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/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>
41 
42 
43 #include "cluster/heartbeat.h"
44 #include "cluster/nodemanager.h"
45 #include "cluster/tcp.h"
46 
47 #include "dlmapi.h"
48 #include "dlmcommon.h"
49 #include "dlmdomain.h"
50 
51 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_THREAD)
52 #include "cluster/masklog.h"
53 
54 static int dlm_thread(void *data);
55 static void dlm_flush_asts(struct dlm_ctxt *dlm);
56 
57 #define dlm_lock_is_remote(dlm, lock)     ((lock)->ml.node != (dlm)->node_num)
58 
59 /* will exit holding res->spinlock, but may drop in function */
60 /* waits until flags are cleared on res->state */
__dlm_wait_on_lockres_flags(struct dlm_lock_resource * res,int flags)61 void __dlm_wait_on_lockres_flags(struct dlm_lock_resource *res, int flags)
62 {
63 	DECLARE_WAITQUEUE(wait, current);
64 
65 	assert_spin_locked(&res->spinlock);
66 
67 	add_wait_queue(&res->wq, &wait);
68 repeat:
69 	set_current_state(TASK_UNINTERRUPTIBLE);
70 	if (res->state & flags) {
71 		spin_unlock(&res->spinlock);
72 		schedule();
73 		spin_lock(&res->spinlock);
74 		goto repeat;
75 	}
76 	remove_wait_queue(&res->wq, &wait);
77 	__set_current_state(TASK_RUNNING);
78 }
79 
__dlm_lockres_has_locks(struct dlm_lock_resource * res)80 int __dlm_lockres_has_locks(struct dlm_lock_resource *res)
81 {
82 	if (list_empty(&res->granted) &&
83 	    list_empty(&res->converting) &&
84 	    list_empty(&res->blocked))
85 		return 0;
86 	return 1;
87 }
88 
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. */
__dlm_lockres_unused(struct dlm_lock_resource * res)93 int __dlm_lockres_unused(struct dlm_lock_resource *res)
94 {
95 	int bit;
96 
97 	assert_spin_locked(&res->spinlock);
98 
99 	if (__dlm_lockres_has_locks(res))
100 		return 0;
101 
102 	/* Locks are in the process of being created */
103 	if (res->inflight_locks)
104 		return 0;
105 
106 	if (!list_empty(&res->dirty) || res->state & DLM_LOCK_RES_DIRTY)
107 		return 0;
108 
109 	if (res->state & (DLM_LOCK_RES_RECOVERING|
110 			DLM_LOCK_RES_RECOVERY_WAITING))
111 		return 0;
112 
113 	/* Another node has this resource with this node as the master */
114 	bit = find_next_bit(res->refmap, O2NM_MAX_NODES, 0);
115 	if (bit < O2NM_MAX_NODES)
116 		return 0;
117 
118 	return 1;
119 }
120 
121 
122 /* Call whenever you may have added or deleted something from one of
123  * the lockres queue's. This will figure out whether it belongs on the
124  * unused list or not and does the appropriate thing. */
__dlm_lockres_calc_usage(struct dlm_ctxt * dlm,struct dlm_lock_resource * res)125 void __dlm_lockres_calc_usage(struct dlm_ctxt *dlm,
126 			      struct dlm_lock_resource *res)
127 {
128 	assert_spin_locked(&dlm->spinlock);
129 	assert_spin_locked(&res->spinlock);
130 
131 	if (__dlm_lockres_unused(res)){
132 		if (list_empty(&res->purge)) {
133 			mlog(0, "%s: Adding res %.*s to purge list\n",
134 			     dlm->name, res->lockname.len, res->lockname.name);
135 
136 			res->last_used = jiffies;
137 			dlm_lockres_get(res);
138 			list_add_tail(&res->purge, &dlm->purge_list);
139 			dlm->purge_count++;
140 		}
141 	} else if (!list_empty(&res->purge)) {
142 		mlog(0, "%s: Removing res %.*s from purge list\n",
143 		     dlm->name, res->lockname.len, res->lockname.name);
144 
145 		list_del_init(&res->purge);
146 		dlm_lockres_put(res);
147 		dlm->purge_count--;
148 	}
149 }
150 
dlm_lockres_calc_usage(struct dlm_ctxt * dlm,struct dlm_lock_resource * res)151 void dlm_lockres_calc_usage(struct dlm_ctxt *dlm,
152 			    struct dlm_lock_resource *res)
153 {
154 	spin_lock(&dlm->spinlock);
155 	spin_lock(&res->spinlock);
156 
157 	__dlm_lockres_calc_usage(dlm, res);
158 
159 	spin_unlock(&res->spinlock);
160 	spin_unlock(&dlm->spinlock);
161 }
162 
163 /*
164  * Do the real purge work:
165  *     unhash the lockres, and
166  *     clear flag DLM_LOCK_RES_DROPPING_REF.
