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 __set_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) && !(res->state & DLM_LOCK_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 static int dlm_purge_lockres(struct dlm_ctxt *dlm,
158 struct dlm_lock_resource *res)
163 spin_lock(&res->spinlock);
164 if (!__dlm_lockres_unused(res)) {
165 spin_unlock(&res->spinlock);
166 mlog(0, "%s:%.*s: tried to purge but not unused\n",
167 dlm->name, res->lockname.len, res->lockname.name);
170 master = (res->owner == dlm->node_num);
172 res->state |= DLM_LOCK_RES_DROPPING_REF;
173 spin_unlock(&res->spinlock);
175 mlog(0, "purging lockres %.*s, master = %d\n", res->lockname.len,
176 res->lockname.name, master);
179 spin_lock(&res->spinlock);
180 /* This ensures that clear refmap is sent after the set */
181 __dlm_wait_on_lockres_flags(res, DLM_LOCK_RES_SETREF_INPROG);
182 spin_unlock(&res->spinlock);
183 /* drop spinlock to do messaging, retake below */
184 spin_unlock(&dlm->spinlock);
185 /* clear our bit from the master's refmap, ignore errors */
186 ret = dlm_drop_lockres_ref(dlm, res);
189 if (!dlm_is_host_down(ret))
192 mlog(0, "%s:%.*s: dlm_deref_lockres returned %d\n",
193 dlm->name, res->lockname.len, res->lockname.name, ret);
194 spin_lock(&dlm->spinlock);
197 if (!list_empty(&res->purge)) {
198 mlog(0, "removing lockres %.*s:%p from purgelist, "
199 "master = %d\n", res->lockname.len, res->lockname.name,
201 list_del_init(&res->purge);
202 dlm_lockres_put(res);
205 __dlm_unhash_lockres(res);
207 /* lockres is not in the hash now. drop the flag and wake up
208 * any processes waiting in dlm_get_lock_resource. */
210 spin_lock(&res->spinlock);
211 res->state &= ~DLM_LOCK_RES_DROPPING_REF;
212 spin_unlock(&res->spinlock);
218 static void dlm_run_purge_list(struct dlm_ctxt *dlm,
221 unsigned int run_max, unused;
222 unsigned long purge_jiffies;
223 struct dlm_lock_resource *lockres;
225 spin_lock(&dlm->spinlock);
226 run_max = dlm->purge_count;
228 while(run_max && !list_empty(&dlm->purge_list)) {
231 lockres = list_entry(dlm->purge_list.next,
232 struct dlm_lock_resource, purge);
234 /* Status of the lockres *might* change so double
235 * check. If the lockres is unused, holding the dlm
236 * spinlock will prevent people from getting and more
237 * refs on it -- there's no need to keep the lockres
239 spin_lock(&lockres->spinlock);
240 unused = __dlm_lockres_unused(lockres);
241 spin_unlock(&lockres->spinlock);
246 purge_jiffies = lockres->last_used +
247 msecs_to_jiffies(DLM_PURGE_INTERVAL_MS);
249 /* Make sure that we want to be processing this guy at
251 if (!purge_now && time_after(purge_jiffies, jiffies)) {
252 /* Since resources are added to the purge list
253 * in tail order, we can stop at the first
254 * unpurgable resource -- anyone added after
255 * him will have a greater last_used value */
259 dlm_lockres_get(lockres);
261 /* This may drop and reacquire the dlm spinlock if it
262 * has to do migration. */
263 if (dlm_purge_lockres(dlm, lockres))
266 dlm_lockres_put(lockres);
268 /* Avoid adding any scheduling latencies */
269 cond_resched_lock(&dlm->spinlock);
272 spin_unlock(&dlm->spinlock);
275 static void dlm_shuffle_lists(struct dlm_ctxt *dlm,
276 struct dlm_lock_resource *res)
278 struct dlm_lock *lock, *target;
279 struct list_head *iter;
280 struct list_head *head;
283 //mlog(0, "res->lockname.len=%d\n", res->lockname.len);
284 //mlog(0, "res->lockname.name=%p\n", res->lockname.name);
285 //mlog(0, "shuffle res %.*s\n", res->lockname.len,
286 // res->lockname.name);
288 /* because this function is called with the lockres
289 * spinlock, and because we know that it is not migrating/
290 * recovering/in-progress, it is fine to reserve asts and
