xref: /linux/fs/ocfs2/dlm/dlmthread.c (revision 4e337adae4e960f64043b9f433c4a825c902616c)
1 /* -*- mode: c; c-basic-offset: 8; -*-
2  * vim: noexpandtab sw=8 ts=8 sts=0:
3  *
4  * dlmthread.c
5  *
6  * standalone DLM module
7  *
8  * Copyright (C) 2004 Oracle.  All rights reserved.
9  *
10  * This program is free software; you can redistribute it and/or
11  * modify it under the terms of the GNU General Public
12  * License as published by the Free Software Foundation; either
13  * version 2 of the License, or (at your option) any later version.
14  *
15  * This program is distributed in the hope that it will be useful,
16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
18  * General Public License for more details.
19  *
20  * You should have received a copy of the GNU General Public
21  * License along with this program; if not, write to the
22  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23  * Boston, MA 021110-1307, USA.
24  *
25  */
26 
27 
28 #include <linux/module.h>
29 #include <linux/fs.h>
30 #include <linux/types.h>
31 #include <linux/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>
43 
44 
45 #include "cluster/heartbeat.h"
46 #include "cluster/nodemanager.h"
47 #include "cluster/tcp.h"
48 
49 #include "dlmapi.h"
50 #include "dlmcommon.h"
51 #include "dlmdomain.h"
52 
53 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_THREAD)
54 #include "cluster/masklog.h"
55 
56 static int dlm_thread(void *data);
57 static void dlm_flush_asts(struct dlm_ctxt *dlm);
58 
59 #define dlm_lock_is_remote(dlm, lock)     ((lock)->ml.node != (dlm)->node_num)
60 
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)
64 {
65 	DECLARE_WAITQUEUE(wait, current);
66 
67 	assert_spin_locked(&res->spinlock);
68 
69 	add_wait_queue(&res->wq, &wait);
70 repeat:
71 	set_current_state(TASK_UNINTERRUPTIBLE);
72 	if (res->state & flags) {
73 		spin_unlock(&res->spinlock);
74 		schedule();
75 		spin_lock(&res->spinlock);
76 		goto repeat;
77 	}
78 	remove_wait_queue(&res->wq, &wait);
79 	current->state = TASK_RUNNING;
80 }
81 
82 int __dlm_lockres_has_locks(struct dlm_lock_resource *res)
83 {
84 	if (list_empty(&res->granted) &&
85 	    list_empty(&res->converting) &&
86 	    list_empty(&res->blocked))
87 		return 0;
88 	return 1;
89 }
90 
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)
96 {
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);
106 			return 1;
107 		}
108 	}
109 	return 0;
110 }
111 
112 
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)
118 {
119 	mlog_entry("%.*s\n", res->lockname.len, res->lockname.name);
120 
121 	assert_spin_locked(&dlm->spinlock);
122 	assert_spin_locked(&res->spinlock);
123 
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);
128 
129 			res->last_used = jiffies;
130 			dlm_lockres_get(res);
131 			list_add_tail(&res->purge, &dlm->purge_list);
132 			dlm->purge_count++;
133 		}
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);
137 
138 		list_del_init(&res->purge);
139 		dlm_lockres_put(res);
140 		dlm->purge_count--;
141 	}
142 }
143 
144 void dlm_lockres_calc_usage(struct dlm_ctxt *dlm,
145 			    struct dlm_lock_resource *res)
146 {
147 	mlog_entry("%.*s\n", res->lockname.len, res->lockname.name);
148 	spin_lock(&dlm->spinlock);
149 	spin_lock(&res->spinlock);
150 
151 	__dlm_lockres_calc_usage(dlm, res);
152 
153 	spin_unlock(&res->spinlock);
154 	spin_unlock(&dlm->spinlock);
155 }
156 
157 static int dlm_purge_lockres(struct dlm_ctxt *dlm,
158 			     struct dlm_lock_resource *res)
159 {
160 	int master;
161 	int ret = 0;
162 
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);
168 		return -ENOTEMPTY;
169 	}
170 	master = (res->owner == dlm->node_num);
171 	if (!master)
