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/highmem.h>
  32 #include <linux/init.h>
  33 #include <linux/sysctl.h>
  34 #include <linux/random.h>
  35 #include <linux/blkdev.h>
  36 #include <linux/socket.h>
  37 #include <linux/inet.h>
  38 #include <linux/timer.h>
  39 #include <linux/kthread.h>
  40 #include <linux/delay.h>
  41
  42
  43 #include "cluster/heartbeat.h"
  44 #include "cluster/nodemanager.h"
  45 #include "cluster/tcp.h"
  46
  47 #include "dlmapi.h"
  48 #include "dlmcommon.h"
  49 #include "dlmdomain.h"
  50
  51 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_THREAD)
  52 #include "cluster/masklog.h"
  53
  54 static int dlm_thread(void *data);
  55 static void dlm_flush_asts(struct dlm_ctxt *dlm);
  56
  57 #define dlm_lock_is_remote(dlm, lock)     ((lock)->ml.node != (dlm)->node_num)
  58
  59 /* will exit holding res->spinlock, but may drop in function */
  60 /* waits until flags are cleared on res->state */
  61 void __dlm_wait_on_lockres_flags(struct dlm_lock_resource *res, int flags)
  62 {
  63         DECLARE_WAITQUEUE(wait, current);
  64
  65         assert_spin_locked(&res->spinlock);
  66
  67         add_wait_queue(&res->wq, &wait);
  68 repeat:
  69         set_current_state(TASK_UNINTERRUPTIBLE);
  70         if (res->state & flags) {
  71                 spin_unlock(&res->spinlock);
  72                 schedule();
  73                 spin_lock(&res->spinlock);
  74                 goto repeat;
  75         }
  76         remove_wait_queue(&res->wq, &wait);
  77         __set_current_state(TASK_RUNNING);
  78 }
  79
  80 int __dlm_lockres_has_locks(struct dlm_lock_resource *res)
  81 {
  82         if (list_empty(&res->granted) &&
  83             list_empty(&res->converting) &&
  84             list_empty(&res->blocked))
  85                 return 0;
  86         return 1;
  87 }
  88
  89 /* "unused": the lockres has no locks, is not on the dirty list,
  90  * has no inflight locks (in the gap between mastery and acquiring
  91  * the first lock), and has no bits in its refmap.
  92  * truly ready to be freed. */
  93 int __dlm_lockres_unused(struct dlm_lock_resource *res)
  94 {
  95         int bit;
  96
  97         assert_spin_locked(&res->spinlock);
  98
  99         if (__dlm_lockres_has_locks(res))
 100                 return 0;
 101
 102         /* Locks are in the process of being created */
 103         if (res->inflight_locks)
 104                 return 0;
 105
 106         if (!list_empty(&res->dirty) || res->state & DLM_LOCK_RES_DIRTY)
 107                 return 0;
 108
 109         if (res->state & DLM_LOCK_RES_RECOVERING)
 110                 return 0;
 111
 112         /* Another node has this resource with this node as the master */
 113         bit = find_next_bit(res->refmap, O2NM_MAX_NODES, 0);
 114         if (bit < O2NM_MAX_NODES)
 115                 return 0;
 116
 117         return 1;
 118 }
 119
 120
 121 /* Call whenever you may have added or deleted something from one of
 122  * the lockres queue's. This will figure out whether it belongs on the
 123  * unused list or not and does the appropriate thing. */
 124 void __dlm_lockres_calc_usage(struct dlm_ctxt *dlm,
 125                               struct dlm_lock_resource *res)
 126 {
 127         assert_spin_locked(&dlm->spinlock);
 128         assert_spin_locked(&res->spinlock);
 129
 130         if (__dlm_lockres_unused(res)){
 131                 if (list_empty(&res->purge)) {
 132                         mlog(0, "%s: Adding res %.*s to purge list\n",
 133                              dlm->name, res->lockname.len, res->lockname.name);
 134
 135                         res->last_used = jiffies;
 136                         dlm_lockres_get(res);
 137                         list_add_tail(&res->purge, &dlm->purge_list);
 138                         dlm->purge_count++;
 139                 }
 140         } else if (!list_empty(&res->purge)) {
 141                 mlog(0, "%s: Removing res %.*s from purge list\n",
 142                      dlm->name, res->lockname.len, res->lockname.name);
 143
 144                 list_del_init(&res->purge);
 145                 dlm_lockres_put(res);
 146                 dlm->purge_count--;
 147         }
 148 }
 149
 150 void dlm_lockres_calc_usage(struct dlm_ctxt *dlm,
 151                             struct dlm_lock_resource *res)
 152 {
 153         spin_lock(&dlm->spinlock);
 154         spin_lock(&res->spinlock);
 155
 156         __dlm_lockres_calc_usage(dlm, res);
 157
