kernel/fs/ocfs2/dlm/dlmthread.c

   1 /* -*- mode: c; c-basic-offset: 8; -*-
   2  * vim: noexpandtab sw=8 ts=8 sts=0:
   3  *
   4  * dlmthread.c
   5  *
   6  * standalone DLM module
   7  *
   8  * Copyright (C) 2004 Oracle.  All rights reserved.
   9  *
  10  * This program is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU General Public
  12  * License as published by the Free Software Foundation; either
  13  * version 2 of the License, or (at your option) any later version.
  14  *
  15  * This program is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  18  * General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU General Public
  21  * License along with this program; if not, write to the
  22  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  23  * Boston, MA 021110-1307, USA.
  24  *
  25  */
  26
  27
  28 #include <linux/module.h>
  29 #include <linux/fs.h>
  30 #include <linux/types.h>
  31 #include <linux/highmem.h>
  32 #include <linux/init.h>
  33 #include <linux/sysctl.h>
  34 #include <linux/random.h>
  35 #include <linux/blkdev.h>
  36 #include <linux/socket.h>
  37 #include <linux/inet.h>
  38 #include <linux/timer.h>
  39 #include <linux/kthread.h>
  40 #include <linux/delay.h>
  41
  42
  43 #include "cluster/heartbeat.h"
  44 #include "cluster/nodemanager.h"
  45 #include "cluster/tcp.h"
  46
  47 #include "dlmapi.h"
  48 #include "dlmcommon.h"
  49 #include "dlmdomain.h"
  50
  51 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_THREAD)
  52 #include "cluster/masklog.h"
  53
  54 static int dlm_thread(void *data);
  55 static void dlm_flush_asts(struct dlm_ctxt *dlm);
  56
  57 #define dlm_lock_is_remote(dlm, lock)     ((lock)->ml.node != (dlm)->node_num)
  58
  59 /* will exit holding res->spinlock, but may drop in function */
  60 /* waits until flags are cleared on res->state */
  61 void __dlm_wait_on_lockres_flags(struct dlm_lock_resource *res, int flags)
  62 {
  63         DECLARE_WAITQUEUE(wait, current);
  64
  65         assert_spin_locked(&res->spinlock);
  66
  67         add_wait_queue(&res->wq, &wait);
  68 repeat:
  69         set_current_state(TASK_UNINTERRUPTIBLE);
  70         if (res->state & flags) {
  71                 spin_unlock(&res->spinlock);
  72                 schedule();
  73                 spin_lock(&res->spinlock);
  74                 goto repeat;
  75         }
  76         remove_wait_queue(&res->wq, &wait);
  77         __set_current_state(TASK_RUNNING);
  78 }
  79
  80 int __dlm_lockres_has_locks(struct dlm_lock_resource *res)
  81 {
  82         if (list_empty(&res->granted) &&
  83             list_empty(&res->converting) &&
  84             list_empty(&res->blocked))
  85                 return 0;
  86         return 1;
  87 }
  88
  89 /* "unused": the lockres has no locks, is not on the dirty list,
  90  * has no inflight locks (in the gap between mastery and acquiring
  91  * the first lock), and has no bits in its refmap.
