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         spin_lock(&res->spinlock);
 216         if (!list_empty(&res->purge)) {
 217                 mlog(0, "removing lockres %.*s:%p from purgelist, "
 218                      "master = %d\n", res->lockname.len, res->lockname.name,
 219                      res, master);
 220                 list_del_init(&res->purge);
 221                 spin_unlock(&res->spinlock);
 222                 dlm_lockres_put(res);
 223                 dlm->purge_count--;
 224         } else
 225                 spin_unlock(&res->spinlock);
 226
 227         __dlm_unhash_lockres(res);
 228
 229         /* lockres is not in the hash now.  drop the flag and wake up
 230          * any processes waiting in dlm_get_lock_resource. */
 231         if (!master) {
 232                 spin_lock(&res->spinlock);
 233                 res->state &= ~DLM_LOCK_RES_DROPPING_REF;
 234                 spin_unlock(&res->spinlock);
 235                 wake_up(&res->wq);
 236         }
 237         return 0;
 238 }
 239
 240 static void dlm_run_purge_list(struct dlm_ctxt *dlm,
 241                                int purge_now)
 242 {
 243         unsigned int run_max, unused;
 244         unsigned long purge_jiffies;
 245         struct dlm_lock_resource *lockres;
 246
 247         spin_lock(&dlm->spinlock);
 248         run_max = dlm->purge_count;
 249
 250         while(run_max && !list_empty(&dlm->purge_list)) {
 251                 run_max--;
 252
 253                 lockres = list_entry(dlm->purge_list.next,
 254                                      struct dlm_lock_resource, purge);
 255
 256                 /* Status of the lockres *might* change so double
 257                  * check. If the lockres is unused, holding the dlm
 258                  * spinlock will prevent people from getting and more
 259                  * refs on it -- there's no need to keep the lockres
 260                  * spinlock. */
 261                 spin_lock(&lockres->spinlock);
 262                 unused = __dlm_lockres_unused(lockres);
 263                 spin_unlock(&lockres->spinlock);
 264
 265                 if (!unused)
 266                         continue;
 267
 268                 purge_jiffies = lockres->last_used +
 269                         msecs_to_jiffies(DLM_PURGE_INTERVAL_MS);
 270
 271                 /* Make sure that we want to be processing this guy at
 272                  * this time. */
 273                 if (!purge_now && time_after(purge_jiffies, jiffies)) {
 274                         /* Since resources are added to the purge list
 275                          * in tail order, we can stop at the first
 276                          * unpurgable resource -- anyone added after
 277                          * him will have a greater last_used value */
 278                         break;
 279                 }
 280
 281                 dlm_lockres_get(lockres);
 282
 283                 /* This may drop and reacquire the dlm spinlock if it
 284                  * has to do migration. */
 285                 if (dlm_purge_lockres(dlm, lockres))
 286                         BUG();
 287
 288                 dlm_lockres_put(lockres);
 289
 290                 /* Avoid adding any scheduling latencies */
 291                 cond_resched_lock(&dlm->spinlock);
 292         }
 293
 294         spin_unlock(&dlm->spinlock);
 295 }
 296
 297 static void dlm_shuffle_lists(struct dlm_ctxt *dlm,
 298                               struct dlm_lock_resource *res)
 299 {
 300         struct dlm_lock *lock, *target;
 301         struct list_head *iter;
 302         struct list_head *head;
 303         int can_grant = 1;
 304
 305         //mlog(0, "res->lockname.len=%d\n", res->lockname.len);
 306         //mlog(0, "res->lockname.name=%p\n", res->lockname.name);
 307         //mlog(0, "shuffle res %.*s\n", res->lockname.len,
 308         //        res->lockname.name);
 309
 310         /* because this function is called with the lockres
 311          * spinlock, and because we know that it is not migrating/
 312          * recovering/in-progress, it is fine to reserve asts and
 313          * basts right before queueing them all throughout */
 314         assert_spin_locked(&res->spinlock);
 315         BUG_ON((res->state & (DLM_LOCK_RES_MIGRATING|
 316                               DLM_LOCK_RES_RECOVERING|
 317                               DLM_LOCK_RES_IN_PROGRESS)));
 318
 319 converting:
 320         if (list_empty(&res->converting))
 321                 goto blocked;
 322         mlog(0, "res %.*s has locks on a convert queue\n", res->lockname.len,
 323              res->lockname.name);
 324
 325         target = list_entry(res->converting.next, struct dlm_lock, list);
 326         if (target->ml.convert_type == LKM_IVMODE) {
 327                 mlog(ML_ERROR, "%.*s: converting a lock with no "
 328                      "convert_type!\n", res->lockname.len, res->lockname.name);
 329                 BUG();
 330         }
