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/slab.h>
32 #include <linux/highmem.h>
33 #include <linux/utsname.h>
34 #include <linux/init.h>
35 #include <linux/sysctl.h>
36 #include <linux/random.h>
37 #include <linux/blkdev.h>
38 #include <linux/socket.h>
39 #include <linux/inet.h>
40 #include <linux/timer.h>
41 #include <linux/kthread.h>
42 #include <linux/delay.h>
45 #include "cluster/heartbeat.h"
46 #include "cluster/nodemanager.h"
47 #include "cluster/tcp.h"
50 #include "dlmcommon.h"
51 #include "dlmdomain.h"
53 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_THREAD)
54 #include "cluster/masklog.h"
56 static int dlm_thread(void *data);
57 static void dlm_flush_asts(struct dlm_ctxt *dlm);
59 #define dlm_lock_is_remote(dlm, lock) ((lock)->ml.node != (dlm)->node_num)
61 /* will exit holding res->spinlock, but may drop in function */
62 /* waits until flags are cleared on res->state */
63 void __dlm_wait_on_lockres_flags(struct dlm_lock_resource *res, int flags)
65 DECLARE_WAITQUEUE(wait, current);
67 assert_spin_locked(&res->spinlock);
69 add_wait_queue(&res->wq, &wait);
71 set_current_state(TASK_UNINTERRUPTIBLE);
72 if (res->state & flags) {
73 spin_unlock(&res->spinlock);
75 spin_lock(&res->spinlock);
78 remove_wait_queue(&res->wq, &wait);
79 current->state = TASK_RUNNING;
82 int __dlm_lockres_has_locks(struct dlm_lock_resource *res)
84 if (list_empty(&res->granted) &&
85 list_empty(&res->converting) &&
86 list_empty(&res->blocked))
91 /* "unused": the lockres has no locks, is not on the dirty list,
92 * has no inflight locks (in the gap between mastery and acquiring
93 * the first lock), and has no bits in its refmap.
94 * truly ready to be freed. */
95 int __dlm_lockres_unused(struct dlm_lock_resource *res)
97 if (!__dlm_lockres_has_locks(res) &&
98 list_empty(&res->dirty)) {
99 /* try not to scan the bitmap unless the first two
100 * conditions are already true */
101 int bit = find_next_bit(res->refmap, O2NM_MAX_NODES, 0);
102 if (bit >= O2NM_MAX_NODES) {
103 /* since the bit for dlm->node_num is not
104 * set, inflight_locks better be zero */
105 BUG_ON(res->inflight_locks != 0);
113 /* Call whenever you may have added or deleted something from one of
114 * the lockres queue's. This will figure out whether it belongs on the
115 * unused list or not and does the appropriate thing. */
116 void __dlm_lockres_calc_usage(struct dlm_ctxt *dlm,
117 struct dlm_lock_resource *res)
119 mlog_entry("%.*s\n", res->lockname.len, res->lockname.name);
121 assert_spin_locked(&dlm->spinlock);
122 assert_spin_locked(&res->spinlock);
124 if (__dlm_lockres_unused(res)){
125 if (list_empty(&res->purge)) {
126 mlog(0, "putting lockres %.*s:%p onto purge list\n",
127 res->lockname.len, res->lockname.name, res);
129 res->last_used = jiffies;
130 dlm_lockres_get(res);
131 list_add_tail(&res->purge, &dlm->purge_list);
134 } else if (!list_empty(&res->purge)) {
135 mlog(0, "removing lockres %.*s:%p from purge list, owner=%u\n",
136 res->lockname.len, res->lockname.name, res, res->owner);
138 list_del_init(&res->purge);
139 dlm_lockres_put(res);
144 void dlm_lockres_calc_usage(struct dlm_ctxt *dlm,
145 struct dlm_lock_resource *res)
147 mlog_entry("%.*s\n", res->lockname.len, res->lockname.name);
148 spin_lock(&dlm->spinlock);
149 spin_lock(&res->spinlock);
151 __dlm_lockres_calc_usage(dlm, res);
153 spin_unlock(&res->spinlock);
154 spin_unlock(&dlm->spinlock);
157 int dlm_purge_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
162 spin_lock(&res->spinlock);
163 if (!__dlm_lockres_unused(res)) {
164 spin_unlock(&res->spinlock);
