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