167  * It requires dlm and lockres spinlock to be taken.
168  */
__dlm_do_purge_lockres(struct dlm_ctxt * dlm,struct dlm_lock_resource * res)169 void __dlm_do_purge_lockres(struct dlm_ctxt *dlm,
170 		struct dlm_lock_resource *res)
171 {
172 	assert_spin_locked(&dlm->spinlock);
173 	assert_spin_locked(&res->spinlock);
174 
175 	if (!list_empty(&res->purge)) {
176 		mlog(0, "%s: Removing res %.*s from purgelist\n",
177 		     dlm->name, res->lockname.len, res->lockname.name);
178 		list_del_init(&res->purge);
179 		dlm_lockres_put(res);
180 		dlm->purge_count--;
181 	}
182 
183 	if (!__dlm_lockres_unused(res)) {
184 		mlog(ML_ERROR, "%s: res %.*s in use after deref\n",
185 		     dlm->name, res->lockname.len, res->lockname.name);
186 		__dlm_print_one_lock_resource(res);
187 		BUG();
188 	}
189 
190 	__dlm_unhash_lockres(dlm, res);
191 
192 	spin_lock(&dlm->track_lock);
193 	if (!list_empty(&res->tracking))
194 		list_del_init(&res->tracking);
195 	else {
196 		mlog(ML_ERROR, "%s: Resource %.*s not on the Tracking list\n",
197 		     dlm->name, res->lockname.len, res->lockname.name);
198 		__dlm_print_one_lock_resource(res);
199 	}
200 	spin_unlock(&dlm->track_lock);
201 
202 	/*
203 	 * lockres is not in the hash now. drop the flag and wake up
204 	 * any processes waiting in dlm_get_lock_resource.
205 	 */
206 	res->state &= ~DLM_LOCK_RES_DROPPING_REF;
207 }
208 
dlm_purge_lockres(struct dlm_ctxt * dlm,struct dlm_lock_resource * res)209 static void dlm_purge_lockres(struct dlm_ctxt *dlm,
210 			     struct dlm_lock_resource *res)
211 {
212 	int master;
213 	int ret = 0;
214 
215 	assert_spin_locked(&dlm->spinlock);
216 	assert_spin_locked(&res->spinlock);
217 
218 	master = (res->owner == dlm->node_num);
219 
220 	mlog(0, "%s: Purging res %.*s, master %d\n", dlm->name,
221 	     res->lockname.len, res->lockname.name, master);
222 
223 	if (!master) {
224 		if (res->state & DLM_LOCK_RES_DROPPING_REF) {
225 			mlog(ML_NOTICE, "%s: res %.*s already in DLM_LOCK_RES_DROPPING_REF state\n",
226 				dlm->name, res->lockname.len, res->lockname.name);
227 			spin_unlock(&res->spinlock);
228 			return;
229 		}
230 
231 		res->state |= DLM_LOCK_RES_DROPPING_REF;
232 		/* drop spinlock...  retake below */
233 		spin_unlock(&res->spinlock);
234 		spin_unlock(&dlm->spinlock);
235 
236 		spin_lock(&res->spinlock);
237 		/* This ensures that clear refmap is sent after the set */
238 		__dlm_wait_on_lockres_flags(res, DLM_LOCK_RES_SETREF_INPROG);
239 		spin_unlock(&res->spinlock);
240 
241 		/* clear our bit from the master's refmap, ignore errors */
242 		ret = dlm_drop_lockres_ref(dlm, res);
243 		if (ret < 0) {
244 			if (!dlm_is_host_down(ret))
245 				BUG();
246 		}
247 		spin_lock(&dlm->spinlock);
248 		spin_lock(&res->spinlock);
249 	}
250 
251 	if (!list_empty(&res->purge)) {
252 		mlog(0, "%s: Removing res %.*s from purgelist, master %d\n",
253 		     dlm->name, res->lockname.len, res->lockname.name, master);
254 		list_del_init(&res->purge);
255 		dlm_lockres_put(res);
256 		dlm->purge_count--;
257 	}
258 