165 mlog(0, "%s:%.*s: tried to purge but not unused\n",
166 dlm->name, res->lockname.len, res->lockname.name);
169 master = (res->owner == dlm->node_num);
171 res->state |= DLM_LOCK_RES_DROPPING_REF;
172 spin_unlock(&res->spinlock);
174 mlog(0, "purging lockres %.*s, master = %d\n", res->lockname.len,
175 res->lockname.name, master);
178 /* drop spinlock to do messaging, retake below */
179 spin_unlock(&dlm->spinlock);
180 /* clear our bit from the master's refmap, ignore errors */
181 ret = dlm_drop_lockres_ref(dlm, res);
184 if (!dlm_is_host_down(ret))
187 mlog(0, "%s:%.*s: dlm_deref_lockres returned %d\n",
188 dlm->name, res->lockname.len, res->lockname.name, ret);
189 spin_lock(&dlm->spinlock);
192 if (!list_empty(&res->purge)) {
193 mlog(0, "removing lockres %.*s:%p from purgelist, "
194 "master = %d\n", res->lockname.len, res->lockname.name,
196 list_del_init(&res->purge);
197 dlm_lockres_put(res);
200 __dlm_unhash_lockres(res);
202 /* lockres is not in the hash now. drop the flag and wake up
203 * any processes waiting in dlm_get_lock_resource. */
205 spin_lock(&res->spinlock);
206 res->state &= ~DLM_LOCK_RES_DROPPING_REF;
207 spin_unlock(&res->spinlock);
213 static void dlm_run_purge_list(struct dlm_ctxt *dlm,
216 unsigned int run_max, unused;
217 unsigned long purge_jiffies;
218 struct dlm_lock_resource *lockres;
220 spin_lock(&dlm->spinlock);
221 run_max = dlm->purge_count;
223 while(run_max && !list_empty(&dlm->purge_list)) {
226 lockres = list_entry(dlm->purge_list.next,
227 struct dlm_lock_resource, purge);
229 /* Status of the lockres *might* change so double
230 * check. If the lockres is unused, holding the dlm
231 * spinlock will prevent people from getting and more
232 * refs on it -- there's no need to keep the lockres
234 spin_lock(&lockres->spinlock);
235 unused = __dlm_lockres_unused(lockres);
236 spin_unlock(&lockres->spinlock);
241 purge_jiffies = lockres->last_used +
242 msecs_to_jiffies(DLM_PURGE_INTERVAL_MS);
244 /* Make sure that we want to be processing this guy at
246 if (!purge_now && time_after(purge_jiffies, jiffies)) {
247 /* Since resources are added to the purge list
248 * in tail order, we can stop at the first
249 * unpurgable resource -- anyone added after
250 * him will have a greater last_used value */
254 mlog(0, "removing lockres %.*s:%p from purgelist\n",
255 lockres->lockname.len, lockres->lockname.name, lockres);
256 list_del_init(&lockres->purge);
257 dlm_lockres_put(lockres);
260 /* This may drop and reacquire the dlm spinlock if it
261 * has to do migration. */
262 mlog(0, "calling dlm_purge_lockres!\n");
263 if (dlm_purge_lockres(dlm, lockres))
265 mlog(0, "DONE calling dlm_purge_lockres!\n");
267 /* Avoid adding any scheduling latencies */
268 cond_resched_lock(&dlm->spinlock);
271 spin_unlock(&dlm->spinlock);
274 static void dlm_shuffle_lists(struct dlm_ctxt *dlm,
275 struct dlm_lock_resource *res)
277 struct dlm_lock *lock, *target;
278 struct list_head *iter;
279 struct list_head *head;
282 //mlog(0, "res->lockname.len=%d\n", res->lockname.len);
283 //mlog(0, "res->lockname.name=%p\n", res->lockname.name);
284 //mlog(0, "shuffle res %.*s\n", res->lockname.len,
285 // res->lockname.name);
287 /* because this function is called with the lockres
288 * spinlock, and because we know that it is not migrating/
289 * recovering/in-progress, it is fine to reserve asts and
290 * basts right before queueing them all throughout */
291 assert_spin_locked(&res->spinlock);
292 BUG_ON((res->state & (DLM_LOCK_RES_MIGRATING|
293 DLM_LOCK_RES_RECOVERING|
294 DLM_LOCK_RES_IN_PROGRESS)));
297 if (list_empty(&res->converting))
