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         if (__dlm_lockres_has_locks(res))
  98                 return 0;
  99
 100         if (!list_empty(&res->dirty) || res->state & DLM_LOCK_RES_DIRTY)
 101                 return 0;
 102
 103         if (res->state & DLM_LOCK_RES_RECOVERING)
 104                 return 0;
 105
 106         bit = find_next_bit(res->refmap, O2NM_MAX_NODES, 0);
 107         if (bit < O2NM_MAX_NODES)
 108                 return 0;
 109
 110         /*
 111          * since the bit for dlm->node_num is not set, inflight_locks better
 112          * be zero
 113          */
 114         BUG_ON(res->inflight_locks != 0);
 115         return 1;
 116 }
 117
 118
 119 /* Call whenever you may have added or deleted something from one of
 120  * the lockres queue's. This will figure out whether it belongs on the
 121  * unused list or not and does the appropriate thing. */
 122 void __dlm_lockres_calc_usage(struct dlm_ctxt *dlm,
 123                               struct dlm_lock_resource *res)
 124 {
 125         assert_spin_locked(&dlm->spinlock);
 126         assert_spin_locked(&res->spinlock);
 127
 128         if (__dlm_lockres_unused(res)){
 129                 if (list_empty(&res->purge)) {
 130                         mlog(0, "%s: Adding res %.*s to purge list\n",
 131                              dlm->name, res->lockname.len, res->lockname.name);
 132
 133                         res->last_used = jiffies;
 134                         dlm_lockres_get(res);
 135                         list_add_tail(&res->purge, &dlm->purge_list);
 136                         dlm->purge_count++;
 137                 }
 138         } else if (!list_empty(&res->purge)) {
 139                 mlog(0, "%s: Removing res %.*s from purge list\n",
 140                      dlm->name, res->lockname.len, res->lockname.name);
 141
 142                 list_del_init(&res->purge);
 143                 dlm_lockres_put(res);
 144                 dlm->purge_count--;
 145         }
 146 }
 147
 148 void dlm_lockres_calc_usage(struct dlm_ctxt *dlm,
 149                             struct dlm_lock_resource *res)
 150 {
 151         spin_lock(&dlm->spinlock);
 152         spin_lock(&res->spinlock);
 153
 154         __dlm_lockres_calc_usage(dlm, res);
 155
 156         spin_unlock(&res->spinlock);
 157         spin_unlock(&dlm->spinlock);
 158 }
 159
 160 static void dlm_purge_lockres(struct dlm_ctxt *dlm,
 161                              struct dlm_lock_resource *res)
 162 {
 163         int master;
 164         int ret = 0;
 165
 166         assert_spin_locked(&dlm->spinlock);
 167         assert_spin_locked(&res->spinlock);
 168
 169         master = (res->owner == dlm->node_num);
 170
 171         mlog(0, "%s: Purging res %.*s, master %d\n", dlm->name,
 172              res->lockname.len, res->lockname.name, master);
 173
 174         if (!master) {
 175                 res->state |= DLM_LOCK_RES_DROPPING_REF;
 176                 /* drop spinlock...  retake below */
 177                 spin_unlock(&res->spinlock);
 178                 spin_unlock(&dlm->spinlock);
 179
 180                 spin_lock(&res->spinlock);
 181                 /* This ensures that clear refmap is sent after the set */
 182                 __dlm_wait_on_lockres_flags(res, DLM_LOCK_RES_SETREF_INPROG);
 183                 spin_unlock(&res->spinlock);
 184
 185                 /* clear our bit from the master's refmap, ignore errors */
 186                 ret = dlm_drop_lockres_ref(dlm, res);
 187                 if (ret < 0) {
 188                         if (!dlm_is_host_down(ret))
 189                                 BUG();
 190                 }
 191                 spin_lock(&dlm->spinlock);
 192                 spin_lock(&res->spinlock);
 193         }
 194
 195         if (!list_empty(&res->purge)) {
 196                 mlog(0, "%s: Removing res %.*s from purgelist, master %d\n",
 197                      dlm->name, res->lockname.len, res->lockname.name, master);
 198                 list_del_init(&res->purge);
 199                 dlm_lockres_put(res);
 200                 dlm->purge_count--;
 201         }
 202
 203         if (!__dlm_lockres_unused(res)) {
 204                 mlog(ML_ERROR, "%s: res %.*s in use after deref\n",
 205                      dlm->name, res->lockname.len, res->lockname.name);
 206                 __dlm_print_one_lock_resource(res);
