2006-01-16 17:50:04 +01:00
|
|
|
/*
|
|
|
|
* Copyright (C) Sistina Software, Inc. 1997-2003 All rights reserved.
|
2006-05-18 21:09:15 +02:00
|
|
|
* Copyright (C) 2004-2006 Red Hat, Inc. All rights reserved.
|
2006-01-16 17:50:04 +01:00
|
|
|
*
|
|
|
|
* This copyrighted material is made available to anyone wishing to use,
|
|
|
|
* modify, copy, or redistribute it subject to the terms and conditions
|
2006-09-01 17:05:15 +02:00
|
|
|
* of the GNU General Public License version 2.
|
2006-01-16 17:50:04 +01:00
|
|
|
*/
|
|
|
|
|
|
|
|
#include <linux/sched.h>
|
|
|
|
#include <linux/slab.h>
|
|
|
|
#include <linux/spinlock.h>
|
|
|
|
#include <linux/completion.h>
|
|
|
|
#include <linux/buffer_head.h>
|
|
|
|
#include <linux/delay.h>
|
|
|
|
#include <linux/sort.h>
|
|
|
|
#include <linux/jhash.h>
|
2006-03-29 21:36:49 +02:00
|
|
|
#include <linux/kallsyms.h>
|
2006-02-27 23:23:27 +01:00
|
|
|
#include <linux/gfs2_ondisk.h>
|
2006-09-12 03:40:30 +02:00
|
|
|
#include <linux/list.h>
|
2006-09-19 07:56:29 +02:00
|
|
|
#include <linux/lm_interface.h>
|
2007-01-17 16:33:23 +01:00
|
|
|
#include <linux/wait.h>
|
2007-03-06 08:10:39 +01:00
|
|
|
#include <linux/module.h>
|
2007-01-29 12:51:45 +01:00
|
|
|
#include <linux/rwsem.h>
|
2006-01-16 17:50:04 +01:00
|
|
|
#include <asm/uaccess.h>
|
2007-03-16 11:26:37 +01:00
|
|
|
#include <linux/seq_file.h>
|
|
|
|
#include <linux/debugfs.h>
|
2007-08-01 14:57:10 +02:00
|
|
|
#include <linux/kthread.h>
|
|
|
|
#include <linux/freezer.h>
|
[GFS2] delay glock demote for a minimum hold time
When a lot of IO, with some distributed mmap IO, is run on a GFS2 filesystem in
a cluster, it will deadlock. The reason is that do_no_page() will repeatedly
call gfs2_sharewrite_nopage(), because each node keeps giving up the glock
too early, and is forced to call unmap_mapping_range(). This bumps the
mapping->truncate_count sequence count, forcing do_no_page() to retry. This
patch institutes a minimum glock hold time a tenth a second. This insures
that even in heavy contention cases, the node has enough time to get some
useful work done before it gives up the glock.
A second issue is that when gfs2_glock_dq() is called from within a page fault
to demote a lock, and the associated page needs to be written out, it will
try to acqire a lock on it, but it has already been locked at a higher level.
This patch puts makes gfs2_glock_dq() use the work queue as well, to avoid this
issue. This is the same patch as Steve Whitehouse originally proposed to fix
this issue, execpt that gfs2_glock_dq() now grabs a reference to the glock
before it queues up the work on it.
Signed-off-by: Benjamin E. Marzinski <bmarzins@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
2007-08-23 20:19:05 +02:00
|
|
|
#include <linux/workqueue.h>
|
|
|
|
#include <linux/jiffies.h>
|
2006-01-16 17:50:04 +01:00
|
|
|
|
|
|
|
#include "gfs2.h"
|
2006-02-27 23:23:27 +01:00
|
|
|
#include "incore.h"
|
2006-01-16 17:50:04 +01:00
|
|
|
#include "glock.h"
|
|
|
|
#include "glops.h"
|
|
|
|
#include "inode.h"
|
|
|
|
#include "lm.h"
|
|
|
|
#include "lops.h"
|
|
|
|
#include "meta_io.h"
|
|
|
|
#include "quota.h"
|
|
|
|
#include "super.h"
|
2006-02-27 23:23:27 +01:00
|
|
|
#include "util.h"
|
2006-01-16 17:50:04 +01:00
|
|
|
|
2006-09-08 19:35:56 +02:00
|
|
|
struct gfs2_gl_hash_bucket {
|
2006-09-12 16:10:01 +02:00
|
|
|
struct hlist_head hb_list;
|
2006-09-08 19:35:56 +02:00
|
|
|
};
|
|
|
|
|
2007-03-16 11:26:37 +01:00
|
|
|
struct glock_iter {
|
|
|
|
int hash; /* hash bucket index */
|
|
|
|
struct gfs2_sbd *sdp; /* incore superblock */
|
|
|
|
struct gfs2_glock *gl; /* current glock struct */
|
|
|
|
struct seq_file *seq; /* sequence file for debugfs */
|
|
|
|
char string[512]; /* scratch space */
|
|
|
|
};
|
|
|
|
|
2006-01-16 17:50:04 +01:00
|
|
|
typedef void (*glock_examiner) (struct gfs2_glock * gl);
|
|
|
|
|
2006-04-28 16:59:12 +02:00
|
|
|
static int gfs2_dump_lockstate(struct gfs2_sbd *sdp);
|
2007-03-23 23:05:15 +01:00
|
|
|
static int dump_glock(struct glock_iter *gi, struct gfs2_glock *gl);
|
2007-03-16 10:40:31 +01:00
|
|
|
static void gfs2_glock_xmote_th(struct gfs2_glock *gl, struct gfs2_holder *gh);
|
2007-01-22 18:15:34 +01:00
|
|
|
static void gfs2_glock_drop_th(struct gfs2_glock *gl);
|
[GFS2] delay glock demote for a minimum hold time
When a lot of IO, with some distributed mmap IO, is run on a GFS2 filesystem in
a cluster, it will deadlock. The reason is that do_no_page() will repeatedly
call gfs2_sharewrite_nopage(), because each node keeps giving up the glock
too early, and is forced to call unmap_mapping_range(). This bumps the
mapping->truncate_count sequence count, forcing do_no_page() to retry. This
patch institutes a minimum glock hold time a tenth a second. This insures
that even in heavy contention cases, the node has enough time to get some
useful work done before it gives up the glock.
A second issue is that when gfs2_glock_dq() is called from within a page fault
to demote a lock, and the associated page needs to be written out, it will
try to acqire a lock on it, but it has already been locked at a higher level.
This patch puts makes gfs2_glock_dq() use the work queue as well, to avoid this
issue. This is the same patch as Steve Whitehouse originally proposed to fix
this issue, execpt that gfs2_glock_dq() now grabs a reference to the glock
before it queues up the work on it.
Signed-off-by: Benjamin E. Marzinski <bmarzins@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
2007-08-23 20:19:05 +02:00
|
|
|
static void run_queue(struct gfs2_glock *gl);
|
|
|
|
|
2007-01-29 12:51:45 +01:00
|
|
|
static DECLARE_RWSEM(gfs2_umount_flush_sem);
|
2007-03-16 11:26:37 +01:00
|
|
|
static struct dentry *gfs2_root;
|
2007-08-01 14:57:10 +02:00
|
|
|
static struct task_struct *scand_process;
|
|
|
|
static unsigned int scand_secs = 5;
|
[GFS2] delay glock demote for a minimum hold time
When a lot of IO, with some distributed mmap IO, is run on a GFS2 filesystem in
a cluster, it will deadlock. The reason is that do_no_page() will repeatedly
call gfs2_sharewrite_nopage(), because each node keeps giving up the glock
too early, and is forced to call unmap_mapping_range(). This bumps the
mapping->truncate_count sequence count, forcing do_no_page() to retry. This
patch institutes a minimum glock hold time a tenth a second. This insures
that even in heavy contention cases, the node has enough time to get some
useful work done before it gives up the glock.
A second issue is that when gfs2_glock_dq() is called from within a page fault
to demote a lock, and the associated page needs to be written out, it will
try to acqire a lock on it, but it has already been locked at a higher level.
This patch puts makes gfs2_glock_dq() use the work queue as well, to avoid this
issue. This is the same patch as Steve Whitehouse originally proposed to fix
this issue, execpt that gfs2_glock_dq() now grabs a reference to the glock
before it queues up the work on it.
