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

// SPDX-License-Identifier: GPL-2.0-only
/******************************************************************************
*******************************************************************************
**
**  Copyright (C) Sistina Software, Inc.  1997-2003  All rights reserved.
**  Copyright (C) 2004-2011 Red Hat, Inc.  All rights reserved.
**
**
*******************************************************************************
******************************************************************************/

#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/* If the start for which we're re-enabling locking (seq) has been superseded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 (1) {
		/*
		 * We call kthread_should_stop() after set_current_state().
		 * This is because it works correctly if kthread_stop() is
		 * called just before set_current_state().
		 */
		set_current_state(TASK_INTERRUPTIBLE);
		if (kthread_should_stop()) {
			set_current_state(TASK_RUNNING);
			break;
		}
		if (!test_bit(LSFL_RECOVER_WORK, &ls->ls_flags) &&
		    !test_bit(LSFL_RECOVER_DOWN, &ls->ls_flags)) {
			if (kthread_should_stop())
				break;
			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
2522fe45	1	// SPDX-License-Identifier: GPL-2.0-only
e7fd4179 DT	2	/******************************************************************************
	3	*******************************************************************************
	4	**
	5	** Copyright (C) Sistina Software, Inc. 1997-2003 All rights reserved.
60f98d18	6	** Copyright (C) 2004-2011 Red Hat, Inc. All rights reserved.
e7fd4179	7	**
e7fd4179 DT	8	**
	9	*******************************************************************************
	10	******************************************************************************/
	11
	12	#include "dlm_internal.h"
	13	#include "lockspace.h"
	14	#include "member.h"
	15	#include "dir.h"
	16	#include "ast.h"
	17	#include "recover.h"
	18	#include "lowcomms.h"
	19	#include "lock.h"
	20	#include "requestqueue.h"
	21	#include "recoverd.h"
	22
	23
	24	/* If the start for which we're re-enabling locking (seq) has been superseded
c36258b5 DT	25	by a newer stop (ls_recover_seq), we need to leave locking disabled.
	26
	27	We suspend dlm_recv threads here to avoid the race where dlm_recv a) sees
	28	locking stopped and b) adds a message to the requestqueue, but dlm_recoverd
	29	enables locking and clears the requestqueue between a and b. */
e7fd4179 DT	30
	31	static int enable_locking(struct dlm_ls *ls, uint64_t seq)
	32	{
	33	int error = -EINTR;
	34
c36258b5 DT	35	down_write(&ls->ls_recv_active);
c36258b5 DT	36
e7fd4179 DT	37	spin_lock(&ls->ls_recover_lock);
	38	if (ls->ls_recover_seq == seq) {
	39	set_bit(LSFL_RUNNING, &ls->ls_flags);
c36258b5	40	/* unblocks processes waiting to enter the dlm */
e7fd4179	41	up_write(&ls->ls_in_recovery);
475f230c	42	clear_bit(LSFL_RECOVER_LOCK, &ls->ls_flags);
e7fd4179 DT	43	error = 0;
	44	}
	45	spin_unlock(&ls->ls_recover_lock);
c36258b5 DT	46
c36258b5 DT	47	up_write(&ls->ls_recv_active);
e7fd4179 DT	48	return error;
	49	}
	50
	51	static int ls_recover(struct dlm_ls ls, struct dlm_recover rv)
	52	{
	53	unsigned long start;
	54	int error, neg = 0;
	55
075f0177	56	log_rinfo(ls, "dlm_recover %llu", (unsigned long long)rv->seq);
e7fd4179	57
90135925	58	mutex_lock(&ls->ls_recoverd_active);
e7fd4179	59
23e8e1aa	60	dlm_callback_suspend(ls);
e7fd4179	61
c04fecb4	62	dlm_clear_toss(ls);
e7fd4179 DT	63
e7fd4179 DT	64	/*
85f0379a DT	65	* This list of root rsb's will be the basis of most of the recovery
85f0379a DT	66	* routines.
