1 /* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
6 * standalone DLM module
8 * Copyright (C) 2004 Oracle. All rights reserved.
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.
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.
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.
28 #include <linux/module.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>
43 #include "cluster/heartbeat.h"
44 #include "cluster/nodemanager.h"
45 #include "cluster/tcp.h"
48 #include "dlmcommon.h"
49 #include "dlmdomain.h"
51 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_THREAD)
52 #include "cluster/masklog.h"
54 static int dlm_thread(void *data);
55 static void dlm_flush_asts(struct dlm_ctxt *dlm);
57 #define dlm_lock_is_remote(dlm, lock) ((lock)->ml.node != (dlm)->node_num)
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)
63 DECLARE_WAITQUEUE(wait, current);
65 assert_spin_locked(&res->spinlock);
67 add_wait_queue(&res->wq, &wait);
69 set_current_state(TASK_UNINTERRUPTIBLE);
70 if (res->state & flags) {
71 spin_unlock(&res->spinlock);
73 spin_lock(&res->spinlock);
76 remove_wait_queue(&res->wq, &wait);
77 __set_current_state(TASK_RUNNING);
80 int __dlm_lockres_has_locks(struct dlm_lock_resource *res)
82 if (list_empty(&res->granted) &&
83 list_empty(&res->converting) &&
84 list_empty(&res->blocked))
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)
97 assert_spin_locked(&res->spinlock);
99 if (__dlm_lockres_has_locks(res))
102 /* Locks are in the process of being created */
103 if (res->inflight_locks)
106 if (!list_empty(&res->dirty) || res->state & DLM_LOCK_RES_DIRTY)
109 if (res->state & DLM_LOCK_RES_RECOVERING)
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)
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)
127 assert_spin_locked(&dlm->spinlock);
128 assert_spin_locked(&res->spinlock);
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);
135 res->last_used = jiffies;
136 dlm_lockres_get(res);
137 list_add_tail(&res->purge, &dlm->purge_list);
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);
144 list_del_init(&res->purge);
145 dlm_lockres_put(res);
150 void dlm_lockres_calc_usage(struct dlm_ctxt *dlm,
151 struct dlm_lock_resource *res)
153 spin_lock(&dlm->spinlock);
154 spin_lock(&res->spinlock);
156 __dlm_lockres_calc_usage(dlm, res);
158 spin_unlock(&res->spinlock);
159 spin_unlock(&dlm->spinlock);
162 static void dlm_purge_lockres(struct dlm_ctxt *dlm,
163 struct dlm_lock_resource *res)
168 assert_spin_locked(&dlm->spinlock);
169 assert_spin_locked(&res->spinlock);
171 master = (res->owner == dlm->node_num);
173 mlog(0, "%s: Purging res %.*s, master %d\n", dlm->name,
174 res->lockname.len, res->lockname.name, master);
177 res->state |= DLM_LOCK_RES_DROPPING_REF;
178 /* drop spinlock... retake below */
179 spin_unlock(&res->spinlock);
180 spin_unlock(&dlm->spinlock);
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);
187 /* clear our bit from the master's refmap, ignore errors */
188 ret = dlm_drop_lockres_ref(dlm, res);
190 if (!dlm_is_host_down(ret))
193 spin_lock(&dlm->spinlock);
194 spin_lock(&res->spinlock);
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);
205 if (!master && ret != 0) {
206 mlog(0, "%s: deref %.*s in progress or master goes down\n",
207 dlm->name, res->lockname.len, res->lockname.name);
208 spin_unlock(&res->spinlock);
212 if (!__dlm_lockres_unused(res)) {
213 mlog(ML_ERROR, "%s: res %.*s in use after deref\n",
214 dlm->name, res->lockname.len, res->lockname.name);
