xref: /openbmc/linux/fs/ocfs2/dlm/dlmthread.c (revision b593bce5)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /* -*- mode: c; c-basic-offset: 8; -*-
3  * vim: noexpandtab sw=8 ts=8 sts=0:
4  *
5  * dlmthread.c
6  *
7  * standalone DLM module
8  *
9  * Copyright (C) 2004 Oracle.  All rights reserved.
10  */
11 
12 
13 #include <linux/module.h>
14 #include <linux/fs.h>
15 #include <linux/types.h>
16 #include <linux/highmem.h>
17 #include <linux/init.h>
18 #include <linux/sysctl.h>
19 #include <linux/random.h>
20 #include <linux/blkdev.h>
21 #include <linux/socket.h>
22 #include <linux/inet.h>
23 #include <linux/timer.h>
24 #include <linux/kthread.h>
25 #include <linux/delay.h>
26 
27 
28 #include "cluster/heartbeat.h"
29 #include "cluster/nodemanager.h"
30 #include "cluster/tcp.h"
31 
32 #include "dlmapi.h"
33 #include "dlmcommon.h"
34 #include "dlmdomain.h"
35 
36 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_THREAD)
37 #include "cluster/masklog.h"
38 
39 static int dlm_thread(void *data);
40 static void dlm_flush_asts(struct dlm_ctxt *dlm);
41 
42 #define dlm_lock_is_remote(dlm, lock)     ((lock)->ml.node != (dlm)->node_num)
43 
44 /* will exit holding res->spinlock, but may drop in function */
45 /* waits until flags are cleared on res->state */
46 void __dlm_wait_on_lockres_flags(struct dlm_lock_resource *res, int flags)
47 {
48 	DECLARE_WAITQUEUE(wait, current);
49 
50 	assert_spin_locked(&res->spinlock);
51 
52 	add_wait_queue(&res->wq, &wait);
53 repeat:
54 	set_current_state(TASK_UNINTERRUPTIBLE);
55 	if (res->state & flags) {
56 		spin_unlock(&res->spinlock);
57 		schedule();
58 		spin_lock(&res->spinlock);
59 		goto repeat;
60 	}
61 	remove_wait_queue(&res->wq, &wait);
62 	__set_current_state(TASK_RUNNING);
63 }
64 
65 int __dlm_lockres_has_locks(struct dlm_lock_resource *res)
66 {
67 	if (list_empty(&res->granted) &&
68 	    list_empty(&res->converting) &&
69 	    list_empty(&res->blocked))
70 		return 0;
71 	return 1;
72 }
73 
74 /* "unused": the lockres has no locks, is not on the dirty list,
75  * has no inflight locks (in the gap between mastery and acquiring
76  * the first lock), and has no bits in its refmap.
77  * truly ready to be freed. */
78 int __dlm_lockres_unused(struct dlm_lock_resource *res)
79 {
80 	int bit;
81 
82 	assert_spin_locked(&res->spinlock);
83 
84 	if (__dlm_lockres_has_locks(res))
85 		return 0;
86 
87 	/* Locks are in the process of being created */
88 	if (res->inflight_locks)
89 		return 0;
90 
91 	if (!list_empty(&res->dirty) || res->state & DLM_LOCK_RES_DIRTY)
92 		return 0;
93 
94 	if (res->state & (DLM_LOCK_RES_RECOVERING|
95 			DLM_LOCK_RES_RECOVERY_WAITING))
96 		return 0;
97 
98 	/* Another node has this resource with this node as the master */
99 	bit = find_next_bit(res->refmap, O2NM_MAX_NODES, 0);
