[linux-2.6-block.git] / fs / dlm / recoverd.c

/******************************************************************************
*******************************************************************************
**
**  Copyright (C) Sistina Software, Inc.  1997-2003  All rights reserved.
**  Copyright (C) 2004-2011 Red Hat, Inc.  All rights reserved.
**
**  This copyrighted material is made available to anyone wishing to use,
**  modify, copy, or redistribute it subject to the terms and conditions
**  of the GNU General Public License v.2.
**
*******************************************************************************
******************************************************************************/

#include "dlm_internal.h"
#include "lockspace.h"
#include "member.h"
#include "dir.h"
#include "ast.h"
#include "recover.h"
#include "lowcomms.h"
#include "lock.h"
#include "requestqueue.h"
#include "recoverd.h"


/* If the start for which we're re-enabling locking (seq) has been superseded
   by a newer stop (ls_recover_seq), we need to leave locking disabled.

   We suspend dlm_recv threads here to avoid the race where dlm_recv a) sees
   locking stopped and b) adds a message to the requestqueue, but dlm_recoverd
   enables locking and clears the requestqueue between a and b. */

static int enable_locking(struct dlm_ls *ls, uint64_t seq)
{
	int error = -EINTR;

	down_write(&ls->ls_recv_active);

	spin_lock(&ls->ls_recover_lock);
	if (ls->ls_recover_seq == seq) {
		set_bit(LSFL_RUNNING, &ls->ls_flags);
		/* unblocks processes waiting to enter the dlm */
		up_write(&ls->ls_in_recovery);
		clear_bit(LSFL_RECOVER_LOCK, &ls->ls_flags);
		error = 0;
	}
	spin_unlock(&ls->ls_recover_lock);

	up_write(&ls->ls_recv_active);
	return error;
}

static int ls_recover(struct dlm_ls *ls, struct dlm_recover *rv)
{
	unsigned long start;
	int error, neg = 0;

	log_rinfo(ls, "dlm_recover %llu", (unsigned long long)rv->seq);

	mutex_lock(&ls->ls_recoverd_active);

	dlm_callback_suspend(ls);

	dlm_clear_toss(ls);

	/*
	 * This list of root rsb's will be the basis of most of the recovery
	 * routines.
	 */

	dlm_create_root_list(ls);

	/*
	 * Add or remove nodes from the lockspace's ls_nodes list.
	 */

	error = dlm_recover_members(ls, rv, &neg);
	if (error) {
		log_rinfo(ls, "dlm_recover_members error %d", error);
		goto fail;
	}

	dlm_recover_dir_nodeid(ls);

	ls->ls_recover_dir_sent_res = 0;
	ls->ls_recover_dir_sent_msg = 0;
	ls->ls_recover_locks_in = 0;

	dlm_set_recover_status(ls, DLM_RS_NODES);

	error = dlm_recover_members_wait(ls);
	if (error) {
		log_rinfo(ls, "dlm_recover_members_wait error %d", error);
		goto fail;
	}

	start = jiffies;

	/*
	 * Rebuild our own share of the directory by collecting from all other
	 * nodes their master rsb names that hash to us.
	 */

	error = dlm_recover_directory(ls);
	if (error) {
		log_rinfo(ls, "dlm_recover_directory error %d", error);
		goto fail;
	}

	dlm_set_recover_status(ls, DLM_RS_DIR);

	error = dlm_recover_directory_wait(ls);
	if (error) {
		log_rinfo(ls, "dlm_recover_directory_wait error %d", error);
		goto fail;
	}

	log_rinfo(ls, "dlm_recover_directory %u out %u messages",
		  ls->ls_recover_dir_sent_res, ls->ls_recover_dir_sent_msg);

	/*
	 * We may have outstanding operations that are waiting for a reply from
	 * a failed node.  Mark these to be resent after recovery.  Unlock and
	 * cancel ops can just be completed.
	 */

	dlm_recover_waiters_pre(ls);

	error = dlm_recovery_stopped(ls);
	if (error)
		goto fail;

	if (neg || dlm_no_directory(ls)) {
		/*
		 * Clear lkb's for departed nodes.
		 */

		dlm_recover_purge(ls);