  92  * truly ready to be freed. */
  93 int __dlm_lockres_unused(struct dlm_lock_resource *res)
  94 {
  95         int bit;
  96
  97         assert_spin_locked(&res->spinlock);
  98
  99         if (__dlm_lockres_has_locks(res))
 100                 return 0;
 101
 102         /* Locks are in the process of being created */
 103         if (res->inflight_locks)
 104                 return 0;
 105
 106         if (!list_empty(&res->dirty) || res->state & DLM_LOCK_RES_DIRTY)
 107                 return 0;
 108
 109         if (res->state & DLM_LOCK_RES_RECOVERING)
 110                 return 0;
 111
 112         /* Another node has this resource with this node as the master */
 113         bit = find_next_bit(res->refmap, O2NM_MAX_NODES, 0);
 114         if (bit < O2NM_MAX_NODES)
 115                 return 0;
 116
 117         return 1;
 118 }
 119
 120
 121 /* Call whenever you may have added or deleted something from one of
 122  * the lockres queue's. This will figure out whether it belongs on the
 123  * unused list or not and does the appropriate thing. */
 124 void __dlm_lockres_calc_usage(struct dlm_ctxt *dlm,
 125                               struct dlm_lock_resource *res)
 126 {
 127         assert_spin_locked(&dlm->spinlock);
 128         assert_spin_locked(&res->spinlock);
 129
 130         if (__dlm_lockres_unused(res)){
 131                 if (list_empty(&res->purge)) {
 132                         mlog(0, "%s: Adding res %.*s to purge list\n",
 133                              dlm->name, res->lockname.len, res->lockname.name);
 134
 135                         res->last_used = jiffies;
 136                         dlm_lockres_get(res);
 137                         list_add_tail(&res->purge, &dlm->purge_list);
 138                         dlm->purge_count++;
 139                 }
 140         } else if (!list_empty(&res->purge)) {
 141                 mlog(0, "%s: Removing res %.*s from purge list\n",
 142                      dlm->name, res->lockname.len, res->lockname.name);
 143
 144                 list_del_init(&res->purge);
 145                 dlm_lockres_put(res);
 146                 dlm->purge_count--;
 147         }
 148 }
 149
 150 void dlm_lockres_calc_usage(struct dlm_ctxt *dlm,
 151                             struct dlm_lock_resource *res)
 152 {
 153         spin_lock(&dlm->spinlock);
 154         spin_lock(&res->spinlock);
 155
 156         __dlm_lockres_calc_usage(dlm, res);
 157
 158         spin_unlock(&res->spinlock);
 159         spin_unlock(&dlm->spinlock);
 160 }
 161
 162 static void dlm_purge_lockres(struct dlm_ctxt *dlm,
 163                              struct dlm_lock_resource *res)
 164 {
 165         int master;
 166         int ret = 0;
 167
 168         assert_spin_locked(&dlm->spinlock);
 169         assert_spin_locked(&res->spinlock);
 170
 171         master = (res->owner == dlm->node_num);
 172
 173         mlog(0, "%s: Purging res %.*s, master %d\n", dlm->name,
 174              res->lockname.len, res->lockname.name, master);
 175
 176         if (!master) {
 177                 res->state |= DLM_LOCK_RES_DROPPING_REF;
 178                 /* drop spinlock...  retake below */
 179                 spin_unlock(&res->spinlock);
 180                 spin_unlock(&dlm->spinlock);
 181
 182                 spin_lock(&res->spinlock);
 183                 /* This ensures that clear refmap is sent after the set */
 184                 __dlm_wait_on_lockres_flags(res, DLM_LOCK_RES_SETREF_INPROG);
 185                 spin_unlock(&res->spinlock);
 186
 187                 /* clear our bit from the master's refmap, ignore errors */
 188                 ret = dlm_drop_lockres_ref(dlm, res);
 189                 if (ret < 0) {
 190                         if (!dlm_is_host_down(ret))
 191                                 BUG();
 192                 }
 193                 spin_lock(&dlm->spinlock);
 194                 spin_lock(&res->spinlock);
 195         }
 196
 197         if (!list_empty(&res->purge)) {
 198                 mlog(0, "%s: Removing res %.*s from purgelist, master %d\n",
 199                      dlm->name, res->lockname.len, res->lockname.name, master);
 200                 list_del_init(&res->purge);
 201                 dlm_lockres_put(res);
 202                 dlm->purge_count--;
 203         }
 204
 205         if (!__dlm_lockres_unused(res)) {
 206                 mlog(ML_ERROR, "%s: res %.*s in use after deref\n",
 207                      dlm->name, res->lockname.len, res->lockname.name);
 208                 __dlm_print_one_lock_resource(res);
 209                 BUG();
 210         }
 211
 212         __dlm_unhash_lockres(dlm, res);
 213
 214         /* lockres is not in the hash now.  drop the flag and wake up
 215          * any processes waiting in dlm_get_lock_resource. */
 216         if (!master) {
 217                 res->state &= ~DLM_LOCK_RES_DROPPING_REF;