100 	if (bit < O2NM_MAX_NODES)
101 		return 0;
102 
103 	return 1;
104 }
105 
106 
107 /* Call whenever you may have added or deleted something from one of
108  * the lockres queue's. This will figure out whether it belongs on the
109  * unused list or not and does the appropriate thing. */
110 void __dlm_lockres_calc_usage(struct dlm_ctxt *dlm,
111 			      struct dlm_lock_resource *res)
112 {
113 	assert_spin_locked(&dlm->spinlock);
114 	assert_spin_locked(&res->spinlock);
115 
116 	if (__dlm_lockres_unused(res)){
117 		if (list_empty(&res->purge)) {
118 			mlog(0, "%s: Adding res %.*s to purge list\n",
119 			     dlm->name, res->lockname.len, res->lockname.name);
120 
121 			res->last_used = jiffies;
122 			dlm_lockres_get(res);
123 			list_add_tail(&res->purge, &dlm->purge_list);
124 			dlm->purge_count++;
125 		}
126 	} else if (!list_empty(&res->purge)) {
127 		mlog(0, "%s: Removing res %.*s from purge list\n",
128 		     dlm->name, res->lockname.len, res->lockname.name);
129 
130 		list_del_init(&res->purge);
131 		dlm_lockres_put(res);
132 		dlm->purge_count--;
133 	}
134 }
135 
136 void dlm_lockres_calc_usage(struct dlm_ctxt *dlm,
137 			    struct dlm_lock_resource *res)
138 {
139 	spin_lock(&dlm->spinlock);
140 	spin_lock(&res->spinlock);
141 
142 	__dlm_lockres_calc_usage(dlm, res);
143 
144 	spin_unlock(&res->spinlock);
145 	spin_unlock(&dlm->spinlock);
146 }
147 
148 /*
149  * Do the real purge work:
150  *     unhash the lockres, and
151  *     clear flag DLM_LOCK_RES_DROPPING_REF.
152  * It requires dlm and lockres spinlock to be taken.
153  */
154 void __dlm_do_purge_lockres(struct dlm_ctxt *dlm,
155 		struct dlm_lock_resource *res)
156 {
157 	assert_spin_locked(&dlm->spinlock);
158 	assert_spin_locked(&res->spinlock);
159 
160 	if (!list_empty(&res->purge)) {
161 		mlog(0, "%s: Removing res %.*s from purgelist\n",
162 		     dlm->name, res->lockname.len, res->lockname.name);
163 		list_del_init(&res->purge);
164 		dlm_lockres_put(res);
165 		dlm->purge_count--;
166 	}
167 
168 	if (!__dlm_lockres_unused(res)) {
169 		mlog(ML_ERROR, "%s: res %.*s in use after deref\n",
170 		     dlm->name, res->lockname.len, res->lockname.name);
171 		__dlm_print_one_lock_resource(res);
172 		BUG();
173 	}
174 
175 	__dlm_unhash_lockres(dlm, res);
176 
177 	spin_lock(&dlm->track_lock);
178 	if (!list_empty(&res->tracking))
179 		list_del_init(&res->tracking);
180 	else {
181 		mlog(ML_ERROR, "%s: Resource %.*s not on the Tracking list\n",
182 		     dlm->name, res->lockname.len, res->lockname.name);
183 		__dlm_print_one_lock_resource(res);
184 	}
185 	spin_unlock(&dlm->track_lock);
186 
187 	/*
188 	 * lockres is not in the hash now. drop the flag and wake up
189 	 * any processes waiting in dlm_get_lock_resource.
