fs/ocfs2/dlm/dlmthread.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * dlmthread.c
   4  *
   5  * standalone DLM module
   6  *
   7  * Copyright (C) 2004 Oracle.  All rights reserved.
   8  */
   9
  10
  11 #include <linux/module.h>
  12 #include <linux/fs.h>
  13 #include <linux/types.h>
  14 #include <linux/highmem.h>
  15 #include <linux/init.h>
  16 #include <linux/sysctl.h>
  17 #include <linux/random.h>
  18 #include <linux/blkdev.h>
  19 #include <linux/socket.h>
  20 #include <linux/inet.h>
  21 #include <linux/timer.h>
  22 #include <linux/kthread.h>
  23 #include <linux/delay.h>
  24
  25
  26 #include "../cluster/heartbeat.h"
  27 #include "../cluster/nodemanager.h"
  28 #include "../cluster/tcp.h"
  29
  30 #include "dlmapi.h"
  31 #include "dlmcommon.h"
  32 #include "dlmdomain.h"
  33
  34 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_THREAD)
  35 #include "../cluster/masklog.h"
  36
  37 static int dlm_thread(void *data);
  38 static void dlm_flush_asts(struct dlm_ctxt *dlm);
  39
  40 /* will exit holding res->spinlock, but may drop in function */
  41 /* waits until flags are cleared on res->state */
  42 void __dlm_wait_on_lockres_flags(struct dlm_lock_resource *res, int flags)
  43 {
  44         DECLARE_WAITQUEUE(wait, current);
  45
  46         assert_spin_locked(&res->spinlock);
  47
  48         add_wait_queue(&res->wq, &wait);
  49 repeat:
  50         set_current_state(TASK_UNINTERRUPTIBLE);
  51         if (res->state & flags) {
  52                 spin_unlock(&res->spinlock);
  53                 schedule();
  54                 spin_lock(&res->spinlock);
  55                 goto repeat;
  56         }
  57         remove_wait_queue(&res->wq, &wait);
  58         __set_current_state(TASK_RUNNING);
  59 }
  60
  61 int __dlm_lockres_has_locks(struct dlm_lock_resource *res)
  62 {
  63         if (list_empty(&res->granted) &&
  64             list_empty(&res->converting) &&
  65             list_empty(&res->blocked))
  66                 return 0;
  67         return 1;
  68 }
  69
  70 /* "unused": the lockres has no locks, is not on the dirty list,
  71  * has no inflight locks (in the gap between mastery and acquiring
  72  * the first lock), and has no bits in its refmap.
  73  * truly ready to be freed. */
  74 int __dlm_lockres_unused(struct dlm_lock_resource *res)
  75 {
  76         int bit;
  77
  78         assert_spin_locked(&res->spinlock);
  79
  80         if (__dlm_lockres_has_locks(res))
  81                 return 0;
  82
  83         /* Locks are in the process of being created */
  84         if (res->inflight_locks)
  85                 return 0;
  86
  87         if (!list_empty(&res->dirty) || res->state & DLM_LOCK_RES_DIRTY)
  88                 return 0;
  89
  90         if (res->state & (DLM_LOCK_RES_RECOVERING|
  91                         DLM_LOCK_RES_RECOVERY_WAITING))
  92                 return 0;
  93
  94         /* Another node has this resource with this node as the master */
  95         bit = find_next_bit(res->refmap, O2NM_MAX_NODES, 0);
  96         if (bit < O2NM_MAX_NODES)
  97                 return 0;
  98
  99         return 1;
 100 }
 101
 102
 103 /* Call whenever you may have added or deleted something from one of
 104  * the lockres queue's. This will figure out whether it belongs on the
 105  * unused list or not and does the appropriate thing. */
 106 void __dlm_lockres_calc_usage(struct dlm_ctxt *dlm,
 107                               struct dlm_lock_resource *res)
 108 {
 109         assert_spin_locked(&dlm->spinlock);
 110         assert_spin_locked(&res->spinlock);
 111
 112         if (__dlm_lockres_unused(res)){
 113                 if (list_empty(&res->purge)) {
 114                         mlog(0, "%s: Adding res %.*s to purge list\n",
 115                              dlm->name, res->lockname.len, res->lockname.name);
 116
 117                         res->last_used = jiffies;
 118                         dlm_lockres_get(res);
 119                         list_add_tail(&res->purge, &dlm->purge_list);
 120                         dlm->purge_count++;
 121                 }
 122         } else if (!list_empty(&res->purge)) {