 218                 spin_unlock(&res->spinlock);
 219                 wake_up(&res->wq);
 220         } else
 221                 spin_unlock(&res->spinlock);
 222 }
 223
 224 static void dlm_run_purge_list(struct dlm_ctxt *dlm,
 225                                int purge_now)
 226 {
 227         unsigned int run_max, unused;
 228         unsigned long purge_jiffies;
 229         struct dlm_lock_resource *lockres;
 230
 231         spin_lock(&dlm->spinlock);
 232         run_max = dlm->purge_count;
 233
 234         while(run_max && !list_empty(&dlm->purge_list)) {
 235                 run_max--;
 236
 237                 lockres = list_entry(dlm->purge_list.next,
 238                                      struct dlm_lock_resource, purge);
 239
 240                 spin_lock(&lockres->spinlock);
 241
 242                 purge_jiffies = lockres->last_used +
 243                         msecs_to_jiffies(DLM_PURGE_INTERVAL_MS);
 244
 245                 /* Make sure that we want to be processing this guy at
 246                  * this time. */
 247                 if (!purge_now && time_after(purge_jiffies, jiffies)) {
 248                         /* Since resources are added to the purge list
 249                          * in tail order, we can stop at the first
 250                          * unpurgable resource -- anyone added after
 251                          * him will have a greater last_used value */
 252                         spin_unlock(&lockres->spinlock);
 253                         break;
 254                 }
 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. */
 260                 unused = __dlm_lockres_unused(lockres);
 261                 if (!unused ||
 262                     (lockres->state & DLM_LOCK_RES_MIGRATING) ||
 263                     (lockres->inflight_assert_workers != 0)) {
 264                         mlog(0, "%s: res %.*s is in use or being remastered, "
 265                              "used %d, state %d, assert master workers %u\n",
 266                              dlm->name, lockres->lockname.len,
 267                              lockres->lockname.name,
 268                              !unused, lockres->state,
 269                              lockres->inflight_assert_workers);
 270                         list_move_tail(&lockres->purge, &dlm->purge_list);
 271                         spin_unlock(&lockres->spinlock);
 272                         continue;
 273                 }
 274
 275                 dlm_lockres_get(lockres);
 276
 277                 dlm_purge_lockres(dlm, lockres);
 278
 279                 dlm_lockres_put(lockres);
 280
 281                 /* Avoid adding any scheduling latencies */
 282                 cond_resched_lock(&dlm->spinlock);
 283         }
 284
 285         spin_unlock(&dlm->spinlock);
 286 }
 287
 288 static void dlm_shuffle_lists(struct dlm_ctxt *dlm,
 289                               struct dlm_lock_resource *res)
 290 {
 291         struct dlm_lock *lock, *target;
 292         int can_grant = 1;
 293
 294         /*
 295          * Because this function is called with the lockres
 296          * spinlock, and because we know that it is not migrating/
 297          * recovering/in-progress, it is fine to reserve asts and
 298          * basts right before queueing them all throughout
 299          */
 300         assert_spin_locked(&dlm->ast_lock);
 301         assert_spin_locked(&res->spinlock);
 302         BUG_ON((res->state & (DLM_LOCK_RES_MIGRATING|
 303                               DLM_LOCK_RES_RECOVERING|
 304                               DLM_LOCK_RES_IN_PROGRESS)));
 305
 306 converting:
 307         if (list_empty(&res->converting))
 308                 goto blocked;
 309         mlog(0, "%s: res %.*s has locks on the convert queue\n", dlm->name,
 310              res->lockname.len, res->lockname.name);
 311
 312         target = list_entry(res->converting.next, struct dlm_lock, list);
 313         if (target->ml.convert_type == LKM_IVMODE) {
 314                 mlog(ML_ERROR, "%s: res %.*s converting lock to invalid mode\n",
 315                      dlm->name, res->lockname.len, res->lockname.name);
 316                 BUG();
 317         }
 318         list_for_each_entry(lock, &res->granted, list) {
 319                 if (lock==target)
 320                         continue;
 321                 if (!dlm_lock_compatible(lock->ml.type,
 322                                          target->ml.convert_type)) {
 323                         can_grant = 0;
 324                         /* queue the BAST if not already */
 325                         if (lock->ml.highest_blocked == LKM_IVMODE) {
 326                                 __dlm_lockres_reserve_ast(res);
 327                                 __dlm_queue_bast(dlm, lock);
 328                         }
 329                         /* update the highest_blocked if needed */