		/*
		 * Get new master nodeid's for rsb's that were mastered on
		 * departed nodes.
		 */

		error = dlm_recover_masters(ls);
		if (error) {
			log_rinfo(ls, "dlm_recover_masters error %d", error);
			goto fail;
		}

		/*
		 * Send our locks on remastered rsb's to the new masters.
		 */

		error = dlm_recover_locks(ls);
		if (error) {
			log_rinfo(ls, "dlm_recover_locks error %d", error);
			goto fail;
		}

		dlm_set_recover_status(ls, DLM_RS_LOCKS);

		error = dlm_recover_locks_wait(ls);
		if (error) {
			log_rinfo(ls, "dlm_recover_locks_wait error %d", error);
			goto fail;
		}

		log_rinfo(ls, "dlm_recover_locks %u in",
			  ls->ls_recover_locks_in);

		/*
		 * Finalize state in master rsb's now that all locks can be
		 * checked.  This includes conversion resolution and lvb
		 * settings.
		 */

		dlm_recover_rsbs(ls);
	} else {
		/*
		 * Other lockspace members may be going through the "neg" steps
		 * while also adding us to the lockspace, in which case they'll
		 * be doing the recover_locks (RS_LOCKS) barrier.
		 */
		dlm_set_recover_status(ls, DLM_RS_LOCKS);

		error = dlm_recover_locks_wait(ls);
		if (error) {
			log_rinfo(ls, "dlm_recover_locks_wait error %d", error);
			goto fail;
		}
	}

	dlm_release_root_list(ls);

	/*
	 * Purge directory-related requests that are saved in requestqueue.
	 * All dir requests from before recovery are invalid now due to the dir
	 * rebuild and will be resent by the requesting nodes.
	 */

	dlm_purge_requestqueue(ls);

	dlm_set_recover_status(ls, DLM_RS_DONE);

	error = dlm_recover_done_wait(ls);
	if (error) {
		log_rinfo(ls, "dlm_recover_done_wait error %d", error);
		goto fail;
	}

	dlm_clear_members_gone(ls);

	dlm_adjust_timeouts(ls);

	dlm_callback_resume(ls);

	error = enable_locking(ls, rv->seq);
	if (error) {
		log_rinfo(ls, "enable_locking error %d", error);
		goto fail;
	}

	error = dlm_process_requestqueue(ls);
	if (error) {
		log_rinfo(ls, "dlm_process_requestqueue error %d", error);
		goto fail;
	}

	error = dlm_recover_waiters_post(ls);
	if (error) {
		log_rinfo(ls, "dlm_recover_waiters_post error %d", error);
		goto fail;
	}

	dlm_recover_grant(ls);

	log_rinfo(ls, "dlm_recover %llu generation %u done: %u ms",
		  (unsigned long long)rv->seq, ls->ls_generation,
		  jiffies_to_msecs(jiffies - start));
	mutex_unlock(&ls->ls_recoverd_active);

	dlm_lsop_recover_done(ls);
	return 0;

 fail:
	dlm_release_root_list(ls);
	log_rinfo(ls, "dlm_recover %llu error %d",
		  (unsigned long long)rv->seq, error);
	mutex_unlock(&ls->ls_recoverd_active);
	return error;
}

/* The dlm_ls_start() that created the rv we take here may already have been
   stopped via dlm_ls_stop(); in that case we need to leave the RECOVERY_STOP
   flag set. */

static void do_ls_recovery(struct dlm_ls *ls)
{
	struct dlm_recover *rv = NULL;

	spin_lock(&ls->ls_recover_lock);
	rv = ls->ls_recover_args;
	ls->ls_recover_args = NULL;
	if (rv && ls->ls_recover_seq == rv->seq)
		clear_bit(LSFL_RECOVER_STOP, &ls->ls_flags);
	spin_unlock(&ls->ls_recover_lock);

	if (rv) {
		ls_recover(ls, rv);
		kfree(rv->nodes);
		kfree(rv);
	}
}

static int dlm_recoverd(void *arg)
{
	struct dlm_ls *ls;

	ls = dlm_find_lockspace_local(arg);
	if (!ls) {
		log_print("dlm_recoverd: no lockspace %p", arg);
		return -1;
	}

