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

   1 /* -*- mode: c; c-basic-offset: 8; -*-
   2  * vim: noexpandtab sw=8 ts=8 sts=0:
   3  *
   4  * dlmthread.c
   5  *
   6  * standalone DLM module
   7  *
   8  * Copyright (C) 2004 Oracle.  All rights reserved.
   9  *
  10  * This program is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU General Public
  12  * License as published by the Free Software Foundation; either
  13  * version 2 of the License, or (at your option) any later version.
  14  *
  15  * This program is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  18  * General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU General Public
  21  * License along with this program; if not, write to the
  22  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  23  * Boston, MA 021110-1307, USA.
  24  *
  25  */
  26
  27
  28 #include <linux/module.h>
  29 #include <linux/fs.h>
  30 #include <linux/types.h>
  31 #include <linux/slab.h>
  32 #include <linux/highmem.h>
  33 #include <linux/utsname.h>
  34 #include <linux/init.h>
  35 #include <linux/sysctl.h>
  36 #include <linux/random.h>
  37 #include <linux/blkdev.h>
  38 #include <linux/socket.h>
  39 #include <linux/inet.h>
  40 #include <linux/timer.h>
  41 #include <linux/kthread.h>
  42 #include <linux/delay.h>
  43
  44
  45 #include "cluster/heartbeat.h"
  46 #include "cluster/nodemanager.h"
  47 #include "cluster/tcp.h"
  48
  49 #include "dlmapi.h"
  50 #include "dlmcommon.h"
  51 #include "dlmdomain.h"
  52
  53 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_THREAD)
  54 #include "cluster/masklog.h"
  55
  56 static int dlm_thread(void *data);
  57 static void dlm_flush_asts(struct dlm_ctxt *dlm);
  58
  59 #define dlm_lock_is_remote(dlm, lock)     ((lock)->ml.node != (dlm)->node_num)
  60
  61 /* will exit holding res->spinlock, but may drop in function */
  62 /* waits until flags are cleared on res->state */
  63 void __dlm_wait_on_lockres_flags(struct dlm_lock_resource *res, int flags)
  64 {
  65         DECLARE_WAITQUEUE(wait, current);
  66
  67         assert_spin_locked(&res->spinlock);
  68
  69         add_wait_queue(&res->wq, &wait);
  70 repeat:
  71         set_current_state(TASK_UNINTERRUPTIBLE);
  72         if (res->state & flags) {
  73                 spin_unlock(&res->spinlock);
  74                 schedule();
  75                 spin_lock(&res->spinlock);
  76                 goto repeat;
  77         }
  78         remove_wait_queue(&res->wq, &wait);
  79         __set_current_state(TASK_RUNNING);
  80 }
  81
  82 int __dlm_lockres_has_locks(struct dlm_lock_resource *res)
  83 {
  84         if (list_empty(&res->granted) &&
  85             list_empty(&res->converting) &&
  86             list_empty(&res->blocked))
  87                 return 0;
  88         return 1;
  89 }
  90
  91 /* "unused": the lockres has no locks, is not on the dirty list,
  92  * has no inflight locks (in the gap between mastery and acquiring
  93  * the first lock), and has no bits in its refmap.
