git.oblomov.eu Git - linux-2.6/blob - fs/ocfs2/dlm/dlmthread.c

   1 /* -*- mode: c; c-basic-offset: 8; -*-
   2  * vim: noexpandtab sw=8 ts=8 sts=0:
   3  *
   4  * dlmthread.c
   5  *
   6  * standalone DLM module
   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
  28 #include <linux/module.h>
  29 #include <linux/fs.h>
  30 #include <linux/types.h>
  31 #include <linux/slab.h>
  32 #include <linux/highmem.h>
  33 #include <linux/utsname.h>
  34 #include <linux/init.h>
  35 #include <linux/sysctl.h>
  36 #include <linux/random.h>
  37 #include <linux/blkdev.h>
  38 #include <linux/socket.h>
  39 #include <linux/inet.h>
  40 #include <linux/timer.h>
  41 #include <linux/kthread.h>
  42 #include <linux/delay.h>
  43
  44
  45 #include "cluster/heartbeat.h"
  46 #include "cluster/nodemanager.h"
  47 #include "cluster/tcp.h"
  48
  49 #include "dlmapi.h"
  50 #include "dlmcommon.h"
  51 #include "dlmdomain.h"
  52
  53 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_THREAD)
  54 #include "cluster/masklog.h"
  55
  56 static int dlm_thread(void *data);
  57 static void dlm_flush_asts(struct dlm_ctxt *dlm);
  58
  59 #define dlm_lock_is_remote(dlm, lock)     ((lock)->ml.node != (dlm)->node_num)
  60
  61 /* will exit holding res->spinlock, but may drop in function */
  62 /* waits until flags are cleared on res->state */
  63 void __dlm_wait_on_lockres_flags(struct dlm_lock_resource *res, int flags)
  64 {
  65         DECLARE_WAITQUEUE(wait, current);
  66
  67         assert_spin_locked(&res->spinlock);
  68
  69         add_wait_queue(&res->wq, &wait);
  70 repeat:
  71         set_current_state(TASK_UNINTERRUPTIBLE);
  72         if (res->state & flags) {
  73                 spin_unlock(&res->spinlock);
  74                 schedule();
  75                 spin_lock(&res->spinlock);
  76                 goto repeat;
  77         }
  78         remove_wait_queue(&res->wq, &wait);
  79         current->state = TASK_RUNNING;
  80 }
  81
  82 int __dlm_lockres_has_locks(struct dlm_lock_resource *res)
  83 {
  84         if (list_empty(&res->granted) &&
  85             list_empty(&res->converting) &&
  86             list_empty(&res->blocked))
  87                 return 0;
  88         return 1;
  89 }
  90
  91 /* "unused": the lockres has no locks, is not on the dirty list,
  92  * has no inflight locks (in the gap between mastery and acquiring
  93  * the first lock), and has no bits in its refmap.
  94  * truly ready to be freed. */
  95 int __dlm_lockres_unused(struct dlm_lock_resource *res)
  96 {
  97         if (!__dlm_lockres_has_locks(res) &&
  98             (list_empty(&res->dirty) && !(res->state & DLM_LOCK_RES_DIRTY))) {
  99                 /* try not to scan the bitmap unless the first two
 100                  * conditions are already true */
 101                 int bit = find_next_bit(res->refmap, O2NM_MAX_NODES, 0);
 102                 if (bit >= O2NM_MAX_NODES) {
 103                         /* since the bit for dlm->node_num is not
 104                          * set, inflight_locks better be zero */
 105                         BUG_ON(res->inflight_locks != 0);
 106                         return 1;
 107                 }
 108         }
 109         return 0;
 110 }
 111
 112
 113 /* Call whenever you may have added or deleted something from one of
 114  * the lockres queue's. This will figure out whether it belongs on the
 115  * unused list or not and does the appropriate thing. */
 116 void __dlm_lockres_calc_usage(struct dlm_ctxt *dlm,
 117                               struct dlm_lock_resource *res)
 118 {
 119         mlog_entry("%.*s\n", res->lockname.len, res->lockname.name);
 120
 121         assert_spin_locked(&dlm->spinlock);
 122         assert_spin_locked(&res->spinlock);
 123
 124         if (__dlm_lockres_unused(res)){
 125                 if (list_empty(&res->purge)) {
 126                         mlog(0, "putting lockres %.*s:%p onto purge list\n",
 127                              res->lockname.len, res->lockname.name, res);
 128
 129                         res->last_used = jiffies;
 130                         dlm_lockres_get(res);
 131                         list_add_tail(&res->purge, &dlm->purge_list);
 132                         dlm->purge_count++;
 133                 }
 134         } else if (!list_empty(&res->purge)) {
 135                 mlog(0, "removing lockres %.*s:%p from purge list, owner=%u\n",