299 mlog(0, "res %.*s has locks on a convert queue\n", res->lockname.len,
302 target = list_entry(res->converting.next, struct dlm_lock, list);
303 if (target->ml.convert_type == LKM_IVMODE) {
304 mlog(ML_ERROR, "%.*s: converting a lock with no "
305 "convert_type!\n", res->lockname.len, res->lockname.name);
308 head = &res->granted;
309 list_for_each(iter, head) {
310 lock = list_entry(iter, struct dlm_lock, list);
313 if (!dlm_lock_compatible(lock->ml.type,
314 target->ml.convert_type)) {
316 /* queue the BAST if not already */
317 if (lock->ml.highest_blocked == LKM_IVMODE) {
318 __dlm_lockres_reserve_ast(res);
319 dlm_queue_bast(dlm, lock);
321 /* update the highest_blocked if needed */
322 if (lock->ml.highest_blocked < target->ml.convert_type)
323 lock->ml.highest_blocked =
324 target->ml.convert_type;
327 head = &res->converting;
328 list_for_each(iter, head) {
329 lock = list_entry(iter, struct dlm_lock, list);
332 if (!dlm_lock_compatible(lock->ml.type,
333 target->ml.convert_type)) {
335 if (lock->ml.highest_blocked == LKM_IVMODE) {
336 __dlm_lockres_reserve_ast(res);
337 dlm_queue_bast(dlm, lock);
339 if (lock->ml.highest_blocked < target->ml.convert_type)
340 lock->ml.highest_blocked =
341 target->ml.convert_type;
345 /* we can convert the lock */
347 spin_lock(&target->spinlock);
348 BUG_ON(target->ml.highest_blocked != LKM_IVMODE);
350 mlog(0, "calling ast for converting lock: %.*s, have: %d, "
351 "granting: %d, node: %u\n", res->lockname.len,
352 res->lockname.name, target->ml.type,
353 target->ml.convert_type, target->ml.node);
355 target->ml.type = target->ml.convert_type;
356 target->ml.convert_type = LKM_IVMODE;
357 list_move_tail(&target->list, &res->granted);
359 BUG_ON(!target->lksb);
360 target->lksb->status = DLM_NORMAL;
362 spin_unlock(&target->spinlock);
364 __dlm_lockres_reserve_ast(res);
365 dlm_queue_ast(dlm, target);
366 /* go back and check for more */
371 if (list_empty(&res->blocked))
373 target = list_entry(res->blocked.next, struct dlm_lock, list);
375 head = &res->granted;
376 list_for_each(iter, head) {
377 lock = list_entry(iter, struct dlm_lock, list);
380 if (!dlm_lock_compatible(lock->ml.type, target->ml.type)) {
382 if (lock->ml.highest_blocked == LKM_IVMODE) {
383 __dlm_lockres_reserve_ast(res);
384 dlm_queue_bast(dlm, lock);
386 if (lock->ml.highest_blocked < target->ml.type)
387 lock->ml.highest_blocked = target->ml.type;
391 head = &res->converting;
392 list_for_each(iter, head) {
393 lock = list_entry(iter, struct dlm_lock, list);
396 if (!dlm_lock_compatible(lock->ml.type, target->ml.type)) {
398 if (lock->ml.highest_blocked == LKM_IVMODE) {
399 __dlm_lockres_reserve_ast(res);
400 dlm_queue_bast(dlm, lock);
402 if (lock->ml.highest_blocked < target->ml.type)
403 lock->ml.highest_blocked = target->ml.type;
407 /* we can grant the blocked lock (only
408 * possible if converting list empty) */
410 spin_lock(&target->spinlock);
411 BUG_ON(target->ml.highest_blocked != LKM_IVMODE);
413 mlog(0, "calling ast for blocked lock: %.*s, granting: %d, "
414 "node: %u\n", res->lockname.len, res->lockname.name,
415 target->ml.type, target->ml.node);
417 // target->ml.type is already correct
418 list_move_tail(&target->list, &res->granted);
420 BUG_ON(!target->lksb);
421 target->lksb->status = DLM_NORMAL;
423 spin_unlock(&target->spinlock);
425 __dlm_lockres_reserve_ast(res);
426 dlm_queue_ast(dlm, target);
427 /* go back and check for more */
435 /* must have NO locks when calling this with res !=NULL * */
436 void dlm_kick_thread(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