Signed-off-by: Benjamin E. Marzinski <bmarzins@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
2007-08-23 20:19:05 +02:00
|
|
|
static struct workqueue_struct *glock_workqueue;
|
2006-04-28 16:59:12 +02:00
|
|
|
|
2006-09-12 16:10:01 +02:00
|
|
|
#define GFS2_GL_HASH_SHIFT 15
|
2006-09-09 22:59:11 +02:00
|
|
|
#define GFS2_GL_HASH_SIZE (1 << GFS2_GL_HASH_SHIFT)
|
|
|
|
#define GFS2_GL_HASH_MASK (GFS2_GL_HASH_SIZE - 1)
|
|
|
|
|
2006-09-07 20:40:21 +02:00
|
|
|
static struct gfs2_gl_hash_bucket gl_hash_table[GFS2_GL_HASH_SIZE];
|
2007-03-23 23:05:15 +01:00
|
|
|
static struct dentry *gfs2_root;
|
2006-09-09 22:59:11 +02:00
|
|
|
|
|
|
|
/*
|
|
|
|
* Despite what you might think, the numbers below are not arbitrary :-)
|
|
|
|
* They are taken from the ipv4 routing hash code, which is well tested
|
|
|
|
* and thus should be nearly optimal. Later on we might tweek the numbers
|
|
|
|
* but for now this should be fine.
|
|
|
|
*
|
|
|
|
* The reason for putting the locks in a separate array from the list heads
|
|
|
|
* is that we can have fewer locks than list heads and save memory. We use
|
|
|
|
* the same hash function for both, but with a different hash mask.
|
|
|
|
*/
|
|
|
|
#if defined(CONFIG_SMP) || defined(CONFIG_DEBUG_SPINLOCK) || \
|
|
|
|
defined(CONFIG_PROVE_LOCKING)
|
|
|
|
|
|
|
|
#ifdef CONFIG_LOCKDEP
|
|
|
|
# define GL_HASH_LOCK_SZ 256
|
|
|
|
#else
|
|
|
|
# if NR_CPUS >= 32
|
|
|
|
# define GL_HASH_LOCK_SZ 4096
|
|
|
|
# elif NR_CPUS >= 16
|
|
|
|
# define GL_HASH_LOCK_SZ 2048
|
|
|
|
# elif NR_CPUS >= 8
|
|
|
|
# define GL_HASH_LOCK_SZ 1024
|
|
|
|
# elif NR_CPUS >= 4
|
|
|
|
# define GL_HASH_LOCK_SZ 512
|
|
|
|
# else
|
|
|
|
# define GL_HASH_LOCK_SZ 256
|
|
|
|
# endif
|
|
|
|
#endif
|
|
|
|
|
|
|
|
/* We never want more locks than chains */
|
|
|
|
#if GFS2_GL_HASH_SIZE < GL_HASH_LOCK_SZ
|
|
|
|
# undef GL_HASH_LOCK_SZ
|
|
|
|
# define GL_HASH_LOCK_SZ GFS2_GL_HASH_SIZE
|
|
|
|
#endif
|
|
|
|
|
|
|
|
static rwlock_t gl_hash_locks[GL_HASH_LOCK_SZ];
|
|
|
|
|
|
|
|
static inline rwlock_t *gl_lock_addr(unsigned int x)
|
|
|
|
{
|
2006-09-10 00:59:27 +02:00
|
|
|
return &gl_hash_locks[x & (GL_HASH_LOCK_SZ-1)];
|
2006-09-09 22:59:11 +02:00
|
|
|
}
|
|
|
|
#else /* not SMP, so no spinlocks required */
|
2006-11-29 07:29:19 +01:00
|
|
|
static inline rwlock_t *gl_lock_addr(unsigned int x)
|
2006-09-09 22:59:11 +02:00
|
|
|
{
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
#endif
|
2006-09-07 20:40:21 +02:00
|
|
|
|
2006-01-16 17:50:04 +01:00
|
|
|
/**
|
|
|
|
* relaxed_state_ok - is a requested lock compatible with the current lock mode?
|
|
|
|
* @actual: the current state of the lock
|
|
|
|
* @requested: the lock state that was requested by the caller
|
|
|
|
* @flags: the modifier flags passed in by the caller
|
|
|
|
*
|
|
|
|
* Returns: 1 if the locks are compatible, 0 otherwise
|
|
|
|
*/
|
|
|
|
|
|
|
|
static inline int relaxed_state_ok(unsigned int actual, unsigned requested,
|
|
|
|
int flags)
|
|
|
|
{
|
|
|
|
if (actual == requested)
|
|
|
|
return 1;
|
|
|
|
|
|
|
|
if (flags & GL_EXACT)
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
if (actual == LM_ST_EXCLUSIVE && requested == LM_ST_SHARED)
|
|
|
|
return 1;
|
|
|
|
|
|
|
|
if (actual != LM_ST_UNLOCKED && (flags & LM_FLAG_ANY))
|
|
|
|
return 1;
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* gl_hash() - Turn glock number into hash bucket number
|
|
|
|
* @lock: The glock number
|
|
|
|
*
|
|
|
|
* Returns: The number of the corresponding hash bucket
|
|
|
|
*/
|
|
|
|
|
2006-09-07 19:12:27 +02:00
|
|
|
static unsigned int gl_hash(const struct gfs2_sbd *sdp,
|
|
|
|
const struct lm_lockname *name)
|
2006-01-16 17:50:04 +01:00
|
|
|
{
|
|
|
|
unsigned int h;
|
|
|
|
|
2006-09-04 18:49:07 +02:00
|
|
|
h = jhash(&name->ln_number, sizeof(u64), 0);
|
2006-01-16 17:50:04 +01:00
|
|
|
h = jhash(&name->ln_type, sizeof(unsigned int), h);
|
2006-09-07 19:12:27 +02:00
|
|
|
h = jhash(&sdp, sizeof(struct gfs2_sbd *), h);
|
2006-01-16 17:50:04 +01:00
|
|
|
h &= GFS2_GL_HASH_MASK;
|
|
|
|
|
|
|
|
return h;
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* glock_free() - Perform a few checks and then release struct gfs2_glock
|
|
|
|
* @gl: The glock to release
|
|
|
|
*
|
|
|
|
* Also calls lock module to release its internal structure for this glock.
|
|
|
|
*
|
|
|
|
*/
|
|
|
|
|
|
|
|
static void glock_free(struct gfs2_glock *gl)
|
|
|
|
{
|
|
|
|
struct gfs2_sbd *sdp = gl->gl_sbd;
|
|
|
|
struct inode *aspace = gl->gl_aspace;
|
|
|
|
|
|
|
|
gfs2_lm_put_lock(sdp, gl->gl_lock);
|
|
|
|
|
|
|
|
if (aspace)
|
|
|
|
gfs2_aspace_put(aspace);
|
|
|
|
|
|
|
|
kmem_cache_free(gfs2_glock_cachep, gl);
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* gfs2_glock_hold() - increment reference count on glock
|
|
|
|
* @gl: The glock to hold
|
|
|
|
*
|
|
|
|
*/
|
|
|
|
|
|
|
|
void gfs2_glock_hold(struct gfs2_glock *gl)
|
|
|
|
{
|
2006-09-13 16:43:37 +02:00
|
|
|
atomic_inc(&gl->gl_ref);
|
2006-01-16 17:50:04 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* gfs2_glock_put() - Decrement reference count on glock
|
|
|
|
* @gl: The glock to put
|
|
|
|
*
|
|
|
|
*/
|
|
|
|
|
|
|
|
int gfs2_glock_put(struct gfs2_glock *gl)
|
|
|
|
{
|
|
|
|
int rv = 0;
|
2006-09-13 16:43:37 +02:00
|
|
|
struct gfs2_sbd *sdp = gl->gl_sbd;
|
2006-01-16 17:50:04 +01:00
|
|
|
|
2006-09-09 22:59:11 +02:00
|
|
|
write_lock(gl_lock_addr(gl->gl_hash));
|
2006-09-13 16:43:37 +02:00
|
|
|
if (atomic_dec_and_test(&gl->gl_ref)) {
|
2006-09-12 16:10:01 +02:00
|
|
|
hlist_del(&gl->gl_list);
|
2006-09-09 22:59:11 +02:00
|
|
|
write_unlock(gl_lock_addr(gl->gl_hash));
|
2006-04-20 22:57:23 +02:00
|
|
|
BUG_ON(spin_is_locked(&gl->gl_spin));
|
2006-09-13 16:43:37 +02:00
|
|
|
gfs2_assert(sdp, gl->gl_state == LM_ST_UNLOCKED);
|
|
|
|
gfs2_assert(sdp, list_empty(&gl->gl_reclaim));
|
|
|
|
gfs2_assert(sdp, list_empty(&gl->gl_holders));
|
|
|
|
gfs2_assert(sdp, list_empty(&gl->gl_waiters1));
|
|
|
|
gfs2_assert(sdp, list_empty(&gl->gl_waiters3));
|
2006-01-16 17:50:04 +01:00
|
|
|
glock_free(gl);
|
|
|
|
rv = 1;
|
|
|
|
goto out;
|
|
|
|
}
|
2006-09-09 22:59:11 +02:00
|
|
|
write_unlock(gl_lock_addr(gl->gl_hash));
|
2006-08-24 23:03:05 +02:00
|
|
|
out:
|
2006-01-16 17:50:04 +01:00
|
|
|
return rv;
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* search_bucket() - Find struct gfs2_glock by lock number
|
|
|
|
* @bucket: the bucket to search
|
|
|
|
* @name: The lock name
|
|
|
|
*
|
|
|
|
* Returns: NULL, or the struct gfs2_glock with the requested number
|
|
|
|
*/
|
|
|
|
|
2006-09-08 19:35:56 +02:00
|
|
|
static struct gfs2_glock *search_bucket(unsigned int hash,
|
2006-08-30 18:50:28 +02:00
|
|
|
const struct gfs2_sbd *sdp,
|
2006-08-30 17:16:23 +02:00
|
|
|
const struct lm_lockname *name)
|
2006-01-16 17:50:04 +01:00
|
|
|
{
|
|
|
|
struct gfs2_glock *gl;
|
2006-09-12 16:10:01 +02:00
|
|
|
struct hlist_node *h;
|
2006-01-16 17:50:04 +01:00
|
|
|
|
2006-09-12 16:10:01 +02:00
|
|
|
hlist_for_each_entry(gl, h, &gl_hash_table[hash].hb_list, gl_list) {
|
2006-01-16 17:50:04 +01:00
|
|
|
if (!lm_name_equal(&gl->gl_name, name))
|
|
|
|
continue;
|
2006-08-30 18:50:28 +02:00
|
|
|
if (gl->gl_sbd != sdp)
|
|
|
|
continue;
|
2006-01-16 17:50:04 +01:00
|
|
|
|
2006-09-13 16:43:37 +02:00
|
|
|
atomic_inc(&gl->gl_ref);
|
2006-01-16 17:50:04 +01:00
|
|
|
|
|
|
|
return gl;
|
|
|
|
}
|
|
|
|
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* gfs2_glock_find() - Find glock by lock number
|
|
|
|
* @sdp: The GFS2 superblock
|
|
|
|
* @name: The lock name
|
|
|
|
*
|
|
|
|
* Returns: NULL, or the struct gfs2_glock with the requested number
|
|
|
|
*/
|
|
|
|
|
2006-09-07 20:40:21 +02:00
|
|
|
static struct gfs2_glock *gfs2_glock_find(const struct gfs2_sbd *sdp,
|
2006-08-30 17:16:23 +02:00
|
|
|
const struct lm_lockname *name)
|
2006-01-16 17:50:04 +01:00
|
|
|
{
|
2006-09-08 19:35:56 +02:00
|
|
|
unsigned int hash = gl_hash(sdp, name);
|
2006-01-16 17:50:04 +01:00
|
|
|
struct gfs2_glock *gl;
|
|
|
|
|
2006-09-09 22:59:11 +02:00
|
|
|
read_lock(gl_lock_addr(hash));
|
2006-09-08 19:35:56 +02:00
|
|
|
gl = search_bucket(hash, sdp, name);
|
2006-09-09 22:59:11 +02:00
|
|
|
read_unlock(gl_lock_addr(hash));
|
2006-01-16 17:50:04 +01:00
|
|
|
|
|
|
|
return gl;
|
|
|
|
}
|
|
|
|
|
[GFS2] delay glock demote for a minimum hold time
When a lot of IO, with some distributed mmap IO, is run on a GFS2 filesystem in
a cluster, it will deadlock. The reason is that do_no_page() will repeatedly
call gfs2_sharewrite_nopage(), because each node keeps giving up the glock
too early, and is forced to call unmap_mapping_range(). This bumps the
mapping->truncate_count sequence count, forcing do_no_page() to retry. This
patch institutes a minimum glock hold time a tenth a second. This insures
that even in heavy contention cases, the node has enough time to get some
useful work done before it gives up the glock.
