fs/ocfs2/dlm/dlmthread.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /* -*- mode: c; c-basic-offset: 8; -*-
   3  * vim: noexpandtab sw=8 ts=8 sts=0:
   4  *
   5  * dlmthread.c
   6  *
   7  * standalone DLM module
   8  *
   9  * Copyright (C) 2004 Oracle.  All rights reserved.
  10  */
  11
  12
  13 #include <linux/module.h>
  14 #include <linux/fs.h>
  15 #include <linux/types.h>
  16 #include <linux/highmem.h>
  17 #include <linux/init.h>
  18 #include <linux/sysctl.h>
  19 #include <linux/random.h>
  20 #include <linux/blkdev.h>
  21 #include <linux/socket.h>
  22 #include <linux/inet.h>
  23 #include <linux/timer.h>
  24 #include <linux/kthread.h>
  25 #include <linux/delay.h>
  26
  27
  28 #include "../cluster/heartbeat.h"
  29 #include "../cluster/nodemanager.h"
  30 #include "../cluster/tcp.h"
  31
  32 #include "dlmapi.h"
  33 #include "dlmcommon.h"
  34 #include "dlmdomain.h"
  35
  36 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_THREAD)
  37 #include "../cluster/masklog.h"
  38
  39 static int dlm_thread(void *data);
  40 static void dlm_flush_asts(struct dlm_ctxt *dlm);
  41
  42 /* will exit holding res->spinlock, but may drop in function */
  43 /* waits until flags are cleared on res->state */
  44 void __dlm_wait_on_lockres_flags(struct dlm_lock_resource *res, int flags)
  45 {
  46         DECLARE_WAITQUEUE(wait, current);
  47
  48         assert_spin_locked(&res->spinlock);
  49
  50         add_wait_queue(&res->wq, &wait);
  51 repeat:
  52         set_current_state(TASK_UNINTERRUPTIBLE);
  53         if (res->state & flags) {
  54                 spin_unlock(&res->spinlock);
  55                 schedule();
  56                 spin_lock(&res->spinlock);
  57                 goto repeat;
  58         }
  59         remove_wait_queue(&res->wq, &wait);
  60         __set_current_state(TASK_RUNNING);
  61 }
  62
  63 int __dlm_lockres_has_locks(struct dlm_lock_resource *res)
  64 {
  65         if (list_empty(&res->granted) &&
  66             list_empty(&res->converting) &&
  67             list_empty(&res->blocked))
  68                 return 0;
  69         return 1;
  70 }
  71
  72 /* "unused": the lockres has no locks, is not on the dirty list,
  73  * has no inflight locks (in the gap between mastery and acquiring
  74  * the first lock), and has no bits in its refmap.
  75  * truly ready to be freed. */
  76 int __dlm_lockres_unused(struct dlm_lock_resource *res)
  77 {
  78         int bit;
  79
  80         assert_spin_locked(&res->spinlock);
  81
  82         if (__dlm_lockres_has_locks(res))
  83                 return 0;
  84
  85         /* Locks are in the process of being created */
  86         if (res->inflight_locks)
  87                 return 0;
  88
  89         if (!list_empty(&res->dirty) || res->state & DLM_LOCK_RES_DIRTY)
  90                 return 0;
  91
  92         if (res->state & (DLM_LOCK_RES_RECOVERING|
  93                         DLM_LOCK_RES_RECOVERY_WAITING))
  94                 return 0;
  95
  96         /* Another node has this resource with this node as the master */
  97         bit = find_next_bit(res->refmap, O2NM_MAX_NODES, 0);
  98         if (bit < O2NM_MAX_NODES)
  99                 return 0;
 100
 101         return 1;
 102 }
 103
 104
 105 /* Call whenever you may have added or deleted something from one of