 158         spin_unlock(&res->spinlock);
 159         spin_unlock(&dlm->spinlock);
 160 }
 161
 162 static void dlm_purge_lockres(struct dlm_ctxt *dlm,
 163                              struct dlm_lock_resource *res)
 164 {
 165         int master;
 166         int ret = 0;
 167
 168         assert_spin_locked(&dlm->spinlock);
 169         assert_spin_locked(&res->spinlock);
 170
 171         master = (res->owner == dlm->node_num);
 172
 173         mlog(0, "%s: Purging res %.*s, master %d\n", dlm->name,
 174              res->lockname.len, res->lockname.name, master);
 175
 176         if (!master) {
 177                 res->state |= DLM_LOCK_RES_DROPPING_REF;
 178                 /* drop spinlock...  retake below */
 179                 spin_unlock(&res->spinlock);
 180                 spin_unlock(&dlm->spinlock);
 181
 182                 spin_lock(&res->spinlock);
 183                 /* This ensures that clear refmap is sent after the set */
 184                 __dlm_wait_on_lockres_flags(res, DLM_LOCK_RES_SETREF_INPROG);
 185                 spin_unlock(&res->spinlock);
 186
 187                 /* clear our bit from the master's refmap, ignore errors */
 188                 ret = dlm_drop_lockres_ref(dlm, res);
 189                 if (ret < 0) {
 190                         if (!dlm_is_host_down(ret))
 191                                 BUG();
 192                 }
 193                 spin_lock(&dlm->spinlock);
 194                 spin_lock(&res->spinlock);
 195         }
 196
 197         if (!list_empty(&res->purge)) {
 198                 mlog(0, "%s: Removing res %.*s from purgelist, master %d\n",
 199                      dlm->name, res->lockname.len, res->lockname.name, master);
 200                 list_del_init(&res->purge);
 201                 dlm_lockres_put(res);
 202                 dlm->purge_count--;
 203         }
 204
 205         if (!master && ret != 0) {
 206                 mlog(0, "%s: deref %.*s in progress or master goes down\n",
 207                         dlm->name, res->lockname.len, res->lockname.name);
 208                 spin_unlock(&res->spinlock);
 209                 return;
 210         }
 211
 212         if (!__dlm_lockres_unused(res)) {
 213                 mlog(ML_ERROR, "%s: res %.*s in use after deref\n",
 214                      dlm->name, res->lockname.len, res->lockname.name);
 215                 __dlm_print_one_lock_resource(res);
 216                 BUG();
 217         }
 218
 219         __dlm_unhash_lockres(dlm, res);
 220
 221         spin_lock(&dlm->track_lock);
 222         if (!list_empty(&res->tracking))
 223                 list_del_init(&res->tracking);
 224         else {
 225                 mlog(ML_ERROR, "Resource %.*s not on the Tracking list\n",
 226                                 res->lockname.len, res->lockname.name);
 227                 __dlm_print_one_lock_resource(res);
 228         }
 229         spin_unlock(&dlm->track_lock);
 230
 231         /* lockres is not in the hash now.  drop the flag and wake up
 232          * any processes waiting in dlm_get_lock_resource. */
 233         if (!master) {
 234                 res->state &= ~DLM_LOCK_RES_DROPPING_REF;
 235                 spin_unlock(&res->spinlock);
 236                 wake_up(&res->wq);
 237         } else
 238                 spin_unlock(&res->spinlock);
 239 }
 240
 241 static void dlm_run_purge_list(struct dlm_ctxt *dlm,
 242                                int purge_now)
 243 {
 244         unsigned int run_max, unused;
 245         unsigned long purge_jiffies;
 246         struct dlm_lock_resource *lockres;
 247
 248         spin_lock(&dlm->spinlock);
 249         run_max = dlm->purge_count;
 250
 251         while(run_max && !list_empty(&dlm->purge_list)) {
 252                 run_max--;
 253
 254                 lockres = list_entry(dlm->purge_list.next,
 255                                      struct dlm_lock_resource, purge);
 256
 257                 spin_lock(&lockres->spinlock);
 258
 259                 purge_jiffies = lockres->last_used +
 260                         msecs_to_jiffies(DLM_PURGE_INTERVAL_MS);
 261
 262                 /* Make sure that we want to be processing this guy at
 263                  * this time. */
 264                 if (!purge_now && time_after(purge_jiffies, jiffies)) {
 265                         /* Since resources are added to the purge list
 266                          * in tail order, we can stop at the first
 267                          * unpurgable resource -- anyone added after