A second issue is that when gfs2_glock_dq() is called from within a page fault
to demote a lock, and the associated page needs to be written out, it will
try to acqire a lock on it, but it has already been locked at a higher level.
This patch puts makes gfs2_glock_dq() use the work queue as well, to avoid this
issue. This is the same patch as Steve Whitehouse originally proposed to fix
this issue, execpt that gfs2_glock_dq() now grabs a reference to the glock
before it queues up the work on it.
Signed-off-by: Benjamin E. Marzinski <bmarzins@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
2007-08-23 20:19:05 +02:00
|
|
|
static void glock_work_func(struct work_struct *work)
|
|
|
|
{
|
|
|
|
struct gfs2_glock *gl = container_of(work, struct gfs2_glock, gl_work.work);
|
|
|
|
|
|
|
|
spin_lock(&gl->gl_spin);
|
|
|
|
if (test_and_clear_bit(GLF_PENDING_DEMOTE, &gl->gl_flags))
|
|
|
|
set_bit(GLF_DEMOTE, &gl->gl_flags);
|
|
|
|
run_queue(gl);
|
|
|
|
spin_unlock(&gl->gl_spin);
|
|
|
|
gfs2_glock_put(gl);
|
|
|
|
}
|
|
|
|
|
2006-01-16 17:50:04 +01:00
|
|
|
/**
|
|
|
|
* gfs2_glock_get() - Get a glock, or create one if one doesn't exist
|
|
|
|
* @sdp: The GFS2 superblock
|
|
|
|
* @number: the lock number
|
|
|
|
* @glops: The glock_operations to use
|
|
|
|
* @create: If 0, don't create the glock if it doesn't exist
|
|
|
|
* @glp: the glock is returned here
|
|
|
|
*
|
|
|
|
* This does not lock a glock, just finds/creates structures for one.
|
|
|
|
*
|
|
|
|
* Returns: errno
|
|
|
|
*/
|
|
|
|
|
2006-09-04 18:49:07 +02:00
|
|
|
int gfs2_glock_get(struct gfs2_sbd *sdp, u64 number,
|
2006-08-30 15:30:00 +02:00
|
|
|
const struct gfs2_glock_operations *glops, int create,
|
2006-01-16 17:50:04 +01:00
|
|
|
struct gfs2_glock **glp)
|
|
|
|
{
|
2006-09-08 19:35:56 +02:00
|
|
|
struct lm_lockname name = { .ln_number = number, .ln_type = glops->go_type };
|
2006-01-16 17:50:04 +01:00
|
|
|
struct gfs2_glock *gl, *tmp;
|
2006-09-08 19:35:56 +02:00
|
|
|
unsigned int hash = gl_hash(sdp, &name);
|
2006-01-16 17:50:04 +01:00
|
|
|
int error;
|
|
|
|
|
2006-09-09 22:59:11 +02:00
|
|
|
read_lock(gl_lock_addr(hash));
|
2006-09-08 19:35:56 +02:00
|
|
|
gl = search_bucket(hash, sdp, &name);
|
2006-09-09 22:59:11 +02:00
|
|
|
read_unlock(gl_lock_addr(hash));
|
2006-01-16 17:50:04 +01:00
|
|
|
|
|
|
|
if (gl || !create) {
|
|
|
|
*glp = gl;
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
gl = kmem_cache_alloc(gfs2_glock_cachep, GFP_KERNEL);
|
|
|
|
if (!gl)
|
|
|
|
return -ENOMEM;
|
|
|
|
|
2006-08-30 16:36:52 +02:00
|
|
|
gl->gl_flags = 0;
|
2006-01-16 17:50:04 +01:00
|
|
|
gl->gl_name = name;
|
2006-09-13 16:43:37 +02:00
|
|
|
atomic_set(&gl->gl_ref, 1);
|
2006-01-16 17:50:04 +01:00
|
|
|
gl->gl_state = LM_ST_UNLOCKED;
|
[GFS2] delay glock demote for a minimum hold time
When a lot of IO, with some distributed mmap IO, is run on a GFS2 filesystem in
a cluster, it will deadlock. The reason is that do_no_page() will repeatedly
call gfs2_sharewrite_nopage(), because each node keeps giving up the glock
too early, and is forced to call unmap_mapping_range(). This bumps the
mapping->truncate_count sequence count, forcing do_no_page() to retry. This
patch institutes a minimum glock hold time a tenth a second. This insures
that even in heavy contention cases, the node has enough time to get some
useful work done before it gives up the glock.
A second issue is that when gfs2_glock_dq() is called from within a page fault
to demote a lock, and the associated page needs to be written out, it will
try to acqire a lock on it, but it has already been locked at a higher level.
This patch puts makes gfs2_glock_dq() use the work queue as well, to avoid this
issue. This is the same patch as Steve Whitehouse originally proposed to fix
this issue, execpt that gfs2_glock_dq() now grabs a reference to the glock
before it queues up the work on it.
Signed-off-by: Benjamin E. Marzinski <bmarzins@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
2007-08-23 20:19:05 +02:00
|
|
|
gl->gl_demote_state = LM_ST_EXCLUSIVE;
|
2006-09-08 19:35:56 +02:00
|
|
|
gl->gl_hash = hash;
|
2007-03-23 23:05:15 +01:00
|
|
|
gl->gl_owner_pid = 0;
|
2006-05-18 22:25:27 +02:00
|
|
|
gl->gl_ip = 0;
|
2006-01-16 17:50:04 +01:00
|
|
|
gl->gl_ops = glops;
|
2006-08-30 16:36:52 +02:00
|
|
|
gl->gl_req_gh = NULL;
|
|
|
|
gl->gl_req_bh = NULL;
|
|
|
|
gl->gl_vn = 0;
|
|
|
|
gl->gl_stamp = jiffies;
|
[GFS2] delay glock demote for a minimum hold time
When a lot of IO, with some distributed mmap IO, is run on a GFS2 filesystem in
a cluster, it will deadlock. The reason is that do_no_page() will repeatedly
call gfs2_sharewrite_nopage(), because each node keeps giving up the glock
too early, and is forced to call unmap_mapping_range(). This bumps the
mapping->truncate_count sequence count, forcing do_no_page() to retry. This
patch institutes a minimum glock hold time a tenth a second. This insures
that even in heavy contention cases, the node has enough time to get some
useful work done before it gives up the glock.
A second issue is that when gfs2_glock_dq() is called from within a page fault
to demote a lock, and the associated page needs to be written out, it will
try to acqire a lock on it, but it has already been locked at a higher level.
This patch puts makes gfs2_glock_dq() use the work queue as well, to avoid this
issue. This is the same patch as Steve Whitehouse originally proposed to fix
this issue, execpt that gfs2_glock_dq() now grabs a reference to the glock
before it queues up the work on it.
Signed-off-by: Benjamin E. Marzinski <bmarzins@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
2007-08-23 20:19:05 +02:00
|
|
|
gl->gl_tchange = jiffies;
|
2006-08-30 16:36:52 +02:00
|
|
|
gl->gl_object = NULL;
|
2006-01-16 17:50:04 +01:00
|
|
|
gl->gl_sbd = sdp;
|
2006-08-30 16:36:52 +02:00
|
|
|
gl->gl_aspace = NULL;
|
2006-01-16 17:50:04 +01:00
|
|
|
lops_init_le(&gl->gl_le, &gfs2_glock_lops);
|
[GFS2] delay glock demote for a minimum hold time
When a lot of IO, with some distributed mmap IO, is run on a GFS2 filesystem in
a cluster, it will deadlock. The reason is that do_no_page() will repeatedly
call gfs2_sharewrite_nopage(), because each node keeps giving up the glock
too early, and is forced to call unmap_mapping_range(). This bumps the
mapping->truncate_count sequence count, forcing do_no_page() to retry. This
patch institutes a minimum glock hold time a tenth a second. This insures
that even in heavy contention cases, the node has enough time to get some
useful work done before it gives up the glock.