259 	if (!master && ret == DLM_DEREF_RESPONSE_INPROG) {
260 		mlog(0, "%s: deref %.*s in progress\n",
261 			dlm->name, res->lockname.len, res->lockname.name);
262 		spin_unlock(&res->spinlock);
263 		return;
264 	}
265 
266 	if (!__dlm_lockres_unused(res)) {
267 		mlog(ML_ERROR, "%s: res %.*s in use after deref\n",
268 		     dlm->name, res->lockname.len, res->lockname.name);
269 		__dlm_print_one_lock_resource(res);
270 		BUG();
271 	}
272 
273 	__dlm_unhash_lockres(dlm, res);
274 
275 	spin_lock(&dlm->track_lock);
276 	if (!list_empty(&res->tracking))
277 		list_del_init(&res->tracking);
278 	else {
279 		mlog(ML_ERROR, "Resource %.*s not on the Tracking list\n",
280 				res->lockname.len, res->lockname.name);
281 		__dlm_print_one_lock_resource(res);
282 	}
283 	spin_unlock(&dlm->track_lock);
284 
285 	/* lockres is not in the hash now.  drop the flag and wake up
286 	 * any processes waiting in dlm_get_lock_resource. */
287 	if (!master) {
288 		res->state &= ~DLM_LOCK_RES_DROPPING_REF;
289 		spin_unlock(&res->spinlock);
290 		wake_up(&res->wq);
291 	} else
292 		spin_unlock(&res->spinlock);
293 }
294 
dlm_run_purge_list(struct dlm_ctxt * dlm,int purge_now)295 static void dlm_run_purge_list(struct dlm_ctxt *dlm,
296 			       int purge_now)
297 {
298 	unsigned int run_max, unused;
299 	unsigned long purge_jiffies;
300 	struct dlm_lock_resource *lockres;
301 
302 	spin_lock(&dlm->spinlock);
303 	run_max = dlm->purge_count;
304 
305 	while(run_max && !list_empty(&dlm->purge_list)) {
306 		run_max--;
307 
308 		lockres = list_entry(dlm->purge_list.next,
309 				     struct dlm_lock_resource, purge);
310 
311 		spin_lock(&lockres->spinlock);
312 
313 		purge_jiffies = lockres->last_used +
314 			msecs_to_jiffies(DLM_PURGE_INTERVAL_MS);
315 
316 		/* Make sure that we want to be processing this guy at
317 		 * this time. */
318 		if (!purge_now && time_after(purge_jiffies, jiffies)) {
319 			/* Since resources are added to the purge list
320 			 * in tail order, we can stop at the first
321 			 * unpurgable resource -- anyone added after
322 			 * him will have a greater last_used value */
323 			spin_unlock(&lockres->spinlock);
324 			break;
325 		}
326 
327 		/* Status of the lockres *might* change so double
328 		 * check. If the lockres is unused, holding the dlm
329 		 * spinlock will prevent people from getting and more
330 		 * refs on it. */
331 		unused = __dlm_lockres_unused(lockres);
332 		if (!unused ||
333 		    (lockres->state & DLM_LOCK_RES_MIGRATING) ||
334 		    (lockres->inflight_assert_workers != 0)) {
335 			mlog(0, "%s: res %.*s is in use or being remastered, "
336 			     "used %d, state %d, assert master workers %u\n",
337 			     dlm->name, lockres->lockname.len,
338 			     lockres->lockname.name,
339 			     !unused, lockres->state,
340 			     lockres->inflight_assert_workers);
341 			list_move_tail(&lockres->purge, &dlm->purge_list);
342 			spin_unlock(&lockres->spinlock);
343 			continue;
344 		}
345 
346 		dlm_lockres_get(lockres);
347 
348 		dlm_purge_lockres(dlm, lockres);
349 
350 		dlm_lockres_put(lockres);