  94  * truly ready to be freed. */
  95 int __dlm_lockres_unused(struct dlm_lock_resource *res)
  96 {
  97         if (!__dlm_lockres_has_locks(res) &&
  98             (list_empty(&res->dirty) && !(res->state & DLM_LOCK_RES_DIRTY))) {
  99                 /* try not to scan the bitmap unless the first two
 100                  * conditions are already true */
 101                 int bit = find_next_bit(res->refmap, O2NM_MAX_NODES, 0);
 102                 if (bit >= O2NM_MAX_NODES) {
 103                         /* since the bit for dlm->node_num is not
 104                          * set, inflight_locks better be zero */
 105                         BUG_ON(res->inflight_locks != 0);
 106                         return 1;
 107                 }
 108         }
 109         return 0;
 110 }
 111
 112
 113 /* Call whenever you may have added or deleted something from one of
 114  * the lockres queue's. This will figure out whether it belongs on the
 115  * unused list or not and does the appropriate thing. */
 116 void __dlm_lockres_calc_usage(struct dlm_ctxt *dlm,
 117                               struct dlm_lock_resource *res)
 118 {
 119         mlog_entry("%.*s\n", res->lockname.len, res->lockname.name);
 120
 121         assert_spin_locked(&dlm->spinlock);
 122         assert_spin_locked(&res->spinlock);
 123
 124         if (__dlm_lockres_unused(res)){
 125                 if (list_empty(&res->purge)) {
 126                         mlog(0, "putting lockres %.*s:%p onto purge list\n",
 127                              res->lockname.len, res->lockname.name, res);
 128
 129                         res->last_used = jiffies;
 130                         dlm_lockres_get(res);
 131                         list_add_tail(&res->purge, &dlm->purge_list);
 132                         dlm->purge_count++;
 133                 }
 134         } else if (!list_empty(&res->purge)) {
 135                 mlog(0, "removing lockres %.*s:%p from purge list, owner=%u\n",
 136                      res->lockname.len, res->lockname.name, res, res->owner);
 137
 138                 list_del_init(&res->purge);
 139                 dlm_lockres_put(res);
 140                 dlm->purge_count--;
 141         }
 142 }
 143
 144 void dlm_lockres_calc_usage(struct dlm_ctxt *dlm,
 145                             struct dlm_lock_resource *res)
 146 {
 147         mlog_entry("%.*s\n", res->lockname.len, res->lockname.name);
 148         spin_lock(&dlm->spinlock);
 149         spin_lock(&res->spinlock);
 150
 151         __dlm_lockres_calc_usage(dlm, res);
 152
 153         spin_unlock(&res->spinlock);
 154         spin_unlock(&dlm->spinlock);
 155 }
 156
 157 static int dlm_purge_lockres(struct dlm_ctxt *dlm,
 158                              struct dlm_lock_resource *res)
 159 {
 160         int master;
 161         int ret = 0;
 162
 163         spin_lock(&res->spinlock);
 164         if (!__dlm_lockres_unused(res)) {
 165                 mlog(0, "%s:%.*s: tried to purge but not unused\n",
 166                      dlm->name, res->lockname.len, res->lockname.name);
 167                 __dlm_print_one_lock_resource(res);
 168                 spin_unlock(&res->spinlock);
 169                 BUG();
 170         }
 171
 172         if (res->state & DLM_LOCK_RES_MIGRATING) {
 173                 mlog(0, "%s:%.*s: Delay dropref as this lockres is "
 174                      "being remastered\n", dlm->name, res->lockname.len,
 175                      res->lockname.name);
 176                 /* Re-add the lockres to the end of the purge list */
 177                 if (!list_empty(&res->purge)) {
 178                         list_del_init(&res->purge);
 179                         list_add_tail(&res->purge, &dlm->purge_list);
 180                 }
 181                 spin_unlock(&res->spinlock);
 182                 return 0;
 183         }
 184
 185         master = (res->owner == dlm->node_num);
 186
 187         if (!master)
 188                 res->state |= DLM_LOCK_RES_DROPPING_REF;
 189         spin_unlock(&res->spinlock);
 190
 191         mlog(0, "purging lockres %.*s, master = %d\n", res->lockname.len,
 192              res->lockname.name, master);
 193
 194         if (!master) {
 195                 /* drop spinlock...  retake below */
 196                 spin_unlock(&dlm->spinlock);
 197
 198                 spin_lock(&res->spinlock);
 199                 /* This ensures that clear refmap is sent after the set */
 200                 __dlm_wait_on_lockres_flags(res, DLM_LOCK_RES_SETREF_INPROG);
 201                 spin_unlock(&res->spinlock);
 202
 203                 /* clear our bit from the master's refmap, ignore errors */
 204                 ret = dlm_drop_lockres_ref(dlm, res);
 205                 if (ret < 0) {
 206                         mlog_errno(ret);
 207                         if (!dlm_is_host_down(ret))
 208                                 BUG();
 209                 }
 210                 mlog(0, "%s:%.*s: dlm_deref_lockres returned %d\n",