A second issue is that when gfs2_glock_dq() is called from within a page fault
to demote a lock, and the associated page needs to be written out, it will
try to acqire a lock on it, but it has already been locked at a higher level.
This patch puts makes gfs2_glock_dq() use the work queue as well, to avoid this
issue. This is the same patch as Steve Whitehouse originally proposed to fix
this issue, execpt that gfs2_glock_dq() now grabs a reference to the glock
before it queues up the work on it.
Signed-off-by: Benjamin E. Marzinski <bmarzins@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
2007-08-23 20:19:05 +02:00
|
|
|
INIT_DELAYED_WORK(&gl->gl_work, glock_work_func);
|
2006-01-16 17:50:04 +01:00
|
|
|
|
|
|
|
/* If this glock protects actual on-disk data or metadata blocks,
|
|
|
|
create a VFS inode to manage the pages/buffers holding them. */
|
2006-09-04 15:49:55 +02:00
|
|
|
if (glops == &gfs2_inode_glops || glops == &gfs2_rgrp_glops) {
|
2006-01-16 17:50:04 +01:00
|
|
|
gl->gl_aspace = gfs2_aspace_get(sdp);
|
|
|
|
if (!gl->gl_aspace) {
|
|
|
|
error = -ENOMEM;
|
|
|
|
goto fail;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
error = gfs2_lm_get_lock(sdp, &name, &gl->gl_lock);
|
|
|
|
if (error)
|
|
|
|
goto fail_aspace;
|
|
|
|
|
2006-09-09 22:59:11 +02:00
|
|
|
write_lock(gl_lock_addr(hash));
|
2006-09-08 19:35:56 +02:00
|
|
|
tmp = search_bucket(hash, sdp, &name);
|
2006-01-16 17:50:04 +01:00
|
|
|
if (tmp) {
|
2006-09-09 22:59:11 +02:00
|
|
|
write_unlock(gl_lock_addr(hash));
|
2006-01-16 17:50:04 +01:00
|
|
|
glock_free(gl);
|
|
|
|
gl = tmp;
|
|
|
|
} else {
|
2006-09-12 16:10:01 +02:00
|
|
|
hlist_add_head(&gl->gl_list, &gl_hash_table[hash].hb_list);
|
2006-09-09 22:59:11 +02:00
|
|
|
write_unlock(gl_lock_addr(hash));
|
2006-01-16 17:50:04 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
*glp = gl;
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
|
2006-08-30 16:36:52 +02:00
|
|
|
fail_aspace:
|
2006-01-16 17:50:04 +01:00
|
|
|
if (gl->gl_aspace)
|
|
|
|
gfs2_aspace_put(gl->gl_aspace);
|
2006-08-30 16:36:52 +02:00
|
|
|
fail:
|
2006-09-25 15:26:04 +02:00
|
|
|
kmem_cache_free(gfs2_glock_cachep, gl);
|
2006-01-16 17:50:04 +01:00
|
|
|
return error;
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* gfs2_holder_init - initialize a struct gfs2_holder in the default way
|
|
|
|
* @gl: the glock
|
|
|
|
* @state: the state we're requesting
|
|
|
|
* @flags: the modifier flags
|
|
|
|
* @gh: the holder structure
|
|
|
|
*
|
|
|
|
*/
|
|
|
|
|
2006-04-20 22:57:23 +02:00
|
|
|
void gfs2_holder_init(struct gfs2_glock *gl, unsigned int state, unsigned flags,
|
2006-01-16 17:50:04 +01:00
|
|
|
struct gfs2_holder *gh)
|
|
|
|
{
|
|
|
|
INIT_LIST_HEAD(&gh->gh_list);
|
|
|
|
gh->gh_gl = gl;
|
2006-03-29 21:36:49 +02:00
|
|
|
gh->gh_ip = (unsigned long)__builtin_return_address(0);
|
2007-03-23 23:05:15 +01:00
|
|
|
gh->gh_owner_pid = current->pid;
|
2006-01-16 17:50:04 +01:00
|
|
|
gh->gh_state = state;
|
|
|
|
gh->gh_flags = flags;
|
|
|
|
gh->gh_error = 0;
|
|
|
|
gh->gh_iflags = 0;
|
|
|
|
gfs2_glock_hold(gl);
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* gfs2_holder_reinit - reinitialize a struct gfs2_holder so we can requeue it
|
|
|
|
* @state: the state we're requesting
|
|
|
|
* @flags: the modifier flags
|
|
|
|
* @gh: the holder structure
|
|
|
|
*
|
|
|
|
* Don't mess with the glock.
|
|
|
|
*
|
|
|
|
*/
|
|
|
|
|
2006-04-20 22:57:23 +02:00
|
|
|
void gfs2_holder_reinit(unsigned int state, unsigned flags, struct gfs2_holder *gh)
|
2006-01-16 17:50:04 +01:00
|
|
|
{
|
|
|
|
gh->gh_state = state;
|
2006-04-26 20:58:26 +02:00
|
|
|
gh->gh_flags = flags;
|
2007-03-16 10:40:31 +01:00
|
|
|
gh->gh_iflags = 0;
|
2006-03-29 21:36:49 +02:00
|
|
|
gh->gh_ip = (unsigned long)__builtin_return_address(0);
|
2006-01-16 17:50:04 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* gfs2_holder_uninit - uninitialize a holder structure (drop glock reference)
|
|
|
|
* @gh: the holder structure
|
|
|
|
*
|
|
|
|
*/
|
|
|
|
|
|
|
|
void gfs2_holder_uninit(struct gfs2_holder *gh)
|
|
|
|
{
|
|
|
|
gfs2_glock_put(gh->gh_gl);
|
|
|
|
gh->gh_gl = NULL;
|
2006-03-29 21:36:49 +02:00
|
|
|
gh->gh_ip = 0;
|
2006-01-16 17:50:04 +01:00
|
|
|
}
|
|
|
|
|
2007-03-16 10:40:31 +01:00
|
|
|
static void gfs2_holder_wake(struct gfs2_holder *gh)
|
2007-01-17 16:33:23 +01:00
|
|
|
{
|
|
|
|
clear_bit(HIF_WAIT, &gh->gh_iflags);
|
2007-06-11 09:22:32 +02:00
|
|
|
smp_mb__after_clear_bit();
|
2007-01-17 16:33:23 +01:00
|
|
|
wake_up_bit(&gh->gh_iflags, HIF_WAIT);
|
|
|
|
}
|
|
|
|
|
2007-06-11 09:22:32 +02:00
|
|
|
static int just_schedule(void *word)
|
2007-01-17 16:33:23 +01:00
|
|
|
{
|
|
|
|
schedule();
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void wait_on_holder(struct gfs2_holder *gh)
|
|
|
|
{
|
|
|
|
might_sleep();
|
2007-06-11 09:22:32 +02:00
|
|
|
wait_on_bit(&gh->gh_iflags, HIF_WAIT, just_schedule, TASK_UNINTERRUPTIBLE);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void gfs2_demote_wake(struct gfs2_glock *gl)
|
|
|
|
{
|
[GFS2] delay glock demote for a minimum hold time
When a lot of IO, with some distributed mmap IO, is run on a GFS2 filesystem in
a cluster, it will deadlock. The reason is that do_no_page() will repeatedly
call gfs2_sharewrite_nopage(), because each node keeps giving up the glock
too early, and is forced to call unmap_mapping_range(). This bumps the
mapping->truncate_count sequence count, forcing do_no_page() to retry. This
patch institutes a minimum glock hold time a tenth a second. This insures
that even in heavy contention cases, the node has enough time to get some
useful work done before it gives up the glock.
A second issue is that when gfs2_glock_dq() is called from within a page fault
to demote a lock, and the associated page needs to be written out, it will
try to acqire a lock on it, but it has already been locked at a higher level.
This patch puts makes gfs2_glock_dq() use the work queue as well, to avoid this
issue. This is the same patch as Steve Whitehouse originally proposed to fix
this issue, execpt that gfs2_glock_dq() now grabs a reference to the glock
before it queues up the work on it.
Signed-off-by: Benjamin E. Marzinski <bmarzins@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
2007-08-23 20:19:05 +02:00
|
|
|
BUG_ON(!spin_is_locked(&gl->gl_spin));
|
|
|
|
gl->gl_demote_state = LM_ST_EXCLUSIVE;
|
2007-06-11 09:22:32 +02:00
|
|
|
clear_bit(GLF_DEMOTE, &gl->gl_flags);
|
|
|
|
smp_mb__after_clear_bit();
|
|
|
|
wake_up_bit(&gl->gl_flags, GLF_DEMOTE);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void wait_on_demote(struct gfs2_glock *gl)
|
|
|
|
{
|
|
|
|
might_sleep();
|
|
|
|
wait_on_bit(&gl->gl_flags, GLF_DEMOTE, just_schedule, TASK_UNINTERRUPTIBLE);
|
2007-01-17 16:33:23 +01:00
|
|
|
}
|
|
|
|
|
2006-01-16 17:50:04 +01:00
|
|
|
/**
|
|
|
|
* rq_mutex - process a mutex request in the queue
|
|
|
|
* @gh: the glock holder
|
|
|
|
*
|
|
|
|
* Returns: 1 if the queue is blocked
|
|
|
|
*/
|
|
|
|
|
|
|
|
static int rq_mutex(struct gfs2_holder *gh)
|
|
|
|
{
|
|
|
|
struct gfs2_glock *gl = gh->gh_gl;
|
|
|
|
|
|
|
|
list_del_init(&gh->gh_list);
|
|
|
|
/* gh->gh_error never examined. */
|
|
|
|
set_bit(GLF_LOCK, &gl->gl_flags);
|
2007-01-23 22:56:36 +01:00
|
|
|
clear_bit(HIF_WAIT, &gh->gh_iflags);
|
2007-01-17 16:33:23 +01:00
|
|
|
smp_mb();
|
|
|
|
wake_up_bit(&gh->gh_iflags, HIF_WAIT);
|
2006-01-16 17:50:04 +01:00
|
|
|
|
|
|
|
return 1;
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* rq_promote - process a promote request in the queue
|
|
|
|
* @gh: the glock holder
|
|
|
|
*
|
|
|
|
* Acquire a new inter-node lock, or change a lock state to more restrictive.