351 
352 		/* Avoid adding any scheduling latencies */
353 		cond_resched_lock(&dlm->spinlock);
354 	}
355 
356 	spin_unlock(&dlm->spinlock);
357 }
358 
dlm_shuffle_lists(struct dlm_ctxt * dlm,struct dlm_lock_resource * res)359 static void dlm_shuffle_lists(struct dlm_ctxt *dlm,
360 			      struct dlm_lock_resource *res)
361 {
362 	struct dlm_lock *lock, *target;
363 	int can_grant = 1;
364 
365 	/*
366 	 * Because this function is called with the lockres
367 	 * spinlock, and because we know that it is not migrating/
368 	 * recovering/in-progress, it is fine to reserve asts and
369 	 * basts right before queueing them all throughout
370 	 */
371 	assert_spin_locked(&dlm->ast_lock);
372 	assert_spin_locked(&res->spinlock);
373 	BUG_ON((res->state & (DLM_LOCK_RES_MIGRATING|
374 			      DLM_LOCK_RES_RECOVERING|
375 			      DLM_LOCK_RES_IN_PROGRESS)));
376 
377 converting:
378 	if (list_empty(&res->converting))
379 		goto blocked;
380 	mlog(0, "%s: res %.*s has locks on the convert queue\n", dlm->name,
381 	     res->lockname.len, res->lockname.name);
382 
383 	target = list_entry(res->converting.next, struct dlm_lock, list);
384 	if (target->ml.convert_type == LKM_IVMODE) {
385 		mlog(ML_ERROR, "%s: res %.*s converting lock to invalid mode\n",
386 		     dlm->name, res->lockname.len, res->lockname.name);
387 		BUG();
388 	}
389 	list_for_each_entry(lock, &res->granted, list) {
390 		if (lock==target)
391 			continue;
392 		if (!dlm_lock_compatible(lock->ml.type,
393 					 target->ml.convert_type)) {
394 			can_grant = 0;
395 			/* queue the BAST if not already */
396 			if (lock->ml.highest_blocked == LKM_IVMODE) {
397 				__dlm_lockres_reserve_ast(res);
398 				__dlm_queue_bast(dlm, lock);
399 			}
400 			/* update the highest_blocked if needed */
401 			if (lock->ml.highest_blocked < target->ml.convert_type)
402 				lock->ml.highest_blocked =
403 					target->ml.convert_type;
404 		}
405 	}
406 
407 	list_for_each_entry(lock, &res->converting, list) {
408 		if (lock==target)
409 			continue;
410 		if (!dlm_lock_compatible(lock->ml.type,
411 					 target->ml.convert_type)) {
412 			can_grant = 0;
413 			if (lock->ml.highest_blocked == LKM_IVMODE) {
414 				__dlm_lockres_reserve_ast(res);
415 				__dlm_queue_bast(dlm, lock);
416 			}
417 			if (lock->ml.highest_blocked < target->ml.convert_type)
418 				lock->ml.highest_blocked =
419 					target->ml.convert_type;
420 		}
421 	}
422 
423 	/* we can convert the lock */
424 	if (can_grant) {
425 		spin_lock(&target->spinlock);
426 		BUG_ON(target->ml.highest_blocked != LKM_IVMODE);
427 
428 		mlog(0, "%s: res %.*s, AST for Converting lock %u:%llu, type "
429 		     "%d => %d, node %u\n", dlm->name, res->lockname.len,
430 		     res->lockname.name,
431 		     dlm_get_lock_cookie_node(be64_to_cpu(target->ml.cookie)),
432 		     dlm_get_lock_cookie_seq(be64_to_cpu(target->ml.cookie)),
433 		     target->ml.type,
434 		     target->ml.convert_type, target->ml.node);
435 
436 		target->ml.type = target->ml.convert_type;
437 		target->ml.convert_type = LKM_IVMODE;