	down_write(&ls->ls_in_recovery);
	set_bit(LSFL_RECOVER_LOCK, &ls->ls_flags);
	wake_up(&ls->ls_recover_lock_wait);

	while (!kthread_should_stop()) {
		set_current_state(TASK_INTERRUPTIBLE);
		if (!test_bit(LSFL_RECOVER_WORK, &ls->ls_flags) &&
		    !test_bit(LSFL_RECOVER_DOWN, &ls->ls_flags))
			schedule();
		set_current_state(TASK_RUNNING);

		if (test_and_clear_bit(LSFL_RECOVER_DOWN, &ls->ls_flags)) {
			down_write(&ls->ls_in_recovery);
			set_bit(LSFL_RECOVER_LOCK, &ls->ls_flags);
			wake_up(&ls->ls_recover_lock_wait);
		}

		if (test_and_clear_bit(LSFL_RECOVER_WORK, &ls->ls_flags))
			do_ls_recovery(ls);
	}

	if (test_bit(LSFL_RECOVER_LOCK, &ls->ls_flags))
		up_write(&ls->ls_in_recovery);

	dlm_put_lockspace(ls);
	return 0;
}

int dlm_recoverd_start(struct dlm_ls *ls)
{
	struct task_struct *p;
	int error = 0;

	p = kthread_run(dlm_recoverd, ls, "dlm_recoverd");
	if (IS_ERR(p))
		error = PTR_ERR(p);
	else
                ls->ls_recoverd_task = p;
	return error;
}

void dlm_recoverd_stop(struct dlm_ls *ls)
{
	kthread_stop(ls->ls_recoverd_task);
}

void dlm_recoverd_suspend(struct dlm_ls *ls)
{
	wake_up(&ls->ls_wait_general);
	mutex_lock(&ls->ls_recoverd_active);
}