190 	 */
191 	res->state &= ~DLM_LOCK_RES_DROPPING_REF;
192 }
193 
194 static void dlm_purge_lockres(struct dlm_ctxt *dlm,
195 			     struct dlm_lock_resource *res)
196 {
197 	int master;
198 	int ret = 0;
199 
200 	assert_spin_locked(&dlm->spinlock);
201 	assert_spin_locked(&res->spinlock);
202 
203 	master = (res->owner == dlm->node_num);
204 
205 	mlog(0, "%s: Purging res %.*s, master %d\n", dlm->name,
206 	     res->lockname.len, res->lockname.name, master);
207 
208 	if (!master) {
209 		if (res->state & DLM_LOCK_RES_DROPPING_REF) {
210 			mlog(ML_NOTICE, "%s: res %.*s already in DLM_LOCK_RES_DROPPING_REF state\n",
211 				dlm->name, res->lockname.len, res->lockname.name);
212 			spin_unlock(&res->spinlock);
213 			return;
214 		}
215 
216 		res->state |= DLM_LOCK_RES_DROPPING_REF;
217 		/* drop spinlock...  retake below */
218 		spin_unlock(&res->spinlock);
219 		spin_unlock(&dlm->spinlock);
220 
221 		spin_lock(&res->spinlock);
222 		/* This ensures that clear refmap is sent after the set */
223 		__dlm_wait_on_lockres_flags(res, DLM_LOCK_RES_SETREF_INPROG);
224 		spin_unlock(&res->spinlock);
225 
226 		/* clear our bit from the master's refmap, ignore errors */
227 		ret = dlm_drop_lockres_ref(dlm, res);
228 		if (ret < 0) {
229 			if (!dlm_is_host_down(ret))
230 				BUG();
231 		}
232 		spin_lock(&dlm->spinlock);
233 		spin_lock(&res->spinlock);
234 	}
235 
236 	if (!list_empty(&res->purge)) {
237 		mlog(0, "%s: Removing res %.*s from purgelist, master %d\n",
238 		     dlm->name, res->lockname.len, res->lockname.name, master);
239 		list_del_init(&res->purge);
240 		dlm_lockres_put(res);
241 		dlm->purge_count--;
242 	}
243 
244 	if (!master && ret == DLM_DEREF_RESPONSE_INPROG) {
245 		mlog(0, "%s: deref %.*s in progress\n",
246 			dlm->name, res->lockname.len, res->lockname.name);
247 		spin_unlock(&res->spinlock);
248 		return;
249 	}
250 
251 	if (!__dlm_lockres_unused(res)) {
252 		mlog(ML_ERROR, "%s: res %.*s in use after deref\n",
253 		     dlm->name, res->lockname.len, res->lockname.name);
254 		__dlm_print_one_lock_resource(res);
255 		BUG();
256 	}
257 
258 	__dlm_unhash_lockres(dlm, res);
259 
260 	spin_lock(&dlm->track_lock);
261 	if (!list_empty(&res->tracking))
262 		list_del_init(&res->tracking);
263 	else {
264 		mlog(ML_ERROR, "Resource %.*s not on the Tracking list\n",
265 				res->lockname.len, res->lockname.name);
266 		__dlm_print_one_lock_resource(res);
267 	}
268 	spin_unlock(&dlm->track_lock);
269 
270 	/* lockres is not in the hash now.  drop the flag and wake up
271 	 * any processes waiting in dlm_get_lock_resource. */
272 	if (!master) {
273 		res->state &= ~DLM_LOCK_RES_DROPPING_REF;
274 		spin_unlock(&res->spinlock);
275 		wake_up(&res->wq);
276 	} else
277 		spin_unlock(&res->spinlock);
278 }
279 
280 static void dlm_run_purge_list(struct dlm_ctxt *dlm,
281 			       int purge_now)
282 {
283 	unsigned int run_max, unused;
284 	unsigned long purge_jiffies;
285 	struct dlm_lock_resource *lockres;
286 
287 	spin_lock(&dlm->spinlock);
288 	run_max = dlm->purge_count;
289 
290 	while(run_max && !list_empty(&dlm->purge_list)) {
291 		run_max--;
292 
293 		lockres = list_entry(dlm->purge_list.next,
294 				     struct dlm_lock_resource, purge);
295 
296 		spin_lock(&lockres->spinlock);
297 
298 		purge_jiffies = lockres->last_used +
299 			msecs_to_jiffies(DLM_PURGE_INTERVAL_MS);
300 