 207                 BUG();
 208         }
 209
 210         __dlm_unhash_lockres(dlm, res);
 211
 212         /* lockres is not in the hash now.  drop the flag and wake up
 213          * any processes waiting in dlm_get_lock_resource. */
 214         if (!master) {
 215                 res->state &= ~DLM_LOCK_RES_DROPPING_REF;
 216                 spin_unlock(&res->spinlock);
 217                 wake_up(&res->wq);
 218         } else
 219                 spin_unlock(&res->spinlock);
 220 }
 221
 222 static void dlm_run_purge_list(struct dlm_ctxt *dlm,
 223                                int purge_now)
 224 {
 225         unsigned int run_max, unused;
 226         unsigned long purge_jiffies;
 227         struct dlm_lock_resource *lockres;
 228
 229         spin_lock(&dlm->spinlock);
 230         run_max = dlm->purge_count;
 231
 232         while(run_max && !list_empty(&dlm->purge_list)) {
 233                 run_max--;
 234
 235                 lockres = list_entry(dlm->purge_list.next,
 236                                      struct dlm_lock_resource, purge);
 237
 238                 spin_lock(&lockres->spinlock);
 239
 240                 purge_jiffies = lockres->last_used +
 241                         msecs_to_jiffies(DLM_PURGE_INTERVAL_MS);
 242
 243                 /* Make sure that we want to be processing this guy at
 244                  * this time. */
 245                 if (!purge_now && time_after(purge_jiffies, jiffies)) {
 246                         /* Since resources are added to the purge list
 247                          * in tail order, we can stop at the first
 248                          * unpurgable resource -- anyone added after
 249                          * him will have a greater last_used value */
 250                         spin_unlock(&lockres->spinlock);
 251                         break;
 252                 }
 253
 254                 /* Status of the lockres *might* change so double
 255                  * check. If the lockres is unused, holding the dlm
 256                  * spinlock will prevent people from getting and more
 257                  * refs on it. */
 258                 unused = __dlm_lockres_unused(lockres);
 259                 if (!unused ||
 260                     (lockres->state & DLM_LOCK_RES_MIGRATING)) {
 261                         mlog(0, "%s: res %.*s is in use or being remastered, "
 262                              "used %d, state %d\n", dlm->name,
 263                              lockres->lockname.len, lockres->lockname.name,
 264                              !unused, lockres->state);
 265                         list_move_tail(&dlm->purge_list, &lockres->purge);
 266                         spin_unlock(&lockres->spinlock);
 267                         continue;
 268                 }
 269
 270                 dlm_lockres_get(lockres);
 271
 272                 dlm_purge_lockres(dlm, lockres);
 273
 274                 dlm_lockres_put(lockres);
 275
 276                 /* Avoid adding any scheduling latencies */
 277                 cond_resched_lock(&dlm->spinlock);
 278         }
 279
 280         spin_unlock(&dlm->spinlock);
 281 }
 282
 283 static void dlm_shuffle_lists(struct dlm_ctxt *dlm,
 284                               struct dlm_lock_resource *res)
 285 {
 286         struct dlm_lock *lock, *target;
 287         struct list_head *iter;
 288         struct list_head *head;
 289         int can_grant = 1;
 290
 291         /*
 292          * Because this function is called with the lockres
 293          * spinlock, and because we know that it is not migrating/
 294          * recovering/in-progress, it is fine to reserve asts and
 295          * basts right before queueing them all throughout
 296          */
 297         assert_spin_locked(&dlm->ast_lock);
 298         assert_spin_locked(&res->spinlock);
 299         BUG_ON((res->state & (DLM_LOCK_RES_MIGRATING|
 300                               DLM_LOCK_RES_RECOVERING|
 301                               DLM_LOCK_RES_IN_PROGRESS)));
 302
 303 converting:
 304         if (list_empty(&res->converting))
 305                 goto blocked;
 306         mlog(0, "%s: res %.*s has locks on the convert queue\n", dlm->name,
 307              res->lockname.len, res->lockname.name);
 308
 309         target = list_entry(res->converting.next, struct dlm_lock, list);
 310         if (target->ml.convert_type == LKM_IVMODE) {