 106  * the lockres queue's. This will figure out whether it belongs on the
 107  * unused list or not and does the appropriate thing. */
 108 void __dlm_lockres_calc_usage(struct dlm_ctxt *dlm,
 109                               struct dlm_lock_resource *res)
 110 {
 111         assert_spin_locked(&dlm->spinlock);
 112         assert_spin_locked(&res->spinlock);
 113
 114         if (__dlm_lockres_unused(res)){
 115                 if (list_empty(&res->purge)) {
 116                         mlog(0, "%s: Adding res %.*s to purge list\n",
 117                              dlm->name, res->lockname.len, res->lockname.name);
 118
 119                         res->last_used = jiffies;
 120                         dlm_lockres_get(res);
 121                         list_add_tail(&res->purge, &dlm->purge_list);
 122                         dlm->purge_count++;
 123                 }
 124         } else if (!list_empty(&res->purge)) {
 125                 mlog(0, "%s: Removing res %.*s from purge list\n",
 126                      dlm->name, res->lockname.len, res->lockname.name);
 127
 128                 list_del_init(&res->purge);
 129                 dlm_lockres_put(res);
 130                 dlm->purge_count--;
 131         }
 132 }
 133
 134 void dlm_lockres_calc_usage(struct dlm_ctxt *dlm,
 135                             struct dlm_lock_resource *res)
 136 {
 137         spin_lock(&dlm->spinlock);
 138         spin_lock(&res->spinlock);
 139
 140         __dlm_lockres_calc_usage(dlm, res);
 141
 142         spin_unlock(&res->spinlock);
 143         spin_unlock(&dlm->spinlock);
 144 }
 145
 146 /*
 147  * Do the real purge work:
 148  *     unhash the lockres, and
 149  *     clear flag DLM_LOCK_RES_DROPPING_REF.
 150  * It requires dlm and lockres spinlock to be taken.
 151  */
 152 void __dlm_do_purge_lockres(struct dlm_ctxt *dlm,
 153                 struct dlm_lock_resource *res)
 154 {
 155         assert_spin_locked(&dlm->spinlock);
 156         assert_spin_locked(&res->spinlock);
 157
 158         if (!list_empty(&res->purge)) {
 159                 mlog(0, "%s: Removing res %.*s from purgelist\n",
 160                      dlm->name, res->lockname.len, res->lockname.name);
 161                 list_del_init(&res->purge);
 162                 dlm_lockres_put(res);
 163                 dlm->purge_count--;
 164         }
 165
 166         if (!__dlm_lockres_unused(res)) {
 167                 mlog(ML_ERROR, "%s: res %.*s in use after deref\n",
 168                      dlm->name, res->lockname.len, res->lockname.name);
 169                 __dlm_print_one_lock_resource(res);
 170                 BUG();
 171         }
 172
 173         __dlm_unhash_lockres(dlm, res);
 174
 175         spin_lock(&dlm->track_lock);
 176         if (!list_empty(&res->tracking))
 177                 list_del_init(&res->tracking);
 178         else {
 179                 mlog(ML_ERROR, "%s: Resource %.*s not on the Tracking list\n",
 180                      dlm->name, res->lockname.len, res->lockname.name);
 181                 __dlm_print_one_lock_resource(res);
 182         }
 183         spin_unlock(&dlm->track_lock);
 184
 185         /*
 186          * lockres is not in the hash now. drop the flag and wake up
 187          * any processes waiting in dlm_get_lock_resource.