172 		res->state |= DLM_LOCK_RES_DROPPING_REF;
173 	spin_unlock(&res->spinlock);
174 
175 	mlog(0, "purging lockres %.*s, master = %d\n", res->lockname.len,
176 	     res->lockname.name, master);
177 
178 	if (!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);
187 		if (ret < 0) {
188 			mlog_errno(ret);
189 			if (!dlm_is_host_down(ret))
190 				BUG();
191 		}
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);
195 	}
196 
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,
200 		     res, master);
201 		list_del_init(&res->purge);
202 		dlm_lockres_put(res);
203 		dlm->purge_count--;
204 	}
205 	__dlm_unhash_lockres(res);
206 
207 	/* lockres is not in the hash now.  drop the flag and wake up
208 	 * any processes waiting in dlm_get_lock_resource. */
209 	if (!master) {
210 		spin_lock(&res->spinlock);
211 		res->state &= ~DLM_LOCK_RES_DROPPING_REF;
212 		spin_unlock(&res->spinlock);
213 		wake_up(&res->wq);
214 	}
215 	return 0;
216 }
217 
218 static void dlm_run_purge_list(struct dlm_ctxt *dlm,
219 			       int purge_now)
220 {
221 	unsigned int run_max, unused;
222 	unsigned long purge_jiffies;
223 	struct dlm_lock_resource *lockres;
224 
225 	spin_lock(&dlm->spinlock);
226 	run_max = dlm->purge_count;
227 
228 	while(run_max && !list_empty(&dlm->purge_list)) {
229 		run_max--;
230 
231 		lockres = list_entry(dlm->purge_list.next,
232 				     struct dlm_lock_resource, purge);
233 
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
238 		 * spinlock. */
239 		spin_lock(&lockres->spinlock);
240 		unused = __dlm_lockres_unused(lockres);
241 		spin_unlock(&lockres->spinlock);
242 
243 		if (!unused)
244 			continue;
245 
246 		purge_jiffies = lockres->last_used +
247 			msecs_to_jiffies(DLM_PURGE_INTERVAL_MS);
248 
249 		/* Make sure that we want to be processing this guy at
250 		 * this time. */
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 */
256 			break;
257 		}
258 
259 		mlog(0, "removing lockres %.*s:%p from purgelist\n",
260 		     lockres->lockname.len, lockres->lockname.name, lockres);
261 		list_del_init(&lockres->purge);
262 		dlm_lockres_put(lockres);
263 		dlm->purge_count--;
264 
265 		/* This may drop and reacquire the dlm spinlock if it
266 		 * has to do migration. */
267 		mlog(0, "calling dlm_purge_lockres!\n");
268 		dlm_lockres_get(lockres);
269 		if (dlm_purge_lockres(dlm, lockres))
270 			BUG();
271 		dlm_lockres_put(lockres);
272 		mlog(0, "DONE calling dlm_purge_lockres!\n");
273 
274 		/* Avoid adding any scheduling latencies */
275 		cond_resched_lock(&dlm->spinlock);
276 	}
277 
278 	spin_unlock(&dlm->spinlock);
279 }
280 
281 static void dlm_shuffle_lists(struct dlm_ctxt *dlm,
282 			      struct dlm_lock_resource *res)
283 {
284 	struct dlm_lock *lock, *target;
285 	struct list_head *iter;
286 	struct list_head *head;
287 	int can_grant = 1;
288 
289 	//mlog(0, "res->lockname.len=%d\n", res->lockname.len);
290 	//mlog(0, "res->lockname.name=%p\n", res->lockname.name);
291 	//mlog(0, "shuffle res %.*s\n", res->lockname.len,
292 	//	  res->lockname.name);
293 
294 	/* because this function is called with the lockres
295 	 * spinlock, and because we know that it is not migrating/
296 	 * recovering/in-progress, it is fine to reserve asts and
297 	 * basts right before queueing them all throughout */
298 	assert_spin_locked(&res->spinlock);
299 	BUG_ON((res->state & (DLM_LOCK_RES_MIGRATING|
300 			      DLM_LOCK_RES_RECOVERING|
301 			      DLM_LOCK_RES_IN_PROGRESS)));
302 
303 converting:
304 	if (list_empty(&res->converting))
305 		goto blocked;
306 	mlog(0, "res %.*s has locks on a convert queue\n", res->lockname.len,
307 	     res->lockname.name);
308 