 211                      dlm->name, res->lockname.len, res->lockname.name, ret);
 212                 spin_lock(&dlm->spinlock);
 213         }
 214
 215         if (!list_empty(&res->purge)) {
 216                 mlog(0, "removing lockres %.*s:%p from purgelist, "
 217                      "master = %d\n", res->lockname.len, res->lockname.name,
 218                      res, master);
 219                 list_del_init(&res->purge);
 220                 dlm_lockres_put(res);
 221                 dlm->purge_count--;
 222         }
 223         __dlm_unhash_lockres(res);
 224
 225         /* lockres is not in the hash now.  drop the flag and wake up
 226          * any processes waiting in dlm_get_lock_resource. */
 227         if (!master) {
 228                 spin_lock(&res->spinlock);
 229                 res->state &= ~DLM_LOCK_RES_DROPPING_REF;
 230                 spin_unlock(&res->spinlock);
 231                 wake_up(&res->wq);
 232         }
 233         return 0;
 234 }
 235
 236 static void dlm_run_purge_list(struct dlm_ctxt *dlm,
 237                                int purge_now)
 238 {
 239         unsigned int run_max, unused;
 240         unsigned long purge_jiffies;
 241         struct dlm_lock_resource *lockres;
 242
 243         spin_lock(&dlm->spinlock);
 244         run_max = dlm->purge_count;
 245
 246         while(run_max && !list_empty(&dlm->purge_list)) {
 247                 run_max--;
 248
 249                 lockres = list_entry(dlm->purge_list.next,
 250                                      struct dlm_lock_resource, purge);
 251
 252                 /* Status of the lockres *might* change so double
 253                  * check. If the lockres is unused, holding the dlm
 254                  * spinlock will prevent people from getting and more
 255                  * refs on it -- there's no need to keep the lockres
 256                  * spinlock. */
 257                 spin_lock(&lockres->spinlock);
 258                 unused = __dlm_lockres_unused(lockres);
 259                 spin_unlock(&lockres->spinlock);
 260
 261                 if (!unused)
 262                         continue;
 263
 264                 purge_jiffies = lockres->last_used +
 265                         msecs_to_jiffies(DLM_PURGE_INTERVAL_MS);
 266
 267                 /* Make sure that we want to be processing this guy at
 268                  * this time. */
 269                 if (!purge_now && time_after(purge_jiffies, jiffies)) {
 270                         /* Since resources are added to the purge list
 271                          * in tail order, we can stop at the first
 272                          * unpurgable resource -- anyone added after
 273                          * him will have a greater last_used value */
 274                         break;
 275                 }
 276
 277                 dlm_lockres_get(lockres);
 278
 279                 /* This may drop and reacquire the dlm spinlock if it
 280                  * has to do migration. */
 281                 if (dlm_purge_lockres(dlm, lockres))
 282                         BUG();
 283
 284                 dlm_lockres_put(lockres);
 285
 286                 /* Avoid adding any scheduling latencies */
 287                 cond_resched_lock(&dlm->spinlock);
 288         }
 289
 290         spin_unlock(&dlm->spinlock);
 291 }
 292
 293 static void dlm_shuffle_lists(struct dlm_ctxt *dlm,
 294                               struct dlm_lock_resource *res)
 295 {
 296         struct dlm_lock *lock, *target;
 297         struct list_head *iter;
 298         struct list_head *head;
 299         int can_grant = 1;
 300
 301         //mlog(0, "res->lockname.len=%d\n", res->lockname.len);
 302         //mlog(0, "res->lockname.name=%p\n", res->lockname.name);
 303         //mlog(0, "shuffle res %.*s\n", res->lockname.len,
 304         //        res->lockname.name);
 305
 306         /* because this function is called with the lockres
 307          * spinlock, and because we know that it is not migrating/
 308          * recovering/in-progress, it is fine to reserve asts and
 309          * basts right before queueing them all throughout */
 310         assert_spin_locked(&res->spinlock);
 311         BUG_ON((res->state & (DLM_LOCK_RES_MIGRATING|
 312                               DLM_LOCK_RES_RECOVERING|
 313                               DLM_LOCK_RES_IN_PROGRESS)));
 314
 315 converting:
 316         if (list_empty(&res->converting))
 317                 goto blocked;
 318         mlog(0, "res %.*s has locks on a convert queue\n", res->lockname.len,
 319              res->lockname.name);
 320
 321         target = list_entry(res->converting.next, struct dlm_lock, list);
 322         if (target->ml.convert_type == LKM_IVMODE) {
 323                 mlog(ML_ERROR, "%.*s: converting a lock with no "
 324                      "convert_type!\n", res->lockname.len, res->lockname.name);
 325                 BUG();
 326         }
 327         head = &res->granted;