438 mlog_entry("dlm=%p, res=%p\n", dlm, res);
440 spin_lock(&dlm->spinlock);
441 spin_lock(&res->spinlock);
442 __dlm_dirty_lockres(dlm, res);
443 spin_unlock(&res->spinlock);
444 spin_unlock(&dlm->spinlock);
446 wake_up(&dlm->dlm_thread_wq);
449 void __dlm_dirty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
451 mlog_entry("dlm=%p, res=%p\n", dlm, res);
453 assert_spin_locked(&dlm->spinlock);
454 assert_spin_locked(&res->spinlock);
456 /* don't shuffle secondary queues */
457 if ((res->owner == dlm->node_num) &&
458 !(res->state & DLM_LOCK_RES_DIRTY)) {
459 /* ref for dirty_list */
460 dlm_lockres_get(res);
461 list_add_tail(&res->dirty, &dlm->dirty_list);
462 res->state |= DLM_LOCK_RES_DIRTY;
467 /* Launch the NM thread for the mounted volume */
468 int dlm_launch_thread(struct dlm_ctxt *dlm)
470 mlog(0, "starting dlm thread...\n");
472 dlm->dlm_thread_task = kthread_run(dlm_thread, dlm, "dlm_thread");
473 if (IS_ERR(dlm->dlm_thread_task)) {
474 mlog_errno(PTR_ERR(dlm->dlm_thread_task));
475 dlm->dlm_thread_task = NULL;
482 void dlm_complete_thread(struct dlm_ctxt *dlm)
484 if (dlm->dlm_thread_task) {
485 mlog(ML_KTHREAD, "waiting for dlm thread to exit\n");
486 kthread_stop(dlm->dlm_thread_task);
487 dlm->dlm_thread_task = NULL;
491 static int dlm_dirty_list_empty(struct dlm_ctxt *dlm)
495 spin_lock(&dlm->spinlock);
496 empty = list_empty(&dlm->dirty_list);
497 spin_unlock(&dlm->spinlock);
502 static void dlm_flush_asts(struct dlm_ctxt *dlm)
505 struct dlm_lock *lock;
506 struct dlm_lock_resource *res;
509 spin_lock(&dlm->ast_lock);
510 while (!list_empty(&dlm->pending_asts)) {
511 lock = list_entry(dlm->pending_asts.next,
512 struct dlm_lock, ast_list);
513 /* get an extra ref on lock */
516 mlog(0, "delivering an ast for this lockres\n");
518 BUG_ON(!lock->ast_pending);
520 /* remove from list (including ref) */
521 list_del_init(&lock->ast_list);
523 spin_unlock(&dlm->ast_lock);
525 if (lock->ml.node != dlm->node_num) {
526 ret = dlm_do_remote_ast(dlm, res, lock);
530 dlm_do_local_ast(dlm, res, lock);
532 spin_lock(&dlm->ast_lock);
534 /* possible that another ast was queued while
535 * we were delivering the last one */
536 if (!list_empty(&lock->ast_list)) {
537 mlog(0, "aha another ast got queued while "
538 "we were finishing the last one. will "
539 "keep the ast_pending flag set.\n");
541 lock->ast_pending = 0;
543 /* drop the extra ref.
544 * this may drop it completely. */
546 dlm_lockres_release_ast(dlm, res);
549 while (!list_empty(&dlm->pending_basts)) {
550 lock = list_entry(dlm->pending_basts.next,
551 struct dlm_lock, bast_list);
552 /* get an extra ref on lock */
556 BUG_ON(!lock->bast_pending);
558 /* get the highest blocked lock, and reset */
559 spin_lock(&lock->spinlock);
560 BUG_ON(lock->ml.highest_blocked <= LKM_IVMODE);
561 hi = lock->ml.highest_blocked;
562 lock->ml.highest_blocked = LKM_IVMODE;
563 spin_unlock(&lock->spinlock);
565 /* remove from list (including ref) */
566 list_del_init(&lock->bast_list);
568 spin_unlock(&dlm->ast_lock);
570 mlog(0, "delivering a bast for this lockres "
571 "(blocked = %d\n", hi);
573 if (lock->ml.node != dlm->node_num) {
574 ret = dlm_send_proxy_bast(dlm, res, lock, hi);
578 dlm_do_local_bast(dlm, res, lock, hi);
580 spin_lock(&dlm->ast_lock);
582 /* possible that another bast was queued while
583 * we were delivering the last one */
584 if (!list_empty(&lock->bast_list)) {
585 mlog(0, "aha another bast got queued while "
586 "we were finishing the last one. will "
587 "keep the bast_pending flag set.\n");
589 lock->bast_pending = 0;
591 /* drop the extra ref.
