972         return 0;
 973 }
 974
 975 static int dlm_match_regions(struct dlm_ctxt *dlm,
 976                              struct dlm_query_region *qr,
 977                              char *local, int locallen)
 978 {
 979         char *remote = qr->qr_regions;
 980         char *l, *r;
 981         int localnr, i, j, foundit;
 982         int status = 0;
 983
 984         if (!o2hb_global_heartbeat_active()) {
 985                 if (qr->qr_numregions) {
 986                         mlog(ML_ERROR, "Domain %s: Joining node %d has global "
 987                              "heartbeat enabled but local node %d does not\n",
 988                              qr->qr_domain, qr->qr_node, dlm->node_num);
 989                         status = -EINVAL;
 990                 }
 991                 goto bail;
 992         }
 993
 994         if (o2hb_global_heartbeat_active() && !qr->qr_numregions) {
 995                 mlog(ML_ERROR, "Domain %s: Local node %d has global "
 996                      "heartbeat enabled but joining node %d does not\n",
 997                      qr->qr_domain, dlm->node_num, qr->qr_node);
 998                 status = -EINVAL;
 999                 goto bail;
1000         }
1001
1002         r = remote;
1003         for (i = 0; i < qr->qr_numregions; ++i) {
1004                 mlog(0, "Region %.*s\n", O2HB_MAX_REGION_NAME_LEN, r);
1005                 r += O2HB_MAX_REGION_NAME_LEN;
1006         }
1007
1008         localnr = min(O2NM_MAX_REGIONS, locallen/O2HB_MAX_REGION_NAME_LEN);
1009         localnr = o2hb_get_all_regions(local, (u8)localnr);
1010
1011         /* compare local regions with remote */
1012         l = local;
1013         for (i = 0; i < localnr; ++i) {
1014                 foundit = 0;
1015                 r = remote;
1016                 for (j = 0; j <= qr->qr_numregions; ++j) {
1017                         if (!memcmp(l, r, O2HB_MAX_REGION_NAME_LEN)) {
1018                                 foundit = 1;
1019                                 break;
1020                         }
1021                         r += O2HB_MAX_REGION_NAME_LEN;
1022                 }
1023                 if (!foundit) {
1024                         status = -EINVAL;
1025                         mlog(ML_ERROR, "Domain %s: Region '%.*s' registered "
1026                              "in local node %d but not in joining node %d\n",
1027                              qr->qr_domain, O2HB_MAX_REGION_NAME_LEN, l,
1028                              dlm->node_num, qr->qr_node);
1029                         goto bail;
1030                 }
1031                 l += O2HB_MAX_REGION_NAME_LEN;
1032         }
1033
1034         /* compare remote with local regions */
1035         r = remote;
1036         for (i = 0; i < qr->qr_numregions; ++i) {
1037                 foundit = 0;
1038                 l = local;
1039                 for (j = 0; j < localnr; ++j) {
1040                         if (!memcmp(r, l, O2HB_MAX_REGION_NAME_LEN)) {
1041                                 foundit = 1;
1042                                 break;
1043                         }
1044                         l += O2HB_MAX_REGION_NAME_LEN;
1045                 }
1046                 if (!foundit) {
1047                         status = -EINVAL;
1048                         mlog(ML_ERROR, "Domain %s: Region '%.*s' registered "
1049                              "in joining node %d but not in local node %d\n",
1050                              qr->qr_domain, O2HB_MAX_REGION_NAME_LEN, r,
1051                              qr->qr_node, dlm->node_num);
1052                         goto bail;
1053                 }
1054                 r += O2HB_MAX_REGION_NAME_LEN;
1055         }
1056
1057 bail:
1058         return status;
1059 }
1060
1061 static int dlm_send_regions(struct dlm_ctxt *dlm, unsigned long *node_map)
1062 {
1063         struct dlm_query_region *qr = NULL;
1064         int status, ret = 0, i;
1065         char *p;
1066
1067         if (find_next_bit(node_map, O2NM_MAX_NODES, 0) >= O2NM_MAX_NODES)
1068                 goto bail;
1069
1070         qr = kzalloc(sizeof(struct dlm_query_region), GFP_KERNEL);
1071         if (!qr) {
1072                 ret = -ENOMEM;
1073                 mlog_errno(ret);
1074                 goto bail;
1075         }
1076
1077         qr->qr_node = dlm->node_num;
1078         qr->qr_namelen = strlen(dlm->name);
1079         memcpy(qr->qr_domain, dlm->name, qr->qr_namelen);
1080         /* if local hb, the numregions will be zero */
1081         if (o2hb_global_heartbeat_active())
1082                 qr->qr_numregions = o2hb_get_all_regions(qr->qr_regions,
1083                                                          O2NM_MAX_REGIONS);
1084
1085         p = qr->qr_regions;
1086         for (i = 0; i < qr->qr_numregions; ++i, p += O2HB_MAX_REGION_NAME_LEN)
1087                 mlog(0, "Region %.*s\n", O2HB_MAX_REGION_NAME_LEN, p);
1088
1089         i = -1;
1090         while ((i = find_next_bit(node_map, O2NM_MAX_NODES,
1091                                   i + 1)) < O2NM_MAX_NODES) {
1092                 if (i == dlm->node_num)
1093                         continue;
1094
1095                 mlog(0, "Sending regions to node %d\n", i);
1096
1097                 ret = o2net_send_message(DLM_QUERY_REGION, DLM_MOD_KEY, qr,
1098                                          sizeof(struct dlm_query_region),
1099                                          i, &status);
1100                 if (ret >= 0)
1101                         ret = status;
1102                 if (ret) {
1103                         mlog(ML_ERROR, "Region mismatch %d, node %d\n",
1104                              ret, i);
1105                         break;
1106                 }
1107         }
1108
1109 bail:
1110         kfree(qr);
1111         return ret;
1112 }
1113
1114 static int dlm_query_region_handler(struct o2net_msg *msg, u32 len,
1115                                     void *data, void **ret_data)
1116 {
1117         struct dlm_query_region *qr;
1118         struct dlm_ctxt *dlm = NULL;
1119         char *local = NULL;
1120         int status = 0;
1121
1122         qr = (struct dlm_query_region *) msg->buf;
1123
1124         mlog(0, "Node %u queries hb regions on domain %s\n", qr->qr_node,
1125              qr->qr_domain);
1126
1127         /* buffer used in dlm_mast_regions() */
1128         local = kmalloc(sizeof(qr->qr_regions), GFP_KERNEL);
1129         if (!local)
1130                 return -ENOMEM;
1131
1132         status = -EINVAL;
1133
1134         spin_lock(&dlm_domain_lock);
1135         dlm = __dlm_lookup_domain_full(qr->qr_domain, qr->qr_namelen);
1136         if (!dlm) {
1137                 mlog(ML_ERROR, "Node %d queried hb regions on domain %s "
1138                      "before join domain\n", qr->qr_node, qr->qr_domain);
1139                 goto out_domain_lock;
1140         }
1141
1142         spin_lock(&dlm->spinlock);
1143         if (dlm->joining_node != qr->qr_node) {
1144                 mlog(ML_ERROR, "Node %d queried hb regions on domain %s "
1145                      "but joining node is %d\n", qr->qr_node, qr->qr_domain,
1146                      dlm->joining_node);
1147                 goto out_dlm_lock;
1148         }
1149