 268                          * him will have a greater last_used value */
 269                         spin_unlock(&lockres->spinlock);
 270                         break;
 271                 }
 272
 273                 /* Status of the lockres *might* change so double
 274                  * check. If the lockres is unused, holding the dlm
 275                  * spinlock will prevent people from getting and more
 276                  * refs on it. */
 277                 unused = __dlm_lockres_unused(lockres);
 278                 if (!unused ||
 279                     (lockres->state & DLM_LOCK_RES_MIGRATING) ||
 280                     (lockres->inflight_assert_workers != 0)) {
 281                         mlog(0, "%s: res %.*s is in use or being remastered, "
 282                              "used %d, state %d, assert master workers %u\n",
 283                              dlm->name, lockres->lockname.len,
 284                              lockres->lockname.name,
 285                              !unused, lockres->state,
 286                              lockres->inflight_assert_workers);
 287                         list_move_tail(&lockres->purge, &dlm->purge_list);
 288                         spin_unlock(&lockres->spinlock);
 289                         continue;
 290                 }
 291
 292                 dlm_lockres_get(lockres);
 293
 294                 dlm_purge_lockres(dlm, lockres);
 295
 296                 dlm_lockres_put(lockres);
 297
 298                 /* Avoid adding any scheduling latencies */
 299                 cond_resched_lock(&dlm->spinlock);
 300         }
 301
 302         spin_unlock(&dlm->spinlock);
 303 }
 304
 305 static void dlm_shuffle_lists(struct dlm_ctxt *dlm,
 306                               struct dlm_lock_resource *res)
 307 {
 308         struct dlm_lock *lock, *target;
 309         int can_grant = 1;
 310
 311         /*
 312          * Because this function is called with the lockres
 313          * spinlock, and because we know that it is not migrating/
 314          * recovering/in-progress, it is fine to reserve asts and
 315          * basts right before queueing them all throughout
 316          */
 317         assert_spin_locked(&dlm->ast_lock);
 318         assert_spin_locked(&res->spinlock);
 319         BUG_ON((res->state & (DLM_LOCK_RES_MIGRATING|
 320                               DLM_LOCK_RES_RECOVERING|
 321                               DLM_LOCK_RES_IN_PROGRESS)));
 322
 323 converting:
 324         if (list_empty(&res->converting))
 325                 goto blocked;
 326         mlog(0, "%s: res %.*s has locks on the convert queue\n", dlm->name,
 327              res->lockname.len, res->lockname.name);
 328
 329         target = list_entry(res->converting.next, struct dlm_lock, list);
 330         if (target->ml.convert_type == LKM_IVMODE) {
 331                 mlog(ML_ERROR, "%s: res %.*s converting lock to invalid mode\n",
 332                      dlm->name, res->lockname.len, res->lockname.name);
 333                 BUG();
 334         }
 335         list_for_each_entry(lock, &res->granted, list) {
 336                 if (lock==target)
 337                         continue;
 338                 if (!dlm_lock_compatible(lock->ml.type,
 339                                          target->ml.convert_type)) {
 340                         can_grant = 0;
 341                         /* queue the BAST if not already */
 342                         if (lock->ml.highest_blocked == LKM_IVMODE) {
 343                                 __dlm_lockres_reserve_ast(res);
 344                                 __dlm_queue_bast(dlm, lock);
 345                         }
 346                         /* update the highest_blocked if needed */
 347                         if (lock->ml.highest_blocked < target->ml.convert_type)
 348                                 lock->ml.highest_blocked =
 349                                         target->ml.convert_type;
 350                 }
 351         }
 352
 353         list_for_each_entry(lock, &res->converting, list) {
 354                 if (lock==target)
 355                         continue;
 356                 if (!dlm_lock_compatible(lock->ml.type,
 357                                          target->ml.convert_type)) {
 358                         can_grant = 0;
 359                         if (lock->ml.highest_blocked == LKM_IVMODE) {
 360                                 __dlm_lockres_reserve_ast(res);
 361                                 __dlm_queue_bast(dlm, lock);