592 * this may drop it completely. */
594 dlm_lockres_release_ast(dlm, res);
596 wake_up(&dlm->ast_wq);
597 spin_unlock(&dlm->ast_lock);
601 #define DLM_THREAD_TIMEOUT_MS (4 * 1000)
602 #define DLM_THREAD_MAX_DIRTY 100
603 #define DLM_THREAD_MAX_ASTS 10
605 static int dlm_thread(void *data)
607 struct dlm_lock_resource *res;
608 struct dlm_ctxt *dlm = data;
609 unsigned long timeout = msecs_to_jiffies(DLM_THREAD_TIMEOUT_MS);
611 mlog(0, "dlm thread running for %s...\n", dlm->name);
613 while (!kthread_should_stop()) {
614 int n = DLM_THREAD_MAX_DIRTY;
616 /* dlm_shutting_down is very point-in-time, but that
617 * doesn't matter as we'll just loop back around if we
618 * get false on the leading edge of a state
620 dlm_run_purge_list(dlm, dlm_shutting_down(dlm));
622 /* We really don't want to hold dlm->spinlock while
623 * calling dlm_shuffle_lists on each lockres that
624 * needs to have its queues adjusted and AST/BASTs
625 * run. So let's pull each entry off the dirty_list
626 * and drop dlm->spinlock ASAP. Once off the list,
627 * res->spinlock needs to be taken again to protect
628 * the queues while calling dlm_shuffle_lists. */
629 spin_lock(&dlm->spinlock);
630 while (!list_empty(&dlm->dirty_list)) {
632 res = list_entry(dlm->dirty_list.next,
633 struct dlm_lock_resource, dirty);
635 /* peel a lockres off, remove it from the list,
636 * unset the dirty flag and drop the dlm lock */
638 dlm_lockres_get(res);
640 spin_lock(&res->spinlock);
641 res->state &= ~DLM_LOCK_RES_DIRTY;
642 list_del_init(&res->dirty);
643 spin_unlock(&res->spinlock);
644 spin_unlock(&dlm->spinlock);
645 /* Drop dirty_list ref */
646 dlm_lockres_put(res);
648 /* lockres can be re-dirtied/re-added to the
649 * dirty_list in this gap, but that is ok */
651 spin_lock(&res->spinlock);
652 if (res->owner != dlm->node_num) {
653 __dlm_print_one_lock_resource(res);
654 mlog(ML_ERROR, "inprog:%s, mig:%s, reco:%s, dirty:%s\n",
655 res->state & DLM_LOCK_RES_IN_PROGRESS ? "yes" : "no",
656 res->state & DLM_LOCK_RES_MIGRATING ? "yes" : "no",
657 res->state & DLM_LOCK_RES_RECOVERING ? "yes" : "no",
658 res->state & DLM_LOCK_RES_DIRTY ? "yes" : "no");
660 BUG_ON(res->owner != dlm->node_num);
662 /* it is now ok to move lockreses in these states
663 * to the dirty list, assuming that they will only be
664 * dirty for a short while. */
665 if (res->state & (DLM_LOCK_RES_IN_PROGRESS |
666 DLM_LOCK_RES_MIGRATING |
667 DLM_LOCK_RES_RECOVERING)) {
668 /* move it to the tail and keep going */
669 spin_unlock(&res->spinlock);
670 mlog(0, "delaying list shuffling for in-"
671 "progress lockres %.*s, state=%d\n",
672 res->lockname.len, res->lockname.name,
678 /* at this point the lockres is not migrating/
679 * recovering/in-progress. we have the lockres
680 * spinlock and do NOT have the dlm lock.
681 * safe to reserve/queue asts and run the lists. */
683 mlog(0, "calling dlm_shuffle_lists with dlm=%s, "
684 "res=%.*s\n", dlm->name,
685 res->lockname.len, res->lockname.name);
687 /* called while holding lockres lock */
688 dlm_shuffle_lists(dlm, res);
689 spin_unlock(&res->spinlock);
691 dlm_lockres_calc_usage(dlm, res);
695 spin_lock(&dlm->spinlock);
696 /* if the lock was in-progress, stick
697 * it on the back of the list */
699 /* ref for dirty_list */
700 dlm_lockres_get(res);
701 spin_lock(&res->spinlock);
702 list_add_tail(&res->dirty, &dlm->dirty_list);
703 res->state |= DLM_LOCK_RES_DIRTY;
704 spin_unlock(&res->spinlock);
706 dlm_lockres_put(res);
708 /* unlikely, but we may need to give time to
711 mlog(0, "throttling dlm_thread\n");
716 spin_unlock(&dlm->spinlock);
719 /* yield and continue right away if there is more work to do */
725 wait_event_interruptible_timeout(dlm->dlm_thread_wq,
726 !dlm_dirty_list_empty(dlm) ||
727 kthread_should_stop(),
731 mlog(0, "quitting DLM thread\n");