438 		list_move_tail(&target->list, &res->granted);
439 
440 		BUG_ON(!target->lksb);
441 		target->lksb->status = DLM_NORMAL;
442 
443 		spin_unlock(&target->spinlock);
444 
445 		__dlm_lockres_reserve_ast(res);
446 		__dlm_queue_ast(dlm, target);
447 		/* go back and check for more */
448 		goto converting;
449 	}
450 
451 blocked:
452 	if (list_empty(&res->blocked))
453 		goto leave;
454 	target = list_entry(res->blocked.next, struct dlm_lock, list);
455 
456 	list_for_each_entry(lock, &res->granted, list) {
457 		if (lock==target)
458 			continue;
459 		if (!dlm_lock_compatible(lock->ml.type, target->ml.type)) {
460 			can_grant = 0;
461 			if (lock->ml.highest_blocked == LKM_IVMODE) {
462 				__dlm_lockres_reserve_ast(res);
463 				__dlm_queue_bast(dlm, lock);
464 			}
465 			if (lock->ml.highest_blocked < target->ml.type)
466 				lock->ml.highest_blocked = target->ml.type;
467 		}
468 	}
469 
470 	list_for_each_entry(lock, &res->converting, list) {
471 		if (lock==target)
472 			continue;
473 		if (!dlm_lock_compatible(lock->ml.type, target->ml.type)) {
474 			can_grant = 0;
475 			if (lock->ml.highest_blocked == LKM_IVMODE) {
476 				__dlm_lockres_reserve_ast(res);
477 				__dlm_queue_bast(dlm, lock);
478 			}
479 			if (lock->ml.highest_blocked < target->ml.type)
480 				lock->ml.highest_blocked = target->ml.type;
481 		}
482 	}
483 
484 	/* we can grant the blocked lock (only
485 	 * possible if converting list empty) */
486 	if (can_grant) {
487 		spin_lock(&target->spinlock);
488 		BUG_ON(target->ml.highest_blocked != LKM_IVMODE);
489 
490 		mlog(0, "%s: res %.*s, AST for Blocked lock %u:%llu, type %d, "
491 		     "node %u\n", dlm->name, res->lockname.len,
492 		     res->lockname.name,
493 		     dlm_get_lock_cookie_node(be64_to_cpu(target->ml.cookie)),
494 		     dlm_get_lock_cookie_seq(be64_to_cpu(target->ml.cookie)),
495 		     target->ml.type, target->ml.node);
496 
497 		/* target->ml.type is already correct */
498 		list_move_tail(&target->list, &res->granted);
499 
500 		BUG_ON(!target->lksb);
501 		target->lksb->status = DLM_NORMAL;
502 
503 		spin_unlock(&target->spinlock);
504 
505 		__dlm_lockres_reserve_ast(res);
506 		__dlm_queue_ast(dlm, target);
507 		/* go back and check for more */
508 		goto converting;
509 	}
510 
511 leave:
512 	return;
513 }
514 
515 /* must have NO locks when calling this with res !=NULL * */
dlm_kick_thread(struct dlm_ctxt * dlm,struct dlm_lock_resource * res)516 void dlm_kick_thread(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
517 {
518 	if (res) {
519 		spin_lock(&dlm->spinlock);
520 		spin_lock(&res->spinlock);
521 		__dlm_dirty_lockres(dlm, res);
522 		spin_unlock(&res->spinlock);
523 		spin_unlock(&dlm->spinlock);
524 	}
525 	wake_up(&dlm->dlm_thread_wq);
526 }
527 
__dlm_dirty_lockres(struct dlm_ctxt * dlm,struct dlm_lock_resource * res)528 void __dlm_dirty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
529 {
530 	assert_spin_locked(&dlm->spinlock);
531 	assert_spin_locked(&res->spinlock);
532 
533 	/* don't shuffle secondary queues */