 311                 mlog(ML_ERROR, "%s: res %.*s converting lock to invalid mode\n",
 312                      dlm->name, res->lockname.len, res->lockname.name);
 313                 BUG();
 314         }
 315         head = &res->granted;
 316         list_for_each(iter, head) {
 317                 lock = list_entry(iter, struct dlm_lock, list);
 318                 if (lock==target)
 319                         continue;
 320                 if (!dlm_lock_compatible(lock->ml.type,
 321                                          target->ml.convert_type)) {
 322                         can_grant = 0;
 323                         /* queue the BAST if not already */
 324                         if (lock->ml.highest_blocked == LKM_IVMODE) {
 325                                 __dlm_lockres_reserve_ast(res);
 326                                 __dlm_queue_bast(dlm, lock);
 327                         }
 328                         /* update the highest_blocked if needed */
 329                         if (lock->ml.highest_blocked < target->ml.convert_type)
 330                                 lock->ml.highest_blocked =
 331                                         target->ml.convert_type;
 332                 }
 333         }
 334         head = &res->converting;
 335         list_for_each(iter, head) {
 336                 lock = list_entry(iter, struct dlm_lock, 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         head = &res->granted;
 386         list_for_each(iter, head) {
 387                 lock = list_entry(iter, struct dlm_lock, list);
 388                 if (lock==target)
 389                         continue;
 390                 if (!dlm_lock_compatible(lock->ml.type, target->ml.type)) {
 391                         can_grant = 0;
 392                         if (lock->ml.highest_blocked == LKM_IVMODE) {
 393                                 __dlm_lockres_reserve_ast(res);
 394                                 __dlm_queue_bast(dlm, lock);
 395                         }
 396                         if (lock->ml.highest_blocked < target->ml.type)
 397                                 lock->ml.highest_blocked = target->ml.type;
 398                 }
 399         }
 400
 401         head = &res->converting;
 402         list_for_each(iter, head) {
 403                 lock = list_entry(iter, struct dlm_lock, list);
 404                 if (lock==target)
 405                         continue;
 406                 if (!dlm_lock_compatible(lock->ml.type, target->ml.type)) {
 407                         can_grant = 0;
 408                         if (lock->ml.highest_blocked == LKM_IVMODE) {
 409                                 __dlm_lockres_reserve_ast(res);
 410                                 __dlm_queue_bast(dlm, lock);
 411                         }
 412                         if (lock->ml.highest_blocked < target->ml.type)
 413                                 lock->ml.highest_blocked = target->ml.type;
 414                 }
 415         }
 416
 417         /* we can grant the blocked lock (only
 418          * possible if converting list empty) */
 419         if (can_grant) {
 420                 spin_lock(&target->spinlock);
 421                 BUG_ON(target->ml.highest_blocked != LKM_IVMODE);
 422
 423                 mlog(0, "%s: res %.*s, AST for Blocked lock %u:%llu, type %d, "
 424                      "node %u\n", dlm->name, res->lockname.len,
 425                      res->lockname.name,
 426                      dlm_get_lock_cookie_node(be64_to_cpu(target->ml.cookie)),
 427                      dlm_get_lock_cookie_seq(be64_to_cpu(target->ml.cookie)),
 428                      target->ml.type, target->ml.node);
 429
 430                 /* target->ml.type is already correct */
 431                 list_move_tail(&target->list, &res->granted);
 432
 433                 BUG_ON(!target->lksb);
 434                 target->lksb->status = DLM_NORMAL;
 435
 436                 spin_unlock(&target->spinlock);
 437
 438                 __dlm_lockres_reserve_ast(res);
 439                 __dlm_queue_ast(dlm, target);
 440                 /* go back and check for more */
 441                 goto converting;
 442         }
 443
 444 leave:
 445         return;
 446 }
 447
 448 /* must have NO locks when calling this with res !=NULL * */
 449 void dlm_kick_thread(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
 450 {
 451         if (res) {