 188          */
 189         res->state &= ~DLM_LOCK_RES_DROPPING_REF;
 190 }
 191
 192 static void dlm_purge_lockres(struct dlm_ctxt *dlm,
 193                              struct dlm_lock_resource *res)
 194 {
 195         int master;
 196         int ret = 0;
 197
 198         assert_spin_locked(&dlm->spinlock);
 199         assert_spin_locked(&res->spinlock);
 200
 201         master = (res->owner == dlm->node_num);
 202
 203         mlog(0, "%s: Purging res %.*s, master %d\n", dlm->name,
 204              res->lockname.len, res->lockname.name, master);
 205
 206         if (!master) {
 207                 if (res->state & DLM_LOCK_RES_DROPPING_REF) {
 208                         mlog(ML_NOTICE, "%s: res %.*s already in DLM_LOCK_RES_DROPPING_REF state\n",
 209                                 dlm->name, res->lockname.len, res->lockname.name);
 210                         spin_unlock(&res->spinlock);
 211                         return;
 212                 }
 213
 214                 res->state |= DLM_LOCK_RES_DROPPING_REF;
 215                 /* drop spinlock...  retake below */
 216                 spin_unlock(&res->spinlock);
 217                 spin_unlock(&dlm->spinlock);
 218
 219                 spin_lock(&res->spinlock);
 220                 /* This ensures that clear refmap is sent after the set */
 221                 __dlm_wait_on_lockres_flags(res, DLM_LOCK_RES_SETREF_INPROG);
 222                 spin_unlock(&res->spinlock);
 223
 224                 /* clear our bit from the master's refmap, ignore errors */
 225                 ret = dlm_drop_lockres_ref(dlm, res);
 226                 if (ret < 0) {
 227                         if (!dlm_is_host_down(ret))
 228                                 BUG();
 229                 }
 230                 spin_lock(&dlm->spinlock);
 231                 spin_lock(&res->spinlock);
 232         }
 233
 234         if (!list_empty(&res->purge)) {
 235                 mlog(0, "%s: Removing res %.*s from purgelist, master %d\n",
 236                      dlm->name, res->lockname.len, res->lockname.name, master);
 237                 list_del_init(&res->purge);
 238                 dlm_lockres_put(res);
 239                 dlm->purge_count--;
 240         }
 241
 242         if (!master && ret == DLM_DEREF_RESPONSE_INPROG) {
 243                 mlog(0, "%s: deref %.*s in progress\n",
 244                         dlm->name, res->lockname.len, res->lockname.name);
 245                 spin_unlock(&res->spinlock);
 246                 return;
 247         }
 248
 249         if (!__dlm_lockres_unused(res)) {
 250                 mlog(ML_ERROR, "%s: res %.*s in use after deref\n",
 251                      dlm->name, res->lockname.len, res->lockname.name);
 252                 __dlm_print_one_lock_resource(res);
 253                 BUG();
 254         }
 255
 256         __dlm_unhash_lockres(dlm, res);
 257
 258         spin_lock(&dlm->track_lock);
 259         if (!list_empty(&res->tracking))
 260                 list_del_init(&res->tracking);
 261         else {
 262                 mlog(ML_ERROR, "Resource %.*s not on the Tracking list\n",
 263                                 res->lockname.len, res->lockname.name);
 264                 __dlm_print_one_lock_resource(res);
 265         }
 266         spin_unlock(&dlm->track_lock);
 267
 268         /* lockres is not in the hash now.  drop the flag and wake up
 269          * any processes waiting in dlm_get_lock_resource. */
 270         if (!master) {
 271                 res->state &= ~DLM_LOCK_RES_DROPPING_REF;
 272                 spin_unlock(&res->spinlock);
 273                 wake_up(&res->wq);
 274         } else
 275                 spin_unlock(&res->spinlock);
 276 }
 277
 278 static void dlm_run_purge_list(struct dlm_ctxt *dlm,
 279                                int purge_now)
 280 {
 281         unsigned int run_max, unused;
 282         unsigned long purge_jiffies;
 283         struct dlm_lock_resource *lockres;
 284
 285         spin_lock(&dlm->spinlock);
 286         run_max = dlm->purge_count;
 287
 288         while(run_max && !list_empty(&dlm->purge_list)) {
 289                 run_max--;
 290
 291                 lockres = list_entry(dlm->purge_list.next,
 292                                      struct dlm_lock_resource, purge);
 293
 294                 spin_lock(&lockres->spinlock);
 295
 296                 purge_jiffies = lockres->last_used +
 297                         msecs_to_jiffies(DLM_PURGE_INTERVAL_MS);
 298
 299                 /* Make sure that we want to be processing this guy at
 300                  * this time. */