void dlm_recoverd_resume(struct dlm_ls *ls)
{
	mutex_unlock(&ls->ls_recoverd_active);
}
Commit	Line	Data
e7fd4179 DT	1	/******************************************************************************
	2	*******************************************************************************
	3	**
	4	** Copyright (C) Sistina Software, Inc. 1997-2003 All rights reserved.
60f98d18	5	** Copyright (C) 2004-2011 Red Hat, Inc. All rights reserved.
e7fd4179 DT	6	**
	7	** This copyrighted material is made available to anyone wishing to use,
	8	** modify, copy, or redistribute it subject to the terms and conditions
	9	** of the GNU General Public License v.2.
	10	**
	11	*******************************************************************************
	12	******************************************************************************/
	13
	14	#include "dlm_internal.h"
	15	#include "lockspace.h"
	16	#include "member.h"
	17	#include "dir.h"
	18	#include "ast.h"
	19	#include "recover.h"
	20	#include "lowcomms.h"
	21	#include "lock.h"
	22	#include "requestqueue.h"
	23	#include "recoverd.h"
	24
	25
	26	/* If the start for which we're re-enabling locking (seq) has been superseded
c36258b5 DT	27	by a newer stop (ls_recover_seq), we need to leave locking disabled.
	28
	29	We suspend dlm_recv threads here to avoid the race where dlm_recv a) sees
	30	locking stopped and b) adds a message to the requestqueue, but dlm_recoverd
	31	enables locking and clears the requestqueue between a and b. */
e7fd4179 DT	32
	33	static int enable_locking(struct dlm_ls *ls, uint64_t seq)
	34	{
	35	int error = -EINTR;
	36
c36258b5 DT	37	down_write(&ls->ls_recv_active);
c36258b5 DT	38
e7fd4179 DT	39	spin_lock(&ls->ls_recover_lock);
	40	if (ls->ls_recover_seq == seq) {
	41	set_bit(LSFL_RUNNING, &ls->ls_flags);
c36258b5	42	/* unblocks processes waiting to enter the dlm */
e7fd4179	43	up_write(&ls->ls_in_recovery);
475f230c	44	clear_bit(LSFL_RECOVER_LOCK, &ls->ls_flags);
e7fd4179 DT	45	error = 0;
	46	}
	47	spin_unlock(&ls->ls_recover_lock);
c36258b5 DT	48
c36258b5 DT	49	up_write(&ls->ls_recv_active);
e7fd4179 DT	50	return error;
	51	}
	52
	53	static int ls_recover(struct dlm_ls ls, struct dlm_recover rv)
	54	{
	55	unsigned long start;
	56	int error, neg = 0;
	57
075f0177	58	log_rinfo(ls, "dlm_recover %llu", (unsigned long long)rv->seq);
e7fd4179	59
90135925	60	mutex_lock(&ls->ls_recoverd_active);
e7fd4179	61
23e8e1aa	62	dlm_callback_suspend(ls);
e7fd4179	63
c04fecb4	64	dlm_clear_toss(ls);
e7fd4179 DT	65
e7fd4179 DT	66	/*
85f0379a DT	67	* This list of root rsb's will be the basis of most of the recovery
85f0379a DT	68	* routines.
e7fd4179 DT	69	*/
e7fd4179 DT	70
85f0379a	71	dlm_create_root_list(ls);
e7fd4179 DT	72
	73	/*
	74	* Add or remove nodes from the lockspace's ls_nodes list.
e7fd4179 DT	75	*/
	76
	77	error = dlm_recover_members(ls, rv, &neg);
	78	if (error) {
075f0177	79	log_rinfo(ls, "dlm_recover_members error %d", error);
e7fd4179 DT	80	goto fail;
e7fd4179 DT	81	}
f95a34c6	82
c04fecb4 DT	83	dlm_recover_dir_nodeid(ls);
	84
	85	ls->ls_recover_dir_sent_res = 0;
	86	ls->ls_recover_dir_sent_msg = 0;
4875647a DT	87	ls->ls_recover_locks_in = 0;
4875647a DT	88
f95a34c6 DT	89	dlm_set_recover_status(ls, DLM_RS_NODES);
	90
	91	error = dlm_recover_members_wait(ls);
	92	if (error) {
075f0177	93	log_rinfo(ls, "dlm_recover_members_wait error %d", error);
f95a34c6 DT	94	goto fail;
	95	}
	96
e7fd4179 DT	97	start = jiffies;
	98
	99	/*
	100	* Rebuild our own share of the directory by collecting from all other
	101	* nodes their master rsb names that hash to us.