|
|
|
|
*
|
|
|
|
* Returns: 1 if the queue is blocked
|
|
|
|
*/
|
|
|
|
|
|
|
|
static int rq_promote(struct gfs2_holder *gh)
|
|
|
|
{
|
|
|
|
struct gfs2_glock *gl = gh->gh_gl;
|
|
|
|
struct gfs2_sbd *sdp = gl->gl_sbd;
|
|
|
|
|
|
|
|
if (!relaxed_state_ok(gl->gl_state, gh->gh_state, gh->gh_flags)) {
|
|
|
|
if (list_empty(&gl->gl_holders)) {
|
|
|
|
gl->gl_req_gh = gh;
|
|
|
|
set_bit(GLF_LOCK, &gl->gl_flags);
|
|
|
|
spin_unlock(&gl->gl_spin);
|
|
|
|
|
|
|
|
if (atomic_read(&sdp->sd_reclaim_count) >
|
|
|
|
gfs2_tune_get(sdp, gt_reclaim_limit) &&
|
|
|
|
!(gh->gh_flags & LM_FLAG_PRIORITY)) {
|
|
|
|
gfs2_reclaim_glock(sdp);
|
|
|
|
gfs2_reclaim_glock(sdp);
|
|
|
|
}
|
|
|
|
|
2007-03-16 10:40:31 +01:00
|
|
|
gfs2_glock_xmote_th(gh->gh_gl, gh);
|
2006-01-16 17:50:04 +01:00
|
|
|
spin_lock(&gl->gl_spin);
|
|
|
|
}
|
|
|
|
return 1;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (list_empty(&gl->gl_holders)) {
|
|
|
|
set_bit(HIF_FIRST, &gh->gh_iflags);
|
|
|
|
set_bit(GLF_LOCK, &gl->gl_flags);
|
|
|
|
} else {
|
|
|
|
struct gfs2_holder *next_gh;
|
2007-01-22 18:10:39 +01:00
|
|
|
if (gh->gh_state == LM_ST_EXCLUSIVE)
|
2006-01-16 17:50:04 +01:00
|
|
|
return 1;
|
|
|
|
next_gh = list_entry(gl->gl_holders.next, struct gfs2_holder,
|
|
|
|
gh_list);
|
2007-01-22 18:10:39 +01:00
|
|
|
if (next_gh->gh_state == LM_ST_EXCLUSIVE)
|
2006-01-16 17:50:04 +01:00
|
|
|
return 1;
|
|
|
|
}
|
|
|
|
|
|
|
|
list_move_tail(&gh->gh_list, &gl->gl_holders);
|
|
|
|
gh->gh_error = 0;
|
|
|
|
set_bit(HIF_HOLDER, &gh->gh_iflags);
|
|
|
|
|
2007-03-16 10:40:31 +01:00
|
|
|
gfs2_holder_wake(gh);
|
2006-01-16 17:50:04 +01:00
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* rq_demote - process a demote request in the queue
|
|
|
|
* @gh: the glock holder
|
|
|
|
*
|
|
|
|
* Returns: 1 if the queue is blocked
|
|
|
|
*/
|
|
|
|
|
2007-03-16 10:40:31 +01:00
|
|
|
static int rq_demote(struct gfs2_glock *gl)
|
2006-01-16 17:50:04 +01:00
|
|
|
{
|
|
|
|
if (!list_empty(&gl->gl_holders))
|
|
|
|
return 1;
|
|
|
|
|
2007-03-16 10:40:31 +01:00
|
|
|
if (gl->gl_state == gl->gl_demote_state ||
|
|
|
|
gl->gl_state == LM_ST_UNLOCKED) {
|
2007-06-11 09:22:32 +02:00
|
|
|
gfs2_demote_wake(gl);
|
2007-03-16 10:40:31 +01:00
|
|
|
return 0;
|
2006-01-16 17:50:04 +01:00
|
|
|
}
|
2007-03-16 10:40:31 +01:00
|
|
|
set_bit(GLF_LOCK, &gl->gl_flags);
|
|
|
|
if (gl->gl_demote_state == LM_ST_UNLOCKED ||
|
2007-07-23 10:54:36 +02:00
|
|
|
gl->gl_state != LM_ST_EXCLUSIVE) {
|
|
|
|
spin_unlock(&gl->gl_spin);
|
2007-03-16 10:40:31 +01:00
|
|
|
gfs2_glock_drop_th(gl);
|
2007-07-23 10:54:36 +02:00
|
|
|
} else {
|
|
|
|
spin_unlock(&gl->gl_spin);
|
2007-03-16 10:40:31 +01:00
|
|
|
gfs2_glock_xmote_th(gl, NULL);
|
2007-07-23 10:54:36 +02:00
|
|
|
}
|
2007-03-16 10:40:31 +01:00
|
|
|
spin_lock(&gl->gl_spin);
|
2006-01-16 17:50:04 +01:00
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* run_queue - process holder structures on a glock
|
|
|
|
* @gl: the glock
|
|
|
|
*
|
|
|
|
*/
|
|
|
|
static void run_queue(struct gfs2_glock *gl)
|
|
|
|
{
|
|
|
|
struct gfs2_holder *gh;
|
|
|
|
int blocked = 1;
|
|
|
|
|
|
|
|
for (;;) {
|
|
|
|
if (test_bit(GLF_LOCK, &gl->gl_flags))
|
|
|
|
break;
|
|
|
|
|
|
|
|
if (!list_empty(&gl->gl_waiters1)) {
|
|
|
|
gh = list_entry(gl->gl_waiters1.next,
|
|
|
|
struct gfs2_holder, gh_list);
|
|
|
|
|
|
|
|
if (test_bit(HIF_MUTEX, &gh->gh_iflags))
|
|
|
|
blocked = rq_mutex(gh);
|
|
|
|
else
|
|
|
|
gfs2_assert_warn(gl->gl_sbd, 0);
|
|
|
|
|
2007-03-16 10:40:31 +01:00
|
|
|
} else if (test_bit(GLF_DEMOTE, &gl->gl_flags)) {
|
|
|
|
blocked = rq_demote(gl);
|
2006-01-16 17:50:04 +01:00
|
|
|
} else if (!list_empty(&gl->gl_waiters3)) {
|
|
|
|
gh = list_entry(gl->gl_waiters3.next,
|
|
|
|
struct gfs2_holder, gh_list);
|
|
|
|
|
|
|
|
if (test_bit(HIF_PROMOTE, &gh->gh_iflags))
|
|
|
|
blocked = rq_promote(gh);
|
|
|
|
else
|
|
|
|
gfs2_assert_warn(gl->gl_sbd, 0);
|
|
|
|
|
|
|
|
} else
|
|
|
|
break;
|
|
|
|
|
|
|
|
if (blocked)
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* gfs2_glmutex_lock - acquire a local lock on a glock
|
|
|
|
* @gl: the glock
|
|
|
|
*
|
|
|
|
* Gives caller exclusive access to manipulate a glock structure.