e7fd4179 DT	67	*/
e7fd4179 DT	68
85f0379a	69	dlm_create_root_list(ls);
e7fd4179 DT	70
	71	/*
	72	* Add or remove nodes from the lockspace's ls_nodes list.
e7fd4179 DT	73	*/
	74
	75	error = dlm_recover_members(ls, rv, &neg);
	76	if (error) {
075f0177	77	log_rinfo(ls, "dlm_recover_members error %d", error);
e7fd4179 DT	78	goto fail;
e7fd4179 DT	79	}
f95a34c6	80
c04fecb4 DT	81	dlm_recover_dir_nodeid(ls);
	82
	83	ls->ls_recover_dir_sent_res = 0;
	84	ls->ls_recover_dir_sent_msg = 0;
4875647a DT	85	ls->ls_recover_locks_in = 0;
4875647a DT	86
f95a34c6 DT	87	dlm_set_recover_status(ls, DLM_RS_NODES);
	88
	89	error = dlm_recover_members_wait(ls);
	90	if (error) {
075f0177	91	log_rinfo(ls, "dlm_recover_members_wait error %d", error);
f95a34c6 DT	92	goto fail;
	93	}
	94
e7fd4179 DT	95	start = jiffies;
	96
	97	/*
	98	* Rebuild our own share of the directory by collecting from all other
	99	* nodes their master rsb names that hash to us.
	100	*/
	101
	102	error = dlm_recover_directory(ls);
	103	if (error) {
075f0177	104	log_rinfo(ls, "dlm_recover_directory error %d", error);
e7fd4179 DT	105	goto fail;
	106	}
	107
f95a34c6	108	dlm_set_recover_status(ls, DLM_RS_DIR);
e7fd4179 DT	109
	110	error = dlm_recover_directory_wait(ls);
	111	if (error) {
075f0177	112	log_rinfo(ls, "dlm_recover_directory_wait error %d", error);
e7fd4179 DT	113	goto fail;
	114	}
	115
075f0177	116	log_rinfo(ls, "dlm_recover_directory %u out %u messages",
c04fecb4 DT	117	ls->ls_recover_dir_sent_res, ls->ls_recover_dir_sent_msg);
c04fecb4 DT	118
e7fd4179 DT	119	/*
	120	* We may have outstanding operations that are waiting for a reply from
	121	* a failed node. Mark these to be resent after recovery. Unlock and
	122	* cancel ops can just be completed.
	123	*/
	124
	125	dlm_recover_waiters_pre(ls);
	126
	127	error = dlm_recovery_stopped(ls);
	128	if (error)
	129	goto fail;
	130
	131	if (neg \|\| dlm_no_directory(ls)) {
	132	/*
	133	* Clear lkb's for departed nodes.
	134	*/
	135
4875647a	136	dlm_recover_purge(ls);
e7fd4179 DT	137
	138	/*
	139	* Get new master nodeid's for rsb's that were mastered on
	140	* departed nodes.
	141	*/
	142
	143	error = dlm_recover_masters(ls);
	144	if (error) {
075f0177	145	log_rinfo(ls, "dlm_recover_masters error %d", error);
e7fd4179 DT	146	goto fail;
	147	}
	148
	149	/*
	150	* Send our locks on remastered rsb's to the new masters.
	151	*/
	152
	153	error = dlm_recover_locks(ls);
	154	if (error) {
075f0177	155	log_rinfo(ls, "dlm_recover_locks error %d", error);
e7fd4179 DT	156	goto fail;
	157	}
	158
f95a34c6 DT	159	dlm_set_recover_status(ls, DLM_RS_LOCKS);
f95a34c6 DT	160
e7fd4179 DT	161	error = dlm_recover_locks_wait(ls);
e7fd4179 DT	162	if (error) {
075f0177	163	log_rinfo(ls, "dlm_recover_locks_wait error %d", error);
e7fd4179 DT	164	goto fail;
	165	}
	166
075f0177	167	log_rinfo(ls, "dlm_recover_locks %u in",
4875647a DT	168	ls->ls_recover_locks_in);
4875647a DT	169
e7fd4179 DT	170	/*
	171	* Finalize state in master rsb's now that all locks can be
	172	* checked. This includes conversion resolution and lvb
	173	* settings.