309 	target = list_entry(res->converting.next, struct dlm_lock, list);
310 	if (target->ml.convert_type == LKM_IVMODE) {
311 		mlog(ML_ERROR, "%.*s: converting a lock with no "
312 		     "convert_type!\n", res->lockname.len, res->lockname.name);
313 		BUG();
314 	}
315 	head = &res->granted;
316 	list_for_each(iter, head) {
317 		lock = list_entry(iter, struct dlm_lock, list);
318 		if (lock==target)
319 			continue;
320 		if (!dlm_lock_compatible(lock->ml.type,
321 					 target->ml.convert_type)) {
322 			can_grant = 0;
323 			/* queue the BAST if not already */
324 			if (lock->ml.highest_blocked == LKM_IVMODE) {
325 				__dlm_lockres_reserve_ast(res);
326 				dlm_queue_bast(dlm, lock);
327 			}
328 			/* update the highest_blocked if needed */
329 			if (lock->ml.highest_blocked < target->ml.convert_type)
330 				lock->ml.highest_blocked =
331 					target->ml.convert_type;
332 		}
333 	}
334 	head = &res->converting;
335 	list_for_each(iter, head) {
336 		lock = list_entry(iter, struct dlm_lock, list);
337 		if (lock==target)
338 			continue;
339 		if (!dlm_lock_compatible(lock->ml.type,
340 					 target->ml.convert_type)) {
341 			can_grant = 0;
342 			if (lock->ml.highest_blocked == LKM_IVMODE) {
343 				__dlm_lockres_reserve_ast(res);
344 				dlm_queue_bast(dlm, lock);
345 			}
346 			if (lock->ml.highest_blocked < target->ml.convert_type)
347 				lock->ml.highest_blocked =
348 					target->ml.convert_type;
349 		}
350 	}
351 
352 	/* we can convert the lock */
353 	if (can_grant) {
354 		spin_lock(&target->spinlock);
355 		BUG_ON(target->ml.highest_blocked != LKM_IVMODE);
356 
357 		mlog(0, "calling ast for converting lock: %.*s, have: %d, "
358 		     "granting: %d, node: %u\n", res->lockname.len,
359 		     res->lockname.name, target->ml.type,
360 		     target->ml.convert_type, target->ml.node);
361 
362 		target->ml.type = target->ml.convert_type;
363 		target->ml.convert_type = LKM_IVMODE;
364 		list_move_tail(&target->list, &res->granted);
365 
366 		BUG_ON(!target->lksb);
367 		target->lksb->status = DLM_NORMAL;
368 
369 		spin_unlock(&target->spinlock);
370 
371 		__dlm_lockres_reserve_ast(res);
372 		dlm_queue_ast(dlm, target);
373 		/* go back and check for more */
374 		goto converting;
375 	}
376 
377 blocked:
378 	if (list_empty(&res->blocked))
379 		goto leave;
380 	target = list_entry(res->blocked.next, struct dlm_lock, list);
381 
382 	head = &res->granted;
383 	list_for_each(iter, head) {
384 		lock = list_entry(iter, struct dlm_lock, list);
385 		if (lock==target)
386 			continue;
387 		if (!dlm_lock_compatible(lock->ml.type, target->ml.type)) {
388 			can_grant = 0;
389 			if (lock->ml.highest_blocked == LKM_IVMODE) {
390 				__dlm_lockres_reserve_ast(res);
391 				dlm_queue_bast(dlm, lock);
392 			}
393 			if (lock->ml.highest_blocked < target->ml.type)
394 				lock->ml.highest_blocked = target->ml.type;
395 		}
396 	}
397 
398 	head = &res->converting;
399 	list_for_each(iter, head) {
400 		lock = list_entry(iter, struct dlm_lock, list);
401 		if (lock==target)
402 			continue;
403 		if (!dlm_lock_compatible(lock->ml.type, target->ml.type)) {
404 			can_grant = 0;
405 			if (lock->ml.highest_blocked == LKM_IVMODE) {
406 				__dlm_lockres_reserve_ast(res);
407 				dlm_queue_bast(dlm, lock);
408 			}
409 			if (lock->ml.highest_blocked < target->ml.type)
410 				lock->ml.highest_blocked = target->ml.type;
411 		}
412 	}
413 
414 	/* we can grant the blocked lock (only
415 	 * possible if converting list empty) */
416 	if (can_grant) {
417 		spin_lock(&target->spinlock);
418 		BUG_ON(target->ml.highest_blocked != LKM_IVMODE);
419 
420 		mlog(0, "calling ast for blocked lock: %.*s, granting: %d, "