534 	if ((res->owner == dlm->node_num)) {
535 		if (res->state & (DLM_LOCK_RES_MIGRATING |
536 				  DLM_LOCK_RES_BLOCK_DIRTY))
537 		    return;
538 
539 		if (list_empty(&res->dirty)) {
540 			/* ref for dirty_list */
541 			dlm_lockres_get(res);
542 			list_add_tail(&res->dirty, &dlm->dirty_list);
543 			res->state |= DLM_LOCK_RES_DIRTY;
544 		}
545 	}
546 
547 	mlog(0, "%s: res %.*s\n", dlm->name, res->lockname.len,
548 	     res->lockname.name);
549 }
550 
551 
552 /* Launch the NM thread for the mounted volume */
dlm_launch_thread(struct dlm_ctxt * dlm)553 int dlm_launch_thread(struct dlm_ctxt *dlm)
554 {
555 	mlog(0, "Starting dlm_thread...\n");
556 
557 	dlm->dlm_thread_task = kthread_run(dlm_thread, dlm, "dlm-%s",
558 			dlm->name);
559 	if (IS_ERR(dlm->dlm_thread_task)) {
560 		mlog_errno(PTR_ERR(dlm->dlm_thread_task));
561 		dlm->dlm_thread_task = NULL;
562 		return -EINVAL;
563 	}
564 
565 	return 0;
566 }
567 
dlm_complete_thread(struct dlm_ctxt * dlm)568 void dlm_complete_thread(struct dlm_ctxt *dlm)
569 {
570 	if (dlm->dlm_thread_task) {
571 		mlog(ML_KTHREAD, "Waiting for dlm thread to exit\n");
572 		kthread_stop(dlm->dlm_thread_task);
573 		dlm->dlm_thread_task = NULL;
574 	}
575 }
576 
dlm_dirty_list_empty(struct dlm_ctxt * dlm)577 static int dlm_dirty_list_empty(struct dlm_ctxt *dlm)
578 {
579 	int empty;
580 
581 	spin_lock(&dlm->spinlock);
582 	empty = list_empty(&dlm->dirty_list);
583 	spin_unlock(&dlm->spinlock);
584 
585 	return empty;
586 }
587 
dlm_flush_asts(struct dlm_ctxt * dlm)588 static void dlm_flush_asts(struct dlm_ctxt *dlm)
589 {
590 	int ret;
591 	struct dlm_lock *lock;
592 	struct dlm_lock_resource *res;
593 	u8 hi;
594 
595 	spin_lock(&dlm->ast_lock);
596 	while (!list_empty(&dlm->pending_asts)) {
597 		lock = list_entry(dlm->pending_asts.next,
598 				  struct dlm_lock, ast_list);
599 		/* get an extra ref on lock */
600 		dlm_lock_get(lock);
601 		res = lock->lockres;
602 		mlog(0, "%s: res %.*s, Flush AST for lock %u:%llu, type %d, "
603 		     "node %u\n", dlm->name, res->lockname.len,
604 		     res->lockname.name,
605 		     dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)),
606 		     dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)),
607 		     lock->ml.type, lock->ml.node);
608 
609 		BUG_ON(!lock->ast_pending);
610 
611 		/* remove from list (including ref) */
612 		list_del_init(&lock->ast_list);
613 		dlm_lock_put(lock);
614 		spin_unlock(&dlm->ast_lock);
615 
616 		if (lock->ml.node != dlm->node_num) {
617 			ret = dlm_do_remote_ast(dlm, res, lock);
618 			if (ret < 0)
619 				mlog_errno(ret);
620 		} else
621 			dlm_do_local_ast(dlm, res, lock);
622 
623 		spin_lock(&dlm->ast_lock);
624 
625 		/* possible that another ast was queued while
626 		 * we were delivering the last one */
627 		if (!list_empty(&lock->ast_list)) {
628 			mlog(0, "%s: res %.*s, AST queued while flushing last "
629 			     "one\n", dlm->name, res->lockname.len,
630 			     res->lockname.name);
631 		} else
632 			lock->ast_pending = 0;
633 
634 		/* drop the extra ref.