|
|
|
|
*/
|
|
|
|
|
2006-06-14 21:32:57 +02:00
|
|
|
static void gfs2_glmutex_lock(struct gfs2_glock *gl)
|
2006-01-16 17:50:04 +01:00
|
|
|
{
|
|
|
|
struct gfs2_holder gh;
|
|
|
|
|
|
|
|
gfs2_holder_init(gl, 0, 0, &gh);
|
|
|
|
set_bit(HIF_MUTEX, &gh.gh_iflags);
|
2007-01-17 16:33:23 +01:00
|
|
|
if (test_and_set_bit(HIF_WAIT, &gh.gh_iflags))
|
|
|
|
BUG();
|
2006-01-16 17:50:04 +01:00
|
|
|
|
|
|
|
spin_lock(&gl->gl_spin);
|
2006-09-07 20:40:21 +02:00
|
|
|
if (test_and_set_bit(GLF_LOCK, &gl->gl_flags)) {
|
2006-01-16 17:50:04 +01:00
|
|
|
list_add_tail(&gh.gh_list, &gl->gl_waiters1);
|
2006-09-07 20:40:21 +02:00
|
|
|
} else {
|
2007-03-23 23:05:15 +01:00
|
|
|
gl->gl_owner_pid = current->pid;
|
2006-05-18 22:25:27 +02:00
|
|
|
gl->gl_ip = (unsigned long)__builtin_return_address(0);
|
2007-01-17 16:33:23 +01:00
|
|
|
clear_bit(HIF_WAIT, &gh.gh_iflags);
|
|
|
|
smp_mb();
|
|
|
|
wake_up_bit(&gh.gh_iflags, HIF_WAIT);
|
2006-05-18 22:25:27 +02:00
|
|
|
}
|
2006-01-16 17:50:04 +01:00
|
|
|
spin_unlock(&gl->gl_spin);
|
|
|
|
|
2007-01-17 16:33:23 +01:00
|
|
|
wait_on_holder(&gh);
|
2006-01-16 17:50:04 +01:00
|
|
|
gfs2_holder_uninit(&gh);
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* gfs2_glmutex_trylock - try to acquire a local lock on a glock
|
|
|
|
* @gl: the glock
|
|
|
|
*
|
|
|
|
* Returns: 1 if the glock is acquired
|
|
|
|
*/
|
|
|
|
|
2006-04-28 16:59:12 +02:00
|
|
|
static int gfs2_glmutex_trylock(struct gfs2_glock *gl)
|
2006-01-16 17:50:04 +01:00
|
|
|
{
|
|
|
|
int acquired = 1;
|
|
|
|
|
|
|
|
spin_lock(&gl->gl_spin);
|
2006-09-07 20:40:21 +02:00
|
|
|
if (test_and_set_bit(GLF_LOCK, &gl->gl_flags)) {
|
2006-01-16 17:50:04 +01:00
|
|
|
acquired = 0;
|
2006-09-07 20:40:21 +02:00
|
|
|
} else {
|
2007-03-23 23:05:15 +01:00
|
|
|
gl->gl_owner_pid = current->pid;
|
2006-05-18 22:25:27 +02:00
|
|
|
gl->gl_ip = (unsigned long)__builtin_return_address(0);
|
|
|
|
}
|
2006-01-16 17:50:04 +01:00
|
|
|
spin_unlock(&gl->gl_spin);
|
|
|
|
|
|
|
|
return acquired;
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* gfs2_glmutex_unlock - release a local lock on a glock
|
|
|
|
* @gl: the glock
|
|
|
|
*
|
|
|
|
*/
|
|
|
|
|
2006-06-14 21:32:57 +02:00
|
|
|
static void gfs2_glmutex_unlock(struct gfs2_glock *gl)
|
2006-01-16 17:50:04 +01:00
|
|
|
{
|
|
|
|
spin_lock(&gl->gl_spin);
|
|
|
|
clear_bit(GLF_LOCK, &gl->gl_flags);
|
2007-03-23 23:05:15 +01:00
|
|
|
gl->gl_owner_pid = 0;
|
2006-05-18 22:25:27 +02:00
|
|
|
gl->gl_ip = 0;
|
2006-01-16 17:50:04 +01:00
|
|
|
run_queue(gl);
|
2006-04-20 22:57:23 +02:00
|
|
|
BUG_ON(!spin_is_locked(&gl->gl_spin));
|
2006-01-16 17:50:04 +01:00
|
|
|
spin_unlock(&gl->gl_spin);
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
2007-03-16 10:40:31 +01:00
|
|
|
* handle_callback - process a demote request
|
2006-01-16 17:50:04 +01:00
|
|
|
* @gl: the glock
|
|
|
|
* @state: the state the caller wants us to change to
|
|
|
|
*
|
2007-03-16 10:40:31 +01:00
|
|
|
* There are only two requests that we are going to see in actual
|
|
|
|
* practise: LM_ST_SHARED and LM_ST_UNLOCKED
|
2006-01-16 17:50:04 +01:00
|
|
|
*/
|
|
|
|
|
[GFS2] delay glock demote for a minimum hold time
When a lot of IO, with some distributed mmap IO, is run on a GFS2 filesystem in
a cluster, it will deadlock. The reason is that do_no_page() will repeatedly
call gfs2_sharewrite_nopage(), because each node keeps giving up the glock
too early, and is forced to call unmap_mapping_range(). This bumps the
mapping->truncate_count sequence count, forcing do_no_page() to retry. This
patch institutes a minimum glock hold time a tenth a second. This insures
that even in heavy contention cases, the node has enough time to get some
useful work done before it gives up the glock.
A second issue is that when gfs2_glock_dq() is called from within a page fault
to demote a lock, and the associated page needs to be written out, it will
try to acqire a lock on it, but it has already been locked at a higher level.
This patch puts makes gfs2_glock_dq() use the work queue as well, to avoid this
issue. This is the same patch as Steve Whitehouse originally proposed to fix
this issue, execpt that gfs2_glock_dq() now grabs a reference to the glock
before it queues up the work on it.
Signed-off-by: Benjamin E. Marzinski <bmarzins@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
2007-08-23 20:19:05 +02:00
|
|
|
static void handle_callback(struct gfs2_glock *gl, unsigned int state,
|
|
|
|
int remote, unsigned long delay)
|
2006-01-16 17:50:04 +01:00
|
|
|
{
|
[GFS2] delay glock demote for a minimum hold time
When a lot of IO, with some distributed mmap IO, is run on a GFS2 filesystem in
a cluster, it will deadlock. The reason is that do_no_page() will repeatedly
call gfs2_sharewrite_nopage(), because each node keeps giving up the glock
too early, and is forced to call unmap_mapping_range(). This bumps the
mapping->truncate_count sequence count, forcing do_no_page() to retry. This
patch institutes a minimum glock hold time a tenth a second. This insures
that even in heavy contention cases, the node has enough time to get some
useful work done before it gives up the glock.
A second issue is that when gfs2_glock_dq() is called from within a page fault
to demote a lock, and the associated page needs to be written out, it will
try to acqire a lock on it, but it has already been locked at a higher level.
This patch puts makes gfs2_glock_dq() use the work queue as well, to avoid this
issue. This is the same patch as Steve Whitehouse originally proposed to fix
this issue, execpt that gfs2_glock_dq() now grabs a reference to the glock
before it queues up the work on it.
Signed-off-by: Benjamin E. Marzinski <bmarzins@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
2007-08-23 20:19:05 +02:00
|
|
|
int bit = delay ? GLF_PENDING_DEMOTE : GLF_DEMOTE;
|
|
|
|
|
2006-01-16 17:50:04 +01:00
|
|
|
spin_lock(&gl->gl_spin);
|
[GFS2] delay glock demote for a minimum hold time
When a lot of IO, with some distributed mmap IO, is run on a GFS2 filesystem in
a cluster, it will deadlock. The reason is that do_no_page() will repeatedly
call gfs2_sharewrite_nopage(), because each node keeps giving up the glock
too early, and is forced to call unmap_mapping_range(). This bumps the
mapping->truncate_count sequence count, forcing do_no_page() to retry. This
patch institutes a minimum glock hold time a tenth a second. This insures
that even in heavy contention cases, the node has enough time to get some
useful work done before it gives up the glock.
A second issue is that when gfs2_glock_dq() is called from within a page fault
to demote a lock, and the associated page needs to be written out, it will
try to acqire a lock on it, but it has already been locked at a higher level.
This patch puts makes gfs2_glock_dq() use the work queue as well, to avoid this
issue. This is the same patch as Steve Whitehouse originally proposed to fix
this issue, execpt that gfs2_glock_dq() now grabs a reference to the glock
before it queues up the work on it.
Signed-off-by: Benjamin E. Marzinski <bmarzins@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
2007-08-23 20:19:05 +02:00
|
|
|
set_bit(bit, &gl->gl_flags);
|
|
|
|
if (gl->gl_demote_state == LM_ST_EXCLUSIVE) {
|
2007-03-16 10:40:31 +01:00
|
|
|
gl->gl_demote_state = state;
|
|
|
|
gl->gl_demote_time = jiffies;
|
2007-06-11 09:22:32 +02:00
|
|
|
if (remote && gl->gl_ops->go_type == LM_TYPE_IOPEN &&
|
|
|
|
gl->gl_object) {
|
|
|
|
struct inode *inode = igrab(gl->gl_object);
|
|
|
|
spin_unlock(&gl->gl_spin);
|
|
|
|
if (inode) {
|
|
|
|
d_prune_aliases(inode);
|
|
|
|
iput(inode);
|
|
|
|
}
|
|
|
|
return;
|
|
|
|
}
|
2007-07-23 11:02:40 +02:00
|
|
|
} else if (gl->gl_demote_state != LM_ST_UNLOCKED &&
|
|
|
|
gl->gl_demote_state != state) {
|
|
|
|
gl->gl_demote_state = LM_ST_UNLOCKED;
|
2006-01-16 17:50:04 +01:00
|
|
|
}
|
|
|
|
spin_unlock(&gl->gl_spin);
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* state_change - record that the glock is now in a different state
|
|
|
|
* @gl: the glock
|
|
|
|
* @new_state the new state
|
|
|
|
*
|
|
|
|
*/
|
|
|
|
|
|
|
|
static void state_change(struct gfs2_glock *gl, unsigned int new_state)
|
|
|
|
{
|
|
|
|
int held1, held2;
|
|
|
|
|
|
|
|
held1 = (gl->gl_state != LM_ST_UNLOCKED);
|
|
|
|
held2 = (new_state != LM_ST_UNLOCKED);
|
|
|
|
|
|
|
|
if (held1 != held2) {
|
2006-02-23 11:11:47 +01:00
|
|
|
if (held2)
|
2006-01-16 17:50:04 +01:00
|
|
|
gfs2_glock_hold(gl);
|
2006-02-23 11:11:47 +01:00
|
|
|
else
|
2006-01-16 17:50:04 +01:00
|
|
|
gfs2_glock_put(gl);
|
|
|
|
}
|
|
|
|
|
|
|
|
gl->gl_state = new_state;
|
[GFS2] delay glock demote for a minimum hold time
When a lot of IO, with some distributed mmap IO, is run on a GFS2 filesystem in
a cluster, it will deadlock. The reason is that do_no_page() will repeatedly
call gfs2_sharewrite_nopage(), because each node keeps giving up the glock
too early, and is forced to call unmap_mapping_range(). This bumps the
mapping->truncate_count sequence count, forcing do_no_page() to retry. This
patch institutes a minimum glock hold time a tenth a second. This insures
that even in heavy contention cases, the node has enough time to get some
useful work done before it gives up the glock.
A second issue is that when gfs2_glock_dq() is called from within a page fault
to demote a lock, and the associated page needs to be written out, it will
try to acqire a lock on it, but it has already been locked at a higher level.
This patch puts makes gfs2_glock_dq() use the work queue as well, to avoid this
issue. This is the same patch as Steve Whitehouse originally proposed to fix
this issue, execpt that gfs2_glock_dq() now grabs a reference to the glock
before it queues up the work on it.