421 		     "node: %u\n", res->lockname.len, res->lockname.name,
422 		     target->ml.type, target->ml.node);
423 
424 		// target->ml.type is already correct
425 		list_move_tail(&target->list, &res->granted);
426 
427 		BUG_ON(!target->lksb);
428 		target->lksb->status = DLM_NORMAL;
429 
430 		spin_unlock(&target->spinlock);
431 
432 		__dlm_lockres_reserve_ast(res);
433 		dlm_queue_ast(dlm, target);
434 		/* go back and check for more */
435 		goto converting;
436 	}
437 
438 leave:
439 	return;
440 }
441 
442 /* must have NO locks when calling this with res !=NULL * */
443 void dlm_kick_thread(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
444 {
445 	mlog_entry("dlm=%p, res=%p\n", dlm, res);
446 	if (res) {
447 		spin_lock(&dlm->spinlock);
448 		spin_lock(&res->spinlock);
449 		__dlm_dirty_lockres(dlm, res);
450 		spin_unlock(&res->spinlock);
451 		spin_unlock(&dlm->spinlock);
452 	}
453 	wake_up(&dlm->dlm_thread_wq);
454 }
455 
456 void __dlm_dirty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
457 {
458 	mlog_entry("dlm=%p, res=%p\n", dlm, res);
459 
460 	assert_spin_locked(&dlm->spinlock);
461 	assert_spin_locked(&res->spinlock);
462 
463 	/* don't shuffle secondary queues */
464 	if ((res->owner == dlm->node_num)) {
465 		if (res->state & (DLM_LOCK_RES_MIGRATING |
466 				  DLM_LOCK_RES_BLOCK_DIRTY))
467 		    return;
468 
469 		if (list_empty(&res->dirty)) {
470 			/* ref for dirty_list */
471 			dlm_lockres_get(res);
472 			list_add_tail(&res->dirty, &dlm->dirty_list);
473 			res->state |= DLM_LOCK_RES_DIRTY;
474 		}
475 	}
476 }
477 
478 
479 /* Launch the NM thread for the mounted volume */
480 int dlm_launch_thread(struct dlm_ctxt *dlm)
481 {
482 	mlog(0, "starting dlm thread...\n");
483 
484 	dlm->dlm_thread_task = kthread_run(dlm_thread, dlm, "dlm_thread");
485 	if (IS_ERR(dlm->dlm_thread_task)) {
486 		mlog_errno(PTR_ERR(dlm->dlm_thread_task));
487 		dlm->dlm_thread_task = NULL;
488 		return -EINVAL;
489 	}
490 
491 	return 0;
492 }
493 
494 void dlm_complete_thread(struct dlm_ctxt *dlm)
495 {
496 	if (dlm->dlm_thread_task) {
497 		mlog(ML_KTHREAD, "waiting for dlm thread to exit\n");
498 		kthread_stop(dlm->dlm_thread_task);
499 		dlm->dlm_thread_task = NULL;
500 	}
501 }
502 
503 static int dlm_dirty_list_empty(struct dlm_ctxt *dlm)
504 {
505 	int empty;
506 
507 	spin_lock(&dlm->spinlock);
508 	empty = list_empty(&dlm->dirty_list);
509 	spin_unlock(&dlm->spinlock);
510 
511 	return empty;
512 }
513 
514 static void dlm_flush_asts(struct dlm_ctxt *dlm)
515 {
516 	int ret;
517 	struct dlm_lock *lock;
518 	struct dlm_lock_resource *res;
519 	u8 hi;
520 
521 	spin_lock(&dlm->ast_lock);
522 	while (!list_empty(&dlm->pending_asts)) {
523 		lock = list_entry(dlm->pending_asts.next,
524 				  struct dlm_lock, ast_list);
525 		/* get an extra ref on lock */
526 		dlm_lock_get(lock);
527 		res = lock->lockres;
528 		mlog(0, "delivering an ast for this lockres\n");
529 
530 		BUG_ON(!lock->ast_pending);
531 
532 		/* remove from list (including ref) */
533 		list_del_init(&lock->ast_list);
534 		dlm_lock_put(lock);
535 		spin_unlock(&dlm->ast_lock);
536 
537 		if (lock->ml.node != dlm->node_num) {
538 			ret = dlm_do_remote_ast(dlm, res, lock);
539 			if (ret < 0)
540 				mlog_errno(ret);
541 		} else
542 			dlm_do_local_ast(dlm, res, lock);
543 
544 		spin_lock(&dlm->ast_lock);
545 
546 		/* possible that another ast was queued while
547 		 * we were delivering the last one */
548 		if (!list_empty(&lock->ast_list)) {
549 			mlog(0, "aha another ast got queued while "
550 			     "we were finishing the last one.  will "
551 			     "keep the ast_pending flag set.\n");
552 		} else
553 			lock->ast_pending = 0;
554 
555 		/* drop the extra ref.