 301                 if (!purge_now && time_after(purge_jiffies, jiffies)) {
 302                         /* Since resources are added to the purge list
 303                          * in tail order, we can stop at the first
 304                          * unpurgable resource -- anyone added after
 305                          * him will have a greater last_used value */
 306                         spin_unlock(&lockres->spinlock);
 307                         break;
 308                 }
 309
 310                 /* Status of the lockres *might* change so double
 311                  * check. If the lockres is unused, holding the dlm
 312                  * spinlock will prevent people from getting and more
 313                  * refs on it. */
 314                 unused = __dlm_lockres_unused(lockres);
 315                 if (!unused ||
 316                     (lockres->state & DLM_LOCK_RES_MIGRATING) ||
 317                     (lockres->inflight_assert_workers != 0)) {
 318                         mlog(0, "%s: res %.*s is in use or being remastered, "
 319                              "used %d, state %d, assert master workers %u\n",
 320                              dlm->name, lockres->lockname.len,
 321                              lockres->lockname.name,
 322                              !unused, lockres->state,
 323                              lockres->inflight_assert_workers);
 324                         list_move_tail(&lockres->purge, &dlm->purge_list);
 325                         spin_unlock(&lockres->spinlock);
 326                         continue;
 327                 }
 328
 329                 dlm_lockres_get(lockres);
 330
 331                 dlm_purge_lockres(dlm, lockres);
 332
 333                 dlm_lockres_put(lockres);
 334
 335                 /* Avoid adding any scheduling latencies */
 336                 cond_resched_lock(&dlm->spinlock);
 337         }
 338
 339         spin_unlock(&dlm->spinlock);
 340 }
 341
 342 static void dlm_shuffle_lists(struct dlm_ctxt *dlm,
 343                               struct dlm_lock_resource *res)
 344 {
 345         struct dlm_lock *lock, *target;
 346         int can_grant = 1;
 347
 348         /*
 349          * Because this function is called with the lockres
 350          * spinlock, and because we know that it is not migrating/
 351          * recovering/in-progress, it is fine to reserve asts and
 352          * basts right before queueing them all throughout
 353          */
 354         assert_spin_locked(&dlm->ast_lock);
 355         assert_spin_locked(&res->spinlock);
 356         BUG_ON((res->state & (DLM_LOCK_RES_MIGRATING|
 357                               DLM_LOCK_RES_RECOVERING|
 358                               DLM_LOCK_RES_IN_PROGRESS)));
 359
 360 converting:
 361         if (list_empty(&res->converting))
 362                 goto blocked;
 363         mlog(0, "%s: res %.*s has locks on the convert queue\n", dlm->name,
 364              res->lockname.len, res->lockname.name);
 365
 366         target = list_entry(res->converting.next, struct dlm_lock, list);
 367         if (target->ml.convert_type == LKM_IVMODE) {
 368                 mlog(ML_ERROR, "%s: res %.*s converting lock to invalid mode\n",
 369                      dlm->name, res->lockname.len, res->lockname.name);
 370                 BUG();
 371         }
 372         list_for_each_entry(lock, &res->granted, list) {
 373                 if (lock==target)
 374                         continue;
 375                 if (!dlm_lock_compatible(lock->ml.type,
 376                                          target->ml.convert_type)) {
 377                         can_grant = 0;
 378                         /* queue the BAST if not already */
 379                         if (lock->ml.highest_blocked == LKM_IVMODE) {
 380                                 __dlm_lockres_reserve_ast(res);
 381                                 __dlm_queue_bast(dlm, lock);
 382                         }
 383                         /* update the highest_blocked if needed */
 384                         if (lock->ml.highest_blocked < target->ml.convert_type)
 385                                 lock->ml.highest_blocked =
 386                                         target->ml.convert_type;
 387                 }
 388         }
 389
 390         list_for_each_entry(lock, &res->converting, list) {
 391                 if (lock==target)
 392                         continue;
 393                 if (!dlm_lock_compatible(lock->ml.type,
 394                                          target->ml.convert_type)) {
 395                         can_grant = 0;