 136                      res->lockname.len, res->lockname.name, res, res->owner);
 137
 138                 list_del_init(&res->purge);
 139                 dlm_lockres_put(res);
 140                 dlm->purge_count--;
 141         }
 142 }
 143
 144 void dlm_lockres_calc_usage(struct dlm_ctxt *dlm,
 145                             struct dlm_lock_resource *res)
 146 {
 147         mlog_entry("%.*s\n", res->lockname.len, res->lockname.name);
 148         spin_lock(&dlm->spinlock);
 149         spin_lock(&res->spinlock);
 150
 151         __dlm_lockres_calc_usage(dlm, res);
 152
 153         spin_unlock(&res->spinlock);
 154         spin_unlock(&dlm->spinlock);
 155 }
 156
 157 static int dlm_purge_lockres(struct dlm_ctxt *dlm,
 158                              struct dlm_lock_resource *res)
 159 {
 160         int master;
 161         int ret = 0;
 162
 163         spin_lock(&res->spinlock);
 164         if (!__dlm_lockres_unused(res)) {
 165                 spin_unlock(&res->spinlock);
 166                 mlog(0, "%s:%.*s: tried to purge but not unused\n",
 167                      dlm->name, res->lockname.len, res->lockname.name);
 168                 return -ENOTEMPTY;
 169         }
 170         master = (res->owner == dlm->node_num);
 171         if (!master)
 172                 res->state |= DLM_LOCK_RES_DROPPING_REF;
 173         spin_unlock(&res->spinlock);
 174
 175         mlog(0, "purging lockres %.*s, master = %d\n", res->lockname.len,
 176              res->lockname.name, master);
 177
 178         if (!master) {
 179                 spin_lock(&res->spinlock);
 180                 /* This ensures that clear refmap is sent after the set */
 181                 __dlm_wait_on_lockres_flags(res, DLM_LOCK_RES_SETREF_INPROG);
 182                 spin_unlock(&res->spinlock);
 183                 /* drop spinlock to do messaging, retake below */
 184                 spin_unlock(&dlm->spinlock);
 185                 /* clear our bit from the master's refmap, ignore errors */
 186                 ret = dlm_drop_lockres_ref(dlm, res);
 187                 if (ret < 0) {
 188                         mlog_errno(ret);
 189                         if (!dlm_is_host_down(ret))
 190                                 BUG();
 191                 }
 192                 mlog(0, "%s:%.*s: dlm_deref_lockres returned %d\n",
 193                      dlm->name, res->lockname.len, res->lockname.name, ret);
 194                 spin_lock(&dlm->spinlock);
 195         }
 196
 197         if (!list_empty(&res->purge)) {
 198                 mlog(0, "removing lockres %.*s:%p from purgelist, "
 199                      "master = %d\n", res->lockname.len, res->lockname.name,
 200                      res, master);
 201                 list_del_init(&res->purge);
 202                 dlm_lockres_put(res);
 203                 dlm->purge_count--;
 204         }
 205         __dlm_unhash_lockres(res);
 206
 207         /* lockres is not in the hash now.  drop the flag and wake up
 208          * any processes waiting in dlm_get_lock_resource. */
 209         if (!master) {
 210                 spin_lock(&res->spinlock);
 211                 res->state &= ~DLM_LOCK_RES_DROPPING_REF;
 212                 spin_unlock(&res->spinlock);
 213                 wake_up(&res->wq);
 214         }
 215         return 0;
 216 }
 217
 218 static void dlm_run_purge_list(struct dlm_ctxt *dlm,
 219                                int purge_now)
 220 {
 221         unsigned int run_max, unused;
 222         unsigned long purge_jiffies;
 223         struct dlm_lock_resource *lockres;
 224
 225         spin_lock(&dlm->spinlock);
 226         run_max = dlm->purge_count;
 227
 228         while(run_max && !list_empty(&dlm->purge_list)) {
 229                 run_max--;
 230
 231                 lockres = list_entry(dlm->purge_list.next,
 232                                      struct dlm_lock_resource, purge);
 233
 234                 /* Status of the lockres *might* change so double
 235                  * check. If the lockres is unused, holding the dlm
 236                  * spinlock will prevent people from getting and more
 237                  * refs on it -- there's no need to keep the lockres
 238                  * spinlock. */
 239                 spin_lock(&lockres->spinlock);
 240                 unused = __dlm_lockres_unused(lockres);
 241                 spin_unlock(&lockres->spinlock);