291 * basts right before queueing them all throughout */
292 assert_spin_locked(&res->spinlock);
293 BUG_ON((res->state & (DLM_LOCK_RES_MIGRATING|
294 DLM_LOCK_RES_RECOVERING|
295 DLM_LOCK_RES_IN_PROGRESS)));
298 if (list_empty(&res->converting))
300 mlog(0, "res %.*s has locks on a convert queue\n", res->lockname.len,
303 target = list_entry(res->converting.next, struct dlm_lock, list);
304 if (target->ml.convert_type == LKM_IVMODE) {
305 mlog(ML_ERROR, "%.*s: converting a lock with no "
306 "convert_type!\n", res->lockname.len, res->lockname.name);
309 head = &res->granted;
310 list_for_each(iter, head) {
311 lock = list_entry(iter, struct dlm_lock, list);
314 if (!dlm_lock_compatible(lock->ml.type,
315 target->ml.convert_type)) {
317 /* queue the BAST if not already */
318 if (lock->ml.highest_blocked == LKM_IVMODE) {
319 __dlm_lockres_reserve_ast(res);
320 dlm_queue_bast(dlm, lock);
322 /* update the highest_blocked if needed */
323 if (lock->ml.highest_blocked < target->ml.convert_type)
324 lock->ml.highest_blocked =
325 target->ml.convert_type;
328 head = &res->converting;
329 list_for_each(iter, head) {
330 lock = list_entry(iter, struct dlm_lock, list);
333 if (!dlm_lock_compatible(lock->ml.type,
334 target->ml.convert_type)) {
336 if (lock->ml.highest_blocked == LKM_IVMODE) {
337 __dlm_lockres_reserve_ast(res);
338 dlm_queue_bast(dlm, lock);
340 if (lock->ml.highest_blocked < target->ml.convert_type)
341 lock->ml.highest_blocked =
342 target->ml.convert_type;
346 /* we can convert the lock */
348 spin_lock(&target->spinlock);
349 BUG_ON(target->ml.highest_blocked != LKM_IVMODE);
351 mlog(0, "calling ast for converting lock: %.*s, have: %d, "
352 "granting: %d, node: %u\n", res->lockname.len,
353 res->lockname.name, target->ml.type,
354 target->ml.convert_type, target->ml.node);
356 target->ml.type = target->ml.convert_type;
357 target->ml.convert_type = LKM_IVMODE;
358 list_move_tail(&target->list, &res->granted);
360 BUG_ON(!target->lksb);
361 target->lksb->status = DLM_NORMAL;
363 spin_unlock(&target->spinlock);
365 __dlm_lockres_reserve_ast(res);
366 dlm_queue_ast(dlm, target);
367 /* go back and check for more */
372 if (list_empty(&res->blocked))
374 target = list_entry(res->blocked.next, struct dlm_lock, list);
376 head = &res->granted;
377 list_for_each(iter, head) {
378 lock = list_entry(iter, struct dlm_lock, list);
381 if (!dlm_lock_compatible(lock->ml.type, target->ml.type)) {
383 if (lock->ml.highest_blocked == LKM_IVMODE) {
384 __dlm_lockres_reserve_ast(res);
385 dlm_queue_bast(dlm, lock);
387 if (lock->ml.highest_blocked < target->ml.type)
388 lock->ml.highest_blocked = target->ml.type;
392 head = &res->converting;
393 list_for_each(iter, head) {
394 lock = list_entry(iter, struct dlm_lock, list);
397 if (!dlm_lock_compatible(lock->ml.type, target->ml.type)) {
399 if (lock->ml.highest_blocked == LKM_IVMODE) {
400 __dlm_lockres_reserve_ast(res);
401 dlm_queue_bast(dlm, lock);
403 if (lock->ml.highest_blocked < target->ml.type)
404 lock->ml.highest_blocked = target->ml.type;
408 /* we can grant the blocked lock (only
409 * possible if converting list empty) */
411 spin_lock(&target->spinlock);
412 BUG_ON(target->ml.highest_blocked != LKM_IVMODE);
414 mlog(0, "calling ast for blocked lock: %.*s, granting: %d, "
415 "node: %u\n", res->lockname.len, res->lockname.name,
416 target->ml.type, target->ml.node);
418 // target->ml.type is already correct
419 list_move_tail(&target->list, &res->granted);
421 BUG_ON(!target->lksb);
422 target->lksb->status = DLM_NORMAL;
424 spin_unlock(&target->spinlock);
426 __dlm_lockres_reserve_ast(res);
427 dlm_queue_ast(dlm, target);
428 /* go back and check for more */
436 /* must have NO locks when calling this with res !=NULL * */