 396                         if (lock->ml.highest_blocked == LKM_IVMODE) {
 397                                 __dlm_lockres_reserve_ast(res);
 398                                 __dlm_queue_bast(dlm, lock);
 399                         }
 400                         if (lock->ml.highest_blocked < target->ml.convert_type)
 401                                 lock->ml.highest_blocked =
 402                                         target->ml.convert_type;
 403                 }
 404         }
 405
 406         /* we can convert the lock */
 407         if (can_grant) {
 408                 spin_lock(&target->spinlock);
 409                 BUG_ON(target->ml.highest_blocked != LKM_IVMODE);
 410
 411                 mlog(0, "%s: res %.*s, AST for Converting lock %u:%llu, type "
 412                      "%d => %d, node %u\n", dlm->name, res->lockname.len,
 413                      res->lockname.name,
 414                      dlm_get_lock_cookie_node(be64_to_cpu(target->ml.cookie)),
 415                      dlm_get_lock_cookie_seq(be64_to_cpu(target->ml.cookie)),
 416                      target->ml.type,
 417                      target->ml.convert_type, target->ml.node);
 418
 419                 target->ml.type = target->ml.convert_type;
 420                 target->ml.convert_type = LKM_IVMODE;
 421                 list_move_tail(&target->list, &res->granted);
 422
 423                 BUG_ON(!target->lksb);
 424                 target->lksb->status = DLM_NORMAL;
 425
 426                 spin_unlock(&target->spinlock);
 427
 428                 __dlm_lockres_reserve_ast(res);
 429                 __dlm_queue_ast(dlm, target);
 430                 /* go back and check for more */
 431                 goto converting;
 432         }
 433
 434 blocked:
 435         if (list_empty(&res->blocked))
 436                 goto leave;
 437         target = list_entry(res->blocked.next, struct dlm_lock, list);
 438
 439         list_for_each_entry(lock, &res->granted, list) {
 440                 if (lock==target)
 441                         continue;
 442                 if (!dlm_lock_compatible(lock->ml.type, target->ml.type)) {
 443                         can_grant = 0;
 444                         if (lock->ml.highest_blocked == LKM_IVMODE) {
 445                                 __dlm_lockres_reserve_ast(res);
 446                                 __dlm_queue_bast(dlm, lock);
 447                         }
 448                         if (lock->ml.highest_blocked < target->ml.type)
 449                                 lock->ml.highest_blocked = target->ml.type;
 450                 }
 451         }
 452
 453         list_for_each_entry(lock, &res->converting, list) {
 454                 if (lock==target)
 455                         continue;
 456                 if (!dlm_lock_compatible(lock->ml.type, target->ml.type)) {
 457                         can_grant = 0;
 458                         if (lock->ml.highest_blocked == LKM_IVMODE) {
 459                                 __dlm_lockres_reserve_ast(res);
 460                                 __dlm_queue_bast(dlm, lock);
 461                         }
 462                         if (lock->ml.highest_blocked < target->ml.type)
 463                                 lock->ml.highest_blocked = target->ml.type;
 464                 }
 465         }
 466
 467         /* we can grant the blocked lock (only
 468          * possible if converting list empty) */
 469         if (can_grant) {
 470                 spin_lock(&target->spinlock);
 471                 BUG_ON(target->ml.highest_blocked != LKM_IVMODE);
 472
 473                 mlog(0, "%s: res %.*s, AST for Blocked lock %u:%llu, type %d, "
 474                      "node %u\n", dlm->name, res->lockname.len,
 475                      res->lockname.name,
 476                      dlm_get_lock_cookie_node(be64_to_cpu(target->ml.cookie)),
 477                      dlm_get_lock_cookie_seq(be64_to_cpu(target->ml.cookie)),
 478                      target->ml.type, target->ml.node);
 479
 480                 /* target->ml.type is already correct */
 481                 list_move_tail(&target->list, &res->granted);
 482
 483                 BUG_ON(!target->lksb);
 484                 target->lksb->status = DLM_NORMAL;
 485
 486                 spin_unlock(&target->spinlock);
 487
 488                 __dlm_lockres_reserve_ast(res);
 489                 __dlm_queue_ast(dlm, target);
 490                 /* go back and check for more */
 491                 goto converting;
 492         }
 493
 494 leave:
 495         return;
 496 }
 497