1150         /* Support for global heartbeat was added in 1.1 */
1151         if (dlm->dlm_locking_proto.pv_major == 1 &&
1152             dlm->dlm_locking_proto.pv_minor == 0) {
1153                 mlog(ML_ERROR, "Node %d queried hb regions on domain %s "
1154                      "but active dlm protocol is %d.%d\n", qr->qr_node,
1155                      qr->qr_domain, dlm->dlm_locking_proto.pv_major,
1156                      dlm->dlm_locking_proto.pv_minor);
1157                 goto out_dlm_lock;
1158         }
1159
1160         status = dlm_match_regions(dlm, qr, local, sizeof(qr->qr_regions));
1161
1162 out_dlm_lock:
1163         spin_unlock(&dlm->spinlock);
1164
1165 out_domain_lock:
1166         spin_unlock(&dlm_domain_lock);
1167
1168         kfree(local);
1169
1170         return status;
1171 }
1172
1173 static int dlm_match_nodes(struct dlm_ctxt *dlm, struct dlm_query_nodeinfo *qn)
1174 {
1175         struct o2nm_node *local;
1176         struct dlm_node_info *remote;
1177         int i, j;
1178         int status = 0;
1179
1180         for (j = 0; j < qn->qn_numnodes; ++j)
1181                 mlog(0, "Node %3d, %pI4:%u\n", qn->qn_nodes[j].ni_nodenum,
1182                      &(qn->qn_nodes[j].ni_ipv4_address),
1183                      ntohs(qn->qn_nodes[j].ni_ipv4_port));
1184
1185         for (i = 0; i < O2NM_MAX_NODES && !status; ++i) {
1186                 local = o2nm_get_node_by_num(i);
1187                 remote = NULL;
1188                 for (j = 0; j < qn->qn_numnodes; ++j) {
1189                         if (qn->qn_nodes[j].ni_nodenum == i) {
1190                                 remote = &(qn->qn_nodes[j]);
1191                                 break;
1192                         }
1193                 }
1194
1195                 if (!local && !remote)
1196                         continue;
1197
1198                 if ((local && !remote) || (!local && remote))
1199                         status = -EINVAL;
1200
1201                 if (!status &&
1202                     ((remote->ni_nodenum != local->nd_num) ||
1203                      (remote->ni_ipv4_port != local->nd_ipv4_port) ||
1204                      (remote->ni_ipv4_address != local->nd_ipv4_address)))
1205                         status = -EINVAL;
1206
1207                 if (status) {
1208                         if (remote && !local)
1209                                 mlog(ML_ERROR, "Domain %s: Node %d (%pI4:%u) "
1210                                      "registered in joining node %d but not in "
1211                                      "local node %d\n", qn->qn_domain,
1212                                      remote->ni_nodenum,
1213                                      &(remote->ni_ipv4_address),
1214                                      ntohs(remote->ni_ipv4_port),
1215                                      qn->qn_nodenum, dlm->node_num);
1216                         if (local && !remote)
1217                                 mlog(ML_ERROR, "Domain %s: Node %d (%pI4:%u) "
1218                                      "registered in local node %d but not in "
1219                                      "joining node %d\n", qn->qn_domain,
1220                                      local->nd_num, &(local->nd_ipv4_address),
1221                                      ntohs(local->nd_ipv4_port),
1222                                      dlm->node_num, qn->qn_nodenum);
1223                         BUG_ON((!local && !remote));
1224                 }
1225
1226                 if (local)
1227                         o2nm_node_put(local);
1228         }
1229
1230         return status;
1231 }
1232
1233 static int dlm_send_nodeinfo(struct dlm_ctxt *dlm, unsigned long *node_map)
1234 {
1235         struct dlm_query_nodeinfo *qn = NULL;
1236         struct o2nm_node *node;
1237         int ret = 0, status, count, i;
1238
1239         if (find_next_bit(node_map, O2NM_MAX_NODES, 0) >= O2NM_MAX_NODES)
1240                 goto bail;
1241
1242         qn = kzalloc(sizeof(struct dlm_query_nodeinfo), GFP_KERNEL);
1243         if (!qn) {
1244                 ret = -ENOMEM;
1245                 mlog_errno(ret);
1246                 goto bail;
1247         }
1248
1249         for (i = 0, count = 0; i < O2NM_MAX_NODES; ++i) {
1250                 node = o2nm_get_node_by_num(i);
1251                 if (!node)
1252                         continue;
1253                 qn->qn_nodes[count].ni_nodenum = node->nd_num;
1254                 qn->qn_nodes[count].ni_ipv4_port = node->nd_ipv4_port;
1255                 qn->qn_nodes[count].ni_ipv4_address = node->nd_ipv4_address;
1256                 mlog(0, "Node %3d, %pI4:%u\n", node->nd_num,
1257                      &(node->nd_ipv4_address), ntohs(node->nd_ipv4_port));
1258                 ++count;
1259                 o2nm_node_put(node);
1260         }
1261
1262         qn->qn_nodenum = dlm->node_num;
1263         qn->qn_numnodes = count;
1264         qn->qn_namelen = strlen(dlm->name);
1265         memcpy(qn->qn_domain, dlm->name, qn->qn_namelen);
1266
1267         i = -1;
1268         while ((i = find_next_bit(node_map, O2NM_MAX_NODES,
1269                                   i + 1)) < O2NM_MAX_NODES) {
1270                 if (i == dlm->node_num)
1271                         continue;
1272
1273                 mlog(0, "Sending nodeinfo to node %d\n", i);
1274
1275                 ret = o2net_send_message(DLM_QUERY_NODEINFO, DLM_MOD_KEY,
1276                                          qn, sizeof(struct dlm_query_nodeinfo),
1277                                          i, &status);
1278                 if (ret >= 0)
1279                         ret = status;
1280                 if (ret) {
1281                         mlog(ML_ERROR, "node mismatch %d, node %d\n", ret, i);
1282                         break;
1283                 }
1284         }
1285
1286 bail:
1287         kfree(qn);
1288         return ret;
1289 }
1290
1291 static int dlm_query_nodeinfo_handler(struct o2net_msg *msg, u32 len,
1292                                       void *data, void **ret_data)
1293 {
1294         struct dlm_query_nodeinfo *qn;
1295         struct dlm_ctxt *dlm = NULL;
1296         int locked = 0, status = -EINVAL;
1297
1298         qn = (struct dlm_query_nodeinfo *) msg->buf;
1299
1300         mlog(0, "Node %u queries nodes on domain %s\n", qn->qn_nodenum,
1301              qn->qn_domain);
1302
1303         spin_lock(&dlm_domain_lock);
1304         dlm = __dlm_lookup_domain_full(qn->qn_domain, qn->qn_namelen);
1305         if (!dlm) {
1306                 mlog(ML_ERROR, "Node %d queried nodes on domain %s before "
1307                      "join domain\n", qn->qn_nodenum, qn->qn_domain);
1308                 goto bail;
1309         }
1310
1311         spin_lock(&dlm->spinlock);
1312         locked = 1;
1313         if (dlm->joining_node != qn->qn_nodenum) {
1314                 mlog(ML_ERROR, "Node %d queried nodes on domain %s but "
1315                      "joining node is %d\n", qn->qn_nodenum, qn->qn_domain,
1316                      dlm->joining_node);
1317                 goto bail;
1318         }
1319
1320         /* Support for node query was added in 1.1 */
1321         if (dlm->dlm_locking_proto.pv_major == 1 &&
1322             dlm->dlm_locking_proto.pv_minor == 0) {