 452                 spin_lock(&dlm->spinlock);
 453                 spin_lock(&res->spinlock);
 454                 __dlm_dirty_lockres(dlm, res);
 455                 spin_unlock(&res->spinlock);
 456                 spin_unlock(&dlm->spinlock);
 457         }
 458         wake_up(&dlm->dlm_thread_wq);
 459 }
 460
 461 void __dlm_dirty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
 462 {
 463         assert_spin_locked(&dlm->spinlock);
 464         assert_spin_locked(&res->spinlock);
 465
 466         /* don't shuffle secondary queues */
 467         if ((res->owner == dlm->node_num)) {
 468                 if (res->state & (DLM_LOCK_RES_MIGRATING |
 469                                   DLM_LOCK_RES_BLOCK_DIRTY))
 470                     return;
 471
 472                 if (list_empty(&res->dirty)) {
 473                         /* ref for dirty_list */
 474                         dlm_lockres_get(res);
 475                         list_add_tail(&res->dirty, &dlm->dirty_list);
 476                         res->state |= DLM_LOCK_RES_DIRTY;
 477                 }
 478         }
 479
 480         mlog(0, "%s: res %.*s\n", dlm->name, res->lockname.len,
 481              res->lockname.name);
 482 }
 483
 484
 485 /* Launch the NM thread for the mounted volume */
 486 int dlm_launch_thread(struct dlm_ctxt *dlm)
 487 {
 488         mlog(0, "Starting dlm_thread...\n");
 489
 490         dlm->dlm_thread_task = kthread_run(dlm_thread, dlm, "dlm_thread");
 491         if (IS_ERR(dlm->dlm_thread_task)) {
 492                 mlog_errno(PTR_ERR(dlm->dlm_thread_task));
 493                 dlm->dlm_thread_task = NULL;
 494                 return -EINVAL;
 495         }
 496
 497         return 0;
 498 }
 499
 500 void dlm_complete_thread(struct dlm_ctxt *dlm)
 501 {
 502         if (dlm->dlm_thread_task) {
 503                 mlog(ML_KTHREAD, "Waiting for dlm thread to exit\n");
 504                 kthread_stop(dlm->dlm_thread_task);
 505                 dlm->dlm_thread_task = NULL;
 506         }
 507 }
 508
 509 static int dlm_dirty_list_empty(struct dlm_ctxt *dlm)
 510 {
 511         int empty;
 512
 513         spin_lock(&dlm->spinlock);
 514         empty = list_empty(&dlm->dirty_list);
 515         spin_unlock(&dlm->spinlock);
 516
 517         return empty;
 518 }
 519
 520 static void dlm_flush_asts(struct dlm_ctxt *dlm)
 521 {
 522         int ret;
 523         struct dlm_lock *lock;
 524         struct dlm_lock_resource *res;
 525         u8 hi;
 526
 527         spin_lock(&dlm->ast_lock);
 528         while (!list_empty(&dlm->pending_asts)) {
 529                 lock = list_entry(dlm->pending_asts.next,
 530                                   struct dlm_lock, ast_list);
 531                 /* get an extra ref on lock */
 532                 dlm_lock_get(lock);
 533                 res = lock->lockres;
 534                 mlog(0, "%s: res %.*s, Flush AST for lock %u:%llu, type %d, "
 535                      "node %u\n", dlm->name, res->lockname.len,
 536                      res->lockname.name,
 537                      dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)),
 538                      dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)),
 539                      lock->ml.type, lock->ml.node);
 540
 541                 BUG_ON(!lock->ast_pending);
 542
 543                 /* remove from list (including ref) */
 544                 list_del_init(&lock->ast_list);
 545                 dlm_lock_put(lock);
 546                 spin_unlock(&dlm->ast_lock);
 547
 548                 if (lock->ml.node != dlm->node_num) {
 549                         ret = dlm_do_remote_ast(dlm, res, lock);
 550                         if (ret < 0)
 551                                 mlog_errno(ret);
 552                 } else
 553                         dlm_do_local_ast(dlm, res, lock);
 554
 555                 spin_lock(&dlm->ast_lock);
 556
 557                 /* possible that another ast was queued while
 558                  * we were delivering the last one */
 559                 if (!list_empty(&lock->ast_list)) {
 560                         mlog(0, "%s: res %.*s, AST queued while flushing last "
 561                              "one\n", dlm->name, res->lockname.len,
 562                              res->lockname.name);
 563                 } else
 564                         lock->ast_pending = 0;
 565
 566                 /* drop the extra ref.