215 __dlm_print_one_lock_resource(res);
219 __dlm_unhash_lockres(dlm, res);
221 spin_lock(&dlm->track_lock);
222 if (!list_empty(&res->tracking))
223 list_del_init(&res->tracking);
225 mlog(ML_ERROR, "Resource %.*s not on the Tracking list\n",
226 res->lockname.len, res->lockname.name);
227 __dlm_print_one_lock_resource(res);
229 spin_unlock(&dlm->track_lock);
231 /* lockres is not in the hash now. drop the flag and wake up
232 * any processes waiting in dlm_get_lock_resource. */
234 res->state &= ~DLM_LOCK_RES_DROPPING_REF;
235 spin_unlock(&res->spinlock);
238 spin_unlock(&res->spinlock);
241 static void dlm_run_purge_list(struct dlm_ctxt *dlm,
244 unsigned int run_max, unused;
245 unsigned long purge_jiffies;
246 struct dlm_lock_resource *lockres;
248 spin_lock(&dlm->spinlock);
249 run_max = dlm->purge_count;
251 while(run_max && !list_empty(&dlm->purge_list)) {
254 lockres = list_entry(dlm->purge_list.next,
255 struct dlm_lock_resource, purge);
257 spin_lock(&lockres->spinlock);
259 purge_jiffies = lockres->last_used +
260 msecs_to_jiffies(DLM_PURGE_INTERVAL_MS);
262 /* Make sure that we want to be processing this guy at
264 if (!purge_now && time_after(purge_jiffies, jiffies)) {
265 /* Since resources are added to the purge list
266 * in tail order, we can stop at the first
267 * unpurgable resource -- anyone added after
268 * him will have a greater last_used value */
269 spin_unlock(&lockres->spinlock);
273 /* Status of the lockres *might* change so double
274 * check. If the lockres is unused, holding the dlm
275 * spinlock will prevent people from getting and more
277 unused = __dlm_lockres_unused(lockres);
279 (lockres->state & DLM_LOCK_RES_MIGRATING) ||
280 (lockres->inflight_assert_workers != 0)) {
281 mlog(0, "%s: res %.*s is in use or being remastered, "
282 "used %d, state %d, assert master workers %u\n",
283 dlm->name, lockres->lockname.len,
284 lockres->lockname.name,
285 !unused, lockres->state,
286 lockres->inflight_assert_workers);
287 list_move_tail(&lockres->purge, &dlm->purge_list);
288 spin_unlock(&lockres->spinlock);
292 dlm_lockres_get(lockres);
294 dlm_purge_lockres(dlm, lockres);
296 dlm_lockres_put(lockres);
298 /* Avoid adding any scheduling latencies */
299 cond_resched_lock(&dlm->spinlock);
302 spin_unlock(&dlm->spinlock);
305 static void dlm_shuffle_lists(struct dlm_ctxt *dlm,
306 struct dlm_lock_resource *res)
308 struct dlm_lock *lock, *target;
312 * Because this function is called with the lockres
313 * spinlock, and because we know that it is not migrating/
314 * recovering/in-progress, it is fine to reserve asts and
315 * basts right before queueing them all throughout
317 assert_spin_locked(&dlm->ast_lock);
318 assert_spin_locked(&res->spinlock);
319 BUG_ON((res->state & (DLM_LOCK_RES_MIGRATING|
320 DLM_LOCK_RES_RECOVERING|
321 DLM_LOCK_RES_IN_PROGRESS)));
324 if (list_empty(&res->converting))
326 mlog(0, "%s: res %.*s has locks on the convert queue\n", dlm->name,
327 res->lockname.len, res->lockname.name);
329 target = list_entry(res->converting.next, struct dlm_lock, list);
330 if (target->ml.convert_type == LKM_IVMODE) {
331 mlog(ML_ERROR, "%s: res %.*s converting lock to invalid mode\n",
332 dlm->name, res->lockname.len, res->lockname.name);
335 list_for_each_entry(lock, &res->granted, list) {
338 if (!dlm_lock_compatible(lock->ml.type,
339 target->ml.convert_type)) {
341 /* queue the BAST if not already */
342 if (lock->ml.highest_blocked == LKM_IVMODE) {
343 __dlm_lockres_reserve_ast(res);