437 void dlm_kick_thread(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
439 mlog_entry("dlm=%p, res=%p\n", dlm, res);
441 spin_lock(&dlm->spinlock);
442 spin_lock(&res->spinlock);
443 __dlm_dirty_lockres(dlm, res);
444 spin_unlock(&res->spinlock);
445 spin_unlock(&dlm->spinlock);
447 wake_up(&dlm->dlm_thread_wq);
450 void __dlm_dirty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
452 mlog_entry("dlm=%p, res=%p\n", dlm, res);
454 assert_spin_locked(&dlm->spinlock);
455 assert_spin_locked(&res->spinlock);
457 /* don't shuffle secondary queues */
458 if ((res->owner == dlm->node_num)) {
459 if (res->state & (DLM_LOCK_RES_MIGRATING |
460 DLM_LOCK_RES_BLOCK_DIRTY))
463 if (list_empty(&res->dirty)) {
464 /* ref for dirty_list */
465 dlm_lockres_get(res);
466 list_add_tail(&res->dirty, &dlm->dirty_list);
467 res->state |= DLM_LOCK_RES_DIRTY;
473 /* Launch the NM thread for the mounted volume */
474 int dlm_launch_thread(struct dlm_ctxt *dlm)
476 mlog(0, "starting dlm thread...\n");
478 dlm->dlm_thread_task = kthread_run(dlm_thread, dlm, "dlm_thread");
479 if (IS_ERR(dlm->dlm_thread_task)) {
480 mlog_errno(PTR_ERR(dlm->dlm_thread_task));
481 dlm->dlm_thread_task = NULL;
488 void dlm_complete_thread(struct dlm_ctxt *dlm)
490 if (dlm->dlm_thread_task) {
491 mlog(ML_KTHREAD, "waiting for dlm thread to exit\n");
492 kthread_stop(dlm->dlm_thread_task);
493 dlm->dlm_thread_task = NULL;
497 static int dlm_dirty_list_empty(struct dlm_ctxt *dlm)
501 spin_lock(&dlm->spinlock);
502 empty = list_empty(&dlm->dirty_list);
503 spin_unlock(&dlm->spinlock);
508 static void dlm_flush_asts(struct dlm_ctxt *dlm)
511 struct dlm_lock *lock;
512 struct dlm_lock_resource *res;
515 spin_lock(&dlm->ast_lock);
516 while (!list_empty(&dlm->pending_asts)) {
517 lock = list_entry(dlm->pending_asts.next,
518 struct dlm_lock, ast_list);
519 /* get an extra ref on lock */
522 mlog(0, "delivering an ast for this lockres\n");
524 BUG_ON(!lock->ast_pending);
526 /* remove from list (including ref) */
527 list_del_init(&lock->ast_list);
529 spin_unlock(&dlm->ast_lock);
531 if (lock->ml.node != dlm->node_num) {
532 ret = dlm_do_remote_ast(dlm, res, lock);
536 dlm_do_local_ast(dlm, res, lock);
538 spin_lock(&dlm->ast_lock);
540 /* possible that another ast was queued while
541 * we were delivering the last one */
542 if (!list_empty(&lock->ast_list)) {
543 mlog(0, "aha another ast got queued while "
544 "we were finishing the last one. will "
545 "keep the ast_pending flag set.\n");
547 lock->ast_pending = 0;
549 /* drop the extra ref.
550 * this may drop it completely. */
552 dlm_lockres_release_ast(dlm, res);
555 while (!list_empty(&dlm->pending_basts)) {
556 lock = list_entry(dlm->pending_basts.next,
557 struct dlm_lock, bast_list);
558 /* get an extra ref on lock */
562 BUG_ON(!lock->bast_pending);
564 /* get the highest blocked lock, and reset */
565 spin_lock(&lock->spinlock);
566 BUG_ON(lock->ml.highest_blocked <= LKM_IVMODE);
567 hi = lock->ml.highest_blocked;
568 lock->ml.highest_blocked = LKM_IVMODE;
569 spin_unlock(&lock->spinlock);
571 /* remove from list (including ref) */
572 list_del_init(&lock->bast_list);
574 spin_unlock(&dlm->ast_lock);
576 mlog(0, "delivering a bast for this lockres "
577 "(blocked = %d\n", hi);
579 if (lock->ml.node != dlm->node_num) {
580 ret = dlm_send_proxy_bast(dlm, res, lock, hi);
584 dlm_do_local_bast(dlm, res, lock, hi);
586 spin_lock(&dlm->ast_lock);
588 /* possible that another bast was queued while
589 * we were delivering the last one */
590 if (!list_empty(&lock->bast_list)) {
591 mlog(0, "aha another bast got queued while "
592 "we were finishing the last one. will "
593 "keep the bast_pending flag set.\n");
595 lock->bast_pending = 0;
597 /* drop the extra ref.