301 		/* Make sure that we want to be processing this guy at
302 		 * this time. */
303 		if (!purge_now && time_after(purge_jiffies, jiffies)) {
304 			/* Since resources are added to the purge list
305 			 * in tail order, we can stop at the first
306 			 * unpurgable resource -- anyone added after
307 			 * him will have a greater last_used value */
308 			spin_unlock(&lockres->spinlock);
309 			break;
310 		}
311 
312 		/* Status of the lockres *might* change so double
313 		 * check. If the lockres is unused, holding the dlm
314 		 * spinlock will prevent people from getting and more
315 		 * refs on it. */
316 		unused = __dlm_lockres_unused(lockres);
317 		if (!unused ||
318 		    (lockres->state & DLM_LOCK_RES_MIGRATING) ||
319 		    (lockres->inflight_assert_workers != 0)) {
320 			mlog(0, "%s: res %.*s is in use or being remastered, "
321 			     "used %d, state %d, assert master workers %u\n",
322 			     dlm->name, lockres->lockname.len,
323 			     lockres->lockname.name,
324 			     !unused, lockres->state,
325 			     lockres->inflight_assert_workers);
326 			list_move_tail(&lockres->purge, &dlm->purge_list);
327 			spin_unlock(&lockres->spinlock);
328 			continue;
329 		}
330 
331 		dlm_lockres_get(lockres);
332 
333 		dlm_purge_lockres(dlm, lockres);
334 
335 		dlm_lockres_put(lockres);
336 
337 		/* Avoid adding any scheduling latencies */
338 		cond_resched_lock(&dlm->spinlock);
339 	}
340 
341 	spin_unlock(&dlm->spinlock);
342 }
343 
344 static void dlm_shuffle_lists(struct dlm_ctxt *dlm,
345 			      struct dlm_lock_resource *res)
346 {
347 	struct dlm_lock *lock, *target;
348 	int can_grant = 1;
349 
350 	/*
351 	 * Because this function is called with the lockres
352 	 * spinlock, and because we know that it is not migrating/
353 	 * recovering/in-progress, it is fine to reserve asts and
354 	 * basts right before queueing them all throughout
355 	 */
356 	assert_spin_locked(&dlm->ast_lock);
357 	assert_spin_locked(&res->spinlock);
358 	BUG_ON((res->state & (DLM_LOCK_RES_MIGRATING|
359 			      DLM_LOCK_RES_RECOVERING|
360 			      DLM_LOCK_RES_IN_PROGRESS)));
361 
362 converting:
363 	if (list_empty(&res->converting))
364 		goto blocked;
365 	mlog(0, "%s: res %.*s has locks on the convert queue\n", dlm->name,
366 	     res->lockname.len, res->lockname.name);
367 
368 	target = list_entry(res->converting.next, struct dlm_lock, list);
369 	if (target->ml.convert_type == LKM_IVMODE) {
370 		mlog(ML_ERROR, "%s: res %.*s converting lock to invalid mode\n",
371 		     dlm->name, res->lockname.len, res->lockname.name);
372 		BUG();
373 	}
374 	list_for_each_entry(lock, &res->granted, list) {
375 		if (lock==target)
376 			continue;
377 		if (!dlm_lock_compatible(lock->ml.type,
378 					 target->ml.convert_type)) {
379 			can_grant = 0;
380 			/* queue the BAST if not already */
381 			if (lock->ml.highest_blocked == LKM_IVMODE) {
382 				__dlm_lockres_reserve_ast(res);
383 				__dlm_queue_bast(dlm, lock);
384 			}
385 			/* update the highest_blocked if needed */
386 			if (lock->ml.highest_blocked < target->ml.convert_type)
387 				lock->ml.highest_blocked =
388 					target->ml.convert_type;
389 		}
390 	}
391 
392 	list_for_each_entry(lock, &res->converting, list) {
393 		if (lock==target)
394 			continue;
395 		if (!dlm_lock_compatible(lock->ml.type,
396 					 target->ml.convert_type)) {