 362                         }
 363                         if (lock->ml.highest_blocked < target->ml.convert_type)
 364                                 lock->ml.highest_blocked =
 365                                         target->ml.convert_type;
 366                 }
 367         }
 368
 369         /* we can convert the lock */
 370         if (can_grant) {
 371                 spin_lock(&target->spinlock);
 372                 BUG_ON(target->ml.highest_blocked != LKM_IVMODE);
 373
 374                 mlog(0, "%s: res %.*s, AST for Converting lock %u:%llu, type "
 375                      "%d => %d, node %u\n", dlm->name, res->lockname.len,
 376                      res->lockname.name,
 377                      dlm_get_lock_cookie_node(be64_to_cpu(target->ml.cookie)),
 378                      dlm_get_lock_cookie_seq(be64_to_cpu(target->ml.cookie)),
 379                      target->ml.type,
 380                      target->ml.convert_type, target->ml.node);
 381
 382                 target->ml.type = target->ml.convert_type;
 383                 target->ml.convert_type = LKM_IVMODE;
 384                 list_move_tail(&target->list, &res->granted);
 385
 386                 BUG_ON(!target->lksb);
 387                 target->lksb->status = DLM_NORMAL;
 388
 389                 spin_unlock(&target->spinlock);
 390
 391                 __dlm_lockres_reserve_ast(res);
 392                 __dlm_queue_ast(dlm, target);
 393                 /* go back and check for more */
 394                 goto converting;
 395         }
 396
 397 blocked:
 398         if (list_empty(&res->blocked))
 399                 goto leave;
 400         target = list_entry(res->blocked.next, struct dlm_lock, list);
 401
 402         list_for_each_entry(lock, &res->granted, list) {
 403                 if (lock==target)
 404                         continue;
 405                 if (!dlm_lock_compatible(lock->ml.type, target->ml.type)) {
 406                         can_grant = 0;
 407                         if (lock->ml.highest_blocked == LKM_IVMODE) {
 408                                 __dlm_lockres_reserve_ast(res);
 409                                 __dlm_queue_bast(dlm, lock);
 410                         }
 411                         if (lock->ml.highest_blocked < target->ml.type)
 412                                 lock->ml.highest_blocked = target->ml.type;
 413                 }
 414         }
 415
 416         list_for_each_entry(lock, &res->converting, list) {
 417                 if (lock==target)
 418                         continue;
 419                 if (!dlm_lock_compatible(lock->ml.type, target->ml.type)) {
 420                         can_grant = 0;
 421                         if (lock->ml.highest_blocked == LKM_IVMODE) {
 422                                 __dlm_lockres_reserve_ast(res);
 423                                 __dlm_queue_bast(dlm, lock);
 424                         }
 425                         if (lock->ml.highest_blocked < target->ml.type)
 426                                 lock->ml.highest_blocked = target->ml.type;
 427                 }
 428         }
 429
 430         /* we can grant the blocked lock (only
 431          * possible if converting list empty) */
 432         if (can_grant) {
 433                 spin_lock(&target->spinlock);
 434                 BUG_ON(target->ml.highest_blocked != LKM_IVMODE);
 435
 436                 mlog(0, "%s: res %.*s, AST for Blocked lock %u:%llu, type %d, "
 437                      "node %u\n", dlm->name, res->lockname.len,
 438                      res->lockname.name,
 439                      dlm_get_lock_cookie_node(be64_to_cpu(target->ml.cookie)),
 440                      dlm_get_lock_cookie_seq(be64_to_cpu(target->ml.cookie)),
 441                      target->ml.type, target->ml.node);
 442
 443                 /* target->ml.type is already correct */
 444                 list_move_tail(&target->list, &res->granted);
 445
 446                 BUG_ON(!target->lksb);
 447                 target->lksb->status = DLM_NORMAL;
 448
 449                 spin_unlock(&target->spinlock);
 450
 451                 __dlm_lockres_reserve_ast(res);
 452                 __dlm_queue_ast(dlm, target);
 453                 /* go back and check for more */
 454                 goto converting;
 455         }
 456
 457 leave:
 458         return;
 459 }
 460
 461 /* must have NO locks when calling this with res !=NULL * */
 462 void dlm_kick_thread(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