 123                 mlog(0, "%s: Removing res %.*s from purge list\n",
 124                      dlm->name, res->lockname.len, res->lockname.name);
 125
 126                 list_del_init(&res->purge);
 127                 dlm_lockres_put(res);
 128                 dlm->purge_count--;
 129         }
 130 }
 131
 132 void dlm_lockres_calc_usage(struct dlm_ctxt *dlm,
 133                             struct dlm_lock_resource *res)
 134 {
 135         spin_lock(&dlm->spinlock);
 136         spin_lock(&res->spinlock);
 137
 138         __dlm_lockres_calc_usage(dlm, res);
 139
 140         spin_unlock(&res->spinlock);
 141         spin_unlock(&dlm->spinlock);
 142 }
 143
 144 /*
 145  * Do the real purge work:
 146  *     unhash the lockres, and
 147  *     clear flag DLM_LOCK_RES_DROPPING_REF.
 148  * It requires dlm and lockres spinlock to be taken.
 149  */
 150 void __dlm_do_purge_lockres(struct dlm_ctxt *dlm,
 151                 struct dlm_lock_resource *res)
 152 {
 153         assert_spin_locked(&dlm->spinlock);
 154         assert_spin_locked(&res->spinlock);
 155
 156         if (!list_empty(&res->purge)) {
 157                 mlog(0, "%s: Removing res %.*s from purgelist\n",
 158                      dlm->name, res->lockname.len, res->lockname.name);
 159                 list_del_init(&res->purge);
 160                 dlm_lockres_put(res);
 161                 dlm->purge_count--;
 162         }
 163
 164         if (!__dlm_lockres_unused(res)) {
 165                 mlog(ML_ERROR, "%s: res %.*s in use after deref\n",
 166                      dlm->name, res->lockname.len, res->lockname.name);
 167                 __dlm_print_one_lock_resource(res);
 168                 BUG();
 169         }
 170
 171         __dlm_unhash_lockres(dlm, res);
 172
 173         spin_lock(&dlm->track_lock);
 174         if (!list_empty(&res->tracking))
 175                 list_del_init(&res->tracking);
 176         else {
 177                 mlog(ML_ERROR, "%s: Resource %.*s not on the Tracking list\n",
 178                      dlm->name, res->lockname.len, res->lockname.name);
 179                 __dlm_print_one_lock_resource(res);
 180         }
 181         spin_unlock(&dlm->track_lock);
 182
 183         /*
 184          * lockres is not in the hash now. drop the flag and wake up
 185          * any processes waiting in dlm_get_lock_resource.
 186          */
 187         res->state &= ~DLM_LOCK_RES_DROPPING_REF;
 188 }
 189
 190 static void dlm_purge_lockres(struct dlm_ctxt *dlm,
 191                              struct dlm_lock_resource *res)
 192 {
 193         int master;
 194         int ret = 0;
 195
 196         assert_spin_locked(&dlm->spinlock);
 197         assert_spin_locked(&res->spinlock);
 198
 199         master = (res->owner == dlm->node_num);
 200
 201         mlog(0, "%s: Purging res %.*s, master %d\n", dlm->name,
 202              res->lockname.len, res->lockname.name, master);
 203
 204         if (!master) {
 205                 if (res->state & DLM_LOCK_RES_DROPPING_REF) {
 206                         mlog(ML_NOTICE, "%s: res %.*s already in DLM_LOCK_RES_DROPPING_REF state\n",
 207                                 dlm->name, res->lockname.len, res->lockname.name);
 208                         spin_unlock(&res->spinlock);
 209                         return;
 210                 }
 211
 212                 res->state |= DLM_LOCK_RES_DROPPING_REF;
 213                 /* drop spinlock...  retake below */
 214                 spin_unlock(&res->spinlock);
 215                 spin_unlock(&dlm->spinlock);
 216
 217                 spin_lock(&res->spinlock);
 218                 /* This ensures that clear refmap is sent after the set */
 219                 __dlm_wait_on_lockres_flags(res, DLM_LOCK_RES_SETREF_INPROG);
 220                 spin_unlock(&res->spinlock);
 221
 222                 /* clear our bit from the master's refmap, ignore errors */
 223                 ret = dlm_drop_lockres_ref(dlm, res);
 224                 if (ret < 0) {
 225                         if (!dlm_is_host_down(ret))
 226                                 BUG();
 227                 }
 228                 spin_lock(&dlm->spinlock);