 242
 243                 if (!unused)
 244                         continue;
 245
 246                 purge_jiffies = lockres->last_used +
 247                         msecs_to_jiffies(DLM_PURGE_INTERVAL_MS);
 248
 249                 /* Make sure that we want to be processing this guy at
 250                  * this time. */
 251                 if (!purge_now && time_after(purge_jiffies, jiffies)) {
 252                         /* Since resources are added to the purge list
 253                          * in tail order, we can stop at the first
 254                          * unpurgable resource -- anyone added after
 255                          * him will have a greater last_used value */
 256                         break;
 257                 }
 258
 259                 mlog(0, "removing lockres %.*s:%p from purgelist\n",
 260                      lockres->lockname.len, lockres->lockname.name, lockres);
 261                 list_del_init(&lockres->purge);
 262                 dlm_lockres_put(lockres);
 263                 dlm->purge_count--;
 264
 265                 /* This may drop and reacquire the dlm spinlock if it
 266                  * has to do migration. */
 267                 mlog(0, "calling dlm_purge_lockres!\n");
 268                 if (dlm_purge_lockres(dlm, lockres))
 269                         BUG();
 270                 mlog(0, "DONE calling dlm_purge_lockres!\n");
 271
 272                 /* Avoid adding any scheduling latencies */
 273                 cond_resched_lock(&dlm->spinlock);
 274         }
 275
 276         spin_unlock(&dlm->spinlock);
 277 }
 278
 279 static void dlm_shuffle_lists(struct dlm_ctxt *dlm,
 280                               struct dlm_lock_resource *res)
 281 {
 282         struct dlm_lock *lock, *target;
 283         struct list_head *iter;
 284         struct list_head *head;
 285         int can_grant = 1;
 286
 287         //mlog(0, "res->lockname.len=%d\n", res->lockname.len);
 288         //mlog(0, "res->lockname.name=%p\n", res->lockname.name);
 289         //mlog(0, "shuffle res %.*s\n", res->lockname.len,
 290         //        res->lockname.name);
 291
 292         /* because this function is called with the lockres
 293          * spinlock, and because we know that it is not migrating/
 294          * recovering/in-progress, it is fine to reserve asts and
 295          * basts right before queueing them all throughout */
 296         assert_spin_locked(&res->spinlock);
 297         BUG_ON((res->state & (DLM_LOCK_RES_MIGRATING|
 298                               DLM_LOCK_RES_RECOVERING|
 299                               DLM_LOCK_RES_IN_PROGRESS)));
 300
 301 converting:
 302         if (list_empty(&res->converting))
 303                 goto blocked;
 304         mlog(0, "res %.*s has locks on a convert queue\n", res->lockname.len,
 305              res->lockname.name);
 306
 307         target = list_entry(res->converting.next, struct dlm_lock, list);
 308         if (target->ml.convert_type == LKM_IVMODE) {
 309                 mlog(ML_ERROR, "%.*s: converting a lock with no "
 310                      "convert_type!\n", res->lockname.len, res->lockname.name);
 311                 BUG();
 312         }
 313         head = &res->granted;
 314         list_for_each(iter, head) {
 315                 lock = list_entry(iter, struct dlm_lock, list);
 316                 if (lock==target)
 317                         continue;
 318                 if (!dlm_lock_compatible(lock->ml.type,
 319                                          target->ml.convert_type)) {
 320                         can_grant = 0;
 321                         /* queue the BAST if not already */
 322                         if (lock->ml.highest_blocked == LKM_IVMODE) {
 323                                 __dlm_lockres_reserve_ast(res);
 324                                 dlm_queue_bast(dlm, lock);
 325                         }
 326                         /* update the highest_blocked if needed */
 327                         if (lock->ml.highest_blocked < target->ml.convert_type)
 328                                 lock->ml.highest_blocked =
 329                                         target->ml.convert_type;
 330                 }
 331         }
 332         head = &res->converting;
 333         list_for_each(iter, head) {
 334                 lock = list_entry(iter, struct dlm_lock, list);