 463 {
 464         if (res) {
 465                 spin_lock(&dlm->spinlock);
 466                 spin_lock(&res->spinlock);
 467                 __dlm_dirty_lockres(dlm, res);
 468                 spin_unlock(&res->spinlock);
 469                 spin_unlock(&dlm->spinlock);
 470         }
 471         wake_up(&dlm->dlm_thread_wq);
 472 }
 473
 474 void __dlm_dirty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
 475 {
 476         assert_spin_locked(&dlm->spinlock);
 477         assert_spin_locked(&res->spinlock);
 478
 479         /* don't shuffle secondary queues */
 480         if ((res->owner == dlm->node_num)) {
 481                 if (res->state & (DLM_LOCK_RES_MIGRATING |
 482                                   DLM_LOCK_RES_BLOCK_DIRTY))
 483                     return;
 484
 485                 if (list_empty(&res->dirty)) {
 486                         /* ref for dirty_list */
 487                         dlm_lockres_get(res);
 488                         list_add_tail(&res->dirty, &dlm->dirty_list);
 489                         res->state |= DLM_LOCK_RES_DIRTY;
 490                 }
 491         }
 492
 493         mlog(0, "%s: res %.*s\n", dlm->name, res->lockname.len,
 494              res->lockname.name);
 495 }
 496
 497
 498 /* Launch the NM thread for the mounted volume */
 499 int dlm_launch_thread(struct dlm_ctxt *dlm)
 500 {
 501         mlog(0, "Starting dlm_thread...\n");
 502
 503         dlm->dlm_thread_task = kthread_run(dlm_thread, dlm, "dlm-%s",
 504                         dlm->name);
 505         if (IS_ERR(dlm->dlm_thread_task)) {
 506                 mlog_errno(PTR_ERR(dlm->dlm_thread_task));
 507                 dlm->dlm_thread_task = NULL;
 508                 return -EINVAL;
 509         }
 510
 511         return 0;
 512 }
 513
 514 void dlm_complete_thread(struct dlm_ctxt *dlm)
 515 {
 516         if (dlm->dlm_thread_task) {
 517                 mlog(ML_KTHREAD, "Waiting for dlm thread to exit\n");
 518                 kthread_stop(dlm->dlm_thread_task);
 519                 dlm->dlm_thread_task = NULL;
 520         }
 521 }
 522
 523 static int dlm_dirty_list_empty(struct dlm_ctxt *dlm)
 524 {
 525         int empty;
 526
 527         spin_lock(&dlm->spinlock);
 528         empty = list_empty(&dlm->dirty_list);
 529         spin_unlock(&dlm->spinlock);
 530
 531         return empty;
 532 }
 533
 534 static void dlm_flush_asts(struct dlm_ctxt *dlm)
 535 {
 536         int ret;
 537         struct dlm_lock *lock;
 538         struct dlm_lock_resource *res;
 539         u8 hi;
 540
 541         spin_lock(&dlm->ast_lock);
 542         while (!list_empty(&dlm->pending_asts)) {
 543                 lock = list_entry(dlm->pending_asts.next,
 544                                   struct dlm_lock, ast_list);
 545                 /* get an extra ref on lock */
 546                 dlm_lock_get(lock);
 547                 res = lock->lockres;
 548                 mlog(0, "%s: res %.*s, Flush AST for lock %u:%llu, type %d, "
 549                      "node %u\n", dlm->name, res->lockname.len,
 550                      res->lockname.name,
 551                      dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)),
 552                      dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)),
 553                      lock->ml.type, lock->ml.node);
 554
 555                 BUG_ON(!lock->ast_pending);
 556
 557                 /* remove from list (including ref) */
 558                 list_del_init(&lock->ast_list);
 559                 dlm_lock_put(lock);
 560                 spin_unlock(&dlm->ast_lock);
 561
 562                 if (lock->ml.node != dlm->node_num) {
 563                         ret = dlm_do_remote_ast(dlm, res, lock);
 564                         if (ret < 0)
 565                                 mlog_errno(ret);
 566                 } else
 567                         dlm_do_local_ast(dlm, res, lock);
 568
 569                 spin_lock(&dlm->ast_lock);
 570
 571                 /* possible that another ast was queued while
 572                  * we were delivering the last one */
 573                 if (!list_empty(&lock->ast_list)) {
 574                         mlog(0, "%s: res %.*s, AST queued while flushing last "
 575                              "one\n", dlm->name, res->lockname.len,
 576                              res->lockname.name);
 577                 } else
 578                         lock->ast_pending = 0;
 579
 580                 /* drop the extra ref.