 328         list_for_each(iter, head) {
 329                 lock = list_entry(iter, struct dlm_lock, list);
 330                 if (lock==target)
 331                         continue;
 332                 if (!dlm_lock_compatible(lock->ml.type,
 333                                          target->ml.convert_type)) {
 334                         can_grant = 0;
 335                         /* queue the BAST if not already */
 336                         if (lock->ml.highest_blocked == LKM_IVMODE) {
 337                                 __dlm_lockres_reserve_ast(res);
 338                                 dlm_queue_bast(dlm, lock);
 339                         }
 340                         /* update the highest_blocked if needed */
 341                         if (lock->ml.highest_blocked < target->ml.convert_type)
 342                                 lock->ml.highest_blocked =
 343                                         target->ml.convert_type;
 344                 }
 345         }
 346         head = &res->converting;
 347         list_for_each(iter, head) {
 348                 lock = list_entry(iter, struct dlm_lock, list);
 349                 if (lock==target)
 350                         continue;
 351                 if (!dlm_lock_compatible(lock->ml.type,
 352                                          target->ml.convert_type)) {
 353                         can_grant = 0;
 354                         if (lock->ml.highest_blocked == LKM_IVMODE) {
 355                                 __dlm_lockres_reserve_ast(res);
 356                                 dlm_queue_bast(dlm, lock);
 357                         }
 358                         if (lock->ml.highest_blocked < target->ml.convert_type)
 359                                 lock->ml.highest_blocked =
 360                                         target->ml.convert_type;
 361                 }
 362         }
 363
 364         /* we can convert the lock */
 365         if (can_grant) {
 366                 spin_lock(&target->spinlock);
 367                 BUG_ON(target->ml.highest_blocked != LKM_IVMODE);
 368
 369                 mlog(0, "calling ast for converting lock: %.*s, have: %d, "
 370                      "granting: %d, node: %u\n", res->lockname.len,
 371                      res->lockname.name, target->ml.type,
 372                      target->ml.convert_type, target->ml.node);
 373
 374                 target->ml.type = target->ml.convert_type;
 375                 target->ml.convert_type = LKM_IVMODE;
 376                 list_move_tail(&target->list, &res->granted);
 377
 378                 BUG_ON(!target->lksb);
 379                 target->lksb->status = DLM_NORMAL;
 380
 381                 spin_unlock(&target->spinlock);
 382
 383                 __dlm_lockres_reserve_ast(res);
 384                 dlm_queue_ast(dlm, target);
 385                 /* go back and check for more */
 386                 goto converting;
 387         }
 388
 389 blocked:
 390         if (list_empty(&res->blocked))
 391                 goto leave;
 392         target = list_entry(res->blocked.next, struct dlm_lock, list);
 393
 394         head = &res->granted;
 395         list_for_each(iter, head) {
 396                 lock = list_entry(iter, struct dlm_lock, list);
 397                 if (lock==target)
 398                         continue;
 399                 if (!dlm_lock_compatible(lock->ml.type, target->ml.type)) {
 400                         can_grant = 0;
 401                         if (lock->ml.highest_blocked == LKM_IVMODE) {
 402                                 __dlm_lockres_reserve_ast(res);
 403                                 dlm_queue_bast(dlm, lock);
 404                         }
 405                         if (lock->ml.highest_blocked < target->ml.type)
 406                                 lock->ml.highest_blocked = target->ml.type;
 407                 }
 408         }
 409
 410         head = &res->converting;
 411         list_for_each(iter, head) {
 412                 lock = list_entry(iter, struct dlm_lock, list);
 413                 if (lock==target)
 414                         continue;
 415                 if (!dlm_lock_compatible(lock->ml.type, target->ml.type)) {
 416                         can_grant = 0;
 417                         if (lock->ml.highest_blocked == LKM_IVMODE) {
 418                                 __dlm_lockres_reserve_ast(res);
 419                                 dlm_queue_bast(dlm, lock);
 420                         }
 421                         if (lock->ml.highest_blocked < target->ml.type)
 422                                 lock->ml.highest_blocked = target->ml.type;
 423                 }
 424         }
 425
 426         /* we can grant the blocked lock (only
 427          * possible if converting list empty) */
 428         if (can_grant) {
 429                 spin_lock(&target->spinlock);
 430                 BUG_ON(target->ml.highest_blocked != LKM_IVMODE);
 431