635 		 * this may drop it completely. */
636 		dlm_lock_put(lock);
637 		dlm_lockres_release_ast(dlm, res);
638 	}
639 
640 	while (!list_empty(&dlm->pending_basts)) {
641 		lock = list_entry(dlm->pending_basts.next,
642 				  struct dlm_lock, bast_list);
643 		/* get an extra ref on lock */
644 		dlm_lock_get(lock);
645 		res = lock->lockres;
646 
647 		BUG_ON(!lock->bast_pending);
648 
649 		/* get the highest blocked lock, and reset */
650 		spin_lock(&lock->spinlock);
651 		BUG_ON(lock->ml.highest_blocked <= LKM_IVMODE);
652 		hi = lock->ml.highest_blocked;
653 		lock->ml.highest_blocked = LKM_IVMODE;
654 		spin_unlock(&lock->spinlock);
655 
656 		/* remove from list (including ref) */
657 		list_del_init(&lock->bast_list);
658 		dlm_lock_put(lock);
659 		spin_unlock(&dlm->ast_lock);
660 
661 		mlog(0, "%s: res %.*s, Flush BAST for lock %u:%llu, "
662 		     "blocked %d, node %u\n",
663 		     dlm->name, res->lockname.len, res->lockname.name,
664 		     dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)),
665 		     dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)),
666 		     hi, lock->ml.node);
667 
668 		if (lock->ml.node != dlm->node_num) {
669 			ret = dlm_send_proxy_bast(dlm, res, lock, hi);
670 			if (ret < 0)
671 				mlog_errno(ret);
672 		} else
673 			dlm_do_local_bast(dlm, res, lock, hi);
674 
675 		spin_lock(&dlm->ast_lock);
676 
677 		/* possible that another bast was queued while
678 		 * we were delivering the last one */
679 		if (!list_empty(&lock->bast_list)) {
680 			mlog(0, "%s: res %.*s, BAST queued while flushing last "
681 			     "one\n", dlm->name, res->lockname.len,
682 			     res->lockname.name);
683 		} else
684 			lock->bast_pending = 0;
685 
686 		/* drop the extra ref.
687 		 * this may drop it completely. */
688 		dlm_lock_put(lock);
689 		dlm_lockres_release_ast(dlm, res);
690 	}
691 	wake_up(&dlm->ast_wq);
692 	spin_unlock(&dlm->ast_lock);
693 }
694 
695 
696 #define DLM_THREAD_TIMEOUT_MS (4 * 1000)
697 #define DLM_THREAD_MAX_DIRTY  100
698 #define DLM_THREAD_MAX_ASTS   10
699 
dlm_thread(void * data)700 static int dlm_thread(void *data)
701 {
702 	struct dlm_lock_resource *res;
703 	struct dlm_ctxt *dlm = data;
704 	unsigned long timeout = msecs_to_jiffies(DLM_THREAD_TIMEOUT_MS);
705 
706 	mlog(0, "dlm thread running for %s...\n", dlm->name);
707 
708 	while (!kthread_should_stop()) {
709 		int n = DLM_THREAD_MAX_DIRTY;
710 
711 		/* dlm_shutting_down is very point-in-time, but that
712 		 * doesn't matter as we'll just loop back around if we
713 		 * get false on the leading edge of a state
714 		 * transition. */
715 		dlm_run_purge_list(dlm, dlm_shutting_down(dlm));
716 
717 		/* We really don't want to hold dlm->spinlock while
718 		 * calling dlm_shuffle_lists on each lockres that
719 		 * needs to have its queues adjusted and AST/BASTs
720 		 * run.  So let's pull each entry off the dirty_list
721 		 * and drop dlm->spinlock ASAP.  Once off the list,
722 		 * res->spinlock needs to be taken again to protect
723 		 * the queues while calling dlm_shuffle_lists.  */
724 		spin_lock(&dlm->spinlock);
725 		while (!list_empty(&dlm->dirty_list)) {
726 			int delay = 0;
727 			res = list_entry(dlm->dirty_list.next,
728 					 struct dlm_lock_resource, dirty);
729 