 229                 spin_lock(&res->spinlock);
 230         }
 231
 232         if (!list_empty(&res->purge)) {
 233                 mlog(0, "%s: Removing res %.*s from purgelist, master %d\n",
 234                      dlm->name, res->lockname.len, res->lockname.name, master);
 235                 list_del_init(&res->purge);
 236                 dlm_lockres_put(res);
 237                 dlm->purge_count--;
 238         }
 239
 240         if (!master && ret == DLM_DEREF_RESPONSE_INPROG) {
 241                 mlog(0, "%s: deref %.*s in progress\n",
 242                         dlm->name, res->lockname.len, res->lockname.name);
 243                 spin_unlock(&res->spinlock);
 244                 return;
 245         }
 246
 247         if (!__dlm_lockres_unused(res)) {
 248                 mlog(ML_ERROR, "%s: res %.*s in use after deref\n",
 249                      dlm->name, res->lockname.len, res->lockname.name);
 250                 __dlm_print_one_lock_resource(res);
 251                 BUG();
 252         }
 253
 254         __dlm_unhash_lockres(dlm, res);
 255
 256         spin_lock(&dlm->track_lock);
 257         if (!list_empty(&res->tracking))
 258                 list_del_init(&res->tracking);
 259         else {
 260                 mlog(ML_ERROR, "Resource %.*s not on the Tracking list\n",
 261                                 res->lockname.len, res->lockname.name);
 262                 __dlm_print_one_lock_resource(res);
 263         }
 264         spin_unlock(&dlm->track_lock);
 265
 266         /* lockres is not in the hash now.  drop the flag and wake up
 267          * any processes waiting in dlm_get_lock_resource. */
 268         if (!master) {
 269                 res->state &= ~DLM_LOCK_RES_DROPPING_REF;
 270                 spin_unlock(&res->spinlock);
 271                 wake_up(&res->wq);
 272         } else
 273                 spin_unlock(&res->spinlock);
 274 }
 275
 276 static void dlm_run_purge_list(struct dlm_ctxt *dlm,
 277                                int purge_now)
 278 {
 279         unsigned int run_max, unused;
 280         unsigned long purge_jiffies;
 281         struct dlm_lock_resource *lockres;
 282
 283         spin_lock(&dlm->spinlock);
 284         run_max = dlm->purge_count;
 285
 286         while(run_max && !list_empty(&dlm->purge_list)) {
 287                 run_max--;
 288
 289                 lockres = list_entry(dlm->purge_list.next,
 290                                      struct dlm_lock_resource, purge);
 291
 292                 spin_lock(&lockres->spinlock);
 293
 294                 purge_jiffies = lockres->last_used +
 295                         msecs_to_jiffies(DLM_PURGE_INTERVAL_MS);
 296
 297                 /* Make sure that we want to be processing this guy at
 298                  * this time. */
 299                 if (!purge_now && time_after(purge_jiffies, jiffies)) {
 300                         /* Since resources are added to the purge list
 301                          * in tail order, we can stop at the first
 302                          * unpurgable resource -- anyone added after
 303                          * him will have a greater last_used value */
 304                         spin_unlock(&lockres->spinlock);
 305                         break;
 306                 }
 307
 308                 /* Status of the lockres *might* change so double
 309                  * check. If the lockres is unused, holding the dlm
 310                  * spinlock will prevent people from getting and more
 311                  * refs on it. */
 312                 unused = __dlm_lockres_unused(lockres);
 313                 if (!unused ||
 314                     (lockres->state & DLM_LOCK_RES_MIGRATING) ||
 315                     (lockres->inflight_assert_workers != 0)) {
 316                         mlog(0, "%s: res %.*s is in use or being remastered, "
 317                              "used %d, state %d, assert master workers %u\n",
 318                              dlm->name, lockres->lockname.len,
 319                              lockres->lockname.name,
 320                              !unused, lockres->state,