397 			can_grant = 0;
398 			if (lock->ml.highest_blocked == LKM_IVMODE) {
399 				__dlm_lockres_reserve_ast(res);
400 				__dlm_queue_bast(dlm, lock);
401 			}
402 			if (lock->ml.highest_blocked < target->ml.convert_type)
403 				lock->ml.highest_blocked =
404 					target->ml.convert_type;
405 		}
406 	}
407 
408 	/* we can convert the lock */
409 	if (can_grant) {
410 		spin_lock(&target->spinlock);
411 		BUG_ON(target->ml.highest_blocked != LKM_IVMODE);
412 
413 		mlog(0, "%s: res %.*s, AST for Converting lock %u:%llu, type "
414 		     "%d => %d, node %u\n", dlm->name, res->lockname.len,
415 		     res->lockname.name,
416 		     dlm_get_lock_cookie_node(be64_to_cpu(target->ml.cookie)),
417 		     dlm_get_lock_cookie_seq(be64_to_cpu(target->ml.cookie)),
418 		     target->ml.type,
419 		     target->ml.convert_type, target->ml.node);
420 
421 		target->ml.type = target->ml.convert_type;
422 		target->ml.convert_type = LKM_IVMODE;
423 		list_move_tail(&target->list, &res->granted);
424 
425 		BUG_ON(!target->lksb);
426 		target->lksb->status = DLM_NORMAL;
427 
428 		spin_unlock(&target->spinlock);
429 
430 		__dlm_lockres_reserve_ast(res);
431 		__dlm_queue_ast(dlm, target);
432 		/* go back and check for more */
433 		goto converting;
434 	}
435 
436 blocked:
437 	if (list_empty(&res->blocked))
438 		goto leave;
439 	target = list_entry(res->blocked.next, struct dlm_lock, list);
440 
441 	list_for_each_entry(lock, &res->granted, list) {
442 		if (lock==target)
443 			continue;
444 		if (!dlm_lock_compatible(lock->ml.type, target->ml.type)) {
445 			can_grant = 0;
446 			if (lock->ml.highest_blocked == LKM_IVMODE) {
447 				__dlm_lockres_reserve_ast(res);
448 				__dlm_queue_bast(dlm, lock);
449 			}
450 			if (lock->ml.highest_blocked < target->ml.type)
451 				lock->ml.highest_blocked = target->ml.type;
452 		}
453 	}
454 
455 	list_for_each_entry(lock, &res->converting, list) {
456 		if (lock==target)
457 			continue;
458 		if (!dlm_lock_compatible(lock->ml.type, target->ml.type)) {
459 			can_grant = 0;
460 			if (lock->ml.highest_blocked == LKM_IVMODE) {
461 				__dlm_lockres_reserve_ast(res);
462 				__dlm_queue_bast(dlm, lock);
463 			}
464 			if (lock->ml.highest_blocked < target->ml.type)
465 				lock->ml.highest_blocked = target->ml.type;
466 		}
467 	}
468 
469 	/* we can grant the blocked lock (only
470 	 * possible if converting list empty) */
471 	if (can_grant) {
472 		spin_lock(&target->spinlock);
473 		BUG_ON(target->ml.highest_blocked != LKM_IVMODE);
474 
475 		mlog(0, "%s: res %.*s, AST for Blocked lock %u:%llu, type %d, "
476 		     "node %u\n", dlm->name, res->lockname.len,
477 		     res->lockname.name,
478 		     dlm_get_lock_cookie_node(be64_to_cpu(target->ml.cookie)),
479 		     dlm_get_lock_cookie_seq(be64_to_cpu(target->ml.cookie)),
480 		     target->ml.type, target->ml.node);
481 
482 		/* target->ml.type is already correct */
483 		list_move_tail(&target->list, &res->granted);
484 
485 		BUG_ON(!target->lksb);
486 		target->lksb->status = DLM_NORMAL;
487 
488 		spin_unlock(&target->spinlock);
489 
490 		__dlm_lockres_reserve_ast(res);
491 		__dlm_queue_ast(dlm, target);
492 		/* go back and check for more */
493 		goto converting;
494 	}
495 
496 leave:
497 	return;
498 }
499 
500 /* must have NO locks when calling this with res !=NULL * */
501 void dlm_kick_thread(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