	102	*/
	103
	104	error = dlm_recover_directory(ls);
	105	if (error) {
075f0177	106	log_rinfo(ls, "dlm_recover_directory error %d", error);
e7fd4179 DT	107	goto fail;
	108	}
	109
f95a34c6	110	dlm_set_recover_status(ls, DLM_RS_DIR);
e7fd4179 DT	111
	112	error = dlm_recover_directory_wait(ls);
	113	if (error) {
075f0177	114	log_rinfo(ls, "dlm_recover_directory_wait error %d", error);
e7fd4179 DT	115	goto fail;
	116	}
	117
075f0177	118	log_rinfo(ls, "dlm_recover_directory %u out %u messages",
c04fecb4 DT	119	ls->ls_recover_dir_sent_res, ls->ls_recover_dir_sent_msg);
c04fecb4 DT	120
e7fd4179 DT	121	/*
	122	* We may have outstanding operations that are waiting for a reply from
	123	* a failed node. Mark these to be resent after recovery. Unlock and
	124	* cancel ops can just be completed.
	125	*/
	126
	127	dlm_recover_waiters_pre(ls);
	128
	129	error = dlm_recovery_stopped(ls);
	130	if (error)
	131	goto fail;
	132
	133	if (neg \|\| dlm_no_directory(ls)) {
	134	/*
	135	* Clear lkb's for departed nodes.
	136	*/
	137
4875647a	138	dlm_recover_purge(ls);
e7fd4179 DT	139
	140	/*
	141	* Get new master nodeid's for rsb's that were mastered on
	142	* departed nodes.
	143	*/
	144
	145	error = dlm_recover_masters(ls);
	146	if (error) {
075f0177	147	log_rinfo(ls, "dlm_recover_masters error %d", error);
e7fd4179 DT	148	goto fail;
	149	}
	150
	151	/*
	152	* Send our locks on remastered rsb's to the new masters.
	153	*/
	154
	155	error = dlm_recover_locks(ls);
	156	if (error) {
075f0177	157	log_rinfo(ls, "dlm_recover_locks error %d", error);
e7fd4179 DT	158	goto fail;
	159	}
	160
f95a34c6 DT	161	dlm_set_recover_status(ls, DLM_RS_LOCKS);
f95a34c6 DT	162
e7fd4179 DT	163	error = dlm_recover_locks_wait(ls);
e7fd4179 DT	164	if (error) {
075f0177	165	log_rinfo(ls, "dlm_recover_locks_wait error %d", error);
e7fd4179 DT	166	goto fail;
	167	}
	168
075f0177	169	log_rinfo(ls, "dlm_recover_locks %u in",
4875647a DT	170	ls->ls_recover_locks_in);
4875647a DT	171
e7fd4179 DT	172	/*
	173	* Finalize state in master rsb's now that all locks can be
	174	* checked. This includes conversion resolution and lvb
	175	* settings.
	176	*/
	177
	178	dlm_recover_rsbs(ls);
91c0dc93 DT	179	} else {
	180	/*
	181	* Other lockspace members may be going through the "neg" steps
	182	* while also adding us to the lockspace, in which case they'll
4b77f2c9	183	* be doing the recover_locks (RS_LOCKS) barrier.
91c0dc93 DT	184	*/
91c0dc93 DT	185	dlm_set_recover_status(ls, DLM_RS_LOCKS);
4b77f2c9 DT	186
	187	error = dlm_recover_locks_wait(ls);
	188	if (error) {
075f0177	189	log_rinfo(ls, "dlm_recover_locks_wait error %d", error);
4b77f2c9 DT	190	goto fail;
4b77f2c9 DT	191	}
e7fd4179 DT	192	}
	193
	194	dlm_release_root_list(ls);
	195
2896ee37 DT	196	/*
	197	* Purge directory-related requests that are saved in requestqueue.
	198	* All dir requests from before recovery are invalid now due to the dir
	199	* rebuild and will be resent by the requesting nodes.
	200	*/
	201
	202	dlm_purge_requestqueue(ls);
	203
e7fd4179	204	dlm_set_recover_status(ls, DLM_RS_DONE);
f95a34c6	205
e7fd4179 DT	206	error = dlm_recover_done_wait(ls);
e7fd4179 DT	207	if (error) {
075f0177	208	log_rinfo(ls, "dlm_recover_done_wait error %d", error);
e7fd4179 DT	209	goto fail;
	210	}
	211
	212	dlm_clear_members_gone(ls);
	213
3ae1acf9 DT	214	dlm_adjust_timeouts(ls);
3ae1acf9 DT	215
23e8e1aa DT	216	dlm_callback_resume(ls);
23e8e1aa DT	217
e7fd4179 DT	218	error = enable_locking(ls, rv->seq);
e7fd4179 DT	219	if (error) {
075f0177	220	log_rinfo(ls, "enable_locking error %d", error);
e7fd4179 DT	221	goto fail;