 581                  * this may drop it completely. */
 582                 dlm_lock_put(lock);
 583                 dlm_lockres_release_ast(dlm, res);
 584         }
 585
 586         while (!list_empty(&dlm->pending_basts)) {
 587                 lock = list_entry(dlm->pending_basts.next,
 588                                   struct dlm_lock, bast_list);
 589                 /* get an extra ref on lock */
 590                 dlm_lock_get(lock);
 591                 res = lock->lockres;
 592
 593                 BUG_ON(!lock->bast_pending);
 594
 595                 /* get the highest blocked lock, and reset */
 596                 spin_lock(&lock->spinlock);
 597                 BUG_ON(lock->ml.highest_blocked <= LKM_IVMODE);
 598                 hi = lock->ml.highest_blocked;
 599                 lock->ml.highest_blocked = LKM_IVMODE;
 600                 spin_unlock(&lock->spinlock);
 601
 602                 /* remove from list (including ref) */
 603                 list_del_init(&lock->bast_list);
 604                 dlm_lock_put(lock);
 605                 spin_unlock(&dlm->ast_lock);
 606
 607                 mlog(0, "%s: res %.*s, Flush BAST for lock %u:%llu, "
 608                      "blocked %d, node %u\n",
 609                      dlm->name, res->lockname.len, res->lockname.name,
 610                      dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)),
 611                      dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)),
 612                      hi, lock->ml.node);
 613
 614                 if (lock->ml.node != dlm->node_num) {
 615                         ret = dlm_send_proxy_bast(dlm, res, lock, hi);
 616                         if (ret < 0)
 617                                 mlog_errno(ret);
 618                 } else
 619                         dlm_do_local_bast(dlm, res, lock, hi);
 620
 621                 spin_lock(&dlm->ast_lock);
 622
 623                 /* possible that another bast was queued while
 624                  * we were delivering the last one */
 625                 if (!list_empty(&lock->bast_list)) {
 626                         mlog(0, "%s: res %.*s, BAST queued while flushing last "
 627                              "one\n", dlm->name, res->lockname.len,
 628                              res->lockname.name);
 629                 } else
 630                         lock->bast_pending = 0;
 631
 632                 /* drop the extra ref.
 633                  * this may drop it completely. */
 634                 dlm_lock_put(lock);
 635                 dlm_lockres_release_ast(dlm, res);
 636         }
 637         wake_up(&dlm->ast_wq);
 638         spin_unlock(&dlm->ast_lock);
 639 }
 640
 641
 642 #define DLM_THREAD_TIMEOUT_MS (4 * 1000)
 643 #define DLM_THREAD_MAX_DIRTY  100
 644 #define DLM_THREAD_MAX_ASTS   10
 645
 646 static int dlm_thread(void *data)
 647 {
 648         struct dlm_lock_resource *res;
 649         struct dlm_ctxt *dlm = data;
 650         unsigned long timeout = msecs_to_jiffies(DLM_THREAD_TIMEOUT_MS);
 651
 652         mlog(0, "dlm thread running for %s...\n", dlm->name);
 653
 654         while (!kthread_should_stop()) {
 655                 int n = DLM_THREAD_MAX_DIRTY;
 656
 657                 /* dlm_shutting_down is very point-in-time, but that
 658                  * doesn't matter as we'll just loop back around if we
 659                  * get false on the leading edge of a state
 660                  * transition. */
 661                 dlm_run_purge_list(dlm, dlm_shutting_down(dlm));
 662
 663                 /* We really don't want to hold dlm->spinlock while
 664                  * calling dlm_shuffle_lists on each lockres that
 665                  * needs to have its queues adjusted and AST/BASTs
 666                  * run.  So let's pull each entry off the dirty_list
 667                  * and drop dlm->spinlock ASAP.  Once off the list,
 668                  * res->spinlock needs to be taken again to protect
 669                  * the queues while calling dlm_shuffle_lists.  */
 670                 spin_lock(&dlm->spinlock);
 671                 while (!list_empty(&dlm->dirty_list)) {
 672                         int delay = 0;
 673                         res = list_entry(dlm->dirty_list.next,
 674                                          struct dlm_lock_resource, dirty);
 675
 676                         /* peel a lockres off, remove it from the list,
 677                          * unset the dirty flag and drop the dlm lock */
 678                         BUG_ON(!res);
 679                         dlm_lockres_get(res);
 680
 681                         spin_lock(&res->spinlock);
 682                         /* We clear the DLM_LOCK_RES_DIRTY state once we shuffle lists below */