556 		 * this may drop it completely. */
557 		dlm_lock_put(lock);
558 		dlm_lockres_release_ast(dlm, res);
559 	}
560 
561 	while (!list_empty(&dlm->pending_basts)) {
562 		lock = list_entry(dlm->pending_basts.next,
563 				  struct dlm_lock, bast_list);
564 		/* get an extra ref on lock */
565 		dlm_lock_get(lock);
566 		res = lock->lockres;
567 
568 		BUG_ON(!lock->bast_pending);
569 
570 		/* get the highest blocked lock, and reset */
571 		spin_lock(&lock->spinlock);
572 		BUG_ON(lock->ml.highest_blocked <= LKM_IVMODE);
573 		hi = lock->ml.highest_blocked;
574 		lock->ml.highest_blocked = LKM_IVMODE;
575 		spin_unlock(&lock->spinlock);
576 
577 		/* remove from list (including ref) */
578 		list_del_init(&lock->bast_list);
579 		dlm_lock_put(lock);
580 		spin_unlock(&dlm->ast_lock);
581 
582 		mlog(0, "delivering a bast for this lockres "
583 		     "(blocked = %d\n", hi);
584 
585 		if (lock->ml.node != dlm->node_num) {
586 			ret = dlm_send_proxy_bast(dlm, res, lock, hi);
587 			if (ret < 0)
588 				mlog_errno(ret);
589 		} else
590 			dlm_do_local_bast(dlm, res, lock, hi);
591 
592 		spin_lock(&dlm->ast_lock);
593 
594 		/* possible that another bast was queued while
595 		 * we were delivering the last one */
596 		if (!list_empty(&lock->bast_list)) {
597 			mlog(0, "aha another bast got queued while "
598 			     "we were finishing the last one.  will "
599 			     "keep the bast_pending flag set.\n");
600 		} else
601 			lock->bast_pending = 0;
602 
603 		/* drop the extra ref.
604 		 * this may drop it completely. */
605 		dlm_lock_put(lock);
606 		dlm_lockres_release_ast(dlm, res);
607 	}
608 	wake_up(&dlm->ast_wq);
609 	spin_unlock(&dlm->ast_lock);
610 }
611 
612 
613 #define DLM_THREAD_TIMEOUT_MS (4 * 1000)
614 #define DLM_THREAD_MAX_DIRTY  100
615 #define DLM_THREAD_MAX_ASTS   10
616 
617 static int dlm_thread(void *data)
618 {
619 	struct dlm_lock_resource *res;
620 	struct dlm_ctxt *dlm = data;
621 	unsigned long timeout = msecs_to_jiffies(DLM_THREAD_TIMEOUT_MS);
622 
623 	mlog(0, "dlm thread running for %s...\n", dlm->name);
624 
625 	while (!kthread_should_stop()) {
626 		int n = DLM_THREAD_MAX_DIRTY;
627 
628 		/* dlm_shutting_down is very point-in-time, but that
629 		 * doesn't matter as we'll just loop back around if we
630 		 * get false on the leading edge of a state
631 		 * transition. */
632 		dlm_run_purge_list(dlm, dlm_shutting_down(dlm));
633 
634 		/* We really don't want to hold dlm->spinlock while
635 		 * calling dlm_shuffle_lists on each lockres that
636 		 * needs to have its queues adjusted and AST/BASTs
637 		 * run.  So let's pull each entry off the dirty_list
638 		 * and drop dlm->spinlock ASAP.  Once off the list,
639 		 * res->spinlock needs to be taken again to protect
640 		 * the queues while calling dlm_shuffle_lists.  */
641 		spin_lock(&dlm->spinlock);
642 		while (!list_empty(&dlm->dirty_list)) {
643 			int delay = 0;
644 			res = list_entry(dlm->dirty_list.next,
645 					 struct dlm_lock_resource, dirty);
646 
647 			/* peel a lockres off, remove it from the list,
648 			 * unset the dirty flag and drop the dlm lock */