 331         head = &res->granted;
 332         list_for_each(iter, head) {
 333                 lock = list_entry(iter, struct dlm_lock, list);
 334                 if (lock==target)
 335                         continue;
 336                 if (!dlm_lock_compatible(lock->ml.type,
 337                                          target->ml.convert_type)) {
 338                         can_grant = 0;
 339                         /* queue the BAST if not already */
 340                         if (lock->ml.highest_blocked == LKM_IVMODE) {
 341                                 __dlm_lockres_reserve_ast(res);
 342                                 dlm_queue_bast(dlm, lock);
 343                         }
 344                         /* update the highest_blocked if needed */
 345                         if (lock->ml.highest_blocked < target->ml.convert_type)
 346                                 lock->ml.highest_blocked =
 347                                         target->ml.convert_type;
 348                 }
 349         }
 350         head = &res->converting;
 351         list_for_each(iter, head) {
 352                 lock = list_entry(iter, struct dlm_lock, list);
 353                 if (lock==target)
 354                         continue;
 355                 if (!dlm_lock_compatible(lock->ml.type,
 356                                          target->ml.convert_type)) {
 357                         can_grant = 0;
 358                         if (lock->ml.highest_blocked == LKM_IVMODE) {
 359                                 __dlm_lockres_reserve_ast(res);
 360                                 dlm_queue_bast(dlm, lock);
 361                         }
 362                         if (lock->ml.highest_blocked < target->ml.convert_type)
 363                                 lock->ml.highest_blocked =
 364                                         target->ml.convert_type;
 365                 }
 366         }
 367
 368         /* we can convert the lock */
 369         if (can_grant) {
 370                 spin_lock(&target->spinlock);
 371                 BUG_ON(target->ml.highest_blocked != LKM_IVMODE);
 372
 373                 mlog(0, "calling ast for converting lock: %.*s, have: %d, "
 374                      "granting: %d, node: %u\n", res->lockname.len,
 375                      res->lockname.name, target->ml.type,
 376                      target->ml.convert_type, target->ml.node);
 377
 378                 target->ml.type = target->ml.convert_type;
 379                 target->ml.convert_type = LKM_IVMODE;
 380                 list_move_tail(&target->list, &res->granted);
 381
 382                 BUG_ON(!target->lksb);
 383                 target->lksb->status = DLM_NORMAL;
 384
 385                 spin_unlock(&target->spinlock);
 386
 387                 __dlm_lockres_reserve_ast(res);
 388                 dlm_queue_ast(dlm, target);
 389                 /* go back and check for more */
 390                 goto converting;
 391         }
 392
 393 blocked:
 394         if (list_empty(&res->blocked))
 395                 goto leave;
 396         target = list_entry(res->blocked.next, struct dlm_lock, list);
 397
 398         head = &res->granted;
 399         list_for_each(iter, head) {
 400                 lock = list_entry(iter, struct dlm_lock, list);
 401                 if (lock==target)
 402                         continue;
 403                 if (!dlm_lock_compatible(lock->ml.type, target->ml.type)) {
 404                         can_grant = 0;
 405                         if (lock->ml.highest_blocked == LKM_IVMODE) {
 406                                 __dlm_lockres_reserve_ast(res);
 407                                 dlm_queue_bast(dlm, lock);
 408                         }
 409                         if (lock->ml.highest_blocked < target->ml.type)
 410                                 lock->ml.highest_blocked = target->ml.type;
 411                 }
 412         }
 413
 414         head = &res->converting;
 415         list_for_each(iter, head) {
 416                 lock = list_entry(iter, struct dlm_lock, list);
 417                 if (lock==target)
 418                         continue;
 419                 if (!dlm_lock_compatible(lock->ml.type, target->ml.type)) {
 420                         can_grant = 0;
 421                         if (lock->ml.highest_blocked == LKM_IVMODE) {
 422                                 __dlm_lockres_reserve_ast(res);
 423                                 dlm_queue_bast(dlm, lock);
 424                         }
 425                         if (lock->ml.highest_blocked < target->ml.type)
 426                                 lock->ml.highest_blocked = target->ml.type;
 427                 }
 428         }
 429
 430         /* we can grant the blocked lock (only
 431          * possible if converting list empty) */
 432         if (can_grant) {
 433                 spin_lock(&target->spinlock);
 434                 BUG_ON(target->ml.highest_blocked != LKM_IVMODE);
 435
