linux/fs/dlm/recoverd.c

381 lines
8.8 KiB
C
Raw Normal View History

// 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/* If the start for which we're re-enabling locking (seq) has been superseded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.
*
* Due to the fact that we must report all membership changes to lsops
* or midcomms layer, it is not permitted to abort ls_recover() until
* this is done.
*/
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;
dlm: fixes for nodir mode The "nodir" mode (statically assign master nodes instead of using the resource directory) has always been highly experimental, and never seriously used. This commit fixes a number of problems, making nodir much more usable. - Major change to recovery: recover all locks and restart all in-progress operations after recovery. In some cases it's not possible to know which in-progess locks to recover, so recover all. (Most require recovery in nodir mode anyway since rehashing changes most master nodes.) - Change the way nodir mode is enabled, from a command line mount arg passed through gfs2, into a sysfs file managed by dlm_controld, consistent with the other config settings. - Allow recovering MSTCPY locks on an rsb that has not yet been turned into a master copy. - Ignore RCOM_LOCK and RCOM_LOCK_REPLY recovery messages from a previous, aborted recovery cycle. Base this on the local recovery status not being in the state where any nodes should be sending LOCK messages for the current recovery cycle. - Hold rsb lock around dlm_purge_mstcpy_locks() because it may run concurrently with dlm_recover_master_copy(). - Maintain highbast on process-copy lkb's (in addition to the master as is usual), because the lkb can switch back and forth between being a master and being a process copy as the master node changes in recovery. - When recovering MSTCPY locks, flag rsb's that have non-empty convert or waiting queues for granting at the end of recovery. (Rename flag from LOCKS_PURGED to RECOVER_GRANT and similar for the recovery function, because it's not only resources with purged locks that need grant a grant attempt.) - Replace a couple of unnecessary assertion panics with error messages. Signed-off-by: David Teigland <teigland@redhat.com>
2012-04-27 04:54:29 +08:00
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);
if (dlm_recovery_stopped(ls)) {
error = -EINTR;
goto fail;
}
if (neg || dlm_no_directory(ls)) {
/*
* Clear lkb's for departed nodes.
*/
dlm: fixes for nodir mode The "nodir" mode (statically assign master nodes instead of using the resource directory) has always been highly experimental, and never seriously used. This commit fixes a number of problems, making nodir much more usable. - Major change to recovery: recover all locks and restart all in-progress operations after recovery. In some cases it's not possible to know which in-progess locks to recover, so recover all. (Most require recovery in nodir mode anyway since rehashing changes most master nodes.) - Change the way nodir mode is enabled, from a command line mount arg passed through gfs2, into a sysfs file managed by dlm_controld, consistent with the other config settings. - Allow recovering MSTCPY locks on an rsb that has not yet been turned into a master copy. - Ignore RCOM_LOCK and RCOM_LOCK_REPLY recovery messages from a previous, aborted recovery cycle. Base this on the local recovery status not being in the state where any nodes should be sending LOCK messages for the current recovery cycle. - Hold rsb lock around dlm_purge_mstcpy_locks() because it may run concurrently with dlm_recover_master_copy(). - Maintain highbast on process-copy lkb's (in addition to the master as is usual), because the lkb can switch back and forth between being a master and being a process copy as the master node changes in recovery. - When recovering MSTCPY locks, flag rsb's that have non-empty convert or waiting queues for granting at the end of recovery. (Rename flag from LOCKS_PURGED to RECOVER_GRANT and similar for the recovery function, because it's not only resources with purged locks that need grant a grant attempt.) - Replace a couple of unnecessary assertion panics with error messages. Signed-off-by: David Teigland <teigland@redhat.com>
2012-04-27 04:54:29 +08:00
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",
dlm: fixes for nodir mode The "nodir" mode (statically assign master nodes instead of using the resource directory) has always been highly experimental, and never seriously used. This commit fixes a number of problems, making nodir much more usable. - Major change to recovery: recover all locks and restart all in-progress operations after recovery. In some cases it's not possible to know which in-progess locks to recover, so recover all. (Most require recovery in nodir mode anyway since rehashing changes most master nodes.) - Change the way nodir mode is enabled, from a command line mount arg passed through gfs2, into a sysfs file managed by dlm_controld, consistent with the other config settings. - Allow recovering MSTCPY locks on an rsb that has not yet been turned into a master copy. - Ignore RCOM_LOCK and RCOM_LOCK_REPLY recovery messages from a previous, aborted recovery cycle. Base this on the local recovery status not being in the state where any nodes should be sending LOCK messages for the current recovery cycle. - Hold rsb lock around dlm_purge_mstcpy_locks() because it may run concurrently with dlm_recover_master_copy(). - Maintain highbast on process-copy lkb's (in addition to the master as is usual), because the lkb can switch back and forth between being a master and being a process copy as the master node changes in recovery. - When recovering MSTCPY locks, flag rsb's that have non-empty convert or waiting queues for granting at the end of recovery. (Rename flag from LOCKS_PURGED to RECOVER_GRANT and similar for the recovery function, because it's not only resources with purged locks that need grant a grant attempt.) - Replace a couple of unnecessary assertion panics with error messages. Signed-off-by: David Teigland <teigland@redhat.com>
2012-04-27 04:54:29 +08:00
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: fixes for nodir mode The "nodir" mode (statically assign master nodes instead of using the resource directory) has always been highly experimental, and never seriously used. This commit fixes a number of problems, making nodir much more usable. - Major change to recovery: recover all locks and restart all in-progress operations after recovery. In some cases it's not possible to know which in-progess locks to recover, so recover all. (Most require recovery in nodir mode anyway since rehashing changes most master nodes.) - Change the way nodir mode is enabled, from a command line mount arg passed through gfs2, into a sysfs file managed by dlm_controld, consistent with the other config settings. - Allow recovering MSTCPY locks on an rsb that has not yet been turned into a master copy. - Ignore RCOM_LOCK and RCOM_LOCK_REPLY recovery messages from a previous, aborted recovery cycle. Base this on the local recovery status not being in the state where any nodes should be sending LOCK messages for the current recovery cycle. - Hold rsb lock around dlm_purge_mstcpy_locks() because it may run concurrently with dlm_recover_master_copy(). - Maintain highbast on process-copy lkb's (in addition to the master as is usual), because the lkb can switch back and forth between being a master and being a process copy as the master node changes in recovery. - When recovering MSTCPY locks, flag rsb's that have non-empty convert or waiting queues for granting at the end of recovery. (Rename flag from LOCKS_PURGED to RECOVER_GRANT and similar for the recovery function, because it's not only resources with purged locks that need grant a grant attempt.) - Replace a couple of unnecessary assertion panics with error messages. Signed-off-by: David Teigland <teigland@redhat.com>
2012-04-27 04:54:29 +08:00
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);
return 0;
fail:
dlm_release_root_list(ls);
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;
int error;
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) {
error = ls_recover(ls, rv);
switch (error) {
case 0:
ls->ls_recovery_result = 0;
complete(&ls->ls_recovery_done);
dlm_lsop_recover_done(ls);
break;
case -EINTR:
/* if recovery was interrupted -EINTR we wait for the next
* ls_recover() iteration until it hopefully succeeds.
*/
log_rinfo(ls, "%s %llu interrupted and should be queued to run again",
__func__, (unsigned long long)rv->seq);
break;
default:
log_rinfo(ls, "%s %llu error %d", __func__,
(unsigned long long)rv->seq, error);
/* let new_lockspace() get aware of critical error */
ls->ls_recovery_result = error;
complete(&ls->ls_recovery_done);
break;
}
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);
}