 432                 mlog(0, "calling ast for blocked lock: %.*s, granting: %d, "
 433                      "node: %u\n", res->lockname.len, res->lockname.name,
 434                      target->ml.type, target->ml.node);
 435
 436                 // target->ml.type is already correct
 437                 list_move_tail(&target->list, &res->granted);
 438
 439                 BUG_ON(!target->lksb);
 440                 target->lksb->status = DLM_NORMAL;
 441
 442                 spin_unlock(&target->spinlock);
 443
 444                 __dlm_lockres_reserve_ast(res);
 445                 dlm_queue_ast(dlm, target);
 446                 /* go back and check for more */
 447                 goto converting;
 448         }
 449
 450 leave:
 451         return;
 452 }
 453
 454 /* must have NO locks when calling this with res !=NULL * */
 455 void dlm_kick_thread(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
 456 {
 457         mlog_entry("dlm=%p, res=%p\n", dlm, res);
 458         if (res) {
 459                 spin_lock(&dlm->spinlock);
 460                 spin_lock(&res->spinlock);
 461                 __dlm_dirty_lockres(dlm, res);
 462                 spin_unlock(&res->spinlock);
 463                 spin_unlock(&dlm->spinlock);
 464         }
 465         wake_up(&dlm->dlm_thread_wq);
 466 }
 467
 468 void __dlm_dirty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
 469 {
 470         mlog_entry("dlm=%p, res=%p\n", dlm, res);
 471
 472         assert_spin_locked(&dlm->spinlock);
 473         assert_spin_locked(&res->spinlock);
 474
 475         /* don't shuffle secondary queues */
 476         if ((res->owner == dlm->node_num)) {
 477                 if (res->state & (DLM_LOCK_RES_MIGRATING |
 478                                   DLM_LOCK_RES_BLOCK_DIRTY))
 479                     return;
 480
 481                 if (list_empty(&res->dirty)) {
 482                         /* ref for dirty_list */
 483                         dlm_lockres_get(res);
 484                         list_add_tail(&res->dirty, &dlm->dirty_list);
 485                         res->state |= DLM_LOCK_RES_DIRTY;
 486                 }
 487         }
 488 }
 489
 490
 491 /* Launch the NM thread for the mounted volume */
 492 int dlm_launch_thread(struct dlm_ctxt *dlm)
 493 {
 494         mlog(0, "starting dlm thread...\n");
 495
 496         dlm->dlm_thread_task = kthread_run(dlm_thread, dlm, "dlm_thread");
 497         if (IS_ERR(dlm->dlm_thread_task)) {
 498                 mlog_errno(PTR_ERR(dlm->dlm_thread_task));
 499                 dlm->dlm_thread_task = NULL;
 500                 return -EINVAL;
 501         }
 502
 503         return 0;
 504 }
 505
 506 void dlm_complete_thread(struct dlm_ctxt *dlm)
 507 {
 508         if (dlm->dlm_thread_task) {
 509                 mlog(ML_KTHREAD, "waiting for dlm thread to exit\n");
 510                 kthread_stop(dlm->dlm_thread_task);
 511                 dlm->dlm_thread_task = NULL;
 512         }
 513 }
 514
 515 static int dlm_dirty_list_empty(struct dlm_ctxt *dlm)
 516 {
 517         int empty;
 518
 519         spin_lock(&dlm->spinlock);
 520         empty = list_empty(&dlm->dirty_list);
 521         spin_unlock(&dlm->spinlock);
 522
 523         return empty;
 524 }
 525
 526 static void dlm_flush_asts(struct dlm_ctxt *dlm)
 527 {
 528         int ret;
 529         struct dlm_lock *lock;
 530         struct dlm_lock_resource *res;
 531         u8 hi;
 532
 533         spin_lock(&dlm->ast_lock);
 534         while (!list_empty(&dlm->pending_asts)) {
 535                 lock = list_entry(dlm->pending_asts.next,
 536                                   struct dlm_lock, ast_list);
 537                 /* get an extra ref on lock */
 538                 dlm_lock_get(lock);
 539                 res = lock->lockres;
 540                 mlog(0, "delivering an ast for this lockres\n");
 541
 542                 BUG_ON(!lock->ast_pending);
 543
 544                 /* remove from list (including ref) */
 545                 list_del_init(&lock->ast_list);
 546                 dlm_lock_put(lock);
 547                 spin_unlock(&dlm->ast_lock);
 548
 549                 if (lock->ml.node != dlm->node_num) {
 550                         ret = dlm_do_remote_ast(dlm, res, lock);
 551                         if (ret < 0)
 552                                 mlog_errno(ret);
 553                 } else
 554                         dlm_do_local_ast(dlm, res, lock);
 555
 556                 spin_lock(&dlm->ast_lock);
 557
 558                 /* possible that another ast was queued while
 559                  * we were delivering the last one */
 560                 if (!list_empty(&lock->ast_list)) {
 561                         mlog(0, "aha another ast got queued while "
 562                              "we were finishing the last one.  will "
 563                              "keep the ast_pending flag set.\n");
 564                 } else
 565                         lock->ast_pending = 0;
 566
 567                 /* drop the extra ref.