	222	}
	223
	224	error = dlm_process_requestqueue(ls);
	225	if (error) {
075f0177	226	log_rinfo(ls, "dlm_process_requestqueue error %d", error);
e7fd4179 DT	227	goto fail;
	228	}
	229
	230	error = dlm_recover_waiters_post(ls);
	231	if (error) {
075f0177	232	log_rinfo(ls, "dlm_recover_waiters_post error %d", error);
e7fd4179 DT	233	goto fail;
	234	}
	235
4875647a	236	dlm_recover_grant(ls);
e7fd4179	237
075f0177	238	log_rinfo(ls, "dlm_recover %llu generation %u done: %u ms",
60f98d18	239	(unsigned long long)rv->seq, ls->ls_generation,
e7fd4179	240	jiffies_to_msecs(jiffies - start));
90135925	241	mutex_unlock(&ls->ls_recoverd_active);
e7fd4179	242
60f98d18	243	dlm_lsop_recover_done(ls);
e7fd4179 DT	244	return 0;
	245
	246	fail:
	247	dlm_release_root_list(ls);
075f0177	248	log_rinfo(ls, "dlm_recover %llu error %d",
57adf7ee	249	(unsigned long long)rv->seq, error);
90135925	250	mutex_unlock(&ls->ls_recoverd_active);
e7fd4179 DT	251	return error;
	252	}
	253
2cdc98aa DT	254	/* The dlm_ls_start() that created the rv we take here may already have been
	255	stopped via dlm_ls_stop(); in that case we need to leave the RECOVERY_STOP
	256	flag set. */
	257
e7fd4179 DT	258	static void do_ls_recovery(struct dlm_ls *ls)
	259	{
	260	struct dlm_recover *rv = NULL;
	261
	262	spin_lock(&ls->ls_recover_lock);
	263	rv = ls->ls_recover_args;
	264	ls->ls_recover_args = NULL;
2cdc98aa	265	if (rv && ls->ls_recover_seq == rv->seq)
475f230c	266	clear_bit(LSFL_RECOVER_STOP, &ls->ls_flags);
e7fd4179 DT	267	spin_unlock(&ls->ls_recover_lock);
	268
	269	if (rv) {
	270	ls_recover(ls, rv);
60f98d18	271	kfree(rv->nodes);
e7fd4179 DT	272	kfree(rv);
	273	}
	274	}
	275
	276	static int dlm_recoverd(void *arg)
	277	{
	278	struct dlm_ls *ls;
	279
	280	ls = dlm_find_lockspace_local(arg);
5f88f1ea DT	281	if (!ls) {
	282	log_print("dlm_recoverd: no lockspace %p", arg);
	283	return -1;
	284	}
e7fd4179	285
475f230c DT	286	down_write(&ls->ls_in_recovery);
	287	set_bit(LSFL_RECOVER_LOCK, &ls->ls_flags);
	288	wake_up(&ls->ls_recover_lock_wait);
	289
e7fd4179 DT	290	while (!kthread_should_stop()) {
e7fd4179 DT	291	set_current_state(TASK_INTERRUPTIBLE);
475f230c DT	292	if (!test_bit(LSFL_RECOVER_WORK, &ls->ls_flags) &&
475f230c DT	293	!test_bit(LSFL_RECOVER_DOWN, &ls->ls_flags))
e7fd4179 DT	294	schedule();
	295	set_current_state(TASK_RUNNING);
	296
475f230c DT	297	if (test_and_clear_bit(LSFL_RECOVER_DOWN, &ls->ls_flags)) {
	298	down_write(&ls->ls_in_recovery);
	299	set_bit(LSFL_RECOVER_LOCK, &ls->ls_flags);
	300	wake_up(&ls->ls_recover_lock_wait);
	301	}
	302
	303	if (test_and_clear_bit(LSFL_RECOVER_WORK, &ls->ls_flags))
e7fd4179 DT	304	do_ls_recovery(ls);
	305	}
	306
475f230c DT	307	if (test_bit(LSFL_RECOVER_LOCK, &ls->ls_flags))
	308	up_write(&ls->ls_in_recovery);
	309
e7fd4179 DT	310	dlm_put_lockspace(ls);
	311	return 0;
	312	}
	313
e7fd4179 DT	314	int dlm_recoverd_start(struct dlm_ls *ls)
	315	{
	316	struct task_struct *p;
	317	int error = 0;
	318
	319	p = kthread_run(dlm_recoverd, ls, "dlm_recoverd");
	320	if (IS_ERR(p))
	321	error = PTR_ERR(p);
	322	else
	323	ls->ls_recoverd_task = p;
	324	return error;
	325	}
	326
	327	void dlm_recoverd_stop(struct dlm_ls *ls)
	328	{
	329	kthread_stop(ls->ls_recoverd_task);
	330	}
	331
	332	void dlm_recoverd_suspend(struct dlm_ls *ls)
	333	{
f6db1b8e	334	wake_up(&ls->ls_wait_general);
90135925	335	mutex_lock(&ls->ls_recoverd_active);
e7fd4179 DT	336	}
	337
	338	void dlm_recoverd_resume(struct dlm_ls *ls)
	339	{
90135925	340	mutex_unlock(&ls->ls_recoverd_active);
e7fd4179 DT	341	}
e7fd4179 DT	342