 436                 mlog(0, "calling ast for blocked lock: %.*s, granting: %d, "
 437                      "node: %u\n", res->lockname.len, res->lockname.name,
 438                      target->ml.type, target->ml.node);
 439
 440                 // target->ml.type is already correct
 441                 list_move_tail(&target->list, &res->granted);
 442
 443                 BUG_ON(!target->lksb);
 444                 target->lksb->status = DLM_NORMAL;
 445
 446                 spin_unlock(&target->spinlock);
 447
 448                 __dlm_lockres_reserve_ast(res);
 449                 dlm_queue_ast(dlm, target);
 450                 /* go back and check for more */
 451                 goto converting;
 452         }
 453
 454 leave:
 455         return;
 456 }
 457
 458 /* must have NO locks when calling this with res !=NULL * */
 459 void dlm_kick_thread(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
 460 {
 461         mlog_entry("dlm=%p, res=%p\n", dlm, res);
 462         if (res) {
 463                 spin_lock(&dlm->spinlock);
 464                 spin_lock(&res->spinlock);
 465                 __dlm_dirty_lockres(dlm, res);
 466                 spin_unlock(&res->spinlock);
 467                 spin_unlock(&dlm->spinlock);
 468         }
 469         wake_up(&dlm->dlm_thread_wq);
 470 }
 471
 472 void __dlm_dirty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
 473 {
 474         mlog_entry("dlm=%p, res=%p\n", dlm, res);
 475
 476         assert_spin_locked(&dlm->spinlock);
 477         assert_spin_locked(&res->spinlock);
 478
 479         /* don't shuffle secondary queues */
 480         if ((res->owner == dlm->node_num)) {
 481                 if (res->state & (DLM_LOCK_RES_MIGRATING |
 482                                   DLM_LOCK_RES_BLOCK_DIRTY))
 483                     return;
 484
 485                 if (list_empty(&res->dirty)) {
 486                         /* ref for dirty_list */
 487                         dlm_lockres_get(res);
 488                         list_add_tail(&res->dirty, &dlm->dirty_list);
 489                         res->state |= DLM_LOCK_RES_DIRTY;
 490                 }
 491         }
 492 }
 493
 494
 495 /* Launch the NM thread for the mounted volume */
 496 int dlm_launch_thread(struct dlm_ctxt *dlm)
 497 {
 498         mlog(0, "starting dlm thread...\n");
 499
 500         dlm->dlm_thread_task = kthread_run(dlm_thread, dlm, "dlm_thread");
 501         if (IS_ERR(dlm->dlm_thread_task)) {
 502                 mlog_errno(PTR_ERR(dlm->dlm_thread_task));
 503                 dlm->dlm_thread_task = NULL;
 504                 return -EINVAL;
 505         }
 506
 507         return 0;
 508 }
 509
 510 void dlm_complete_thread(struct dlm_ctxt *dlm)
 511 {
 512         if (dlm->dlm_thread_task) {
 513                 mlog(ML_KTHREAD, "waiting for dlm thread to exit\n");
 514                 kthread_stop(dlm->dlm_thread_task);
 515                 dlm->dlm_thread_task = NULL;
 516         }
 517 }
 518
 519 static int dlm_dirty_list_empty(struct dlm_ctxt *dlm)
 520 {
 521         int empty;
 522
 523         spin_lock(&dlm->spinlock);
 524         empty = list_empty(&dlm->dirty_list);
 525         spin_unlock(&dlm->spinlock);
 526
 527         return empty;
 528 }
 529
 530 static void dlm_flush_asts(struct dlm_ctxt *dlm)
 531 {
 532         int ret;
 533         struct dlm_lock *lock;
 534         struct dlm_lock_resource *res;
 535         u8 hi;
 536
 537         spin_lock(&dlm->ast_lock);
 538         while (!list_empty(&dlm->pending_asts)) {
 539                 lock = list_entry(dlm->pending_asts.next,
 540                                   struct dlm_lock, ast_list);
 541                 /* get an extra ref on lock */
 542                 dlm_lock_get(lock);
 543                 res = lock->lockres;
 544                 mlog(0, "delivering an ast for this lockres\n");
 545
 546                 BUG_ON(!lock->ast_pending);
 547
 548                 /* remove from list (including ref) */
 549                 list_del_init(&lock->ast_list);
 550                 dlm_lock_put(lock);
 551                 spin_unlock(&dlm->ast_lock);
 552
 553                 if (lock->ml.node != dlm->node_num) {
 554                         ret = dlm_do_remote_ast(dlm, res, lock);
 555                         if (ret < 0)
 556                                 mlog_errno(ret);
 557                 } else
 558                         dlm_do_local_ast(dlm, res, lock);
 559
 560                 spin_lock(&dlm->ast_lock);
 561
 562                 /* possible that another ast was queued while
 563                  * we were delivering the last one */
 564                 if (!list_empty(&lock->ast_list)) {
 565                         mlog(0, "aha another ast got queued while "
 566                              "we were finishing the last one.  will "
 567                              "keep the ast_pending flag set.\n");
 568                 } else
 569                         lock->ast_pending = 0;
 570
 571                 /* drop the extra ref.