 568                  * this may drop it completely. */
 569                 dlm_lock_put(lock);
 570                 dlm_lockres_release_ast(dlm, res);
 571         }
 572
 573         while (!list_empty(&dlm->pending_basts)) {
 574                 lock = list_entry(dlm->pending_basts.next,
 575                                   struct dlm_lock, bast_list);
 576                 /* get an extra ref on lock */
 577                 dlm_lock_get(lock);
 578                 res = lock->lockres;
 579
 580                 BUG_ON(!lock->bast_pending);
 581
 582                 /* get the highest blocked lock, and reset */
 583                 spin_lock(&lock->spinlock);
 584                 BUG_ON(lock->ml.highest_blocked <= LKM_IVMODE);
 585                 hi = lock->ml.highest_blocked;
 586                 lock->ml.highest_blocked = LKM_IVMODE;
 587                 spin_unlock(&lock->spinlock);
 588
 589                 /* remove from list (including ref) */
 590                 list_del_init(&lock->bast_list);
 591                 dlm_lock_put(lock);
 592                 spin_unlock(&dlm->ast_lock);
 593
 594                 mlog(0, "delivering a bast for this lockres "
 595                      "(blocked = %d\n", hi);
 596
 597                 if (lock->ml.node != dlm->node_num) {
 598                         ret = dlm_send_proxy_bast(dlm, res, lock, hi);
 599                         if (ret < 0)
 600                                 mlog_errno(ret);
 601                 } else
 602                         dlm_do_local_bast(dlm, res, lock, hi);
 603
 604                 spin_lock(&dlm->ast_lock);
 605
 606                 /* possible that another bast was queued while
 607                  * we were delivering the last one */
 608                 if (!list_empty(&lock->bast_list)) {
 609                         mlog(0, "aha another bast got queued while "
 610                              "we were finishing the last one.  will "
 611                              "keep the bast_pending flag set.\n");
 612                 } else
 613                         lock->bast_pending = 0;
 614
 615                 /* drop the extra ref.
 616                  * this may drop it completely. */
 617                 dlm_lock_put(lock);
 618                 dlm_lockres_release_ast(dlm, res);
 619         }
 620         wake_up(&dlm->ast_wq);
 621         spin_unlock(&dlm->ast_lock);
 622 }
 623
 624
 625 #define DLM_THREAD_TIMEOUT_MS (4 * 1000)
 626 #define DLM_THREAD_MAX_DIRTY  100
 627 #define DLM_THREAD_MAX_ASTS   10
 628
 629 static int dlm_thread(void *data)
 630 {
 631         struct dlm_lock_resource *res;
 632         struct dlm_ctxt *dlm = data;
 633         unsigned long timeout = msecs_to_jiffies(DLM_THREAD_TIMEOUT_MS);
 634
 635         mlog(0, "dlm thread running for %s...\n", dlm->name);
 636
 637         while (!kthread_should_stop()) {
 638                 int n = DLM_THREAD_MAX_DIRTY;
 639
 640                 /* dlm_shutting_down is very point-in-time, but that
 641                  * doesn't matter as we'll just loop back around if we
 642                  * get false on the leading edge of a state
 643                  * transition. */
 644                 dlm_run_purge_list(dlm, dlm_shutting_down(dlm));
 645
 646                 /* We really don't want to hold dlm->spinlock while
 647                  * calling dlm_shuffle_lists on each lockres that
 648                  * needs to have its queues adjusted and AST/BASTs
 649                  * run.  So let's pull each entry off the dirty_list
 650                  * and drop dlm->spinlock ASAP.  Once off the list,
 651                  * res->spinlock needs to be taken again to protect
 652                  * the queues while calling dlm_shuffle_lists.  */
 653                 spin_lock(&dlm->spinlock);
 654                 while (!list_empty(&dlm->dirty_list)) {
 655                         int delay = 0;
 656                         res = list_entry(dlm->dirty_list.next,
 657                                          struct dlm_lock_resource, dirty);
 658
 659                         /* peel a lockres off, remove it from the list,
 660                          * unset the dirty flag and drop the dlm lock */
 661                         BUG_ON(!res);
 662                         dlm_lockres_get(res);
 663
 664                         spin_lock(&res->spinlock);
 665                         /* We clear the DLM_LOCK_RES_DIRTY state once we shuffle lists below */
 666                         list_del_init(&res->dirty);
 667                         spin_unlock(&res->spinlock);