 567                  * this may drop it completely. */
 568                 dlm_lock_put(lock);
 569                 dlm_lockres_release_ast(dlm, res);
 570         }
 571
 572         while (!list_empty(&dlm->pending_basts)) {
 573                 lock = list_entry(dlm->pending_basts.next,
 574                                   struct dlm_lock, bast_list);
 575                 /* get an extra ref on lock */
 576                 dlm_lock_get(lock);
 577                 res = lock->lockres;
 578
 579                 BUG_ON(!lock->bast_pending);
 580
 581                 /* get the highest blocked lock, and reset */
 582                 spin_lock(&lock->spinlock);
 583                 BUG_ON(lock->ml.highest_blocked <= LKM_IVMODE);
 584                 hi = lock->ml.highest_blocked;
 585                 lock->ml.highest_blocked = LKM_IVMODE;
 586                 spin_unlock(&lock->spinlock);
 587
 588                 /* remove from list (including ref) */
 589                 list_del_init(&lock->bast_list);
 590                 dlm_lock_put(lock);
 591                 spin_unlock(&dlm->ast_lock);
 592
 593                 mlog(0, "%s: res %.*s, Flush BAST for lock %u:%llu, "
 594                      "blocked %d, node %u\n",
 595                      dlm->name, res->lockname.len, res->lockname.name,
 596                      dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)),
 597                      dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)),
 598                      hi, lock->ml.node);
 599
 600                 if (lock->ml.node != dlm->node_num) {
 601                         ret = dlm_send_proxy_bast(dlm, res, lock, hi);
 602                         if (ret < 0)
 603                                 mlog_errno(ret);
 604                 } else
 605                         dlm_do_local_bast(dlm, res, lock, hi);
 606
 607                 spin_lock(&dlm->ast_lock);
 608
 609                 /* possible that another bast was queued while
 610                  * we were delivering the last one */
 611                 if (!list_empty(&lock->bast_list)) {
 612                         mlog(0, "%s: res %.*s, BAST queued while flushing last "
 613                              "one\n", dlm->name, res->lockname.len,
 614                              res->lockname.name);
 615                 } else
 616                         lock->bast_pending = 0;
 617
 618                 /* drop the extra ref.
 619                  * this may drop it completely. */
 620                 dlm_lock_put(lock);
 621                 dlm_lockres_release_ast(dlm, res);
 622         }
 623         wake_up(&dlm->ast_wq);
 624         spin_unlock(&dlm->ast_lock);
 625 }
 626
 627
 628 #define DLM_THREAD_TIMEOUT_MS (4 * 1000)
 629 #define DLM_THREAD_MAX_DIRTY  100
 630 #define DLM_THREAD_MAX_ASTS   10
 631
 632 static int dlm_thread(void *data)
 633 {
 634         struct dlm_lock_resource *res;
 635         struct dlm_ctxt *dlm = data;
 636         unsigned long timeout = msecs_to_jiffies(DLM_THREAD_TIMEOUT_MS);
 637
 638         mlog(0, "dlm thread running for %s...\n", dlm->name);
 639
 640         while (!kthread_should_stop()) {
 641                 int n = DLM_THREAD_MAX_DIRTY;
 642
 643                 /* dlm_shutting_down is very point-in-time, but that
 644                  * doesn't matter as we'll just loop back around if we
 645                  * get false on the leading edge of a state
 646                  * transition. */
 647                 dlm_run_purge_list(dlm, dlm_shutting_down(dlm));
 648
 649                 /* We really don't want to hold dlm->spinlock while
 650                  * calling dlm_shuffle_lists on each lockres that
 651                  * needs to have its queues adjusted and AST/BASTs
 652                  * run.  So let's pull each entry off the dirty_list
 653                  * and drop dlm->spinlock ASAP.  Once off the list,
 654                  * res->spinlock needs to be taken again to protect
 655                  * the queues while calling dlm_shuffle_lists.  */
 656                 spin_lock(&dlm->spinlock);
 657                 while (!list_empty(&dlm->dirty_list)) {
 658                         int delay = 0;
 659                         res = list_entry(dlm->dirty_list.next,
 660                                          struct dlm_lock_resource, dirty);
 661