 321                              lockres->inflight_assert_workers);
 322                         list_move_tail(&lockres->purge, &dlm->purge_list);
 323                         spin_unlock(&lockres->spinlock);
 324                         continue;
 325                 }
 326
 327                 dlm_lockres_get(lockres);
 328
 329                 dlm_purge_lockres(dlm, lockres);
 330
 331                 dlm_lockres_put(lockres);
 332
 333                 /* Avoid adding any scheduling latencies */
 334                 cond_resched_lock(&dlm->spinlock);
 335         }
 336
 337         spin_unlock(&dlm->spinlock);
 338 }
 339
 340 static void dlm_shuffle_lists(struct dlm_ctxt *dlm,
 341                               struct dlm_lock_resource *res)
 342 {
 343         struct dlm_lock *lock, *target;
 344         int can_grant = 1;
 345
 346         /*
 347          * Because this function is called with the lockres
 348          * spinlock, and because we know that it is not migrating/
 349          * recovering/in-progress, it is fine to reserve asts and
 350          * basts right before queueing them all throughout
 351          */
 352         assert_spin_locked(&dlm->ast_lock);
 353         assert_spin_locked(&res->spinlock);
 354         BUG_ON((res->state & (DLM_LOCK_RES_MIGRATING|
 355                               DLM_LOCK_RES_RECOVERING|
 356                               DLM_LOCK_RES_IN_PROGRESS)));
 357
 358 converting:
 359         if (list_empty(&res->converting))
 360                 goto blocked;
 361         mlog(0, "%s: res %.*s has locks on the convert queue\n", dlm->name,
 362              res->lockname.len, res->lockname.name);
 363
 364         target = list_entry(res->converting.next, struct dlm_lock, list);
 365         if (target->ml.convert_type == LKM_IVMODE) {
 366                 mlog(ML_ERROR, "%s: res %.*s converting lock to invalid mode\n",
 367                      dlm->name, res->lockname.len, res->lockname.name);
 368                 BUG();
 369         }
 370         list_for_each_entry(lock, &res->granted, list) {
 371                 if (lock==target)
 372                         continue;
 373                 if (!dlm_lock_compatible(lock->ml.type,
 374                                          target->ml.convert_type)) {
 375                         can_grant = 0;
 376                         /* queue the BAST if not already */
 377                         if (lock->ml.highest_blocked == LKM_IVMODE) {
 378                                 __dlm_lockres_reserve_ast(res);
 379                                 __dlm_queue_bast(dlm, lock);
 380                         }
 381                         /* update the highest_blocked if needed */
 382                         if (lock->ml.highest_blocked < target->ml.convert_type)
 383                                 lock->ml.highest_blocked =
 384                                         target->ml.convert_type;
 385                 }
 386         }
 387
 388         list_for_each_entry(lock, &res->converting, list) {
 389                 if (lock==target)
 390                         continue;
 391                 if (!dlm_lock_compatible(lock->ml.type,
 392                                          target->ml.convert_type)) {
 393                         can_grant = 0;
 394                         if (lock->ml.highest_blocked == LKM_IVMODE) {
 395                                 __dlm_lockres_reserve_ast(res);
 396                                 __dlm_queue_bast(dlm, lock);
 397                         }
 398                         if (lock->ml.highest_blocked < target->ml.convert_type)
 399                                 lock->ml.highest_blocked =
 400                                         target->ml.convert_type;
 401                 }
 402         }
 403
 404         /* we can convert the lock */
 405         if (can_grant) {
 406                 spin_lock(&target->spinlock);
 407                 BUG_ON(target->ml.highest_blocked != LKM_IVMODE);
 408
 409                 mlog(0, "%s: res %.*s, AST for Converting lock %u:%llu, type "
 410                      "%d => %d, node %u\n", dlm->name, res->lockname.len,
 411                      res->lockname.name,