 572                  * this may drop it completely. */
 573                 dlm_lock_put(lock);
 574                 dlm_lockres_release_ast(dlm, res);
 575         }
 576
 577         while (!list_empty(&dlm->pending_basts)) {
 578                 lock = list_entry(dlm->pending_basts.next,
 579                                   struct dlm_lock, bast_list);
 580                 /* get an extra ref on lock */
 581                 dlm_lock_get(lock);
 582                 res = lock->lockres;
 583
 584                 BUG_ON(!lock->bast_pending);
 585
 586                 /* get the highest blocked lock, and reset */
 587                 spin_lock(&lock->spinlock);
 588                 BUG_ON(lock->ml.highest_blocked <= LKM_IVMODE);
 589                 hi = lock->ml.highest_blocked;
 590                 lock->ml.highest_blocked = LKM_IVMODE;
 591                 spin_unlock(&lock->spinlock);
 592
 593                 /* remove from list (including ref) */
 594                 list_del_init(&lock->bast_list);
 595                 dlm_lock_put(lock);
 596                 spin_unlock(&dlm->ast_lock);
 597
 598                 mlog(0, "delivering a bast for this lockres "
 599                      "(blocked = %d\n", hi);
 600
 601                 if (lock->ml.node != dlm->node_num) {
 602                         ret = dlm_send_proxy_bast(dlm, res, lock, hi);
 603                         if (ret < 0)
 604                                 mlog_errno(ret);
 605                 } else
 606                         dlm_do_local_bast(dlm, res, lock, hi);
 607
 608                 spin_lock(&dlm->ast_lock);
 609
 610                 /* possible that another bast was queued while
 611                  * we were delivering the last one */
 612                 if (!list_empty(&lock->bast_list)) {
 613                         mlog(0, "aha another bast got queued while "
 614                              "we were finishing the last one.  will "
 615                              "keep the bast_pending flag set.\n");
 616                 } else
 617                         lock->bast_pending = 0;
 618
 619                 /* drop the extra ref.
 620                  * this may drop it completely. */
 621                 dlm_lock_put(lock);
 622                 dlm_lockres_release_ast(dlm, res);
 623         }
 624         wake_up(&dlm->ast_wq);
 625         spin_unlock(&dlm->ast_lock);
 626 }
 627
 628
 629 #define DLM_THREAD_TIMEOUT_MS (4 * 1000)
 630 #define DLM_THREAD_MAX_DIRTY  100
 631 #define DLM_THREAD_MAX_ASTS   10
 632
 633 static int dlm_thread(void *data)
 634 {
 635         struct dlm_lock_resource *res;
 636         struct dlm_ctxt *dlm = data;
 637         unsigned long timeout = msecs_to_jiffies(DLM_THREAD_TIMEOUT_MS);
 638
 639         mlog(0, "dlm thread running for %s...\n", dlm->name);
 640
 641         while (!kthread_should_stop()) {
 642                 int n = DLM_THREAD_MAX_DIRTY;
 643
 644                 /* dlm_shutting_down is very point-in-time, but that
 645                  * doesn't matter as we'll just loop back around if we
 646                  * get false on the leading edge of a state
 647                  * transition. */
 648                 dlm_run_purge_list(dlm, dlm_shutting_down(dlm));
 649
 650                 /* We really don't want to hold dlm->spinlock while
 651                  * calling dlm_shuffle_lists on each lockres that
 652                  * needs to have its queues adjusted and AST/BASTs
 653                  * run.  So let's pull each entry off the dirty_list
 654                  * and drop dlm->spinlock ASAP.  Once off the list,
 655                  * res->spinlock needs to be taken again to protect
 656                  * the queues while calling dlm_shuffle_lists.  */
 657                 spin_lock(&dlm->spinlock);
 658                 while (!list_empty(&dlm->dirty_list)) {
 659                         int delay = 0;
 660                         res = list_entry(dlm->dirty_list.next,
 661                                          struct dlm_lock_resource, dirty);
 662
 663                         /* peel a lockres off, remove it from the list,
 664                          * unset the dirty flag and drop the dlm lock */
 665                         BUG_ON(!res);
 666                         dlm_lockres_get(res);
 667
 668                         spin_lock(&res->spinlock);
 669                         /* We clear the DLM_LOCK_RES_DIRTY state once we shuffle lists below */
 670                         list_del_init(&res->dirty);
 671                         spin_unlock(&res->spinlock);