1323                 mlog(ML_ERROR, "Node %d queried nodes on domain %s "
1324                      "but active dlm protocol is %d.%d\n", qn->qn_nodenum,
1325                      qn->qn_domain, dlm->dlm_locking_proto.pv_major,
1326                      dlm->dlm_locking_proto.pv_minor);
1327                 goto bail;
1328         }
1329
1330         status = dlm_match_nodes(dlm, qn);
1331
1332 bail:
1333         if (locked)
1334                 spin_unlock(&dlm->spinlock);
1335         spin_unlock(&dlm_domain_lock);
1336
1337         return status;
1338 }
1339
1340 static int dlm_cancel_join_handler(struct o2net_msg *msg, u32 len, void *data,
1341                                    void **ret_data)
1342 {
1343         struct dlm_cancel_join *cancel;
1344         struct dlm_ctxt *dlm = NULL;
1345
1346         cancel = (struct dlm_cancel_join *) msg->buf;
1347
1348         mlog(0, "node %u cancels join on domain %s\n", cancel->node_idx,
1349                   cancel->domain);
1350
1351         spin_lock(&dlm_domain_lock);
1352         dlm = __dlm_lookup_domain_full(cancel->domain, cancel->name_len);
1353
1354         if (dlm) {
1355                 spin_lock(&dlm->spinlock);
1356
1357                 /* Yikes, this guy wants to cancel his join. No
1358                  * problem, we simply cleanup our join state. */
1359                 BUG_ON(dlm->joining_node != cancel->node_idx);
1360                 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
1361
1362                 spin_unlock(&dlm->spinlock);
1363         }
1364         spin_unlock(&dlm_domain_lock);
1365
1366         return 0;
1367 }
1368
1369 static int dlm_send_one_join_cancel(struct dlm_ctxt *dlm,
1370                                     unsigned int node)
1371 {
1372         int status;
1373         struct dlm_cancel_join cancel_msg;
1374
1375         memset(&cancel_msg, 0, sizeof(cancel_msg));
1376         cancel_msg.node_idx = dlm->node_num;
1377         cancel_msg.name_len = strlen(dlm->name);
1378         memcpy(cancel_msg.domain, dlm->name, cancel_msg.name_len);
1379
1380         status = o2net_send_message(DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY,
1381                                     &cancel_msg, sizeof(cancel_msg), node,
1382                                     NULL);
1383         if (status < 0) {
1384                 mlog(ML_ERROR, "Error %d when sending message %u (key 0x%x) to "
1385                      "node %u\n", status, DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY,
1386                      node);
1387                 goto bail;
1388         }
1389
1390 bail:
1391         return status;
1392 }
1393
1394 /* map_size should be in bytes. */
1395 static int dlm_send_join_cancels(struct dlm_ctxt *dlm,
1396                                  unsigned long *node_map,
1397                                  unsigned int map_size)
1398 {
1399         int status, tmpstat;
1400         int node;
1401
1402         if (map_size != (BITS_TO_LONGS(O2NM_MAX_NODES) *
1403                          sizeof(unsigned long))) {
1404                 mlog(ML_ERROR,
1405                      "map_size %u != BITS_TO_LONGS(O2NM_MAX_NODES) %u\n",
1406                      map_size, (unsigned)BITS_TO_LONGS(O2NM_MAX_NODES));
1407                 return -EINVAL;
1408         }
1409
1410         status = 0;
1411         node = -1;
1412         while ((node = find_next_bit(node_map, O2NM_MAX_NODES,
1413                                      node + 1)) < O2NM_MAX_NODES) {
1414                 if (node == dlm->node_num)
1415                         continue;
1416
1417                 tmpstat = dlm_send_one_join_cancel(dlm, node);
1418                 if (tmpstat) {
1419                         mlog(ML_ERROR, "Error return %d cancelling join on "
1420                              "node %d\n", tmpstat, node);
1421                         if (!status)
1422                                 status = tmpstat;
1423                 }
1424         }
1425
1426         if (status)
1427                 mlog_errno(status);
1428         return status;
1429 }
1430
1431 static int dlm_request_join(struct dlm_ctxt *dlm,
1432                             int node,
1433                             enum dlm_query_join_response_code *response)
1434 {
1435         int status;
1436         struct dlm_query_join_request join_msg;
1437         struct dlm_query_join_packet packet;
1438         u32 join_resp;
1439
1440         mlog(0, "querying node %d\n", node);
1441
1442         memset(&join_msg, 0, sizeof(join_msg));
1443         join_msg.node_idx = dlm->node_num;
1444         join_msg.name_len = strlen(dlm->name);
1445         memcpy(join_msg.domain, dlm->name, join_msg.name_len);
1446         join_msg.dlm_proto = dlm->dlm_locking_proto;
1447         join_msg.fs_proto = dlm->fs_locking_proto;
1448
1449         /* copy live node map to join message */
1450         byte_copymap(join_msg.node_map, dlm->live_nodes_map, O2NM_MAX_NODES);
1451
1452         status = o2net_send_message(DLM_QUERY_JOIN_MSG, DLM_MOD_KEY, &join_msg,
1453                                     sizeof(join_msg), node, &join_resp);
1454         if (status < 0 && status != -ENOPROTOOPT) {
1455                 mlog(ML_ERROR, "Error %d when sending message %u (key 0x%x) to "
1456                      "node %u\n", status, DLM_QUERY_JOIN_MSG, DLM_MOD_KEY,
1457                      node);
1458                 goto bail;
1459         }
1460         dlm_query_join_wire_to_packet(join_resp, &packet);
1461
1462         /* -ENOPROTOOPT from the net code means the other side isn't
1463             listening for our message type -- that's fine, it means
1464             his dlm isn't up, so we can consider him a 'yes' but not
1465             joined into the domain.  */
1466         if (status == -ENOPROTOOPT) {
1467                 status = 0;
1468                 *response = JOIN_OK_NO_MAP;
1469         } else {
1470                 *response = packet.code;
1471                 switch (packet.code) {
1472                 case JOIN_DISALLOW:
1473                 case JOIN_OK_NO_MAP:
1474                         break;
1475                 case JOIN_PROTOCOL_MISMATCH:
1476                         mlog(ML_NOTICE,
1477                              "This node requested DLM locking protocol %u.%u and "
1478                              "filesystem locking protocol %u.%u.  At least one of "
1479                              "the protocol versions on node %d is not compatible, "
1480                              "disconnecting\n",
1481                              dlm->dlm_locking_proto.pv_major,
1482                              dlm->dlm_locking_proto.pv_minor,
1483                              dlm->fs_locking_proto.pv_major,
1484                              dlm->fs_locking_proto.pv_minor,
1485                              node);
1486                         status = -EPROTO;
1487                         break;
1488                 case JOIN_OK:
1489                         /* Use the same locking protocol as the remote node */
1490                         dlm->dlm_locking_proto.pv_minor = packet.dlm_minor;
1491                         dlm->fs_locking_proto.pv_minor = packet.fs_minor;
1492                         mlog(0,
1493                              "Node %d responds JOIN_OK with DLM locking protocol "