Signed-off-by: Benjamin E. Marzinski <bmarzins@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
2007-08-23 20:19:05 +02:00
|
|
|
gl->gl_tchange = jiffies;
|
2006-01-16 17:50:04 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* xmote_bh - Called after the lock module is done acquiring a lock
|
|
|
|
* @gl: The glock in question
|
|
|
|
* @ret: the int returned from the lock module
|
|
|
|
*
|
|
|
|
*/
|
|
|
|
|
|
|
|
static void xmote_bh(struct gfs2_glock *gl, unsigned int ret)
|
|
|
|
{
|
|
|
|
struct gfs2_sbd *sdp = gl->gl_sbd;
|
2006-08-30 15:30:00 +02:00
|
|
|
const struct gfs2_glock_operations *glops = gl->gl_ops;
|
2006-01-16 17:50:04 +01:00
|
|
|
struct gfs2_holder *gh = gl->gl_req_gh;
|
|
|
|
int prev_state = gl->gl_state;
|
|
|
|
int op_done = 1;
|
|
|
|
|
|
|
|
gfs2_assert_warn(sdp, test_bit(GLF_LOCK, &gl->gl_flags));
|
2007-01-22 19:09:04 +01:00
|
|
|
gfs2_assert_warn(sdp, list_empty(&gl->gl_holders));
|
2006-01-16 17:50:04 +01:00
|
|
|
gfs2_assert_warn(sdp, !(ret & LM_OUT_ASYNC));
|
|
|
|
|
|
|
|
state_change(gl, ret & LM_OUT_ST_MASK);
|
|
|
|
|
|
|
|
if (prev_state != LM_ST_UNLOCKED && !(ret & LM_OUT_CACHEABLE)) {
|
|
|
|
if (glops->go_inval)
|
2006-11-20 16:37:45 +01:00
|
|
|
glops->go_inval(gl, DIO_METADATA);
|
2006-01-16 17:50:04 +01:00
|
|
|
} else if (gl->gl_state == LM_ST_DEFERRED) {
|
|
|
|
/* We might not want to do this here.
|
|
|
|
Look at moving to the inode glops. */
|
|
|
|
if (glops->go_inval)
|
2006-11-20 16:37:45 +01:00
|
|
|
glops->go_inval(gl, 0);
|
2006-01-16 17:50:04 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
/* Deal with each possible exit condition */
|
|
|
|
|
2007-03-16 10:40:31 +01:00
|
|
|
if (!gh) {
|
2006-01-16 17:50:04 +01:00
|
|
|
gl->gl_stamp = jiffies;
|
2007-07-23 10:54:36 +02:00
|
|
|
if (ret & LM_OUT_CANCELED) {
|
2007-03-16 10:40:31 +01:00
|
|
|
op_done = 0;
|
2007-07-23 10:54:36 +02:00
|
|
|
} else {
|
|
|
|
spin_lock(&gl->gl_spin);
|
|
|
|
if (gl->gl_state != gl->gl_demote_state) {
|
|
|
|
gl->gl_req_bh = NULL;
|
|
|
|
spin_unlock(&gl->gl_spin);
|
|
|
|
gfs2_glock_drop_th(gl);
|
|
|
|
gfs2_glock_put(gl);
|
|
|
|
return;
|
|
|
|
}
|
2007-06-11 09:22:32 +02:00
|
|
|
gfs2_demote_wake(gl);
|
2007-07-23 10:54:36 +02:00
|
|
|
spin_unlock(&gl->gl_spin);
|
|
|
|
}
|
2007-03-16 10:40:31 +01:00
|
|
|
} else {
|
2006-01-16 17:50:04 +01:00
|
|
|
spin_lock(&gl->gl_spin);
|
|
|
|
list_del_init(&gh->gh_list);
|
|
|
|
gh->gh_error = -EIO;
|
2007-03-16 10:40:31 +01:00
|
|
|
if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
|
|
|
|
goto out;
|
|
|
|
gh->gh_error = GLR_CANCELED;
|
|
|
|
if (ret & LM_OUT_CANCELED)
|
|
|
|
goto out;
|
|
|
|
if (relaxed_state_ok(gl->gl_state, gh->gh_state, gh->gh_flags)) {
|
|
|
|
list_add_tail(&gh->gh_list, &gl->gl_holders);
|
2006-01-16 17:50:04 +01:00
|
|
|
gh->gh_error = 0;
|
2007-03-16 10:40:31 +01:00
|
|
|
set_bit(HIF_HOLDER, &gh->gh_iflags);
|
|
|
|
set_bit(HIF_FIRST, &gh->gh_iflags);
|
|
|
|
op_done = 0;
|
|
|
|
goto out;
|
2006-01-16 17:50:04 +01:00
|
|
|
}
|
|
|
|
gh->gh_error = GLR_TRYFAILED;
|
2007-03-16 10:40:31 +01:00
|
|
|
if (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB))
|
|
|
|
goto out;
|
|
|
|
gh->gh_error = -EINVAL;
|
2006-01-16 17:50:04 +01:00
|
|
|
if (gfs2_assert_withdraw(sdp, 0) == -1)
|
|
|
|
fs_err(sdp, "ret = 0x%.8X\n", ret);
|
2007-03-16 10:40:31 +01:00
|
|
|
out:
|
|
|
|
spin_unlock(&gl->gl_spin);
|
2006-01-16 17:50:04 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
if (glops->go_xmote_bh)
|
|
|
|
glops->go_xmote_bh(gl);
|
|
|
|
|
|
|
|
if (op_done) {
|
|
|
|
spin_lock(&gl->gl_spin);
|
|
|
|
gl->gl_req_gh = NULL;
|
|
|
|
gl->gl_req_bh = NULL;
|
|
|
|
clear_bit(GLF_LOCK, &gl->gl_flags);
|
|
|
|
spin_unlock(&gl->gl_spin);
|
|
|
|
}
|
|
|
|
|
|
|
|
gfs2_glock_put(gl);
|
|
|
|
|
2007-01-17 16:33:23 +01:00
|
|
|
if (gh)
|
2007-03-16 10:40:31 +01:00
|
|
|
gfs2_holder_wake(gh);
|
2006-01-16 17:50:04 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* gfs2_glock_xmote_th - Call into the lock module to acquire or change a glock
|
|
|
|
* @gl: The glock in question
|
|
|
|
* @state: the requested state
|
|
|
|
* @flags: modifier flags to the lock call
|
|
|
|
*
|
|
|
|
*/
|
|
|
|
|
2007-07-23 10:54:36 +02:00
|
|
|
static void gfs2_glock_xmote_th(struct gfs2_glock *gl, struct gfs2_holder *gh)
|
2006-01-16 17:50:04 +01:00
|
|
|
{
|
|
|
|
struct gfs2_sbd *sdp = gl->gl_sbd;
|
2007-03-16 10:40:31 +01:00
|
|
|
int flags = gh ? gh->gh_flags : 0;
|
|
|
|
unsigned state = gh ? gh->gh_state : gl->gl_demote_state;
|
2006-08-30 15:30:00 +02:00
|
|
|
const struct gfs2_glock_operations *glops = gl->gl_ops;
|
2006-01-16 17:50:04 +01:00
|
|
|
int lck_flags = flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB |
|
|
|
|
LM_FLAG_NOEXP | LM_FLAG_ANY |
|
|
|
|
LM_FLAG_PRIORITY);
|
|
|
|
unsigned int lck_ret;
|
|
|
|
|
2007-01-22 18:15:34 +01:00
|
|
|
if (glops->go_xmote_th)
|
|
|
|
glops->go_xmote_th(gl);
|
|
|
|
|
2006-01-16 17:50:04 +01:00
|
|
|
gfs2_assert_warn(sdp, test_bit(GLF_LOCK, &gl->gl_flags));
|
2007-01-22 19:09:04 +01:00
|
|
|
gfs2_assert_warn(sdp, list_empty(&gl->gl_holders));
|
2006-01-16 17:50:04 +01:00
|
|
|
gfs2_assert_warn(sdp, state != LM_ST_UNLOCKED);
|
|
|
|
gfs2_assert_warn(sdp, state != gl->gl_state);
|
|
|
|
|
|
|
|
gfs2_glock_hold(gl);
|
|
|
|
gl->gl_req_bh = xmote_bh;
|
|
|
|
|
2006-08-30 16:36:52 +02:00
|
|
|
lck_ret = gfs2_lm_lock(sdp, gl->gl_lock, gl->gl_state, state, lck_flags);
|
2006-01-16 17:50:04 +01:00
|
|
|
|
|
|
|
if (gfs2_assert_withdraw(sdp, !(lck_ret & LM_OUT_ERROR)))
|
|
|
|
return;
|
|
|
|
|
|
|
|
if (lck_ret & LM_OUT_ASYNC)
|
|
|
|
gfs2_assert_warn(sdp, lck_ret == LM_OUT_ASYNC);
|
|
|
|
else
|
|
|
|
xmote_bh(gl, lck_ret);
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* drop_bh - Called after a lock module unlock completes
|
|
|
|
* @gl: the glock
|
|
|
|
* @ret: the return status
|
|
|
|
*
|
|
|
|
* Doesn't wake up the process waiting on the struct gfs2_holder (if any)
|
|
|
|
* Doesn't drop the reference on the glock the top half took out
|
|
|
|
*
|
|
|
|
*/
|
|
|
|
|
|
|
|
static void drop_bh(struct gfs2_glock *gl, unsigned int ret)
|
|
|
|
{
|
|
|
|
struct gfs2_sbd *sdp = gl->gl_sbd;
|
2006-08-30 15:30:00 +02:00
|
|
|
const struct gfs2_glock_operations *glops = gl->gl_ops;
|
2006-01-16 17:50:04 +01:00
|
|
|
struct gfs2_holder *gh = gl->gl_req_gh;
|
|
|
|
|
|
|
|
gfs2_assert_warn(sdp, test_bit(GLF_LOCK, &gl->gl_flags));
|
2007-01-22 19:09:04 +01:00
|
|
|
gfs2_assert_warn(sdp, list_empty(&gl->gl_holders));
|
2006-01-16 17:50:04 +01:00
|
|
|
gfs2_assert_warn(sdp, !ret);
|
|
|
|
|
|
|
|
state_change(gl, LM_ST_UNLOCKED);
|
|
|
|
|
|
|
|
if (glops->go_inval)
|
2006-11-20 16:37:45 +01:00
|
|
|
glops->go_inval(gl, DIO_METADATA);
|
2006-01-16 17:50:04 +01:00
|
|
|
|
|
|
|
if (gh) {
|
|
|
|
spin_lock(&gl->gl_spin);
|
|
|
|
list_del_init(&gh->gh_list);
|
|
|
|
gh->gh_error = 0;
|
|
|
|
spin_unlock(&gl->gl_spin);
|
|
|
|
}
|
|
|
|
|
|
|
|
spin_lock(&gl->gl_spin);
|
[GFS2] delay glock demote for a minimum hold time
When a lot of IO, with some distributed mmap IO, is run on a GFS2 filesystem in
a cluster, it will deadlock. The reason is that do_no_page() will repeatedly
call gfs2_sharewrite_nopage(), because each node keeps giving up the glock
too early, and is forced to call unmap_mapping_range(). This bumps the
mapping->truncate_count sequence count, forcing do_no_page() to retry. This
patch institutes a minimum glock hold time a tenth a second. This insures
that even in heavy contention cases, the node has enough time to get some
useful work done before it gives up the glock.