502 {
503 	if (res) {
504 		spin_lock(&dlm->spinlock);
505 		spin_lock(&res->spinlock);
506 		__dlm_dirty_lockres(dlm, res);
507 		spin_unlock(&res->spinlock);
508 		spin_unlock(&dlm->spinlock);
509 	}
510 	wake_up(&dlm->dlm_thread_wq);
511 }
512 
513 void __dlm_dirty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
514 {
515 	assert_spin_locked(&dlm->spinlock);
516 	assert_spin_locked(&res->spinlock);
517 
518 	/* don't shuffle secondary queues */
519 	if (res->owner == dlm->node_num) {
520 		if (res->state & (DLM_LOCK_RES_MIGRATING |
521 				  DLM_LOCK_RES_BLOCK_DIRTY))
522 		    return;
523 
524 		if (list_empty(&res->dirty)) {
525 			/* ref for dirty_list */
526 			dlm_lockres_get(res);
527 			list_add_tail(&res->dirty, &dlm->dirty_list);
528 			res->state |= DLM_LOCK_RES_DIRTY;
529 		}
530 	}
531 
532 	mlog(0, "%s: res %.*s\n", dlm->name, res->lockname.len,
533 	     res->lockname.name);
534 }
535 
536 
537 /* Launch the NM thread for the mounted volume */
538 int dlm_launch_thread(struct dlm_ctxt *dlm)
539 {
540 	mlog(0, "Starting dlm_thread...\n");
541 
542 	dlm->dlm_thread_task = kthread_run(dlm_thread, dlm, "dlm-%s",
543 			dlm->name);
544 	if (IS_ERR(dlm->dlm_thread_task)) {
545 		mlog_errno(PTR_ERR(dlm->dlm_thread_task));
546 		dlm->dlm_thread_task = NULL;
547 		return -EINVAL;
548 	}
549 
550 	return 0;
551 }
552 
553 void dlm_complete_thread(struct dlm_ctxt *dlm)
554 {
555 	if (dlm->dlm_thread_task) {
556 		mlog(ML_KTHREAD, "Waiting for dlm thread to exit\n");
557 		kthread_stop(dlm->dlm_thread_task);
558 		dlm->dlm_thread_task = NULL;
559 	}
560 }
561 
562 static int dlm_dirty_list_empty(struct dlm_ctxt *dlm)
563 {
564 	int empty;
565 
566 	spin_lock(&dlm->spinlock);
567 	empty = list_empty(&dlm->dirty_list);
568 	spin_unlock(&dlm->spinlock);
569 
570 	return empty;
571 }
572 
573 static void dlm_flush_asts(struct dlm_ctxt *dlm)
574 {
575 	int ret;
576 	struct dlm_lock *lock;
577 	struct dlm_lock_resource *res;
578 	u8 hi;
579 
580 	spin_lock(&dlm->ast_lock);
581 	while (!list_empty(&dlm->pending_asts)) {
582 		lock = list_entry(dlm->pending_asts.next,
583 				  struct dlm_lock, ast_list);
584 		/* get an extra ref on lock */
585 		dlm_lock_get(lock);
586 		res = lock->lockres;
587 		mlog(0, "%s: res %.*s, Flush AST for lock %u:%llu, type %d, "
588 		     "node %u\n", dlm->name, res->lockname.len,
589 		     res->lockname.name,
590 		     dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)),
591 		     dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)),
592 		     lock->ml.type, lock->ml.node);
593 
594 		BUG_ON(!lock->ast_pending);
595 
596 		/* remove from list (including ref) */
597 		list_del_init(&lock->ast_list);
598 		dlm_lock_put(lock);
599 		spin_unlock(&dlm->ast_lock);
600 
601 		if (lock->ml.node != dlm->node_num) {
602 			ret = dlm_do_remote_ast(dlm, res, lock);
603 			if (ret < 0)
604 				mlog_errno(ret);
605 		} else
606 			dlm_do_local_ast(dlm, res, lock);
607 
608 		spin_lock(&dlm->ast_lock);
609 
610 		/* possible that another ast was queued while
611 		 * we were delivering the last one */
612 		if (!list_empty(&lock->ast_list)) {
613 			mlog(0, "%s: res %.*s, AST queued while flushing last "
614 			     "one\n", dlm->name, res->lockname.len,
615 			     res->lockname.name);
616 		} else
617 			lock->ast_pending = 0;
618 
619 		/* drop the extra ref.