 335                 if (lock==target)
 336                         continue;
 337                 if (!dlm_lock_compatible(lock->ml.type,
 338                                          target->ml.convert_type)) {
 339                         can_grant = 0;
 340                         if (lock->ml.highest_blocked == LKM_IVMODE) {
 341                                 __dlm_lockres_reserve_ast(res);
 342                                 dlm_queue_bast(dlm, lock);
 343                         }
 344                         if (lock->ml.highest_blocked < target->ml.convert_type)
 345                                 lock->ml.highest_blocked =
 346                                         target->ml.convert_type;
 347                 }
 348         }
 349
 350         /* we can convert the lock */
 351         if (can_grant) {
 352                 spin_lock(&target->spinlock);
 353                 BUG_ON(target->ml.highest_blocked != LKM_IVMODE);
 354
 355                 mlog(0, "calling ast for converting lock: %.*s, have: %d, "
 356                      "granting: %d, node: %u\n", res->lockname.len,
 357                      res->lockname.name, target->ml.type,
 358                      target->ml.convert_type, target->ml.node);
 359
 360                 target->ml.type = target->ml.convert_type;
 361                 target->ml.convert_type = LKM_IVMODE;
 362                 list_move_tail(&target->list, &res->granted);
 363
 364                 BUG_ON(!target->lksb);
 365                 target->lksb->status = DLM_NORMAL;
 366
 367                 spin_unlock(&target->spinlock);
 368
 369                 __dlm_lockres_reserve_ast(res);
 370                 dlm_queue_ast(dlm, target);
 371                 /* go back and check for more */
 372                 goto converting;
 373         }
 374
 375 blocked:
 376         if (list_empty(&res->blocked))
 377                 goto leave;
 378         target = list_entry(res->blocked.next, struct dlm_lock, list);
 379
 380         head = &res->granted;
 381         list_for_each(iter, head) {
 382                 lock = list_entry(iter, struct dlm_lock, list);
 383                 if (lock==target)
 384                         continue;
 385                 if (!dlm_lock_compatible(lock->ml.type, target->ml.type)) {
 386                         can_grant = 0;
 387                         if (lock->ml.highest_blocked == LKM_IVMODE) {
 388                                 __dlm_lockres_reserve_ast(res);
 389                                 dlm_queue_bast(dlm, lock);
 390                         }
 391                         if (lock->ml.highest_blocked < target->ml.type)
 392                                 lock->ml.highest_blocked = target->ml.type;
 393                 }
 394         }
 395
 396         head = &res->converting;
 397         list_for_each(iter, head) {
 398                 lock = list_entry(iter, struct dlm_lock, list);
 399                 if (lock==target)
 400                         continue;
 401                 if (!dlm_lock_compatible(lock->ml.type, target->ml.type)) {
 402                         can_grant = 0;
 403                         if (lock->ml.highest_blocked == LKM_IVMODE) {
 404                                 __dlm_lockres_reserve_ast(res);
 405                                 dlm_queue_bast(dlm, lock);
 406                         }
 407                         if (lock->ml.highest_blocked < target->ml.type)
 408                                 lock->ml.highest_blocked = target->ml.type;
 409                 }
 410         }
 411
 412         /* we can grant the blocked lock (only
 413          * possible if converting list empty) */
 414         if (can_grant) {
 415                 spin_lock(&target->spinlock);
 416                 BUG_ON(target->ml.highest_blocked != LKM_IVMODE);
 417
 418                 mlog(0, "calling ast for blocked lock: %.*s, granting: %d, "
 419                      "node: %u\n", res->lockname.len, res->lockname.name,
 420                      target->ml.type, target->ml.node);
 421
 422                 // target->ml.type is already correct
 423                 list_move_tail(&target->list, &res->granted);
 424
 425                 BUG_ON(!target->lksb);
 426                 target->lksb->status = DLM_NORMAL;
 427
 428                 spin_unlock(&target->spinlock);
 429
 430                 __dlm_lockres_reserve_ast(res);
 431                 dlm_queue_ast(dlm, target);