 330                         if (lock->ml.highest_blocked < target->ml.convert_type)
 331                                 lock->ml.highest_blocked =
 332                                         target->ml.convert_type;
 333                 }
 334         }
 335
 336         list_for_each_entry(lock, &res->converting, list) {
 337                 if (lock==target)
 338                         continue;
 339                 if (!dlm_lock_compatible(lock->ml.type,
 340                                          target->ml.convert_type)) {
 341                         can_grant = 0;
 342                         if (lock->ml.highest_blocked == LKM_IVMODE) {
 343                                 __dlm_lockres_reserve_ast(res);
 344                                 __dlm_queue_bast(dlm, lock);
 345                         }
 346                         if (lock->ml.highest_blocked < target->ml.convert_type)
 347                                 lock->ml.highest_blocked =
 348                                         target->ml.convert_type;
 349                 }
 350         }
 351
 352         /* we can convert the lock */
 353         if (can_grant) {
 354                 spin_lock(&target->spinlock);
 355                 BUG_ON(target->ml.highest_blocked != LKM_IVMODE);
 356
 357                 mlog(0, "%s: res %.*s, AST for Converting lock %u:%llu, type "
 358                      "%d => %d, node %u\n", dlm->name, res->lockname.len,
 359                      res->lockname.name,
 360                      dlm_get_lock_cookie_node(be64_to_cpu(target->ml.cookie)),
 361                      dlm_get_lock_cookie_seq(be64_to_cpu(target->ml.cookie)),
 362                      target->ml.type,
 363                      target->ml.convert_type, target->ml.node);
 364
 365                 target->ml.type = target->ml.convert_type;
 366                 target->ml.convert_type = LKM_IVMODE;
 367                 list_move_tail(&target->list, &res->granted);
 368
 369                 BUG_ON(!target->lksb);
 370                 target->lksb->status = DLM_NORMAL;
 371
 372                 spin_unlock(&target->spinlock);
 373
 374                 __dlm_lockres_reserve_ast(res);
 375                 __dlm_queue_ast(dlm, target);
 376                 /* go back and check for more */
 377                 goto converting;
 378         }
 379
 380 blocked:
 381         if (list_empty(&res->blocked))
 382                 goto leave;
 383         target = list_entry(res->blocked.next, struct dlm_lock, list);
 384
 385         list_for_each_entry(lock, &res->granted, list) {
 386                 if (lock==target)
 387                         continue;
 388                 if (!dlm_lock_compatible(lock->ml.type, target->ml.type)) {
 389                         can_grant = 0;
 390                         if (lock->ml.highest_blocked == LKM_IVMODE) {
 391                                 __dlm_lockres_reserve_ast(res);
 392                                 __dlm_queue_bast(dlm, lock);
 393                         }
 394                         if (lock->ml.highest_blocked < target->ml.type)
 395                                 lock->ml.highest_blocked = target->ml.type;
 396                 }
 397         }
 398
 399         list_for_each_entry(lock, &res->converting, list) {
 400                 if (lock==target)
 401                         continue;
 402                 if (!dlm_lock_compatible(lock->ml.type, target->ml.type)) {
 403                         can_grant = 0;
 404                         if (lock->ml.highest_blocked == LKM_IVMODE) {
 405                                 __dlm_lockres_reserve_ast(res);
 406                                 __dlm_queue_bast(dlm, lock);
 407                         }
 408                         if (lock->ml.highest_blocked < target->ml.type)
 409                                 lock->ml.highest_blocked = target->ml.type;
 410                 }
 411         }
 412
 413         /* we can grant the blocked lock (only
 414          * possible if converting list empty) */
 415         if (can_grant) {
 416                 spin_lock(&target->spinlock);
 417                 BUG_ON(target->ml.highest_blocked != LKM_IVMODE);
 418
 419                 mlog(0, "%s: res %.*s, AST for Blocked lock %u:%llu, type %d, "
 420                      "node %u\n", dlm->name, res->lockname.len,
 421                      res->lockname.name,
 422                      dlm_get_lock_cookie_node(be64_to_cpu(target->ml.cookie)),
 423                      dlm_get_lock_cookie_seq(be64_to_cpu(target->ml.cookie)),
 424                      target->ml.type, target->ml.node);
 425
 426                 /* target->ml.type is already correct */
 427                 list_move_tail(&target->list, &res->granted);
 428
 429                 BUG_ON(!target->lksb);
 430                 target->lksb->status = DLM_NORMAL;
 431
 432                 spin_unlock(&target->spinlock);