	174	*/
	175
	176	dlm_recover_rsbs(ls);
91c0dc93 DT	177	} else {
	178	/*
	179	* Other lockspace members may be going through the "neg" steps
	180	* while also adding us to the lockspace, in which case they'll
4b77f2c9	181	* be doing the recover_locks (RS_LOCKS) barrier.
91c0dc93 DT	182	*/
91c0dc93 DT	183	dlm_set_recover_status(ls, DLM_RS_LOCKS);
4b77f2c9 DT	184
	185	error = dlm_recover_locks_wait(ls);
	186	if (error) {
075f0177	187	log_rinfo(ls, "dlm_recover_locks_wait error %d", error);
4b77f2c9 DT	188	goto fail;
4b77f2c9 DT	189	}
e7fd4179 DT	190	}
	191
	192	dlm_release_root_list(ls);
	193
2896ee37 DT	194	/*
	195	* Purge directory-related requests that are saved in requestqueue.
	196	* All dir requests from before recovery are invalid now due to the dir
	197	* rebuild and will be resent by the requesting nodes.
	198	*/
	199
	200	dlm_purge_requestqueue(ls);
	201
e7fd4179	202	dlm_set_recover_status(ls, DLM_RS_DONE);
f95a34c6	203
e7fd4179 DT	204	error = dlm_recover_done_wait(ls);
e7fd4179 DT	205	if (error) {
075f0177	206	log_rinfo(ls, "dlm_recover_done_wait error %d", error);
e7fd4179 DT	207	goto fail;
	208	}
	209
	210	dlm_clear_members_gone(ls);
	211
3ae1acf9 DT	212	dlm_adjust_timeouts(ls);
3ae1acf9 DT	213
23e8e1aa DT	214	dlm_callback_resume(ls);
23e8e1aa DT	215
e7fd4179 DT	216	error = enable_locking(ls, rv->seq);
e7fd4179 DT	217	if (error) {
075f0177	218	log_rinfo(ls, "enable_locking error %d", error);
e7fd4179 DT	219	goto fail;
	220	}
	221
	222	error = dlm_process_requestqueue(ls);
	223	if (error) {
075f0177	224	log_rinfo(ls, "dlm_process_requestqueue error %d", error);
e7fd4179 DT	225	goto fail;
	226	}
	227
	228	error = dlm_recover_waiters_post(ls);
	229	if (error) {
075f0177	230	log_rinfo(ls, "dlm_recover_waiters_post error %d", error);
e7fd4179 DT	231	goto fail;
	232	}
	233
4875647a	234	dlm_recover_grant(ls);
e7fd4179	235
075f0177	236	log_rinfo(ls, "dlm_recover %llu generation %u done: %u ms",
60f98d18	237	(unsigned long long)rv->seq, ls->ls_generation,
e7fd4179	238	jiffies_to_msecs(jiffies - start));
90135925	239	mutex_unlock(&ls->ls_recoverd_active);
e7fd4179	240
60f98d18	241	dlm_lsop_recover_done(ls);
e7fd4179 DT	242	return 0;
	243
	244	fail:
	245	dlm_release_root_list(ls);
075f0177	246	log_rinfo(ls, "dlm_recover %llu error %d",
57adf7ee	247	(unsigned long long)rv->seq, error);
90135925	248	mutex_unlock(&ls->ls_recoverd_active);
e7fd4179 DT	249	return error;
	250	}
	251
2cdc98aa DT	252	/* The dlm_ls_start() that created the rv we take here may already have been
	253	stopped via dlm_ls_stop(); in that case we need to leave the RECOVERY_STOP
	254	flag set. */
	255
e7fd4179 DT	256	static void do_ls_recovery(struct dlm_ls *ls)