 432                 /* go back and check for more */
 433                 goto converting;
 434         }
 435
 436 leave:
 437         return;
 438 }
 439
 440 /* must have NO locks when calling this with res !=NULL * */
 441 void dlm_kick_thread(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
 442 {
 443         mlog_entry("dlm=%p, res=%p\n", dlm, res);
 444         if (res) {
 445                 spin_lock(&dlm->spinlock);
 446                 spin_lock(&res->spinlock);
 447                 __dlm_dirty_lockres(dlm, res);
 448                 spin_unlock(&res->spinlock);
 449                 spin_unlock(&dlm->spinlock);
 450         }
 451         wake_up(&dlm->dlm_thread_wq);
 452 }
 453
 454 void __dlm_dirty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
 455 {
 456         mlog_entry("dlm=%p, res=%p\n", dlm, res);
 457
 458         assert_spin_locked(&dlm->spinlock);
 459         assert_spin_locked(&res->spinlock);
 460
 461         /* don't shuffle secondary queues */
 462         if ((res->owner == dlm->node_num)) {
 463                 if (res->state & (DLM_LOCK_RES_MIGRATING |
 464                                   DLM_LOCK_RES_BLOCK_DIRTY))
 465                     return;
 466
 467                 if (list_empty(&res->dirty)) {
 468                         /* ref for dirty_list */
 469                         dlm_lockres_get(res);
 470                         list_add_tail(&res->dirty, &dlm->dirty_list);
 471                         res->state |= DLM_LOCK_RES_DIRTY;
 472                 }
 473         }
 474 }
 475
 476
 477 /* Launch the NM thread for the mounted volume */
 478 int dlm_launch_thread(struct dlm_ctxt *dlm)
 479 {
 480         mlog(0, "starting dlm thread...\n");
 481
 482         dlm->dlm_thread_task = kthread_run(dlm_thread, dlm, "dlm_thread");
 483         if (IS_ERR(dlm->dlm_thread_task)) {
 484                 mlog_errno(PTR_ERR(dlm->dlm_thread_task));
 485                 dlm->dlm_thread_task = NULL;
 486                 return -EINVAL;
 487         }
 488
 489         return 0;
 490 }
 491
 492 void dlm_complete_thread(struct dlm_ctxt *dlm)
 493 {
 494         if (dlm->dlm_thread_task) {
 495                 mlog(ML_KTHREAD, "waiting for dlm thread to exit\n");
 496                 kthread_stop(dlm->dlm_thread_task);
 497                 dlm->dlm_thread_task = NULL;
 498         }
 499 }
 500
 501 static int dlm_dirty_list_empty(struct dlm_ctxt *dlm)
 502 {
 503         int empty;
 504
 505         spin_lock(&dlm->spinlock);
 506         empty = list_empty(&dlm->dirty_list);
 507         spin_unlock(&dlm->spinlock);
 508
 509         return empty;
 510 }
 511
 512 static void dlm_flush_asts(struct dlm_ctxt *dlm)
 513 {
 514         int ret;
 515         struct dlm_lock *lock;
 516         struct dlm_lock_resource *res;
 517         u8 hi;
 518
 519         spin_lock(&dlm->ast_lock);
 520         while (!list_empty(&dlm->pending_asts)) {
 521                 lock = list_entry(dlm->pending_asts.next,
 522                                   struct dlm_lock, ast_list);
 523                 /* get an extra ref on lock */
 524                 dlm_lock_get(lock);
 525                 res = lock->lockres;
 526                 mlog(0, "delivering an ast for this lockres\n");
 527
 528                 BUG_ON(!lock->ast_pending);
 529
 530                 /* remove from list (including ref) */
 531                 list_del_init(&lock->ast_list);
 532                 dlm_lock_put(lock);
 533                 spin_unlock(&dlm->ast_lock);
 534
 535                 if (lock->ml.node != dlm->node_num) {
 536                         ret = dlm_do_remote_ast(dlm, res, lock);
 537                         if (ret < 0)
 538                                 mlog_errno(ret);
 539                 } else
 540                         dlm_do_local_ast(dlm, res, lock);
 541
 542                 spin_lock(&dlm->ast_lock);
 543
 544                 /* possible that another ast was queued while
 545                  * we were delivering the last one */
 546                 if (!list_empty(&lock->ast_list)) {
 547                         mlog(0, "aha another ast got queued while "
 548                              "we were finishing the last one.  will "
 549                              "keep the ast_pending flag set.\n");
 550                 } else
 551                         lock->ast_pending = 0;
 552
 553                 /* drop the extra ref.