344 __dlm_queue_bast(dlm, lock);
346 /* update the highest_blocked if needed */
347 if (lock->ml.highest_blocked < target->ml.convert_type)
348 lock->ml.highest_blocked =
349 target->ml.convert_type;
353 list_for_each_entry(lock, &res->converting, list) {
356 if (!dlm_lock_compatible(lock->ml.type,
357 target->ml.convert_type)) {
359 if (lock->ml.highest_blocked == LKM_IVMODE) {
360 __dlm_lockres_reserve_ast(res);
361 __dlm_queue_bast(dlm, lock);
363 if (lock->ml.highest_blocked < target->ml.convert_type)
364 lock->ml.highest_blocked =
365 target->ml.convert_type;
369 /* we can convert the lock */
371 spin_lock(&target->spinlock);
372 BUG_ON(target->ml.highest_blocked != LKM_IVMODE);
374 mlog(0, "%s: res %.*s, AST for Converting lock %u:%llu, type "
375 "%d => %d, node %u\n", dlm->name, res->lockname.len,
377 dlm_get_lock_cookie_node(be64_to_cpu(target->ml.cookie)),
378 dlm_get_lock_cookie_seq(be64_to_cpu(target->ml.cookie)),
380 target->ml.convert_type, target->ml.node);
382 target->ml.type = target->ml.convert_type;
383 target->ml.convert_type = LKM_IVMODE;
384 list_move_tail(&target->list, &res->granted);
386 BUG_ON(!target->lksb);
387 target->lksb->status = DLM_NORMAL;
389 spin_unlock(&target->spinlock);
391 __dlm_lockres_reserve_ast(res);
392 __dlm_queue_ast(dlm, target);
393 /* go back and check for more */
398 if (list_empty(&res->blocked))
400 target = list_entry(res->blocked.next, struct dlm_lock, list);
402 list_for_each_entry(lock, &res->granted, list) {
405 if (!dlm_lock_compatible(lock->ml.type, target->ml.type)) {
407 if (lock->ml.highest_blocked == LKM_IVMODE) {
408 __dlm_lockres_reserve_ast(res);
409 __dlm_queue_bast(dlm, lock);
411 if (lock->ml.highest_blocked < target->ml.type)
412 lock->ml.highest_blocked = target->ml.type;
416 list_for_each_entry(lock, &res->converting, list) {
419 if (!dlm_lock_compatible(lock->ml.type, target->ml.type)) {
421 if (lock->ml.highest_blocked == LKM_IVMODE) {
422 __dlm_lockres_reserve_ast(res);
423 __dlm_queue_bast(dlm, lock);
425 if (lock->ml.highest_blocked < target->ml.type)
426 lock->ml.highest_blocked = target->ml.type;
430 /* we can grant the blocked lock (only
431 * possible if converting list empty) */
433 spin_lock(&target->spinlock);
434 BUG_ON(target->ml.highest_blocked != LKM_IVMODE);
436 mlog(0, "%s: res %.*s, AST for Blocked lock %u:%llu, type %d, "
437 "node %u\n", dlm->name, res->lockname.len,
439 dlm_get_lock_cookie_node(be64_to_cpu(target->ml.cookie)),
440 dlm_get_lock_cookie_seq(be64_to_cpu(target->ml.cookie)),
441 target->ml.type, target->ml.node);
443 /* target->ml.type is already correct */
444 list_move_tail(&target->list, &res->granted);
446 BUG_ON(!target->lksb);
447 target->lksb->status = DLM_NORMAL;
449 spin_unlock(&target->spinlock);
451 __dlm_lockres_reserve_ast(res);
452 __dlm_queue_ast(dlm, target);
453 /* go back and check for more */
461 /* must have NO locks when calling this with res !=NULL * */
462 void dlm_kick_thread(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
465 spin_lock(&dlm->spinlock);
466 spin_lock(&res->spinlock);
467 __dlm_dirty_lockres(dlm, res);
468 spin_unlock(&res->spinlock);
469 spin_unlock(&dlm->spinlock);
471 wake_up(&dlm->dlm_thread_wq);
474 void __dlm_dirty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
476 assert_spin_locked(&dlm->spinlock);
477 assert_spin_locked(&res->spinlock);
479 /* don't shuffle secondary queues */
480 if ((res->owner == dlm->node_num)) {
481 if (res->state & (DLM_LOCK_RES_MIGRATING |