 433
 434                 __dlm_lockres_reserve_ast(res);
 435                 __dlm_queue_ast(dlm, target);
 436                 /* go back and check for more */
 437                 goto converting;
 438         }
 439
 440 leave:
 441         return;
 442 }
 443
 444 /* must have NO locks when calling this with res !=NULL * */
 445 void dlm_kick_thread(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
 446 {
 447         if (res) {
 448                 spin_lock(&dlm->spinlock);
 449                 spin_lock(&res->spinlock);
 450                 __dlm_dirty_lockres(dlm, res);
 451                 spin_unlock(&res->spinlock);
 452                 spin_unlock(&dlm->spinlock);
 453         }
 454         wake_up(&dlm->dlm_thread_wq);
 455 }
 456
 457 void __dlm_dirty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
 458 {
 459         assert_spin_locked(&dlm->spinlock);
 460         assert_spin_locked(&res->spinlock);
 461
 462         /* don't shuffle secondary queues */
 463         if ((res->owner == dlm->node_num)) {
 464                 if (res->state & (DLM_LOCK_RES_MIGRATING |
 465                                   DLM_LOCK_RES_BLOCK_DIRTY))
 466                     return;
 467
 468                 if (list_empty(&res->dirty)) {
 469                         /* ref for dirty_list */
 470                         dlm_lockres_get(res);
 471                         list_add_tail(&res->dirty, &dlm->dirty_list);
 472                         res->state |= DLM_LOCK_RES_DIRTY;
 473                 }
 474         }
 475
 476         mlog(0, "%s: res %.*s\n", dlm->name, res->lockname.len,
 477              res->lockname.name);
 478 }
 479
 480
 481 /* Launch the NM thread for the mounted volume */
 482 int dlm_launch_thread(struct dlm_ctxt *dlm)
 483 {
 484         mlog(0, "Starting dlm_thread...\n");
 485
 486         dlm->dlm_thread_task = kthread_run(dlm_thread, dlm, "dlm_thread");
 487         if (IS_ERR(dlm->dlm_thread_task)) {
 488                 mlog_errno(PTR_ERR(dlm->dlm_thread_task));
 489                 dlm->dlm_thread_task = NULL;
 490                 return -EINVAL;
 491         }
 492
 493         return 0;
 494 }
 495
 496 void dlm_complete_thread(struct dlm_ctxt *dlm)
 497 {
 498         if (dlm->dlm_thread_task) {
 499                 mlog(ML_KTHREAD, "Waiting for dlm thread to exit\n");
 500                 kthread_stop(dlm->dlm_thread_task);
 501                 dlm->dlm_thread_task = NULL;
 502         }
 503 }
 504
 505 static int dlm_dirty_list_empty(struct dlm_ctxt *dlm)
 506 {
 507         int empty;
 508
 509         spin_lock(&dlm->spinlock);
 510         empty = list_empty(&dlm->dirty_list);
 511         spin_unlock(&dlm->spinlock);
 512
 513         return empty;
 514 }
 515
 516 static void dlm_flush_asts(struct dlm_ctxt *dlm)
 517 {
 518         int ret;
 519         struct dlm_lock *lock;
 520         struct dlm_lock_resource *res;
 521         u8 hi;
 522
 523         spin_lock(&dlm->ast_lock);
 524         while (!list_empty(&dlm->pending_asts)) {
 525                 lock = list_entry(dlm->pending_asts.next,
 526                                   struct dlm_lock, ast_list);
 527                 /* get an extra ref on lock */
 528                 dlm_lock_get(lock);
 529                 res = lock->lockres;
 530                 mlog(0, "%s: res %.*s, Flush AST for lock %u:%llu, type %d, "
 531                      "node %u\n", dlm->name, res->lockname.len,
 532                      res->lockname.name,
 533                      dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)),
 534                      dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)),
 535                      lock->ml.type, lock->ml.node);
 536
 537                 BUG_ON(!lock->ast_pending);
 538
 539                 /* remove from list (including ref) */
 540                 list_del_init(&lock->ast_list);
 541                 dlm_lock_put(lock);
 542                 spin_unlock(&dlm->ast_lock);
 543
 544                 if (lock->ml.node != dlm->node_num) {
 545                         ret = dlm_do_remote_ast(dlm, res, lock);
 546                         if (ret < 0)
 547                                 mlog_errno(ret);
 548                 } else
 549                         dlm_do_local_ast(dlm, res, lock);
 550
 551                 spin_lock(&dlm->ast_lock);
 552
 553                 /* possible that another ast was queued while
 554                  * we were delivering the last one */
 555                 if (!list_empty(&lock->ast_list)) {
 556                         mlog(0, "%s: res %.*s, AST queued while flushing last "
 557                              "one\n", dlm->name, res->lockname.len,
 558                              res->lockname.name);
 559                 } else
 560                         lock->ast_pending = 0;
 561
 562                 /* drop the extra ref.