649 			BUG_ON(!res);
650 			dlm_lockres_get(res);
651 
652 			spin_lock(&res->spinlock);
653 			/* We clear the DLM_LOCK_RES_DIRTY state once we shuffle lists below */
654 			list_del_init(&res->dirty);
655 			spin_unlock(&res->spinlock);
656 			spin_unlock(&dlm->spinlock);
657 			/* Drop dirty_list ref */
658 			dlm_lockres_put(res);
659 
660 		 	/* lockres can be re-dirtied/re-added to the
661 			 * dirty_list in this gap, but that is ok */
662 
663 			spin_lock(&res->spinlock);
664 			if (res->owner != dlm->node_num) {
665 				__dlm_print_one_lock_resource(res);
666 				mlog(ML_ERROR, "inprog:%s, mig:%s, reco:%s, dirty:%s\n",
667 				     res->state & DLM_LOCK_RES_IN_PROGRESS ? "yes" : "no",
668 				     res->state & DLM_LOCK_RES_MIGRATING ? "yes" : "no",
669 				     res->state & DLM_LOCK_RES_RECOVERING ? "yes" : "no",
670 				     res->state & DLM_LOCK_RES_DIRTY ? "yes" : "no");
671 			}
672 			BUG_ON(res->owner != dlm->node_num);
673 
674 			/* it is now ok to move lockreses in these states
675 			 * to the dirty list, assuming that they will only be
676 			 * dirty for a short while. */
677 			BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
678 			if (res->state & (DLM_LOCK_RES_IN_PROGRESS |
679 					  DLM_LOCK_RES_RECOVERING)) {
680 				/* move it to the tail and keep going */
681 				res->state &= ~DLM_LOCK_RES_DIRTY;
682 				spin_unlock(&res->spinlock);
683 				mlog(0, "delaying list shuffling for in-"
684 				     "progress lockres %.*s, state=%d\n",
685 				     res->lockname.len, res->lockname.name,
686 				     res->state);
687 				delay = 1;
688 				goto in_progress;
689 			}
690 
691 			/* at this point the lockres is not migrating/
692 			 * recovering/in-progress.  we have the lockres
693 			 * spinlock and do NOT have the dlm lock.
694 			 * safe to reserve/queue asts and run the lists. */
695 
696 			mlog(0, "calling dlm_shuffle_lists with dlm=%s, "
697 			     "res=%.*s\n", dlm->name,
698 			     res->lockname.len, res->lockname.name);
699 
700 			/* called while holding lockres lock */
701 			dlm_shuffle_lists(dlm, res);
702 			res->state &= ~DLM_LOCK_RES_DIRTY;
703 			spin_unlock(&res->spinlock);
704 
705 			dlm_lockres_calc_usage(dlm, res);
706 
707 in_progress:
708 
709 			spin_lock(&dlm->spinlock);
710 			/* if the lock was in-progress, stick
711 			 * it on the back of the list */
712 			if (delay) {
713 				spin_lock(&res->spinlock);
714 				__dlm_dirty_lockres(dlm, res);
715 				spin_unlock(&res->spinlock);
716 			}
717 			dlm_lockres_put(res);
718 
719 			/* unlikely, but we may need to give time to
720 			 * other tasks */
721 			if (!--n) {
722 				mlog(0, "throttling dlm_thread\n");
723 				break;
724 			}
725 		}
726 
727 		spin_unlock(&dlm->spinlock);
728 		dlm_flush_asts(dlm);
729 
730 		/* yield and continue right away if there is more work to do */
731 		if (!n) {
732 			cond_resched();
733 			continue;
734 		}
735 
736 		wait_event_interruptible_timeout(dlm->dlm_thread_wq,
737 						 !dlm_dirty_list_empty(dlm) ||
738 						 kthread_should_stop(),
739 						 timeout);
740 	}
741 
742 	mlog(0, "quitting DLM thread\n");
743 	return 0;
744 }
745