 412                      dlm_get_lock_cookie_node(be64_to_cpu(target->ml.cookie)),
 413                      dlm_get_lock_cookie_seq(be64_to_cpu(target->ml.cookie)),
 414                      target->ml.type,
 415                      target->ml.convert_type, target->ml.node);
 416
 417                 target->ml.type = target->ml.convert_type;
 418                 target->ml.convert_type = LKM_IVMODE;
 419                 list_move_tail(&target->list, &res->granted);
 420
 421                 BUG_ON(!target->lksb);
 422                 target->lksb->status = DLM_NORMAL;
 423
 424                 spin_unlock(&target->spinlock);
 425
 426                 __dlm_lockres_reserve_ast(res);
 427                 __dlm_queue_ast(dlm, target);
 428                 /* go back and check for more */
 429                 goto converting;
 430         }
 431
 432 blocked:
 433         if (list_empty(&res->blocked))
 434                 goto leave;
 435         target = list_entry(res->blocked.next, struct dlm_lock, list);
 436
 437         list_for_each_entry(lock, &res->granted, list) {
 438                 if (lock==target)
 439                         continue;
 440                 if (!dlm_lock_compatible(lock->ml.type, target->ml.type)) {
 441                         can_grant = 0;
 442                         if (lock->ml.highest_blocked == LKM_IVMODE) {
 443                                 __dlm_lockres_reserve_ast(res);
 444                                 __dlm_queue_bast(dlm, lock);
 445                         }
 446                         if (lock->ml.highest_blocked < target->ml.type)
 447                                 lock->ml.highest_blocked = target->ml.type;
 448                 }
 449         }
 450
 451         list_for_each_entry(lock, &res->converting, list) {
 452                 if (lock==target)
 453                         continue;
 454                 if (!dlm_lock_compatible(lock->ml.type, target->ml.type)) {
 455                         can_grant = 0;
 456                         if (lock->ml.highest_blocked == LKM_IVMODE) {
 457                                 __dlm_lockres_reserve_ast(res);
 458                                 __dlm_queue_bast(dlm, lock);
 459                         }
 460                         if (lock->ml.highest_blocked < target->ml.type)
 461                                 lock->ml.highest_blocked = target->ml.type;
 462                 }
 463         }
 464
 465         /* we can grant the blocked lock (only
 466          * possible if converting list empty) */
 467         if (can_grant) {
 468                 spin_lock(&target->spinlock);
 469                 BUG_ON(target->ml.highest_blocked != LKM_IVMODE);
 470
 471                 mlog(0, "%s: res %.*s, AST for Blocked lock %u:%llu, type %d, "
 472                      "node %u\n", dlm->name, res->lockname.len,
 473                      res->lockname.name,
 474                      dlm_get_lock_cookie_node(be64_to_cpu(target->ml.cookie)),
 475                      dlm_get_lock_cookie_seq(be64_to_cpu(target->ml.cookie)),
 476                      target->ml.type, target->ml.node);
 477
 478                 /* target->ml.type is already correct */
 479                 list_move_tail(&target->list, &res->granted);
 480
 481                 BUG_ON(!target->lksb);
 482                 target->lksb->status = DLM_NORMAL;
 483
 484                 spin_unlock(&target->spinlock);
 485
 486                 __dlm_lockres_reserve_ast(res);
 487                 __dlm_queue_ast(dlm, target);
 488                 /* go back and check for more */
 489                 goto converting;
 490         }
 491
 492 leave:
 493         return;
 494 }
 495
 496 /* must have NO locks when calling this with res !=NULL * */
 497 void dlm_kick_thread(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
 498 {
 499         if (res) {
 500                 spin_lock(&dlm->spinlock);
 501                 spin_lock(&res->spinlock);
 502                 __dlm_dirty_lockres(dlm, res);
 503                 spin_unlock(&res->spinlock);
 504                 spin_unlock(&dlm->spinlock);
 505         }
 506         wake_up(&dlm->dlm_thread_wq);