 554                  * this may drop it completely. */
 555                 dlm_lock_put(lock);
 556                 dlm_lockres_release_ast(dlm, res);
 557         }
 558
 559         while (!list_empty(&dlm->pending_basts)) {
 560                 lock = list_entry(dlm->pending_basts.next,
 561                                   struct dlm_lock, bast_list);
 562                 /* get an extra ref on lock */
 563                 dlm_lock_get(lock);
 564                 res = lock->lockres;
 565
 566                 BUG_ON(!lock->bast_pending);
 567
 568                 /* get the highest blocked lock, and reset */
 569                 spin_lock(&lock->spinlock);
 570                 BUG_ON(lock->ml.highest_blocked <= LKM_IVMODE);
 571                 hi = lock->ml.highest_blocked;
 572                 lock->ml.highest_blocked = LKM_IVMODE;
 573                 spin_unlock(&lock->spinlock);
 574
 575                 /* remove from list (including ref) */
 576                 list_del_init(&lock->bast_list);
 577                 dlm_lock_put(lock);
 578                 spin_unlock(&dlm->ast_lock);
 579
 580                 mlog(0, "delivering a bast for this lockres "
 581                      "(blocked = %d\n", hi);
 582
 583                 if (lock->ml.node != dlm->node_num) {
 584                         ret = dlm_send_proxy_bast(dlm, res, lock, hi);
 585                         if (ret < 0)
 586                                 mlog_errno(ret);
 587                 } else
 588                         dlm_do_local_bast(dlm, res, lock, hi);
 589
 590                 spin_lock(&dlm->ast_lock);
 591
 592                 /* possible that another bast was queued while
 593                  * we were delivering the last one */
 594                 if (!list_empty(&lock->bast_list)) {
 595                         mlog(0, "aha another bast got queued while "
 596                              "we were finishing the last one.  will "
 597                              "keep the bast_pending flag set.\n");
 598                 } else
 599                         lock->bast_pending = 0;
 600
 601                 /* drop the extra ref.
 602                  * this may drop it completely. */
 603                 dlm_lock_put(lock);
 604                 dlm_lockres_release_ast(dlm, res);
 605         }
 606         wake_up(&dlm->ast_wq);
 607         spin_unlock(&dlm->ast_lock);
 608 }
 609
 610
 611 #define DLM_THREAD_TIMEOUT_MS (4 * 1000)
 612 #define DLM_THREAD_MAX_DIRTY  100
 613 #define DLM_THREAD_MAX_ASTS   10
 614
 615 static int dlm_thread(void *data)
 616 {
 617         struct dlm_lock_resource *res;
 618         struct dlm_ctxt *dlm = data;
 619         unsigned long timeout = msecs_to_jiffies(DLM_THREAD_TIMEOUT_MS);
 620
 621         mlog(0, "dlm thread running for %s...\n", dlm->name);
 622
 623         while (!kthread_should_stop()) {
 624                 int n = DLM_THREAD_MAX_DIRTY;
 625
 626                 /* dlm_shutting_down is very point-in-time, but that
 627                  * doesn't matter as we'll just loop back around if we
 628                  * get false on the leading edge of a state
 629                  * transition. */
 630                 dlm_run_purge_list(dlm, dlm_shutting_down(dlm));
 631
 632                 /* We really don't want to hold dlm->spinlock while
 633                  * calling dlm_shuffle_lists on each lockres that
 634                  * needs to have its queues adjusted and AST/BASTs
 635                  * run.  So let's pull each entry off the dirty_list
 636                  * and drop dlm->spinlock ASAP.  Once off the list,
 637                  * res->spinlock needs to be taken again to protect
 638                  * the queues while calling dlm_shuffle_lists.  */
 639                 spin_lock(&dlm->spinlock);
 640                 while (!list_empty(&dlm->dirty_list)) {
 641                         int delay = 0;
 642                         res = list_entry(dlm->dirty_list.next,
 643                                          struct dlm_lock_resource, dirty);
 644
 645                         /* peel a lockres off, remove it from the list,
 646                          * unset the dirty flag and drop the dlm lock */
 647                         BUG_ON(!res);
 648                         dlm_lockres_get(res);
 649
 650                         spin_lock(&res->spinlock);
 651                         /* We clear the DLM_LOCK_RES_DIRTY state once we shuffle lists below */
 652                         list_del_init(&res->dirty);
 653                         spin_unlock(&res->spinlock);