620 		 * this may drop it completely. */
621 		dlm_lock_put(lock);
622 		dlm_lockres_release_ast(dlm, res);
623 	}
624 
625 	while (!list_empty(&dlm->pending_basts)) {
626 		lock = list_entry(dlm->pending_basts.next,
627 				  struct dlm_lock, bast_list);
628 		/* get an extra ref on lock */
629 		dlm_lock_get(lock);
630 		res = lock->lockres;
631 
632 		BUG_ON(!lock->bast_pending);
633 
634 		/* get the highest blocked lock, and reset */
635 		spin_lock(&lock->spinlock);
636 		BUG_ON(lock->ml.highest_blocked <= LKM_IVMODE);
637 		hi = lock->ml.highest_blocked;
638 		lock->ml.highest_blocked = LKM_IVMODE;
639 		spin_unlock(&lock->spinlock);
640 
641 		/* remove from list (including ref) */
642 		list_del_init(&lock->bast_list);
643 		dlm_lock_put(lock);
644 		spin_unlock(&dlm->ast_lock);
645 
646 		mlog(0, "%s: res %.*s, Flush BAST for lock %u:%llu, "
647 		     "blocked %d, node %u\n",
648 		     dlm->name, res->lockname.len, res->lockname.name,
649 		     dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)),
650 		     dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)),
651 		     hi, lock->ml.node);
652 
653 		if (lock->ml.node != dlm->node_num) {
654 			ret = dlm_send_proxy_bast(dlm, res, lock, hi);
655 			if (ret < 0)
656 				mlog_errno(ret);
657 		} else
658 			dlm_do_local_bast(dlm, res, lock, hi);
659 
660 		spin_lock(&dlm->ast_lock);
661 
662 		/* possible that another bast was queued while
663 		 * we were delivering the last one */
664 		if (!list_empty(&lock->bast_list)) {
665 			mlog(0, "%s: res %.*s, BAST queued while flushing last "
666 			     "one\n", dlm->name, res->lockname.len,
667 			     res->lockname.name);
668 		} else
669 			lock->bast_pending = 0;
670 
671 		/* drop the extra ref.
672 		 * this may drop it completely. */
673 		dlm_lock_put(lock);
674 		dlm_lockres_release_ast(dlm, res);
675 	}
676 	wake_up(&dlm->ast_wq);
677 	spin_unlock(&dlm->ast_lock);
678 }
679 
680 
681 #define DLM_THREAD_TIMEOUT_MS (4 * 1000)
682 #define DLM_THREAD_MAX_DIRTY  100
683 #define DLM_THREAD_MAX_ASTS   10
684 
685 static int dlm_thread(void *data)
686 {
687 	struct dlm_lock_resource *res;
688 	struct dlm_ctxt *dlm = data;
689 	unsigned long timeout = msecs_to_jiffies(DLM_THREAD_TIMEOUT_MS);
690 
691 	mlog(0, "dlm thread running for %s...\n", dlm->name);
692 
693 	while (!kthread_should_stop()) {
694 		int n = DLM_THREAD_MAX_DIRTY;
695 
696 		/* dlm_shutting_down is very point-in-time, but that
697 		 * doesn't matter as we'll just loop back around if we
698 		 * get false on the leading edge of a state
699 		 * transition. */
700 		dlm_run_purge_list(dlm, dlm_shutting_down(dlm));
701 
702 		/* We really don't want to hold dlm->spinlock while
703 		 * calling dlm_shuffle_lists on each lockres that
704 		 * needs to have its queues adjusted and AST/BASTs
705 		 * run.  So let's pull each entry off the dirty_list
706 		 * and drop dlm->spinlock ASAP.  Once off the list,
707 		 * res->spinlock needs to be taken again to protect
708 		 * the queues while calling dlm_shuffle_lists.  */
709 		spin_lock(&dlm->spinlock);
710 		while (!list_empty(&dlm->dirty_list)) {
711 			int delay = 0;
712 			res = list_entry(dlm->dirty_list.next,
713 					 struct dlm_lock_resource, dirty);
714 