1494                              "%u.%u and fs locking protocol %u.%u\n",
1495                              node,
1496                              dlm->dlm_locking_proto.pv_major,
1497                              dlm->dlm_locking_proto.pv_minor,
1498                              dlm->fs_locking_proto.pv_major,
1499                              dlm->fs_locking_proto.pv_minor);
1500                         break;
1501                 default:
1502                         status = -EINVAL;
1503                         mlog(ML_ERROR, "invalid response %d from node %u\n",
1504                              packet.code, node);
1505                         /* Reset response to JOIN_DISALLOW */
1506                         *response = JOIN_DISALLOW;
1507                         break;
1508                 }
1509         }
1510
1511         mlog(0, "status %d, node %d response is %d\n", status, node,
1512              *response);
1513
1514 bail:
1515         return status;
1516 }
1517
1518 static int dlm_send_one_join_assert(struct dlm_ctxt *dlm,
1519                                     unsigned int node)
1520 {
1521         int status;
1522         int ret;
1523         struct dlm_assert_joined assert_msg;
1524
1525         mlog(0, "Sending join assert to node %u\n", node);
1526
1527         memset(&assert_msg, 0, sizeof(assert_msg));
1528         assert_msg.node_idx = dlm->node_num;
1529         assert_msg.name_len = strlen(dlm->name);
1530         memcpy(assert_msg.domain, dlm->name, assert_msg.name_len);
1531
1532         status = o2net_send_message(DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY,
1533                                     &assert_msg, sizeof(assert_msg), node,
1534                                     &ret);
1535         if (status < 0)
1536                 mlog(ML_ERROR, "Error %d when sending message %u (key 0x%x) to "
1537                      "node %u\n", status, DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY,
1538                      node);
1539         else
1540                 status = ret;
1541
1542         return status;
1543 }
1544
1545 static void dlm_send_join_asserts(struct dlm_ctxt *dlm,
1546                                   unsigned long *node_map)
1547 {
1548         int status, node, live;
1549
1550         status = 0;
1551         node = -1;
1552         while ((node = find_next_bit(node_map, O2NM_MAX_NODES,
1553                                      node + 1)) < O2NM_MAX_NODES) {
1554                 if (node == dlm->node_num)
1555                         continue;
1556
1557                 do {
1558                         /* It is very important that this message be
1559                          * received so we spin until either the node
1560                          * has died or it gets the message. */
1561                         status = dlm_send_one_join_assert(dlm, node);
1562
1563                         spin_lock(&dlm->spinlock);
1564                         live = test_bit(node, dlm->live_nodes_map);
1565                         spin_unlock(&dlm->spinlock);
1566
1567                         if (status) {
1568                                 mlog(ML_ERROR, "Error return %d asserting "
1569                                      "join on node %d\n", status, node);
1570
1571                                 /* give us some time between errors... */
1572                                 if (live)
1573                                         msleep(DLM_DOMAIN_BACKOFF_MS);
1574                         }
1575                 } while (status && live);
1576         }
1577 }
1578
1579 struct domain_join_ctxt {
1580         unsigned long live_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
1581         unsigned long yes_resp_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
1582 };
1583
1584 static int dlm_should_restart_join(struct dlm_ctxt *dlm,
1585                                    struct domain_join_ctxt *ctxt,
1586                                    enum dlm_query_join_response_code response)
1587 {
1588         int ret;
1589
1590         if (response == JOIN_DISALLOW) {
1591                 mlog(0, "Latest response of disallow -- should restart\n");
1592                 return 1;
1593         }
1594
1595         spin_lock(&dlm->spinlock);
1596         /* For now, we restart the process if the node maps have
1597          * changed at all */
1598         ret = memcmp(ctxt->live_map, dlm->live_nodes_map,
1599                      sizeof(dlm->live_nodes_map));
1600         spin_unlock(&dlm->spinlock);
1601
1602         if (ret)
1603                 mlog(0, "Node maps changed -- should restart\n");
1604
1605         return ret;
1606 }
1607
1608 static int dlm_try_to_join_domain(struct dlm_ctxt *dlm)
1609 {
1610         int status = 0, tmpstat, node;
1611         struct domain_join_ctxt *ctxt;
1612         enum dlm_query_join_response_code response = JOIN_DISALLOW;
1613
1614         mlog(0, "%p", dlm);
1615
1616         ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL);
1617         if (!ctxt) {
1618                 status = -ENOMEM;
1619                 mlog_errno(status);
1620                 goto bail;
1621         }
1622
1623         /* group sem locking should work for us here -- we're already
1624          * registered for heartbeat events so filling this should be
1625          * atomic wrt getting those handlers called. */
1626         o2hb_fill_node_map(dlm->live_nodes_map, sizeof(dlm->live_nodes_map));
1627
1628         spin_lock(&dlm->spinlock);
1629         memcpy(ctxt->live_map, dlm->live_nodes_map, sizeof(ctxt->live_map));
1630
1631         __dlm_set_joining_node(dlm, dlm->node_num);
1632
1633         spin_unlock(&dlm->spinlock);
1634
1635         node = -1;
1636         while ((node = find_next_bit(ctxt->live_map, O2NM_MAX_NODES,
1637                                      node + 1)) < O2NM_MAX_NODES) {
1638                 if (node == dlm->node_num)
1639                         continue;
1640
1641                 status = dlm_request_join(dlm, node, &response);
1642                 if (status < 0) {
1643                         mlog_errno(status);
1644                         goto bail;
1645                 }
1646
1647                 /* Ok, either we got a response or the node doesn't have a
1648                  * dlm up. */
1649                 if (response == JOIN_OK)
1650                         set_bit(node, ctxt->yes_resp_map);
1651
1652                 if (dlm_should_restart_join(dlm, ctxt, response)) {
1653                         status = -EAGAIN;
1654                         goto bail;
1655                 }
1656         }
1657
1658         mlog(0, "Yay, done querying nodes!\n");
1659
1660         /* Yay, everyone agree's we can join the domain. My domain is
1661          * comprised of all nodes who were put in the
1662          * yes_resp_map. Copy that into our domain map and send a join
1663          * assert message to clean up everyone elses state. */
1664         spin_lock(&dlm->spinlock);
1665         memcpy(dlm->domain_map, ctxt->yes_resp_map,
1666                sizeof(ctxt->yes_resp_map));
1667         set_bit(dlm->node_num, dlm->domain_map);
1668         spin_unlock(&dlm->spinlock);
1669
1670         /* Support for global heartbeat and node info was added in 1.1 */