 563                  * this may drop it completely. */
 564                 dlm_lock_put(lock);
 565                 dlm_lockres_release_ast(dlm, res);
 566         }
 567
 568         while (!list_empty(&dlm->pending_basts)) {
 569                 lock = list_entry(dlm->pending_basts.next,
 570                                   struct dlm_lock, bast_list);
 571                 /* get an extra ref on lock */
 572                 dlm_lock_get(lock);
 573                 res = lock->lockres;
 574
 575                 BUG_ON(!lock->bast_pending);
 576
 577                 /* get the highest blocked lock, and reset */
 578                 spin_lock(&lock->spinlock);
 579                 BUG_ON(lock->ml.highest_blocked <= LKM_IVMODE);
 580                 hi = lock->ml.highest_blocked;
 581                 lock->ml.highest_blocked = LKM_IVMODE;
 582                 spin_unlock(&lock->spinlock);
 583
 584                 /* remove from list (including ref) */
 585                 list_del_init(&lock->bast_list);
 586                 dlm_lock_put(lock);
 587                 spin_unlock(&dlm->ast_lock);
 588
 589                 mlog(0, "%s: res %.*s, Flush BAST for lock %u:%llu, "
 590                      "blocked %d, node %u\n",
 591                      dlm->name, res->lockname.len, res->lockname.name,
 592                      dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)),
 593                      dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)),
 594                      hi, lock->ml.node);
 595
 596                 if (lock->ml.node != dlm->node_num) {
 597                         ret = dlm_send_proxy_bast(dlm, res, lock, hi);
 598                         if (ret < 0)
 599                                 mlog_errno(ret);
 600                 } else
 601                         dlm_do_local_bast(dlm, res, lock, hi);
 602
 603                 spin_lock(&dlm->ast_lock);
 604
 605                 /* possible that another bast was queued while
 606                  * we were delivering the last one */
 607                 if (!list_empty(&lock->bast_list)) {
 608                         mlog(0, "%s: res %.*s, BAST queued while flushing last "
 609                              "one\n", dlm->name, res->lockname.len,
 610                              res->lockname.name);
 611                 } else
 612                         lock->bast_pending = 0;
 613
 614                 /* drop the extra ref.