 662                         /* peel a lockres off, remove it from the list,
 663                          * unset the dirty flag and drop the dlm lock */
 664                         BUG_ON(!res);
 665                         dlm_lockres_get(res);
 666
 667                         spin_lock(&res->spinlock);
 668                         /* We clear the DLM_LOCK_RES_DIRTY state once we shuffle lists below */
 669                         list_del_init(&res->dirty);
 670                         spin_unlock(&res->spinlock);
 671                         spin_unlock(&dlm->spinlock);
 672                         /* Drop dirty_list ref */
 673                         dlm_lockres_put(res);
 674
 675                         /* lockres can be re-dirtied/re-added to the
 676                          * dirty_list in this gap, but that is ok */
 677
 678                         spin_lock(&dlm->ast_lock);
 679                         spin_lock(&res->spinlock);
 680                         if (res->owner != dlm->node_num) {
 681                                 __dlm_print_one_lock_resource(res);
 682                                 mlog(ML_ERROR, "%s: inprog %d, mig %d, reco %d,"
 683                                      " dirty %d\n", dlm->name,
 684                                      !!(res->state & DLM_LOCK_RES_IN_PROGRESS),
 685                                      !!(res->state & DLM_LOCK_RES_MIGRATING),
 686                                      !!(res->state & DLM_LOCK_RES_RECOVERING),
 687                                      !!(res->state & DLM_LOCK_RES_DIRTY));
 688                         }
 689                         BUG_ON(res->owner != dlm->node_num);
 690
 691                         /* it is now ok to move lockreses in these states
 692                          * to the dirty list, assuming that they will only be
 693                          * dirty for a short while. */
 694                         BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
 695                         if (res->state & (DLM_LOCK_RES_IN_PROGRESS |
 696                                           DLM_LOCK_RES_RECOVERING)) {
 697                                 /* move it to the tail and keep going */
 698                                 res->state &= ~DLM_LOCK_RES_DIRTY;
 699                                 spin_unlock(&res->spinlock);
 700                                 spin_unlock(&dlm->ast_lock);
 701                                 mlog(0, "%s: res %.*s, inprogress, delay list "
 702                                      "shuffle, state %d\n", dlm->name,
 703                                      res->lockname.len, res->lockname.name,
 704                                      res->state);
 705                                 delay = 1;
 706                                 goto in_progress;
 707                         }
 708
 709                         /* at this point the lockres is not migrating/
 710                          * recovering/in-progress.  we have the lockres
 711                          * spinlock and do NOT have the dlm lock.
 712                          * safe to reserve/queue asts and run the lists. */
 713
 714                         /* called while holding lockres lock */
 715                         dlm_shuffle_lists(dlm, res);
 716                         res->state &= ~DLM_LOCK_RES_DIRTY;
 717                         spin_unlock(&res->spinlock);
 718                         spin_unlock(&dlm->ast_lock);
 719
 720                         dlm_lockres_calc_usage(dlm, res);
 721
 722 in_progress:
 723
 724                         spin_lock(&dlm->spinlock);
 725                         /* if the lock was in-progress, stick
 726                          * it on the back of the list */
 727                         if (delay) {
 728                                 spin_lock(&res->spinlock);
 729                                 __dlm_dirty_lockres(dlm, res);
 730                                 spin_unlock(&res->spinlock);
 731                         }
 732                         dlm_lockres_put(res);
 733
 734                         /* unlikely, but we may need to give time to
 735                          * other tasks */
 736                         if (!--n) {
 737                                 mlog(0, "%s: Throttling dlm thread\n",
 738                                      dlm->name);
 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 }