1671         if (dlm->dlm_locking_proto.pv_major > 1 ||
1672             dlm->dlm_locking_proto.pv_minor > 0) {
1673                 status = dlm_send_nodeinfo(dlm, ctxt->yes_resp_map);
1674                 if (status) {
1675                         mlog_errno(status);
1676                         goto bail;
1677                 }
1678                 status = dlm_send_regions(dlm, ctxt->yes_resp_map);
1679                 if (status) {
1680                         mlog_errno(status);
1681                         goto bail;
1682                 }
1683         }
1684
1685         dlm_send_join_asserts(dlm, ctxt->yes_resp_map);
1686
1687         /* Joined state *must* be set before the joining node
1688          * information, otherwise the query_join handler may read no
1689          * current joiner but a state of NEW and tell joining nodes
1690          * we're not in the domain. */
1691         spin_lock(&dlm_domain_lock);
1692         dlm->dlm_state = DLM_CTXT_JOINED;
1693         dlm->num_joins++;
1694         spin_unlock(&dlm_domain_lock);
1695
1696 bail:
1697         spin_lock(&dlm->spinlock);
1698         __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
1699         if (!status) {
1700                 printk(KERN_NOTICE "o2dlm: Joining domain %s ", dlm->name);
1701                 __dlm_print_nodes(dlm);
1702         }
1703         spin_unlock(&dlm->spinlock);
1704
1705         if (ctxt) {
1706                 /* Do we need to send a cancel message to any nodes? */
1707                 if (status < 0) {
1708                         tmpstat = dlm_send_join_cancels(dlm,
1709                                                         ctxt->yes_resp_map,
1710                                                         sizeof(ctxt->yes_resp_map));
1711                         if (tmpstat < 0)
1712                                 mlog_errno(tmpstat);
1713                 }
1714                 kfree(ctxt);
1715         }
1716
1717         mlog(0, "returning %d\n", status);
1718         return status;
1719 }
1720
1721 static void dlm_unregister_domain_handlers(struct dlm_ctxt *dlm)
1722 {
1723         o2hb_unregister_callback(dlm->name, &dlm->dlm_hb_up);
1724         o2hb_unregister_callback(dlm->name, &dlm->dlm_hb_down);
1725         o2net_unregister_handler_list(&dlm->dlm_domain_handlers);
1726 }
1727
1728 static int dlm_register_domain_handlers(struct dlm_ctxt *dlm)
1729 {
1730         int status;
1731
1732         mlog(0, "registering handlers.\n");
1733
1734         o2hb_setup_callback(&dlm->dlm_hb_down, O2HB_NODE_DOWN_CB,
1735                             dlm_hb_node_down_cb, dlm, DLM_HB_NODE_DOWN_PRI);
1736         o2hb_setup_callback(&dlm->dlm_hb_up, O2HB_NODE_UP_CB,
1737                             dlm_hb_node_up_cb, dlm, DLM_HB_NODE_UP_PRI);
1738
1739         status = o2hb_register_callback(dlm->name, &dlm->dlm_hb_down);
1740         if (status)
1741                 goto bail;
1742
1743         status = o2hb_register_callback(dlm->name, &dlm->dlm_hb_up);
1744         if (status)
1745                 goto bail;
1746
1747         status = o2net_register_handler(DLM_MASTER_REQUEST_MSG, dlm->key,
1748                                         sizeof(struct dlm_master_request),
1749                                         dlm_master_request_handler,
1750                                         dlm, NULL, &dlm->dlm_domain_handlers);
1751         if (status)
1752                 goto bail;
1753
1754         status = o2net_register_handler(DLM_ASSERT_MASTER_MSG, dlm->key,
1755                                         sizeof(struct dlm_assert_master),
1756                                         dlm_assert_master_handler,
1757                                         dlm, dlm_assert_master_post_handler,
1758                                         &dlm->dlm_domain_handlers);
1759         if (status)
1760                 goto bail;
1761
1762         status = o2net_register_handler(DLM_CREATE_LOCK_MSG, dlm->key,
1763                                         sizeof(struct dlm_create_lock),
1764                                         dlm_create_lock_handler,
1765                                         dlm, NULL, &dlm->dlm_domain_handlers);
1766         if (status)
1767                 goto bail;
1768
1769         status = o2net_register_handler(DLM_CONVERT_LOCK_MSG, dlm->key,
1770                                         DLM_CONVERT_LOCK_MAX_LEN,
1771                                         dlm_convert_lock_handler,
1772                                         dlm, NULL, &dlm->dlm_domain_handlers);
1773         if (status)
1774                 goto bail;
1775
1776         status = o2net_register_handler(DLM_UNLOCK_LOCK_MSG, dlm->key,
1777                                         DLM_UNLOCK_LOCK_MAX_LEN,
1778                                         dlm_unlock_lock_handler,
1779                                         dlm, NULL, &dlm->dlm_domain_handlers);
1780         if (status)
1781                 goto bail;
1782
1783         status = o2net_register_handler(DLM_PROXY_AST_MSG, dlm->key,
1784                                         DLM_PROXY_AST_MAX_LEN,
1785                                         dlm_proxy_ast_handler,
1786                                         dlm, NULL, &dlm->dlm_domain_handlers);
1787         if (status)
1788                 goto bail;
1789
1790         status = o2net_register_handler(DLM_EXIT_DOMAIN_MSG, dlm->key,
1791                                         sizeof(struct dlm_exit_domain),
1792                                         dlm_exit_domain_handler,
1793                                         dlm, NULL, &dlm->dlm_domain_handlers);
1794         if (status)
1795                 goto bail;
1796
1797         status = o2net_register_handler(DLM_DEREF_LOCKRES_MSG, dlm->key,
1798                                         sizeof(struct dlm_deref_lockres),
1799                                         dlm_deref_lockres_handler,
1800                                         dlm, NULL, &dlm->dlm_domain_handlers);
1801         if (status)
1802                 goto bail;
1803
1804         status = o2net_register_handler(DLM_MIGRATE_REQUEST_MSG, dlm->key,
1805                                         sizeof(struct dlm_migrate_request),
1806                                         dlm_migrate_request_handler,
1807                                         dlm, NULL, &dlm->dlm_domain_handlers);
1808         if (status)
1809                 goto bail;
1810
1811         status = o2net_register_handler(DLM_MIG_LOCKRES_MSG, dlm->key,
1812                                         DLM_MIG_LOCKRES_MAX_LEN,
1813                                         dlm_mig_lockres_handler,
1814                                         dlm, NULL, &dlm->dlm_domain_handlers);
1815         if (status)
1816                 goto bail;
1817
1818         status = o2net_register_handler(DLM_MASTER_REQUERY_MSG, dlm->key,
1819                                         sizeof(struct dlm_master_requery),
1820                                         dlm_master_requery_handler,
1821                                         dlm, NULL, &dlm->dlm_domain_handlers);
1822         if (status)
1823                 goto bail;
1824
1825         status = o2net_register_handler(DLM_LOCK_REQUEST_MSG, dlm->key,