482 DLM_LOCK_RES_BLOCK_DIRTY))
485 if (list_empty(&res->dirty)) {
486 /* ref for dirty_list */
487 dlm_lockres_get(res);
488 list_add_tail(&res->dirty, &dlm->dirty_list);
489 res->state |= DLM_LOCK_RES_DIRTY;
493 mlog(0, "%s: res %.*s\n", dlm->name, res->lockname.len,
498 /* Launch the NM thread for the mounted volume */
499 int dlm_launch_thread(struct dlm_ctxt *dlm)
501 mlog(0, "Starting dlm_thread...\n");
503 dlm->dlm_thread_task = kthread_run(dlm_thread, dlm, "dlm-%s",
505 if (IS_ERR(dlm->dlm_thread_task)) {
506 mlog_errno(PTR_ERR(dlm->dlm_thread_task));
507 dlm->dlm_thread_task = NULL;
514 void dlm_complete_thread(struct dlm_ctxt *dlm)
516 if (dlm->dlm_thread_task) {
517 mlog(ML_KTHREAD, "Waiting for dlm thread to exit\n");
518 kthread_stop(dlm->dlm_thread_task);
519 dlm->dlm_thread_task = NULL;
523 static int dlm_dirty_list_empty(struct dlm_ctxt *dlm)
527 spin_lock(&dlm->spinlock);
528 empty = list_empty(&dlm->dirty_list);
529 spin_unlock(&dlm->spinlock);
534 static void dlm_flush_asts(struct dlm_ctxt *dlm)
537 struct dlm_lock *lock;
538 struct dlm_lock_resource *res;
541 spin_lock(&dlm->ast_lock);
542 while (!list_empty(&dlm->pending_asts)) {
543 lock = list_entry(dlm->pending_asts.next,
544 struct dlm_lock, ast_list);
545 /* get an extra ref on lock */
548 mlog(0, "%s: res %.*s, Flush AST for lock %u:%llu, type %d, "
549 "node %u\n", dlm->name, res->lockname.len,
551 dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)),
552 dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)),
553 lock->ml.type, lock->ml.node);
555 BUG_ON(!lock->ast_pending);
557 /* remove from list (including ref) */
558 list_del_init(&lock->ast_list);
560 spin_unlock(&dlm->ast_lock);
562 if (lock->ml.node != dlm->node_num) {
563 ret = dlm_do_remote_ast(dlm, res, lock);
567 dlm_do_local_ast(dlm, res, lock);
569 spin_lock(&dlm->ast_lock);
571 /* possible that another ast was queued while
572 * we were delivering the last one */
573 if (!list_empty(&lock->ast_list)) {
574 mlog(0, "%s: res %.*s, AST queued while flushing last "
575 "one\n", dlm->name, res->lockname.len,
578 lock->ast_pending = 0;
580 /* drop the extra ref.
581 * this may drop it completely. */
583 dlm_lockres_release_ast(dlm, res);
586 while (!list_empty(&dlm->pending_basts)) {
587 lock = list_entry(dlm->pending_basts.next,
588 struct dlm_lock, bast_list);
589 /* get an extra ref on lock */
593 BUG_ON(!lock->bast_pending);
595 /* get the highest blocked lock, and reset */
596 spin_lock(&lock->spinlock);
597 BUG_ON(lock->ml.highest_blocked <= LKM_IVMODE);
598 hi = lock->ml.highest_blocked;
599 lock->ml.highest_blocked = LKM_IVMODE;
600 spin_unlock(&lock->spinlock);
602 /* remove from list (including ref) */
603 list_del_init(&lock->bast_list);
605 spin_unlock(&dlm->ast_lock);
607 mlog(0, "%s: res %.*s, Flush BAST for lock %u:%llu, "
608 "blocked %d, node %u\n",
609 dlm->name, res->lockname.len, res->lockname.name,
610 dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)),
611 dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)),
614 if (lock->ml.node != dlm->node_num) {
615 ret = dlm_send_proxy_bast(dlm, res, lock, hi);
619 dlm_do_local_bast(dlm, res, lock, hi);
621 spin_lock(&dlm->ast_lock);
623 /* possible that another bast was queued while
624 * we were delivering the last one */
625 if (!list_empty(&lock->bast_list)) {
626 mlog(0, "%s: res %.*s, BAST queued while flushing last "
627 "one\n", dlm->name, res->lockname.len,
630 lock->bast_pending = 0;
632 /* drop the extra ref.