 498 /* must have NO locks when calling this with res !=NULL * */
 499 void dlm_kick_thread(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
 500 {
 501         if (res) {
 502                 spin_lock(&dlm->spinlock);
 503                 spin_lock(&res->spinlock);
 504                 __dlm_dirty_lockres(dlm, res);
 505                 spin_unlock(&res->spinlock);
 506                 spin_unlock(&dlm->spinlock);
 507         }
 508         wake_up(&dlm->dlm_thread_wq);
 509 }
 510
 511 void __dlm_dirty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
 512 {
 513         assert_spin_locked(&dlm->spinlock);
 514         assert_spin_locked(&res->spinlock);
 515
 516         /* don't shuffle secondary queues */
 517         if (res->owner == dlm->node_num) {
 518                 if (res->state & (DLM_LOCK_RES_MIGRATING |
 519                                   DLM_LOCK_RES_BLOCK_DIRTY))
 520                     return;
 521
 522                 if (list_empty(&res->dirty)) {
 523                         /* ref for dirty_list */
 524                         dlm_lockres_get(res);
 525                         list_add_tail(&res->dirty, &dlm->dirty_list);
 526                         res->state |= DLM_LOCK_RES_DIRTY;
 527                 }
 528         }
 529
 530         mlog(0, "%s: res %.*s\n", dlm->name, res->lockname.len,
 531              res->lockname.name);
 532 }
 533
 534
 535 /* Launch the NM thread for the mounted volume */
 536 int dlm_launch_thread(struct dlm_ctxt *dlm)
 537 {
 538         mlog(0, "Starting dlm_thread...\n");
 539
 540         dlm->dlm_thread_task = kthread_run(dlm_thread, dlm, "dlm-%s",
 541                         dlm->name);
 542         if (IS_ERR(dlm->dlm_thread_task)) {
 543                 mlog_errno(PTR_ERR(dlm->dlm_thread_task));
 544                 dlm->dlm_thread_task = NULL;
 545                 return -EINVAL;
 546         }
 547
 548         return 0;
 549 }
 550
 551 void dlm_complete_thread(struct dlm_ctxt *dlm)
 552 {
 553         if (dlm->dlm_thread_task) {
 554                 mlog(ML_KTHREAD, "Waiting for dlm thread to exit\n");
 555                 kthread_stop(dlm->dlm_thread_task);
 556                 dlm->dlm_thread_task = NULL;
 557         }
 558 }
 559
 560 static int dlm_dirty_list_empty(struct dlm_ctxt *dlm)
 561 {
 562         int empty;
 563
 564         spin_lock(&dlm->spinlock);
 565         empty = list_empty(&dlm->dirty_list);
 566         spin_unlock(&dlm->spinlock);
 567
 568         return empty;
 569 }
 570
 571 static void dlm_flush_asts(struct dlm_ctxt *dlm)
 572 {
 573         int ret;
 574         struct dlm_lock *lock;
 575         struct dlm_lock_resource *res;
 576         u8 hi;
 577
 578         spin_lock(&dlm->ast_lock);
 579         while (!list_empty(&dlm->pending_asts)) {
 580                 lock = list_entry(dlm->pending_asts.next,
 581                                   struct dlm_lock, ast_list);
 582                 /* get an extra ref on lock */
 583                 dlm_lock_get(lock);
 584                 res = lock->lockres;
 585                 mlog(0, "%s: res %.*s, Flush AST for lock %u:%llu, type %d, "
 586                      "node %u\n", dlm->name, res->lockname.len,
 587                      res->lockname.name,
 588                      dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)),
 589                      dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)),
 590                      lock->ml.type, lock->ml.node);
 591
 592                 BUG_ON(!lock->ast_pending);
 593
 594                 /* remove from list (including ref) */
 595                 list_del_init(&lock->ast_list);
 596                 dlm_lock_put(lock);
 597                 spin_unlock(&dlm->ast_lock);
 598
 599                 if (lock->ml.node != dlm->node_num) {
 600                         ret = dlm_do_remote_ast(dlm, res, lock);
 601                         if (ret < 0)
 602                                 mlog_errno(ret);
 603                 } else
 604                         dlm_do_local_ast(dlm, res, lock);
 605
 606                 spin_lock(&dlm->ast_lock);
 607
 608                 /* possible that another ast was queued while
 609                  * we were delivering the last one */
 610                 if (!list_empty(&lock->ast_list)) {
 611                         mlog(0, "%s: res %.*s, AST queued while flushing last "
 612                              "one\n", dlm->name, res->lockname.len,
 613                              res->lockname.name);
 614                 } else
 615                         lock->ast_pending = 0;
 616
 617                 /* drop the extra ref.