1826                                         sizeof(struct dlm_lock_request),
1827                                         dlm_request_all_locks_handler,
1828                                         dlm, NULL, &dlm->dlm_domain_handlers);
1829         if (status)
1830                 goto bail;
1831
1832         status = o2net_register_handler(DLM_RECO_DATA_DONE_MSG, dlm->key,
1833                                         sizeof(struct dlm_reco_data_done),
1834                                         dlm_reco_data_done_handler,
1835                                         dlm, NULL, &dlm->dlm_domain_handlers);
1836         if (status)
1837                 goto bail;
1838
1839         status = o2net_register_handler(DLM_BEGIN_RECO_MSG, dlm->key,
1840                                         sizeof(struct dlm_begin_reco),
1841                                         dlm_begin_reco_handler,
1842                                         dlm, NULL, &dlm->dlm_domain_handlers);
1843         if (status)
1844                 goto bail;
1845
1846         status = o2net_register_handler(DLM_FINALIZE_RECO_MSG, dlm->key,
1847                                         sizeof(struct dlm_finalize_reco),
1848                                         dlm_finalize_reco_handler,
1849                                         dlm, NULL, &dlm->dlm_domain_handlers);
1850         if (status)
1851                 goto bail;
1852
1853         status = o2net_register_handler(DLM_BEGIN_EXIT_DOMAIN_MSG, dlm->key,
1854                                         sizeof(struct dlm_exit_domain),
1855                                         dlm_begin_exit_domain_handler,
1856                                         dlm, NULL, &dlm->dlm_domain_handlers);
1857         if (status)
1858                 goto bail;
1859
1860         status = o2net_register_handler(DLM_DEREF_LOCKRES_DONE, dlm->key,
1861                                         sizeof(struct dlm_deref_lockres_done),
1862                                         dlm_deref_lockres_done_handler,
1863                                         dlm, NULL, &dlm->dlm_domain_handlers);
1864 bail:
1865         if (status)
1866                 dlm_unregister_domain_handlers(dlm);
1867
1868         return status;
1869 }
1870
1871 static int dlm_join_domain(struct dlm_ctxt *dlm)
1872 {
1873         int status;
1874         unsigned int backoff;
1875         unsigned int total_backoff = 0;
1876         char wq_name[O2NM_MAX_NAME_LEN];
1877
1878         BUG_ON(!dlm);
1879
1880         mlog(0, "Join domain %s\n", dlm->name);
1881
1882         status = dlm_register_domain_handlers(dlm);
1883         if (status) {
1884                 mlog_errno(status);
1885                 goto bail;
1886         }
1887
1888         status = dlm_launch_thread(dlm);
1889         if (status < 0) {
1890                 mlog_errno(status);
1891                 goto bail;
1892         }
1893
1894         status = dlm_launch_recovery_thread(dlm);
1895         if (status < 0) {
1896                 mlog_errno(status);
1897                 goto bail;
1898         }
1899
1900         status = dlm_debug_init(dlm);
1901         if (status < 0) {
1902                 mlog_errno(status);
1903                 goto bail;
1904         }
1905
1906         snprintf(wq_name, O2NM_MAX_NAME_LEN, "dlm_wq-%s", dlm->name);
1907         dlm->dlm_worker = create_singlethread_workqueue(wq_name);
1908         if (!dlm->dlm_worker) {
1909                 status = -ENOMEM;
1910                 mlog_errno(status);
1911                 goto bail;
1912         }
1913
1914         do {
1915                 status = dlm_try_to_join_domain(dlm);
1916
1917                 /* If we're racing another node to the join, then we
1918                  * need to back off temporarily and let them
1919                  * complete. */
1920 #define DLM_JOIN_TIMEOUT_MSECS  90000
1921                 if (status == -EAGAIN) {
1922                         if (signal_pending(current)) {
1923                                 status = -ERESTARTSYS;
1924                                 goto bail;
1925                         }
1926
1927                         if (total_backoff > DLM_JOIN_TIMEOUT_MSECS) {
1928                                 status = -ERESTARTSYS;
1929                                 mlog(ML_NOTICE, "Timed out joining dlm domain "
1930                                      "%s after %u msecs\n", dlm->name,
1931                                      total_backoff);
1932                                 goto bail;
1933                         }
1934
1935                         /*
1936                          * <chip> After you!
1937                          * <dale> No, after you!
1938                          * <chip> I insist!
1939                          * <dale> But you first!
1940                          * ...
1941                          */
1942                         backoff = (unsigned int)(jiffies & 0x3);
1943                         backoff *= DLM_DOMAIN_BACKOFF_MS;
1944                         total_backoff += backoff;
1945                         mlog(0, "backoff %d\n", backoff);
1946                         msleep(backoff);
1947                 }
1948         } while (status == -EAGAIN);
1949
1950         if (status < 0) {
1951                 mlog_errno(status);
1952                 goto bail;
1953         }
1954
1955         status = 0;
1956 bail:
1957         wake_up(&dlm_domain_events);
1958
1959         if (status) {
1960                 dlm_unregister_domain_handlers(dlm);
1961                 dlm_debug_shutdown(dlm);
1962                 dlm_complete_thread(dlm);
1963                 dlm_complete_recovery_thread(dlm);
1964                 dlm_destroy_dlm_worker(dlm);
1965         }
1966
1967         return status;
1968 }
1969
1970 static struct dlm_ctxt *dlm_alloc_ctxt(const char *domain,
1971                                 u32 key)
1972 {
1973         int i;
1974         int ret;
1975         struct dlm_ctxt *dlm = NULL;
1976
1977         dlm = kzalloc(sizeof(*dlm), GFP_KERNEL);
1978         if (!dlm) {
1979                 ret = -ENOMEM;
1980                 mlog_errno(ret);
1981                 goto leave;
1982         }
1983
1984         dlm->name = kstrdup(domain, GFP_KERNEL);
1985         if (dlm->name == NULL) {
1986                 ret = -ENOMEM;
1987                 mlog_errno(ret);
1988                 goto leave;
1989         }
1990
1991         dlm->lockres_hash = (struct hlist_head **)dlm_alloc_pagevec(DLM_HASH_PAGES);
1992         if (!dlm->lockres_hash) {
1993                 ret = -ENOMEM;
1994                 mlog_errno(ret);
1995                 goto leave;
1996         }
1997
1998         for (i = 0; i < DLM_HASH_BUCKETS; i++)
1999                 INIT_HLIST_HEAD(dlm_lockres_hash(dlm, i));
2000
2001         dlm->master_hash = (struct hlist_head **)
2002                                 dlm_alloc_pagevec(DLM_HASH_PAGES);
2003         if (!dlm->master_hash) {
2004                 ret = -ENOMEM;
2005                 mlog_errno(ret);
2006                 goto leave;
2007         }
2008
2009         for (i = 0; i < DLM_HASH_BUCKETS; i++)
2010                 INIT_HLIST_HEAD(dlm_master_hash(dlm, i));
2011
2012         dlm->key = key;
2013         dlm->node_num = o2nm_this_node();