 683                         list_del_init(&res->dirty);
 684                         spin_unlock(&res->spinlock);
 685                         spin_unlock(&dlm->spinlock);
 686                         /* Drop dirty_list ref */
 687                         dlm_lockres_put(res);
 688
 689                         /* lockres can be re-dirtied/re-added to the
 690                          * dirty_list in this gap, but that is ok */
 691
 692                         spin_lock(&dlm->ast_lock);
 693                         spin_lock(&res->spinlock);
 694                         if (res->owner != dlm->node_num) {
 695                                 __dlm_print_one_lock_resource(res);
 696                                 mlog(ML_ERROR, "%s: inprog %d, mig %d, reco %d,"
 697                                      " dirty %d\n", dlm->name,
 698                                      !!(res->state & DLM_LOCK_RES_IN_PROGRESS),
 699                                      !!(res->state & DLM_LOCK_RES_MIGRATING),
 700                                      !!(res->state & DLM_LOCK_RES_RECOVERING),
 701                                      !!(res->state & DLM_LOCK_RES_DIRTY));
 702                         }
 703                         BUG_ON(res->owner != dlm->node_num);
 704
 705                         /* it is now ok to move lockreses in these states
 706                          * to the dirty list, assuming that they will only be
 707                          * dirty for a short while. */
 708                         BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
 709                         if (res->state & (DLM_LOCK_RES_IN_PROGRESS |
 710                                           DLM_LOCK_RES_RECOVERING)) {
 711                                 /* move it to the tail and keep going */
 712                                 res->state &= ~DLM_LOCK_RES_DIRTY;
 713                                 spin_unlock(&res->spinlock);
 714                                 spin_unlock(&dlm->ast_lock);
 715                                 mlog(0, "%s: res %.*s, inprogress, delay list "
 716                                      "shuffle, state %d\n", dlm->name,
 717                                      res->lockname.len, res->lockname.name,
 718                                      res->state);
 719                                 delay = 1;
 720                                 goto in_progress;
 721                         }
 722
 723                         /* at this point the lockres is not migrating/
 724                          * recovering/in-progress.  we have the lockres
 725                          * spinlock and do NOT have the dlm lock.
 726                          * safe to reserve/queue asts and run the lists. */
 727
 728                         /* called while holding lockres lock */
 729                         dlm_shuffle_lists(dlm, res);
 730                         res->state &= ~DLM_LOCK_RES_DIRTY;
 731                         spin_unlock(&res->spinlock);
 732                         spin_unlock(&dlm->ast_lock);
 733
 734                         dlm_lockres_calc_usage(dlm, res);
 735
 736 in_progress:
 737
 738                         spin_lock(&dlm->spinlock);
 739                         /* if the lock was in-progress, stick
 740                          * it on the back of the list */
 741                         if (delay) {
 742                                 spin_lock(&res->spinlock);
 743                                 __dlm_dirty_lockres(dlm, res);
 744                                 spin_unlock(&res->spinlock);
 745                         }
 746                         dlm_lockres_put(res);
 747
 748                         /* unlikely, but we may need to give time to
 749                          * other tasks */
 750                         if (!--n) {
 751                                 mlog(0, "%s: Throttling dlm thread\n",
 752                                      dlm->name);
 753                                 break;
 754                         }
 755                 }
 756
 757                 spin_unlock(&dlm->spinlock);
 758                 dlm_flush_asts(dlm);
 759
 760                 /* yield and continue right away if there is more work to do */
 761                 if (!n) {
 762                         cond_resched();
 763                         continue;
 764                 }
 765
 766                 wait_event_interruptible_timeout(dlm->dlm_thread_wq,
 767                                                  !dlm_dirty_list_empty(dlm) ||
 768                                                  kthread_should_stop(),
 769                                                  timeout);
 770         }
 771
 772         mlog(0, "quitting DLM thread\n");
 773         return 0;
 774 }