	257	{
	258	struct dlm_recover *rv = NULL;
	259
	260	spin_lock(&ls->ls_recover_lock);
	261	rv = ls->ls_recover_args;
	262	ls->ls_recover_args = NULL;
2cdc98aa	263	if (rv && ls->ls_recover_seq == rv->seq)
475f230c	264	clear_bit(LSFL_RECOVER_STOP, &ls->ls_flags);
e7fd4179 DT	265	spin_unlock(&ls->ls_recover_lock);
	266
	267	if (rv) {
	268	ls_recover(ls, rv);
60f98d18	269	kfree(rv->nodes);
e7fd4179 DT	270	kfree(rv);
	271	}
	272	}
	273
	274	static int dlm_recoverd(void *arg)
	275	{
	276	struct dlm_ls *ls;
	277
	278	ls = dlm_find_lockspace_local(arg);
5f88f1ea DT	279	if (!ls) {
	280	log_print("dlm_recoverd: no lockspace %p", arg);
	281	return -1;
	282	}
e7fd4179	283
475f230c DT	284	down_write(&ls->ls_in_recovery);
	285	set_bit(LSFL_RECOVER_LOCK, &ls->ls_flags);
	286	wake_up(&ls->ls_recover_lock_wait);
	287
e412f920	288	while (1) {
	289	/*
	290	* We call kthread_should_stop() after set_current_state().
	291	* This is because it works correctly if kthread_stop() is
	292	* called just before set_current_state().
	293	*/
e7fd4179	294	set_current_state(TASK_INTERRUPTIBLE);
e412f920	295	if (kthread_should_stop()) {
	296	set_current_state(TASK_RUNNING);
	297	break;
	298	}
475f230c	299	if (!test_bit(LSFL_RECOVER_WORK, &ls->ls_flags) &&
9e1b0211 GJ	300	!test_bit(LSFL_RECOVER_DOWN, &ls->ls_flags)) {
	301	if (kthread_should_stop())
	302	break;
e7fd4179	303	schedule();
9e1b0211	304	}
e7fd4179 DT	305	set_current_state(TASK_RUNNING);
e7fd4179 DT	306
475f230c DT	307	if (test_and_clear_bit(LSFL_RECOVER_DOWN, &ls->ls_flags)) {
	308	down_write(&ls->ls_in_recovery);
	309	set_bit(LSFL_RECOVER_LOCK, &ls->ls_flags);
	310	wake_up(&ls->ls_recover_lock_wait);
	311	}
	312
	313	if (test_and_clear_bit(LSFL_RECOVER_WORK, &ls->ls_flags))
e7fd4179 DT	314	do_ls_recovery(ls);
	315	}
	316
475f230c DT	317	if (test_bit(LSFL_RECOVER_LOCK, &ls->ls_flags))
	318	up_write(&ls->ls_in_recovery);
	319
e7fd4179 DT	320	dlm_put_lockspace(ls);
	321	return 0;
	322	}
	323
e7fd4179 DT	324	int dlm_recoverd_start(struct dlm_ls *ls)
	325	{
	326	struct task_struct *p;
	327	int error = 0;
	328
	329	p = kthread_run(dlm_recoverd, ls, "dlm_recoverd");
	330	if (IS_ERR(p))
	331	error = PTR_ERR(p);
	332	else
	333	ls->ls_recoverd_task = p;
	334	return error;
	335	}
	336
	337	void dlm_recoverd_stop(struct dlm_ls *ls)
	338	{
	339	kthread_stop(ls->ls_recoverd_task);
	340	}
	341
	342	void dlm_recoverd_suspend(struct dlm_ls *ls)
	343	{
f6db1b8e	344	wake_up(&ls->ls_wait_general);
90135925	345	mutex_lock(&ls->ls_recoverd_active);
e7fd4179 DT	346	}
	347
	348	void dlm_recoverd_resume(struct dlm_ls *ls)
	349	{
90135925	350	mutex_unlock(&ls->ls_recoverd_active);
e7fd4179 DT	351	}
e7fd4179 DT	352