2014
2015         ret = dlm_create_debugfs_subroot(dlm);
2016         if (ret < 0)
2017                 goto leave;
2018
2019         spin_lock_init(&dlm->spinlock);
2020         spin_lock_init(&dlm->master_lock);
2021         spin_lock_init(&dlm->ast_lock);
2022         spin_lock_init(&dlm->track_lock);
2023         INIT_LIST_HEAD(&dlm->list);
2024         INIT_LIST_HEAD(&dlm->dirty_list);
2025         INIT_LIST_HEAD(&dlm->reco.resources);
2026         INIT_LIST_HEAD(&dlm->reco.node_data);
2027         INIT_LIST_HEAD(&dlm->purge_list);
2028         INIT_LIST_HEAD(&dlm->dlm_domain_handlers);
2029         INIT_LIST_HEAD(&dlm->tracking_list);
2030         dlm->reco.state = 0;
2031
2032         INIT_LIST_HEAD(&dlm->pending_asts);
2033         INIT_LIST_HEAD(&dlm->pending_basts);
2034
2035         mlog(0, "dlm->recovery_map=%p, &(dlm->recovery_map[0])=%p\n",
2036                   dlm->recovery_map, &(dlm->recovery_map[0]));
2037
2038         memset(dlm->recovery_map, 0, sizeof(dlm->recovery_map));
2039         memset(dlm->live_nodes_map, 0, sizeof(dlm->live_nodes_map));
2040         memset(dlm->domain_map, 0, sizeof(dlm->domain_map));
2041
2042         dlm->dlm_thread_task = NULL;
2043         dlm->dlm_reco_thread_task = NULL;
2044         dlm->dlm_worker = NULL;
2045         init_waitqueue_head(&dlm->dlm_thread_wq);
2046         init_waitqueue_head(&dlm->dlm_reco_thread_wq);
2047         init_waitqueue_head(&dlm->reco.event);
2048         init_waitqueue_head(&dlm->ast_wq);
2049         init_waitqueue_head(&dlm->migration_wq);
2050         INIT_LIST_HEAD(&dlm->mle_hb_events);
2051
2052         dlm->joining_node = DLM_LOCK_RES_OWNER_UNKNOWN;
2053         init_waitqueue_head(&dlm->dlm_join_events);
2054
2055         dlm->reco.new_master = O2NM_INVALID_NODE_NUM;
2056         dlm->reco.dead_node = O2NM_INVALID_NODE_NUM;
2057
2058         atomic_set(&dlm->res_tot_count, 0);
2059         atomic_set(&dlm->res_cur_count, 0);
2060         for (i = 0; i < DLM_MLE_NUM_TYPES; ++i) {
2061                 atomic_set(&dlm->mle_tot_count[i], 0);
2062                 atomic_set(&dlm->mle_cur_count[i], 0);
2063         }
2064
2065         spin_lock_init(&dlm->work_lock);
2066         INIT_LIST_HEAD(&dlm->work_list);
2067         INIT_WORK(&dlm->dispatched_work, dlm_dispatch_work);
2068
2069         kref_init(&dlm->dlm_refs);
2070         dlm->dlm_state = DLM_CTXT_NEW;
2071
2072         INIT_LIST_HEAD(&dlm->dlm_eviction_callbacks);
2073
2074         mlog(0, "context init: refcount %u\n",
2075                   atomic_read(&dlm->dlm_refs.refcount));
2076
2077 leave:
2078         if (ret < 0 && dlm) {
2079                 if (dlm->master_hash)
2080                         dlm_free_pagevec((void **)dlm->master_hash,
2081                                         DLM_HASH_PAGES);
2082
2083                 if (dlm->lockres_hash)
2084                         dlm_free_pagevec((void **)dlm->lockres_hash,
2085                                         DLM_HASH_PAGES);
2086
2087                 kfree(dlm->name);
2088                 kfree(dlm);
2089                 dlm = NULL;
2090         }
2091         return dlm;
2092 }
2093
2094 /*
2095  * Compare a requested locking protocol version against the current one.
2096  *
2097  * If the major numbers are different, they are incompatible.
2098  * If the current minor is greater than the request, they are incompatible.
2099  * If the current minor is less than or equal to the request, they are
2100  * compatible, and the requester should run at the current minor version.
2101  */
2102 static int dlm_protocol_compare(struct dlm_protocol_version *existing,
2103                                 struct dlm_protocol_version *request)
2104 {
2105         if (existing->pv_major != request->pv_major)
2106                 return 1;
2107
2108         if (existing->pv_minor > request->pv_minor)
2109                 return 1;
2110
2111         if (existing->pv_minor < request->pv_minor)
2112                 request->pv_minor = existing->pv_minor;
2113
2114         return 0;
2115 }
2116
2117 /*
2118  * dlm_register_domain: one-time setup per "domain".
2119  *
2120  * The filesystem passes in the requested locking version via proto.
2121  * If registration was successful, proto will contain the negotiated
2122  * locking protocol.
2123  */
2124 struct dlm_ctxt * dlm_register_domain(const char *domain,
2125                                u32 key,
2126                                struct dlm_protocol_version *fs_proto)
2127 {
2128         int ret;
2129         struct dlm_ctxt *dlm = NULL;
2130         struct dlm_ctxt *new_ctxt = NULL;
2131
2132         if (strlen(domain) >= O2NM_MAX_NAME_LEN) {
2133                 ret = -ENAMETOOLONG;
2134                 mlog(ML_ERROR, "domain name length too long\n");
2135                 goto leave;
2136         }
2137
2138         mlog(0, "register called for domain \"%s\"\n", domain);
2139
2140 retry:
2141         dlm = NULL;
2142         if (signal_pending(current)) {
2143                 ret = -ERESTARTSYS;
2144                 mlog_errno(ret);
2145                 goto leave;
2146         }
2147
2148         spin_lock(&dlm_domain_lock);
2149
2150         dlm = __dlm_lookup_domain(domain);
2151         if (dlm) {
2152                 if (dlm->dlm_state != DLM_CTXT_JOINED) {
2153                         spin_unlock(&dlm_domain_lock);
2154
2155                         mlog(0, "This ctxt is not joined yet!\n");
2156                         wait_event_interruptible(dlm_domain_events,
2157                                                  dlm_wait_on_domain_helper(
2158                                                          domain));
2159                         goto retry;
2160                 }
2161
2162                 if (dlm_protocol_compare(&dlm->fs_locking_proto, fs_proto)) {
2163                         spin_unlock(&dlm_domain_lock);
2164                         mlog(ML_ERROR,
2165                              "Requested locking protocol version is not "
2166                              "compatible with already registered domain "
2167                              "\"%s\"\n", domain);
2168                         ret = -EPROTO;
2169                         goto leave;
2170                 }
2171
2172                 __dlm_get(dlm);
2173                 dlm->num_joins++;
2174
2175                 spin_unlock(&dlm_domain_lock);
2176
2177                 ret = 0;
2178                 goto leave;
2179         }
2180
2181         /* doesn't exist */
2182         if (!new_ctxt) {
2183                 spin_unlock(&dlm_domain_lock);
2184
2185                 new_ctxt = dlm_alloc_ctxt(domain, key);
2186                 if (new_ctxt)
2187                         goto retry;
2188
2189                 ret = -ENOMEM;
2190                 mlog_errno(ret);
2191                 goto leave;
2192         }
2193
2194         /* a little variable switch-a-roo here... */
2195         dlm = new_ctxt;
2196         new_ctxt = NULL;
2197
2198         /* add the new domain */
2199         list_add_tail(&dlm->list, &dlm_domains);
2200         spin_unlock(&dlm_domain_lock);
2201
2202         /*
2203          * Pass the locking protocol version into the join.  If the join
2204          * succeeds, it will have the negotiated protocol set.