 507 }
 508
 509 void __dlm_dirty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
 510 {
 511         assert_spin_locked(&dlm->spinlock);
 512         assert_spin_locked(&res->spinlock);
 513
 514         /* don't shuffle secondary queues */
 515         if (res->owner == dlm->node_num) {
 516                 if (res->state & (DLM_LOCK_RES_MIGRATING |
 517                                   DLM_LOCK_RES_BLOCK_DIRTY))
 518                     return;
 519
 520                 if (list_empty(&res->dirty)) {
 521                         /* ref for dirty_list */
 522                         dlm_lockres_get(res);
 523                         list_add_tail(&res->dirty, &dlm->dirty_list);
 524                         res->state |= DLM_LOCK_RES_DIRTY;
 525                 }
 526         }
 527
 528         mlog(0, "%s: res %.*s\n", dlm->name, res->lockname.len,
 529              res->lockname.name);
 530 }
 531
 532
 533 /* Launch the NM thread for the mounted volume */
 534 int dlm_launch_thread(struct dlm_ctxt *dlm)
 535 {
 536         mlog(0, "Starting dlm_thread...\n");
 537
 538         dlm->dlm_thread_task = kthread_run(dlm_thread, dlm, "dlm-%s",
 539                         dlm->name);
 540         if (IS_ERR(dlm->dlm_thread_task)) {
 541                 mlog_errno(PTR_ERR(dlm->dlm_thread_task));
 542                 dlm->dlm_thread_task = NULL;
 543                 return -EINVAL;
 544         }
 545
 546         return 0;
 547 }
 548
 549 void dlm_complete_thread(struct dlm_ctxt *dlm)
 550 {
 551         if (dlm->dlm_thread_task) {
 552                 mlog(ML_KTHREAD, "Waiting for dlm thread to exit\n");
 553                 kthread_stop(dlm->dlm_thread_task);
 554                 dlm->dlm_thread_task = NULL;
 555         }
 556 }
 557
 558 static int dlm_dirty_list_empty(struct dlm_ctxt *dlm)
 559 {
 560         int empty;
 561
 562         spin_lock(&dlm->spinlock);
 563         empty = list_empty(&dlm->dirty_list);
 564         spin_unlock(&dlm->spinlock);
 565
 566         return empty;
 567 }
 568
 569 static void dlm_flush_asts(struct dlm_ctxt *dlm)
 570 {
 571         int ret;
 572         struct dlm_lock *lock;
 573         struct dlm_lock_resource *res;
 574         u8 hi;
 575
 576         spin_lock(&dlm->ast_lock);
 577         while (!list_empty(&dlm->pending_asts)) {
 578                 lock = list_entry(dlm->pending_asts.next,
 579                                   struct dlm_lock, ast_list);
 580                 /* get an extra ref on lock */
 581                 dlm_lock_get(lock);
 582                 res = lock->lockres;
 583                 mlog(0, "%s: res %.*s, Flush AST for lock %u:%llu, type %d, "
 584                      "node %u\n", dlm->name, res->lockname.len,
 585                      res->lockname.name,
 586                      dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)),
 587                      dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)),
 588                      lock->ml.type, lock->ml.node);
 589
 590                 BUG_ON(!lock->ast_pending);
 591
 592                 /* remove from list (including ref) */
 593                 list_del_init(&lock->ast_list);
 594                 dlm_lock_put(lock);
 595                 spin_unlock(&dlm->ast_lock);
 596
 597                 if (lock->ml.node != dlm->node_num) {
 598                         ret = dlm_do_remote_ast(dlm, res, lock);
 599                         if (ret < 0)
 600                                 mlog_errno(ret);
 601                 } else
 602                         dlm_do_local_ast(dlm, res, lock);
 603
 604                 spin_lock(&dlm->ast_lock);
 605
 606                 /* possible that another ast was queued while
 607                  * we were delivering the last one */
 608                 if (!list_empty(&lock->ast_list)) {
 609                         mlog(0, "%s: res %.*s, AST queued while flushing last "
 610                              "one\n", dlm->name, res->lockname.len,
 611                              res->lockname.name);
 612                 } else
 613                         lock->ast_pending = 0;
 614
 615                 /* drop the extra ref.