 668                         spin_unlock(&dlm->spinlock);
 669                         /* Drop dirty_list ref */
 670                         dlm_lockres_put(res);
 671
 672                         /* lockres can be re-dirtied/re-added to the
 673                          * dirty_list in this gap, but that is ok */
 674
 675                         spin_lock(&res->spinlock);
 676                         if (res->owner != dlm->node_num) {
 677                                 __dlm_print_one_lock_resource(res);
 678                                 mlog(ML_ERROR, "inprog:%s, mig:%s, reco:%s, dirty:%s\n",
 679                                      res->state & DLM_LOCK_RES_IN_PROGRESS ? "yes" : "no",
 680                                      res->state & DLM_LOCK_RES_MIGRATING ? "yes" : "no",
 681                                      res->state & DLM_LOCK_RES_RECOVERING ? "yes" : "no",
 682                                      res->state & DLM_LOCK_RES_DIRTY ? "yes" : "no");
 683                         }
 684                         BUG_ON(res->owner != dlm->node_num);
 685
 686                         /* it is now ok to move lockreses in these states
 687                          * to the dirty list, assuming that they will only be
 688                          * dirty for a short while. */
 689                         BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
 690                         if (res->state & (DLM_LOCK_RES_IN_PROGRESS |
 691                                           DLM_LOCK_RES_RECOVERING)) {
 692                                 /* move it to the tail and keep going */
 693                                 res->state &= ~DLM_LOCK_RES_DIRTY;
 694                                 spin_unlock(&res->spinlock);
 695                                 mlog(0, "delaying list shuffling for in-"
 696                                      "progress lockres %.*s, state=%d\n",
 697                                      res->lockname.len, res->lockname.name,
 698                                      res->state);
 699                                 delay = 1;
 700                                 goto in_progress;
 701                         }
 702
 703                         /* at this point the lockres is not migrating/
 704                          * recovering/in-progress.  we have the lockres
 705                          * spinlock and do NOT have the dlm lock.
 706                          * safe to reserve/queue asts and run the lists. */
 707
 708                         mlog(0, "calling dlm_shuffle_lists with dlm=%s, "
 709                              "res=%.*s\n", dlm->name,
 710                              res->lockname.len, res->lockname.name);
 711
 712                         /* called while holding lockres lock */
 713                         dlm_shuffle_lists(dlm, res);
 714                         res->state &= ~DLM_LOCK_RES_DIRTY;
 715                         spin_unlock(&res->spinlock);
 716
 717                         dlm_lockres_calc_usage(dlm, res);
 718
 719 in_progress:
 720
 721                         spin_lock(&dlm->spinlock);
 722                         /* if the lock was in-progress, stick
 723                          * it on the back of the list */
 724                         if (delay) {
 725                                 spin_lock(&res->spinlock);
 726                                 __dlm_dirty_lockres(dlm, res);
 727                                 spin_unlock(&res->spinlock);
 728                         }
 729                         dlm_lockres_put(res);
 730
 731                         /* unlikely, but we may need to give time to
 732                          * other tasks */
 733                         if (!--n) {
 734                                 mlog(0, "throttling dlm_thread\n");
 735                                 break;
 736                         }
 737                 }
 738
 739                 spin_unlock(&dlm->spinlock);
 740                 dlm_flush_asts(dlm);
 741
 742                 /* yield and continue right away if there is more work to do */
 743                 if (!n) {
 744                         cond_resched();
 745                         continue;
 746                 }
 747
 748                 wait_event_interruptible_timeout(dlm->dlm_thread_wq,
 749                                                  !dlm_dirty_list_empty(dlm) ||
 750                                                  kthread_should_stop(),
 751                                                  timeout);
 752         }
 753
 754         mlog(0, "quitting DLM thread\n");
 755         return 0;
 756 }