 654                         spin_unlock(&dlm->spinlock);
 655                         /* Drop dirty_list ref */
 656                         dlm_lockres_put(res);
 657
 658                         /* lockres can be re-dirtied/re-added to the
 659                          * dirty_list in this gap, but that is ok */
 660
 661                         spin_lock(&res->spinlock);
 662                         if (res->owner != dlm->node_num) {
 663                                 __dlm_print_one_lock_resource(res);
 664                                 mlog(ML_ERROR, "inprog:%s, mig:%s, reco:%s, dirty:%s\n",
 665                                      res->state & DLM_LOCK_RES_IN_PROGRESS ? "yes" : "no",
 666                                      res->state & DLM_LOCK_RES_MIGRATING ? "yes" : "no",
 667                                      res->state & DLM_LOCK_RES_RECOVERING ? "yes" : "no",
 668                                      res->state & DLM_LOCK_RES_DIRTY ? "yes" : "no");
 669                         }
 670                         BUG_ON(res->owner != dlm->node_num);
 671
 672                         /* it is now ok to move lockreses in these states
 673                          * to the dirty list, assuming that they will only be
 674                          * dirty for a short while. */
 675                         BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
 676                         if (res->state & (DLM_LOCK_RES_IN_PROGRESS |
 677                                           DLM_LOCK_RES_RECOVERING)) {
 678                                 /* move it to the tail and keep going */
 679                                 res->state &= ~DLM_LOCK_RES_DIRTY;
 680                                 spin_unlock(&res->spinlock);
 681                                 mlog(0, "delaying list shuffling for in-"
 682                                      "progress lockres %.*s, state=%d\n",
 683                                      res->lockname.len, res->lockname.name,
 684                                      res->state);
 685                                 delay = 1;
 686                                 goto in_progress;
 687                         }
 688
 689                         /* at this point the lockres is not migrating/
 690                          * recovering/in-progress.  we have the lockres
 691                          * spinlock and do NOT have the dlm lock.
 692                          * safe to reserve/queue asts and run the lists. */
 693
 694                         mlog(0, "calling dlm_shuffle_lists with dlm=%s, "
 695                              "res=%.*s\n", dlm->name,
 696                              res->lockname.len, res->lockname.name);
 697
 698                         /* called while holding lockres lock */
 699                         dlm_shuffle_lists(dlm, res);
 700                         res->state &= ~DLM_LOCK_RES_DIRTY;
 701                         spin_unlock(&res->spinlock);
 702
 703                         dlm_lockres_calc_usage(dlm, res);
 704
 705 in_progress:
 706
 707                         spin_lock(&dlm->spinlock);
 708                         /* if the lock was in-progress, stick
 709                          * it on the back of the list */
 710                         if (delay) {
 711                                 spin_lock(&res->spinlock);
 712                                 __dlm_dirty_lockres(dlm, res);
 713                                 spin_unlock(&res->spinlock);
 714                         }
 715                         dlm_lockres_put(res);
 716
 717                         /* unlikely, but we may need to give time to
 718                          * other tasks */
 719                         if (!--n) {
 720                                 mlog(0, "throttling dlm_thread\n");
 721                                 break;
 722                         }
 723                 }
 724
 725                 spin_unlock(&dlm->spinlock);
 726                 dlm_flush_asts(dlm);
 727
 728                 /* yield and continue right away if there is more work to do */
 729                 if (!n) {
 730                         cond_resched();
 731                         continue;
 732                 }
 733
 734                 wait_event_interruptible_timeout(dlm->dlm_thread_wq,
 735                                                  !dlm_dirty_list_empty(dlm) ||
 736                                                  kthread_should_stop(),
 737                                                  timeout);
 738         }
 739
 740         mlog(0, "quitting DLM thread\n");
 741         return 0;
 742 }