 618                  * this may drop it completely. */
 619                 dlm_lock_put(lock);
 620                 dlm_lockres_release_ast(dlm, res);
 621         }
 622
 623         while (!list_empty(&dlm->pending_basts)) {
 624                 lock = list_entry(dlm->pending_basts.next,
 625                                   struct dlm_lock, bast_list);
 626                 /* get an extra ref on lock */
 627                 dlm_lock_get(lock);
 628                 res = lock->lockres;
 629
 630                 BUG_ON(!lock->bast_pending);
 631
 632                 /* get the highest blocked lock, and reset */
 633                 spin_lock(&lock->spinlock);
 634                 BUG_ON(lock->ml.highest_blocked <= LKM_IVMODE);
 635                 hi = lock->ml.highest_blocked;
 636                 lock->ml.highest_blocked = LKM_IVMODE;
 637                 spin_unlock(&lock->spinlock);
 638
 639                 /* remove from list (including ref) */
 640                 list_del_init(&lock->bast_list);
 641                 dlm_lock_put(lock);
 642                 spin_unlock(&dlm->ast_lock);
 643
 644                 mlog(0, "%s: res %.*s, Flush BAST for lock %u:%llu, "
 645                      "blocked %d, node %u\n",
 646                      dlm->name, res->lockname.len, res->lockname.name,
 647                      dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)),
 648                      dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)),
 649                      hi, lock->ml.node);
 650
 651                 if (lock->ml.node != dlm->node_num) {
 652                         ret = dlm_send_proxy_bast(dlm, res, lock, hi);
 653                         if (ret < 0)
 654                                 mlog_errno(ret);
 655                 } else
 656                         dlm_do_local_bast(dlm, res, lock, hi);
 657
 658                 spin_lock(&dlm->ast_lock);
 659
 660                 /* possible that another bast was queued while
 661                  * we were delivering the last one */
 662                 if (!list_empty(&lock->bast_list)) {
 663                         mlog(0, "%s: res %.*s, BAST queued while flushing last "
 664                              "one\n", dlm->name, res->lockname.len,
 665                              res->lockname.name);
 666                 } else
 667                         lock->bast_pending = 0;
 668
 669                 /* drop the extra ref.
 670                  * this may drop it completely. */
 671                 dlm_lock_put(lock);
 672                 dlm_lockres_release_ast(dlm, res);
 673         }
 674         wake_up(&dlm->ast_wq);
 675         spin_unlock(&dlm->ast_lock);
 676 }
 677
 678
 679 #define DLM_THREAD_TIMEOUT_MS (4 * 1000)
 680 #define DLM_THREAD_MAX_DIRTY  100
 681
 682 static int dlm_thread(void *data)
 683 {
 684         struct dlm_lock_resource *res;
 685         struct dlm_ctxt *dlm = data;
 686         unsigned long timeout = msecs_to_jiffies(DLM_THREAD_TIMEOUT_MS);
 687
 688         mlog(0, "dlm thread running for %s...\n", dlm->name);
 689
 690         while (!kthread_should_stop()) {
 691                 int n = DLM_THREAD_MAX_DIRTY;
 692
 693                 /* dlm_shutting_down is very point-in-time, but that
 694                  * doesn't matter as we'll just loop back around if we
 695                  * get false on the leading edge of a state
 696                  * transition. */
 697                 dlm_run_purge_list(dlm, dlm_shutting_down(dlm));
 698
 699                 /* We really don't want to hold dlm->spinlock while
 700                  * calling dlm_shuffle_lists on each lockres that
 701                  * needs to have its queues adjusted and AST/BASTs
 702                  * run.  So let's pull each entry off the dirty_list
 703                  * and drop dlm->spinlock ASAP.  Once off the list,
 704                  * res->spinlock needs to be taken again to protect
 705                  * the queues while calling dlm_shuffle_lists.  */
 706                 spin_lock(&dlm->spinlock);
 707                 while (!list_empty(&dlm->dirty_list)) {
 708                         int delay = 0;
 709                         res = list_entry(dlm->dirty_list.next,
 710                                          struct dlm_lock_resource, dirty);
 711
 712                         /* peel a lockres off, remove it from the list,
 713                          * unset the dirty flag and drop the dlm lock */
 714                         BUG_ON(!res);
 715                         dlm_lockres_get(res);
 716
 717                         spin_lock(&res->spinlock);
 718                         /* We clear the DLM_LOCK_RES_DIRTY state once we shuffle lists below */
 719                         list_del_init(&res->dirty);