2205          */
2206         dlm->dlm_locking_proto = dlm_protocol;
2207         dlm->fs_locking_proto = *fs_proto;
2208
2209         ret = dlm_join_domain(dlm);
2210         if (ret) {
2211                 mlog_errno(ret);
2212                 dlm_put(dlm);
2213                 goto leave;
2214         }
2215
2216         /* Tell the caller what locking protocol we negotiated */
2217         *fs_proto = dlm->fs_locking_proto;
2218
2219         ret = 0;
2220 leave:
2221         if (new_ctxt)
2222                 dlm_free_ctxt_mem(new_ctxt);
2223
2224         if (ret < 0)
2225                 dlm = ERR_PTR(ret);
2226
2227         return dlm;
2228 }
2229 EXPORT_SYMBOL_GPL(dlm_register_domain);
2230
2231 static LIST_HEAD(dlm_join_handlers);
2232
2233 static void dlm_unregister_net_handlers(void)
2234 {
2235         o2net_unregister_handler_list(&dlm_join_handlers);
2236 }
2237
2238 static int dlm_register_net_handlers(void)
2239 {
2240         int status = 0;
2241
2242         status = o2net_register_handler(DLM_QUERY_JOIN_MSG, DLM_MOD_KEY,
2243                                         sizeof(struct dlm_query_join_request),
2244                                         dlm_query_join_handler,
2245                                         NULL, NULL, &dlm_join_handlers);
2246         if (status)
2247                 goto bail;
2248
2249         status = o2net_register_handler(DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY,
2250                                         sizeof(struct dlm_assert_joined),
2251                                         dlm_assert_joined_handler,
2252                                         NULL, NULL, &dlm_join_handlers);
2253         if (status)
2254                 goto bail;
2255
2256         status = o2net_register_handler(DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY,
2257                                         sizeof(struct dlm_cancel_join),
2258                                         dlm_cancel_join_handler,
2259                                         NULL, NULL, &dlm_join_handlers);
2260         if (status)
2261                 goto bail;
2262
2263         status = o2net_register_handler(DLM_QUERY_REGION, DLM_MOD_KEY,
2264                                         sizeof(struct dlm_query_region),
2265                                         dlm_query_region_handler,
2266                                         NULL, NULL, &dlm_join_handlers);
2267
2268         if (status)
2269                 goto bail;
2270
2271         status = o2net_register_handler(DLM_QUERY_NODEINFO, DLM_MOD_KEY,
2272                                         sizeof(struct dlm_query_nodeinfo),
2273                                         dlm_query_nodeinfo_handler,
2274                                         NULL, NULL, &dlm_join_handlers);
2275 bail:
2276         if (status < 0)
2277                 dlm_unregister_net_handlers();
2278
2279         return status;
2280 }
2281
2282 /* Domain eviction callback handling.
2283  *
2284  * The file system requires notification of node death *before* the
2285  * dlm completes it's recovery work, otherwise it may be able to
2286  * acquire locks on resources requiring recovery. Since the dlm can
2287  * evict a node from it's domain *before* heartbeat fires, a similar
2288  * mechanism is required. */
2289
2290 /* Eviction is not expected to happen often, so a per-domain lock is
2291  * not necessary. Eviction callbacks are allowed to sleep for short
2292  * periods of time. */
2293 static DECLARE_RWSEM(dlm_callback_sem);
2294
2295 void dlm_fire_domain_eviction_callbacks(struct dlm_ctxt *dlm,
2296                                         int node_num)
2297 {
2298         struct dlm_eviction_cb *cb;
2299
2300         down_read(&dlm_callback_sem);
2301         list_for_each_entry(cb, &dlm->dlm_eviction_callbacks, ec_item) {
2302                 cb->ec_func(node_num, cb->ec_data);
2303         }
2304         up_read(&dlm_callback_sem);
2305 }
2306
2307 void dlm_setup_eviction_cb(struct dlm_eviction_cb *cb,
2308                            dlm_eviction_func *f,
2309                            void *data)
2310 {
2311         INIT_LIST_HEAD(&cb->ec_item);
2312         cb->ec_func = f;
2313         cb->ec_data = data;
2314 }
2315 EXPORT_SYMBOL_GPL(dlm_setup_eviction_cb);
2316
2317 void dlm_register_eviction_cb(struct dlm_ctxt *dlm,
2318                               struct dlm_eviction_cb *cb)
2319 {
2320         down_write(&dlm_callback_sem);
2321         list_add_tail(&cb->ec_item, &dlm->dlm_eviction_callbacks);
2322         up_write(&dlm_callback_sem);
2323 }
2324 EXPORT_SYMBOL_GPL(dlm_register_eviction_cb);
2325
2326 void dlm_unregister_eviction_cb(struct dlm_eviction_cb *cb)
2327 {
2328         down_write(&dlm_callback_sem);
2329         list_del_init(&cb->ec_item);
2330         up_write(&dlm_callback_sem);
2331 }
2332 EXPORT_SYMBOL_GPL(dlm_unregister_eviction_cb);
2333
2334 static int __init dlm_init(void)
2335 {
2336         int status;
2337
2338         status = dlm_init_mle_cache();
2339         if (status) {
2340                 mlog(ML_ERROR, "Could not create o2dlm_mle slabcache\n");
2341                 goto error;
2342         }
2343
2344         status = dlm_init_master_caches();
2345         if (status) {
2346                 mlog(ML_ERROR, "Could not create o2dlm_lockres and "
2347                      "o2dlm_lockname slabcaches\n");
2348                 goto error;
2349         }
2350
2351         status = dlm_init_lock_cache();
2352         if (status) {
2353                 mlog(ML_ERROR, "Count not create o2dlm_lock slabcache\n");
2354                 goto error;
2355         }
2356
2357         status = dlm_register_net_handlers();
2358         if (status) {
2359                 mlog(ML_ERROR, "Unable to register network handlers\n");
2360                 goto error;
2361         }
2362
2363         status = dlm_create_debugfs_root();
2364         if (status)
2365                 goto error;
2366
2367         return 0;
2368 error:
2369         dlm_unregister_net_handlers();
2370         dlm_destroy_lock_cache();
2371         dlm_destroy_master_caches();
2372         dlm_destroy_mle_cache();
2373         return -1;
2374 }
2375
2376 static void __exit dlm_exit (void)
2377 {
2378         dlm_destroy_debugfs_root();
2379         dlm_unregister_net_handlers();
2380         dlm_destroy_lock_cache();
2381         dlm_destroy_master_caches();
2382         dlm_destroy_mle_cache();
2383 }
2384
2385 MODULE_AUTHOR("Oracle");
2386 MODULE_LICENSE("GPL");
2387 MODULE_DESCRIPTION("OCFS2 Distributed Lock Management");
2388
2389 module_init(dlm_init);
2390 module_exit(dlm_exit);