730 			/* peel a lockres off, remove it from the list,
731 			 * unset the dirty flag and drop the dlm lock */
732 			BUG_ON(!res);
733 			dlm_lockres_get(res);
734 
735 			spin_lock(&res->spinlock);
736 			/* We clear the DLM_LOCK_RES_DIRTY state once we shuffle lists below */
737 			list_del_init(&res->dirty);
738 			spin_unlock(&res->spinlock);
739 			spin_unlock(&dlm->spinlock);
740 			/* Drop dirty_list ref */
741 			dlm_lockres_put(res);
742 
743 		 	/* lockres can be re-dirtied/re-added to the
744 			 * dirty_list in this gap, but that is ok */
745 
746 			spin_lock(&dlm->ast_lock);
747 			spin_lock(&res->spinlock);
748 			if (res->owner != dlm->node_num) {
749 				__dlm_print_one_lock_resource(res);
750 				mlog(ML_ERROR, "%s: inprog %d, mig %d, reco %d,"
751 				     " dirty %d\n", dlm->name,
752 				     !!(res->state & DLM_LOCK_RES_IN_PROGRESS),
753 				     !!(res->state & DLM_LOCK_RES_MIGRATING),
754 				     !!(res->state & DLM_LOCK_RES_RECOVERING),
755 				     !!(res->state & DLM_LOCK_RES_DIRTY));
756 			}
757 			BUG_ON(res->owner != dlm->node_num);
758 
759 			/* it is now ok to move lockreses in these states
760 			 * to the dirty list, assuming that they will only be
761 			 * dirty for a short while. */
762 			BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
763 			if (res->state & (DLM_LOCK_RES_IN_PROGRESS |
764 					  DLM_LOCK_RES_RECOVERING |
765 					  DLM_LOCK_RES_RECOVERY_WAITING)) {
766 				/* move it to the tail and keep going */
767 				res->state &= ~DLM_LOCK_RES_DIRTY;
768 				spin_unlock(&res->spinlock);
769 				spin_unlock(&dlm->ast_lock);
770 				mlog(0, "%s: res %.*s, inprogress, delay list "
771 				     "shuffle, state %d\n", dlm->name,
772 				     res->lockname.len, res->lockname.name,
773 				     res->state);
774 				delay = 1;
775 				goto in_progress;
776 			}
777 
778 			/* at this point the lockres is not migrating/
779 			 * recovering/in-progress.  we have the lockres
780 			 * spinlock and do NOT have the dlm lock.
781 			 * safe to reserve/queue asts and run the lists. */
782 
783 			/* called while holding lockres lock */
784 			dlm_shuffle_lists(dlm, res);
785 			res->state &= ~DLM_LOCK_RES_DIRTY;
786 			spin_unlock(&res->spinlock);
787 			spin_unlock(&dlm->ast_lock);
788 
789 			dlm_lockres_calc_usage(dlm, res);
790 
791 in_progress:
792 
793 			spin_lock(&dlm->spinlock);
794 			/* if the lock was in-progress, stick
795 			 * it on the back of the list */
796 			if (delay) {
797 				spin_lock(&res->spinlock);
798 				__dlm_dirty_lockres(dlm, res);
799 				spin_unlock(&res->spinlock);
800 			}
801 			dlm_lockres_put(res);
802 
803 			/* unlikely, but we may need to give time to
804 			 * other tasks */
805 			if (!--n) {
806 				mlog(0, "%s: Throttling dlm thread\n",
807 				     dlm->name);
808 				break;
809 			}
810 		}
811 
812 		spin_unlock(&dlm->spinlock);
813 		dlm_flush_asts(dlm);
814 
815 		/* yield and continue right away if there is more work to do */
816 		if (!n) {
817 			cond_resched();
818 			continue;
819 		}
820 
821 		wait_event_interruptible_timeout(dlm->dlm_thread_wq,
822 						 !dlm_dirty_list_empty(dlm) ||
823 						 kthread_should_stop(),
824 						 timeout);
825 	}
826 
827 	mlog(0, "quitting DLM thread\n");
828 	return 0;
829 }
830