633 * this may drop it completely. */
635 dlm_lockres_release_ast(dlm, res);
637 wake_up(&dlm->ast_wq);
638 spin_unlock(&dlm->ast_lock);
642 #define DLM_THREAD_TIMEOUT_MS (4 * 1000)
643 #define DLM_THREAD_MAX_DIRTY 100
644 #define DLM_THREAD_MAX_ASTS 10
646 static int dlm_thread(void *data)
648 struct dlm_lock_resource *res;
649 struct dlm_ctxt *dlm = data;
650 unsigned long timeout = msecs_to_jiffies(DLM_THREAD_TIMEOUT_MS);
652 mlog(0, "dlm thread running for %s...\n", dlm->name);
654 while (!kthread_should_stop()) {
655 int n = DLM_THREAD_MAX_DIRTY;
657 /* dlm_shutting_down is very point-in-time, but that
658 * doesn't matter as we'll just loop back around if we
659 * get false on the leading edge of a state
661 dlm_run_purge_list(dlm, dlm_shutting_down(dlm));
663 /* We really don't want to hold dlm->spinlock while
664 * calling dlm_shuffle_lists on each lockres that
665 * needs to have its queues adjusted and AST/BASTs
666 * run. So let's pull each entry off the dirty_list
667 * and drop dlm->spinlock ASAP. Once off the list,
668 * res->spinlock needs to be taken again to protect
669 * the queues while calling dlm_shuffle_lists. */
670 spin_lock(&dlm->spinlock);
671 while (!list_empty(&dlm->dirty_list)) {
673 res = list_entry(dlm->dirty_list.next,
674 struct dlm_lock_resource, dirty);
676 /* peel a lockres off, remove it from the list,
677 * unset the dirty flag and drop the dlm lock */
679 dlm_lockres_get(res);
681 spin_lock(&res->spinlock);
682 /* We clear the DLM_LOCK_RES_DIRTY state once we shuffle lists below */
683 list_del_init(&res->dirty);
684 spin_unlock(&res->spinlock);
685 spin_unlock(&dlm->spinlock);
686 /* Drop dirty_list ref */
687 dlm_lockres_put(res);
689 /* lockres can be re-dirtied/re-added to the
690 * dirty_list in this gap, but that is ok */
692 spin_lock(&dlm->ast_lock);
693 spin_lock(&res->spinlock);
694 if (res->owner != dlm->node_num) {
695 __dlm_print_one_lock_resource(res);
696 mlog(ML_ERROR, "%s: inprog %d, mig %d, reco %d,"
697 " dirty %d\n", dlm->name,
698 !!(res->state & DLM_LOCK_RES_IN_PROGRESS),
699 !!(res->state & DLM_LOCK_RES_MIGRATING),
700 !!(res->state & DLM_LOCK_RES_RECOVERING),
701 !!(res->state & DLM_LOCK_RES_DIRTY));
703 BUG_ON(res->owner != dlm->node_num);
705 /* it is now ok to move lockreses in these states
706 * to the dirty list, assuming that they will only be
707 * dirty for a short while. */
708 BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
709 if (res->state & (DLM_LOCK_RES_IN_PROGRESS |
710 DLM_LOCK_RES_RECOVERING)) {
711 /* move it to the tail and keep going */
712 res->state &= ~DLM_LOCK_RES_DIRTY;
713 spin_unlock(&res->spinlock);
714 spin_unlock(&dlm->ast_lock);
715 mlog(0, "%s: res %.*s, inprogress, delay list "
716 "shuffle, state %d\n", dlm->name,
717 res->lockname.len, res->lockname.name,
723 /* at this point the lockres is not migrating/
724 * recovering/in-progress. we have the lockres
725 * spinlock and do NOT have the dlm lock.
726 * safe to reserve/queue asts and run the lists. */
728 /* called while holding lockres lock */
729 dlm_shuffle_lists(dlm, res);
730 res->state &= ~DLM_LOCK_RES_DIRTY;
731 spin_unlock(&res->spinlock);
732 spin_unlock(&dlm->ast_lock);
734 dlm_lockres_calc_usage(dlm, res);
738 spin_lock(&dlm->spinlock);
739 /* if the lock was in-progress, stick
740 * it on the back of the list */
742 spin_lock(&res->spinlock);
743 __dlm_dirty_lockres(dlm, res);
744 spin_unlock(&res->spinlock);
746 dlm_lockres_put(res);
748 /* unlikely, but we may need to give time to
751 mlog(0, "%s: Throttling dlm thread\n",
757 spin_unlock(&dlm->spinlock);
760 /* yield and continue right away if there is more work to do */
766 wait_event_interruptible_timeout(dlm->dlm_thread_wq,
767 !dlm_dirty_list_empty(dlm) ||
768 kthread_should_stop(),
772 mlog(0, "quitting DLM thread\n");