 615                  * this may drop it completely. */
 616                 dlm_lock_put(lock);
 617                 dlm_lockres_release_ast(dlm, res);
 618         }
 619         wake_up(&dlm->ast_wq);
 620         spin_unlock(&dlm->ast_lock);
 621 }
 622
 623
 624 #define DLM_THREAD_TIMEOUT_MS (4 * 1000)
 625 #define DLM_THREAD_MAX_DIRTY  100
 626 #define DLM_THREAD_MAX_ASTS   10
 627
 628 static int dlm_thread(void *data)
 629 {
 630         struct dlm_lock_resource *res;
 631         struct dlm_ctxt *dlm = data;
 632         unsigned long timeout = msecs_to_jiffies(DLM_THREAD_TIMEOUT_MS);
 633
 634         mlog(0, "dlm thread running for %s...\n", dlm->name);
 635
 636         while (!kthread_should_stop()) {
 637                 int n = DLM_THREAD_MAX_DIRTY;
 638
 639                 /* dlm_shutting_down is very point-in-time, but that
 640                  * doesn't matter as we'll just loop back around if we
 641                  * get false on the leading edge of a state
 642                  * transition. */
 643                 dlm_run_purge_list(dlm, dlm_shutting_down(dlm));
 644
 645                 /* We really don't want to hold dlm->spinlock while
 646                  * calling dlm_shuffle_lists on each lockres that
 647                  * needs to have its queues adjusted and AST/BASTs
 648                  * run.  So let's pull each entry off the dirty_list
 649                  * and drop dlm->spinlock ASAP.  Once off the list,
 650                  * res->spinlock needs to be taken again to protect
 651                  * the queues while calling dlm_shuffle_lists.  */
 652                 spin_lock(&dlm->spinlock);
 653                 while (!list_empty(&dlm->dirty_list)) {
 654                         int delay = 0;
 655                         res = list_entry(dlm->dirty_list.next,
 656                                          struct dlm_lock_resource, dirty);
 657
 658                         /* peel a lockres off, remove it from the list,
 659                          * unset the dirty flag and drop the dlm lock */
 660                         BUG_ON(!res);
 661                         dlm_lockres_get(res);
 662
 663                         spin_lock(&res->spinlock);
 664                         /* We clear the DLM_LOCK_RES_DIRTY state once we shuffle lists below */
 665                         list_del_init(&res->dirty);
 666                         spin_unlock(&res->spinlock);
 667                         spin_unlock(&dlm->spinlock);
 668                         /* Drop dirty_list ref */
 669                         dlm_lockres_put(res);
 670
 671                         /* lockres can be re-dirtied/re-added to the
 672                          * dirty_list in this gap, but that is ok */
 673
 674                         spin_lock(&dlm->ast_lock);
 675                         spin_lock(&res->spinlock);
 676                         if (res->owner != dlm->node_num) {
 677                                 __dlm_print_one_lock_resource(res);
 678                                 mlog(ML_ERROR, "%s: inprog %d, mig %d, reco %d,"
 679                                      " dirty %d\n", dlm->name,
 680                                      !!(res->state & DLM_LOCK_RES_IN_PROGRESS),
 681                                      !!(res->state & DLM_LOCK_RES_MIGRATING),
 682                                      !!(res->state & DLM_LOCK_RES_RECOVERING),
 683                                      !!(res->state & DLM_LOCK_RES_DIRTY));
 684                         }
 685                         BUG_ON(res->owner != dlm->node_num);
 686
 687                         /* it is now ok to move lockreses in these states
 688                          * to the dirty list, assuming that they will only be
 689                          * dirty for a short while. */
 690                         BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
 691                         if (res->state & (DLM_LOCK_RES_IN_PROGRESS |
 692                                           DLM_LOCK_RES_RECOVERING)) {
 693                                 /* move it to the tail and keep going */
 694                                 res->state &= ~DLM_LOCK_RES_DIRTY;
 695                                 spin_unlock(&res->spinlock);
 696                                 spin_unlock(&dlm->ast_lock);
 697                                 mlog(0, "%s: res %.*s, inprogress, delay list "
 698                                      "shuffle, state %d\n", dlm->name,
 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                         /* called while holding lockres lock */
 711                         dlm_shuffle_lists(dlm, res);
 712                         res->state &= ~DLM_LOCK_RES_DIRTY;
 713                         spin_unlock(&res->spinlock);
 714                         spin_unlock(&dlm->ast_lock);
 715
 716                         dlm_lockres_calc_usage(dlm, res);
 717
 718 in_progress:
 719
 720                         spin_lock(&dlm->spinlock);
 721                         /* if the lock was in-progress, stick
 722                          * it on the back of the list */
 723                         if (delay) {
 724                                 spin_lock(&res->spinlock);
 725                                 __dlm_dirty_lockres(dlm, res);
 726                                 spin_unlock(&res->spinlock);
 727                         }
 728                         dlm_lockres_put(res);
 729
 730                         /* unlikely, but we may need to give time to
 731                          * other tasks */
 732                         if (!--n) {
 733                                 mlog(0, "%s: Throttling dlm thread\n",
 734                                      dlm->name);
 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 }