 720                         spin_unlock(&res->spinlock);
 721                         spin_unlock(&dlm->spinlock);
 722                         /* Drop dirty_list ref */
 723                         dlm_lockres_put(res);
 724
 725                         /* lockres can be re-dirtied/re-added to the
 726                          * dirty_list in this gap, but that is ok */
 727
 728                         spin_lock(&dlm->ast_lock);
 729                         spin_lock(&res->spinlock);
 730                         if (res->owner != dlm->node_num) {
 731                                 __dlm_print_one_lock_resource(res);
 732                                 mlog(ML_ERROR, "%s: inprog %d, mig %d, reco %d,"
 733                                      " dirty %d\n", dlm->name,
 734                                      !!(res->state & DLM_LOCK_RES_IN_PROGRESS),
 735                                      !!(res->state & DLM_LOCK_RES_MIGRATING),
 736                                      !!(res->state & DLM_LOCK_RES_RECOVERING),
 737                                      !!(res->state & DLM_LOCK_RES_DIRTY));
 738                         }
 739                         BUG_ON(res->owner != dlm->node_num);
 740
 741                         /* it is now ok to move lockreses in these states
 742                          * to the dirty list, assuming that they will only be
 743                          * dirty for a short while. */
 744                         BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
 745                         if (res->state & (DLM_LOCK_RES_IN_PROGRESS |
 746                                           DLM_LOCK_RES_RECOVERING |
 747                                           DLM_LOCK_RES_RECOVERY_WAITING)) {
 748                                 /* move it to the tail and keep going */
 749                                 res->state &= ~DLM_LOCK_RES_DIRTY;
 750                                 spin_unlock(&res->spinlock);
 751                                 spin_unlock(&dlm->ast_lock);
 752                                 mlog(0, "%s: res %.*s, inprogress, delay list "
 753                                      "shuffle, state %d\n", dlm->name,
 754                                      res->lockname.len, res->lockname.name,
 755                                      res->state);
 756                                 delay = 1;
 757                                 goto in_progress;
 758                         }
 759
 760                         /* at this point the lockres is not migrating/
 761                          * recovering/in-progress.  we have the lockres
 762                          * spinlock and do NOT have the dlm lock.
 763                          * safe to reserve/queue asts and run the lists. */
 764
 765                         /* called while holding lockres lock */
 766                         dlm_shuffle_lists(dlm, res);
 767                         res->state &= ~DLM_LOCK_RES_DIRTY;
 768                         spin_unlock(&res->spinlock);
 769                         spin_unlock(&dlm->ast_lock);
 770
 771                         dlm_lockres_calc_usage(dlm, res);
 772
 773 in_progress:
 774
 775                         spin_lock(&dlm->spinlock);
 776                         /* if the lock was in-progress, stick
 777                          * it on the back of the list */
 778                         if (delay) {
 779                                 spin_lock(&res->spinlock);
 780                                 __dlm_dirty_lockres(dlm, res);
 781                                 spin_unlock(&res->spinlock);
 782                         }
 783                         dlm_lockres_put(res);
 784
 785                         /* unlikely, but we may need to give time to
 786                          * other tasks */
 787                         if (!--n) {
 788                                 mlog(0, "%s: Throttling dlm thread\n",
 789                                      dlm->name);
 790                                 break;
 791                         }
 792                 }
 793
 794                 spin_unlock(&dlm->spinlock);
 795                 dlm_flush_asts(dlm);
 796
 797                 /* yield and continue right away if there is more work to do */
 798                 if (!n) {
 799                         cond_resched();
 800                         continue;
 801                 }
 802
 803                 wait_event_interruptible_timeout(dlm->dlm_thread_wq,
 804                                                  !dlm_dirty_list_empty(dlm) ||
 805                                                  kthread_should_stop(),
 806                                                  timeout);
 807         }
 808
 809         mlog(0, "quitting DLM thread\n");
 810         return 0;
 811 }