715 			/* peel a lockres off, remove it from the list,
716 			 * unset the dirty flag and drop the dlm lock */
717 			BUG_ON(!res);
718 			dlm_lockres_get(res);
719 
720 			spin_lock(&res->spinlock);
721 			/* We clear the DLM_LOCK_RES_DIRTY state once we shuffle lists below */
722 			list_del_init(&res->dirty);
723 			spin_unlock(&res->spinlock);
724 			spin_unlock(&dlm->spinlock);
725 			/* Drop dirty_list ref */
726 			dlm_lockres_put(res);
727 
728 		 	/* lockres can be re-dirtied/re-added to the
729 			 * dirty_list in this gap, but that is ok */
730 
731 			spin_lock(&dlm->ast_lock);
732 			spin_lock(&res->spinlock);
733 			if (res->owner != dlm->node_num) {
734 				__dlm_print_one_lock_resource(res);
735 				mlog(ML_ERROR, "%s: inprog %d, mig %d, reco %d,"
736 				     " dirty %d\n", dlm->name,
737 				     !!(res->state & DLM_LOCK_RES_IN_PROGRESS),
738 				     !!(res->state & DLM_LOCK_RES_MIGRATING),
739 				     !!(res->state & DLM_LOCK_RES_RECOVERING),
740 				     !!(res->state & DLM_LOCK_RES_DIRTY));
741 			}
742 			BUG_ON(res->owner != dlm->node_num);
743 
744 			/* it is now ok to move lockreses in these states
745 			 * to the dirty list, assuming that they will only be
746 			 * dirty for a short while. */
747 			BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
748 			if (res->state & (DLM_LOCK_RES_IN_PROGRESS |
749 					  DLM_LOCK_RES_RECOVERING |
750 					  DLM_LOCK_RES_RECOVERY_WAITING)) {
751 				/* move it to the tail and keep going */
752 				res->state &= ~DLM_LOCK_RES_DIRTY;
753 				spin_unlock(&res->spinlock);
754 				spin_unlock(&dlm->ast_lock);
755 				mlog(0, "%s: res %.*s, inprogress, delay list "
756 				     "shuffle, state %d\n", dlm->name,
757 				     res->lockname.len, res->lockname.name,
758 				     res->state);
759 				delay = 1;
760 				goto in_progress;
761 			}
762 
763 			/* at this point the lockres is not migrating/
764 			 * recovering/in-progress.  we have the lockres
765 			 * spinlock and do NOT have the dlm lock.
766 			 * safe to reserve/queue asts and run the lists. */
767 
768 			/* called while holding lockres lock */
769 			dlm_shuffle_lists(dlm, res);
770 			res->state &= ~DLM_LOCK_RES_DIRTY;
771 			spin_unlock(&res->spinlock);
772 			spin_unlock(&dlm->ast_lock);
773 
774 			dlm_lockres_calc_usage(dlm, res);
775 
776 in_progress:
777 
778 			spin_lock(&dlm->spinlock);
779 			/* if the lock was in-progress, stick
780 			 * it on the back of the list */
781 			if (delay) {
782 				spin_lock(&res->spinlock);
783 				__dlm_dirty_lockres(dlm, res);
784 				spin_unlock(&res->spinlock);
785 			}
786 			dlm_lockres_put(res);
787 
788 			/* unlikely, but we may need to give time to
789 			 * other tasks */
790 			if (!--n) {
791 				mlog(0, "%s: Throttling dlm thread\n",
792 				     dlm->name);
793 				break;
794 			}
795 		}
796 
797 		spin_unlock(&dlm->spinlock);
798 		dlm_flush_asts(dlm);
799 
800 		/* yield and continue right away if there is more work to do */
801 		if (!n) {
802 			cond_resched();
803 			continue;
804 		}
805 
806 		wait_event_interruptible_timeout(dlm->dlm_thread_wq,
807 						 !dlm_dirty_list_empty(dlm) ||
808 						 kthread_should_stop(),
809 						 timeout);
810 	}
811 
812 	mlog(0, "quitting DLM thread\n");
813 	return 0;
814 }
815