 616                  * this may drop it completely. */
 617                 dlm_lock_put(lock);
 618                 dlm_lockres_release_ast(dlm, res);
 619         }
 620
 621         while (!list_empty(&dlm->pending_basts)) {
 622                 lock = list_entry(dlm->pending_basts.next,
 623                                   struct dlm_lock, bast_list);
 624                 /* get an extra ref on lock */
 625                 dlm_lock_get(lock);
 626                 res = lock->lockres;
 627
 628                 BUG_ON(!lock->bast_pending);
 629
 630                 /* get the highest blocked lock, and reset */
 631                 spin_lock(&lock->spinlock);
 632                 BUG_ON(lock->ml.highest_blocked <= LKM_IVMODE);
 633                 hi = lock->ml.highest_blocked;
 634                 lock->ml.highest_blocked = LKM_IVMODE;
 635                 spin_unlock(&lock->spinlock);
 636
 637                 /* remove from list (including ref) */
 638                 list_del_init(&lock->bast_list);
 639                 dlm_lock_put(lock);
 640                 spin_unlock(&dlm->ast_lock);
 641
 642                 mlog(0, "%s: res %.*s, Flush BAST for lock %u:%llu, "
 643                      "blocked %d, node %u\n",
 644                      dlm->name, res->lockname.len, res->lockname.name,
 645                      dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)),
 646                      dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)),
 647                      hi, lock->ml.node);
 648
 649                 if (lock->ml.node != dlm->node_num) {
 650                         ret = dlm_send_proxy_bast(dlm, res, lock, hi);
 651                         if (ret < 0)
 652                                 mlog_errno(ret);
 653                 } else
 654                         dlm_do_local_bast(dlm, res, lock, hi);
 655
 656                 spin_lock(&dlm->ast_lock);
 657
 658                 /* possible that another bast was queued while
 659                  * we were delivering the last one */
 660                 if (!list_empty(&lock->bast_list)) {
 661                         mlog(0, "%s: res %.*s, BAST queued while flushing last "
 662                              "one\n", dlm->name, res->lockname.len,
 663                              res->lockname.name);
 664                 } else
 665                         lock->bast_pending = 0;
 666
 667                 /* drop the extra ref.
 668                  * this may drop it completely. */
 669                 dlm_lock_put(lock);
 670                 dlm_lockres_release_ast(dlm, res);
 671         }
 672         wake_up(&dlm->ast_wq);
 673         spin_unlock(&dlm->ast_lock);
 674 }
 675
 676
 677 #define DLM_THREAD_TIMEOUT_MS (4 * 1000)
 678 #define DLM_THREAD_MAX_DIRTY  100
 679
 680 static int dlm_thread(void *data)
 681 {
 682         struct dlm_lock_resource *res;
 683         struct dlm_ctxt *dlm = data;
 684         unsigned long timeout = msecs_to_jiffies(DLM_THREAD_TIMEOUT_MS);
 685
 686         mlog(0, "dlm thread running for %s...\n", dlm->name);
 687
 688         while (!kthread_should_stop()) {
 689                 int n = DLM_THREAD_MAX_DIRTY;
 690
 691                 /* dlm_shutting_down is very point-in-time, but that
 692                  * doesn't matter as we'll just loop back around if we
 693                  * get false on the leading edge of a state
 694                  * transition. */
 695                 dlm_run_purge_list(dlm, dlm_shutting_down(dlm));
 696
 697                 /* We really don't want to hold dlm->spinlock while
 698                  * calling dlm_shuffle_lists on each lockres that
 699                  * needs to have its queues adjusted and AST/BASTs
 700                  * run.  So let's pull each entry off the dirty_list
 701                  * and drop dlm->spinlock ASAP.  Once off the list,
 702                  * res->spinlock needs to be taken again to protect
 703                  * the queues while calling dlm_shuffle_lists.  */
 704                 spin_lock(&dlm->spinlock);
 705                 while (!list_empty(&dlm->dirty_list)) {
 706                         int delay = 0;
 707                         res = list_entry(dlm->dirty_list.next,
 708                                          struct dlm_lock_resource, dirty);
 709
 710                         /* peel a lockres off, remove it from the list,
 711                          * unset the dirty flag and drop the dlm lock */
 712                         BUG_ON(!res);
 713                         dlm_lockres_get(res);
 714
 715                         spin_lock(&res->spinlock);
 716                         /* We clear the DLM_LOCK_RES_DIRTY state once we shuffle lists below */
 717                         list_del_init(&res->dirty);