 672                         spin_unlock(&dlm->spinlock);
 673                         /* Drop dirty_list ref */
 674                         dlm_lockres_put(res);
 675
 676                         /* lockres can be re-dirtied/re-added to the
 677                          * dirty_list in this gap, but that is ok */
 678
 679                         spin_lock(&res->spinlock);
 680                         if (res->owner != dlm->node_num) {
 681                                 __dlm_print_one_lock_resource(res);
 682                                 mlog(ML_ERROR, "inprog:%s, mig:%s, reco:%s, dirty:%s\n",
 683                                      res->state & DLM_LOCK_RES_IN_PROGRESS ? "yes" : "no",
 684                                      res->state & DLM_LOCK_RES_MIGRATING ? "yes" : "no",
 685                                      res->state & DLM_LOCK_RES_RECOVERING ? "yes" : "no",
 686                                      res->state & DLM_LOCK_RES_DIRTY ? "yes" : "no");
 687                         }
 688                         BUG_ON(res->owner != dlm->node_num);
 689
 690                         /* it is now ok to move lockreses in these states
 691                          * to the dirty list, assuming that they will only be
 692                          * dirty for a short while. */
 693                         BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
 694                         if (res->state & (DLM_LOCK_RES_IN_PROGRESS |
 695                                           DLM_LOCK_RES_RECOVERING)) {
 696                                 /* move it to the tail and keep going */
 697                                 res->state &= ~DLM_LOCK_RES_DIRTY;
 698                                 spin_unlock(&res->spinlock);
 699                                 mlog(0, "delaying list shuffling for in-"
 700                                      "progress lockres %.*s, state=%d\n",
 701                                      res->lockname.len, res->lockname.name,
 702                                      res->state);
 703                                 delay = 1;
 704                                 goto in_progress;
 705                         }
 706
 707                         /* at this point the lockres is not migrating/
 708                          * recovering/in-progress.  we have the lockres
 709                          * spinlock and do NOT have the dlm lock.
 710                          * safe to reserve/queue asts and run the lists. */
 711
 712                         mlog(0, "calling dlm_shuffle_lists with dlm=%s, "
 713                              "res=%.*s\n", dlm->name,
 714                              res->lockname.len, res->lockname.name);
 715
 716                         /* called while holding lockres lock */
 717                         dlm_shuffle_lists(dlm, res);
 718                         res->state &= ~DLM_LOCK_RES_DIRTY;
 719                         spin_unlock(&res->spinlock);
 720
 721                         dlm_lockres_calc_usage(dlm, res);
 722
 723 in_progress:
 724
 725                         spin_lock(&dlm->spinlock);
 726                         /* if the lock was in-progress, stick
 727                          * it on the back of the list */
 728                         if (delay) {
 729                                 spin_lock(&res->spinlock);
 730                                 __dlm_dirty_lockres(dlm, res);
 731                                 spin_unlock(&res->spinlock);
 732                         }
 733                         dlm_lockres_put(res);
 734
 735                         /* unlikely, but we may need to give time to
 736                          * other tasks */
 737                         if (!--n) {
 738                                 mlog(0, "throttling dlm_thread\n");
 739                                 break;
 740                         }
 741                 }
 742
 743                 spin_unlock(&dlm->spinlock);
 744                 dlm_flush_asts(dlm);
 745
 746                 /* yield and continue right away if there is more work to do */
 747                 if (!n) {
 748                         cond_resched();
 749                         continue;
 750                 }
 751
 752                 wait_event_interruptible_timeout(dlm->dlm_thread_wq,
 753                                                  !dlm_dirty_list_empty(dlm) ||
 754                                                  kthread_should_stop(),
 755                                                  timeout);
 756         }
 757
 758         mlog(0, "quitting DLM thread\n");
 759         return 0;
 760 }