598 * this may drop it completely. */
600 dlm_lockres_release_ast(dlm, res);
602 wake_up(&dlm->ast_wq);
603 spin_unlock(&dlm->ast_lock);
607 #define DLM_THREAD_TIMEOUT_MS (4 * 1000)
608 #define DLM_THREAD_MAX_DIRTY 100
609 #define DLM_THREAD_MAX_ASTS 10
611 static int dlm_thread(void *data)
613 struct dlm_lock_resource *res;
614 struct dlm_ctxt *dlm = data;
615 unsigned long timeout = msecs_to_jiffies(DLM_THREAD_TIMEOUT_MS);
617 mlog(0, "dlm thread running for %s...\n", dlm->name);
619 while (!kthread_should_stop()) {
620 int n = DLM_THREAD_MAX_DIRTY;
622 /* dlm_shutting_down is very point-in-time, but that
623 * doesn't matter as we'll just loop back around if we
624 * get false on the leading edge of a state
626 dlm_run_purge_list(dlm, dlm_shutting_down(dlm));
628 /* We really don't want to hold dlm->spinlock while
629 * calling dlm_shuffle_lists on each lockres that
630 * needs to have its queues adjusted and AST/BASTs
631 * run. So let's pull each entry off the dirty_list
632 * and drop dlm->spinlock ASAP. Once off the list,
633 * res->spinlock needs to be taken again to protect
634 * the queues while calling dlm_shuffle_lists. */
635 spin_lock(&dlm->spinlock);
636 while (!list_empty(&dlm->dirty_list)) {
638 res = list_entry(dlm->dirty_list.next,
639 struct dlm_lock_resource, dirty);
641 /* peel a lockres off, remove it from the list,
642 * unset the dirty flag and drop the dlm lock */
644 dlm_lockres_get(res);
646 spin_lock(&res->spinlock);
647 /* We clear the DLM_LOCK_RES_DIRTY state once we shuffle lists below */
648 list_del_init(&res->dirty);
649 spin_unlock(&res->spinlock);
650 spin_unlock(&dlm->spinlock);
651 /* Drop dirty_list ref */
652 dlm_lockres_put(res);
654 /* lockres can be re-dirtied/re-added to the
655 * dirty_list in this gap, but that is ok */
657 spin_lock(&res->spinlock);
658 if (res->owner != dlm->node_num) {
659 __dlm_print_one_lock_resource(res);
660 mlog(ML_ERROR, "inprog:%s, mig:%s, reco:%s, dirty:%s\n",
661 res->state & DLM_LOCK_RES_IN_PROGRESS ? "yes" : "no",
662 res->state & DLM_LOCK_RES_MIGRATING ? "yes" : "no",
663 res->state & DLM_LOCK_RES_RECOVERING ? "yes" : "no",
664 res->state & DLM_LOCK_RES_DIRTY ? "yes" : "no");
666 BUG_ON(res->owner != dlm->node_num);
668 /* it is now ok to move lockreses in these states
669 * to the dirty list, assuming that they will only be
670 * dirty for a short while. */
671 BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
672 if (res->state & (DLM_LOCK_RES_IN_PROGRESS |
673 DLM_LOCK_RES_RECOVERING)) {
674 /* move it to the tail and keep going */
675 res->state &= ~DLM_LOCK_RES_DIRTY;
676 spin_unlock(&res->spinlock);
677 mlog(0, "delaying list shuffling for in-"
678 "progress lockres %.*s, state=%d\n",
679 res->lockname.len, res->lockname.name,
685 /* at this point the lockres is not migrating/
686 * recovering/in-progress. we have the lockres
687 * spinlock and do NOT have the dlm lock.
688 * safe to reserve/queue asts and run the lists. */
690 mlog(0, "calling dlm_shuffle_lists with dlm=%s, "
691 "res=%.*s\n", dlm->name,
692 res->lockname.len, res->lockname.name);
694 /* called while holding lockres lock */
695 dlm_shuffle_lists(dlm, res);
696 res->state &= ~DLM_LOCK_RES_DIRTY;
697 spin_unlock(&res->spinlock);
699 dlm_lockres_calc_usage(dlm, res);
703 spin_lock(&dlm->spinlock);
704 /* if the lock was in-progress, stick
705 * it on the back of the list */
707 spin_lock(&res->spinlock);
708 __dlm_dirty_lockres(dlm, res);
709 spin_unlock(&res->spinlock);
711 dlm_lockres_put(res);
713 /* unlikely, but we may need to give time to
716 mlog(0, "throttling dlm_thread\n");
721 spin_unlock(&dlm->spinlock);
724 /* yield and continue right away if there is more work to do */
730 wait_event_interruptible_timeout(dlm->dlm_thread_wq,
731 !dlm_dirty_list_empty(dlm) ||
732 kthread_should_stop(),
736 mlog(0, "quitting DLM thread\n");