 718                         spin_unlock(&res->spinlock);
 719                         spin_unlock(&dlm->spinlock);
 720                         /* Drop dirty_list ref */
 721                         dlm_lockres_put(res);
 722
 723                         /* lockres can be re-dirtied/re-added to the
 724                          * dirty_list in this gap, but that is ok */
 725
 726                         spin_lock(&dlm->ast_lock);
 727                         spin_lock(&res->spinlock);
 728                         if (res->owner != dlm->node_num) {
 729                                 __dlm_print_one_lock_resource(res);
 730                                 mlog(ML_ERROR, "%s: inprog %d, mig %d, reco %d,"
 731                                      " dirty %d\n", dlm->name,
 732                                      !!(res->state & DLM_LOCK_RES_IN_PROGRESS),
 733                                      !!(res->state & DLM_LOCK_RES_MIGRATING),
 734                                      !!(res->state & DLM_LOCK_RES_RECOVERING),
 735                                      !!(res->state & DLM_LOCK_RES_DIRTY));
 736                         }
 737                         BUG_ON(res->owner != dlm->node_num);
 738
 739                         /* it is now ok to move lockreses in these states
 740                          * to the dirty list, assuming that they will only be
 741                          * dirty for a short while. */
 742                         BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
 743                         if (res->state & (DLM_LOCK_RES_IN_PROGRESS |
 744                                           DLM_LOCK_RES_RECOVERING |
 745                                           DLM_LOCK_RES_RECOVERY_WAITING)) {
 746                                 /* move it to the tail and keep going */
 747                                 res->state &= ~DLM_LOCK_RES_DIRTY;
 748                                 spin_unlock(&res->spinlock);
 749                                 spin_unlock(&dlm->ast_lock);
 750                                 mlog(0, "%s: res %.*s, inprogress, delay list "
 751                                      "shuffle, state %d\n", dlm->name,
 752                                      res->lockname.len, res->lockname.name,
 753                                      res->state);
 754                                 delay = 1;
 755                                 goto in_progress;
 756                         }
 757
 758                         /* at this point the lockres is not migrating/
 759                          * recovering/in-progress.  we have the lockres
 760                          * spinlock and do NOT have the dlm lock.
 761                          * safe to reserve/queue asts and run the lists. */
 762
 763                         /* called while holding lockres lock */
 764                         dlm_shuffle_lists(dlm, res);
 765                         res->state &= ~DLM_LOCK_RES_DIRTY;
 766                         spin_unlock(&res->spinlock);
 767                         spin_unlock(&dlm->ast_lock);
 768
 769                         dlm_lockres_calc_usage(dlm, res);
 770
 771 in_progress:
 772
 773                         spin_lock(&dlm->spinlock);
 774                         /* if the lock was in-progress, stick
 775                          * it on the back of the list */
 776                         if (delay) {
 777                                 spin_lock(&res->spinlock);
 778                                 __dlm_dirty_lockres(dlm, res);
 779                                 spin_unlock(&res->spinlock);
 780                         }
 781                         dlm_lockres_put(res);
 782
 783                         /* unlikely, but we may need to give time to
 784                          * other tasks */
 785                         if (!--n) {
 786                                 mlog(0, "%s: Throttling dlm thread\n",
 787                                      dlm->name);
 788                                 break;
 789                         }
 790                 }
 791
 792                 spin_unlock(&dlm->spinlock);
 793                 dlm_flush_asts(dlm);
 794
 795                 /* yield and continue right away if there is more work to do */
 796                 if (!n) {
 797                         cond_resched();
 798                         continue;
 799                 }
 800
 801                 wait_event_interruptible_timeout(dlm->dlm_thread_wq,
 802                                                  !dlm_dirty_list_empty(dlm) ||
 803                                                  kthread_should_stop(),
 804                                                  timeout);
 805         }
 806
 807         mlog(0, "quitting DLM thread\n");
 808         return 0;
 809 }