A second issue is that when gfs2_glock_dq() is called from within a page fault
to demote a lock, and the associated page needs to be written out, it will
try to acqire a lock on it, but it has already been locked at a higher level.
This patch puts makes gfs2_glock_dq() use the work queue as well, to avoid this
issue. This is the same patch as Steve Whitehouse originally proposed to fix
this issue, execpt that gfs2_glock_dq() now grabs a reference to the glock
before it queues up the work on it.
Signed-off-by: Benjamin E. Marzinski <bmarzins@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
2007-08-23 20:19:05 +02:00
|
|
|
gfs2_demote_wake(gl);
|
2006-01-16 17:50:04 +01:00
|
|
|
gl->gl_req_gh = NULL;
|
|
|
|
gl->gl_req_bh = NULL;
|
|
|
|
clear_bit(GLF_LOCK, &gl->gl_flags);
|
|
|
|
spin_unlock(&gl->gl_spin);
|
|
|
|
|
|
|
|
gfs2_glock_put(gl);
|
|
|
|
|
2007-01-17 16:33:23 +01:00
|
|
|
if (gh)
|
2007-03-16 10:40:31 +01:00
|
|
|
gfs2_holder_wake(gh);
|
2006-01-16 17:50:04 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* gfs2_glock_drop_th - call into the lock module to unlock a lock
|
|
|
|
* @gl: the glock
|
|
|
|
*
|
|
|
|
*/
|
|
|
|
|
2007-01-22 18:15:34 +01:00
|
|
|
static void gfs2_glock_drop_th(struct gfs2_glock *gl)
|
2006-01-16 17:50:04 +01:00
|
|
|
{
|
|
|
|
struct gfs2_sbd *sdp = gl->gl_sbd;
|
2006-08-30 15:30:00 +02:00
|
|
|
const struct gfs2_glock_operations *glops = gl->gl_ops;
|
2006-01-16 17:50:04 +01:00
|
|
|
unsigned int ret;
|
|
|
|
|
2007-01-22 18:15:34 +01:00
|
|
|
if (glops->go_drop_th)
|
|
|
|
glops->go_drop_th(gl);
|
|
|
|
|
2006-01-16 17:50:04 +01:00
|
|
|
gfs2_assert_warn(sdp, test_bit(GLF_LOCK, &gl->gl_flags));
|
2007-01-22 19:09:04 +01:00
|
|
|
gfs2_assert_warn(sdp, list_empty(&gl->gl_holders));
|
2006-01-16 17:50:04 +01:00
|
|
|
gfs2_assert_warn(sdp, gl->gl_state != LM_ST_UNLOCKED);
|
|
|
|
|
|
|
|
gfs2_glock_hold(gl);
|
|
|
|
gl->gl_req_bh = drop_bh;
|
|
|
|
|
|
|
|
ret = gfs2_lm_unlock(sdp, gl->gl_lock, gl->gl_state);
|
|
|
|
|
|
|
|
if (gfs2_assert_withdraw(sdp, !(ret & LM_OUT_ERROR)))
|
|
|
|
return;
|
|
|
|
|
|
|
|
if (!ret)
|
|
|
|
drop_bh(gl, ret);
|
|
|
|
else
|
|
|
|
gfs2_assert_warn(sdp, ret == LM_OUT_ASYNC);
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* do_cancels - cancel requests for locks stuck waiting on an expire flag
|
|
|
|
* @gh: the LM_FLAG_PRIORITY holder waiting to acquire the lock
|
|
|
|
*
|
|
|
|
* Don't cancel GL_NOCANCEL requests.
|
|
|
|
*/
|
|
|
|
|
|
|
|
static void do_cancels(struct gfs2_holder *gh)
|
|
|
|
{
|
|
|
|
struct gfs2_glock *gl = gh->gh_gl;
|
|
|
|
|
|
|
|
spin_lock(&gl->gl_spin);
|
|
|
|
|
|
|
|
while (gl->gl_req_gh != gh &&
|
|
|
|
!test_bit(HIF_HOLDER, &gh->gh_iflags) &&
|
|
|
|
!list_empty(&gh->gh_list)) {
|
2006-09-04 15:49:55 +02:00
|
|
|
if (gl->gl_req_bh && !(gl->gl_req_gh &&
|
|
|
|
(gl->gl_req_gh->gh_flags & GL_NOCANCEL))) {
|
2006-01-16 17:50:04 +01:00
|
|
|
spin_unlock(&gl->gl_spin);
|
|
|
|
gfs2_lm_cancel(gl->gl_sbd, gl->gl_lock);
|
|
|
|
msleep(100);
|
|
|
|
spin_lock(&gl->gl_spin);
|
|
|
|
} else {
|
|
|
|
spin_unlock(&gl->gl_spin);
|
|
|
|
msleep(100);
|
|
|
|
spin_lock(&gl->gl_spin);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
spin_unlock(&gl->gl_spin);
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* glock_wait_internal - wait on a glock acquisition
|
|
|
|
* @gh: the glock holder
|
|
|
|
*
|
|
|
|
* Returns: 0 on success
|
|
|
|
*/
|
|
|
|
|
|
|
|
static int glock_wait_internal(struct gfs2_holder *gh)
|
|
|
|
{
|
|
|
|
struct gfs2_glock *gl = gh->gh_gl;
|
|
|
|
struct gfs2_sbd *sdp = gl->gl_sbd;
|
2006-08-30 15:30:00 +02:00
|
|
|
const struct gfs2_glock_operations *glops = gl->gl_ops;
|
2006-01-16 17:50:04 +01:00
|
|
|
|
|
|
|
if (test_bit(HIF_ABORTED, &gh->gh_iflags))
|
|
|
|
return -EIO;
|
|
|
|
|
|
|
|
if (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)) {
|
|
|
|
spin_lock(&gl->gl_spin);
|
|
|
|
if (gl->gl_req_gh != gh &&
|
|
|
|
!test_bit(HIF_HOLDER, &gh->gh_iflags) &&
|
|
|
|
!list_empty(&gh->gh_list)) {
|
|
|
|
list_del_init(&gh->gh_list);
|
|
|
|
gh->gh_error = GLR_TRYFAILED;
|
|
|
|
run_queue(gl);
|
|
|
|
spin_unlock(&gl->gl_spin);
|
|
|
|
return gh->gh_error;
|
|
|
|
}
|
|
|
|
spin_unlock(&gl->gl_spin);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (gh->gh_flags & LM_FLAG_PRIORITY)
|
|
|
|
do_cancels(gh);
|
|
|
|
|
2007-01-17 16:33:23 +01:00
|
|
|
wait_on_holder(gh);
|
2006-01-16 17:50:04 +01:00
|
|
|
if (gh->gh_error)
|
|
|
|
return gh->gh_error;
|
|
|
|
|
|
|
|
gfs2_assert_withdraw(sdp, test_bit(HIF_HOLDER, &gh->gh_iflags));
|
2006-09-07 20:40:21 +02:00
|
|
|
gfs2_assert_withdraw(sdp, relaxed_state_ok(gl->gl_state, gh->gh_state,
|
2006-01-16 17:50:04 +01:00
|
|
|
gh->gh_flags));
|
|
|
|
|
|
|
|
if (test_bit(HIF_FIRST, &gh->gh_iflags)) {
|
|
|
|
gfs2_assert_warn(sdp, test_bit(GLF_LOCK, &gl->gl_flags));
|
|
|
|
|
|
|
|
if (glops->go_lock) {
|
|
|
|
gh->gh_error = glops->go_lock(gh);
|
|
|
|
if (gh->gh_error) {
|
|
|
|
spin_lock(&gl->gl_spin);
|
|
|
|
list_del_init(&gh->gh_list);
|
|
|
|
spin_unlock(&gl->gl_spin);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
spin_lock(&gl->gl_spin);
|
|
|
|
gl->gl_req_gh = NULL;
|
|
|
|
gl->gl_req_bh = NULL;
|
|
|
|
clear_bit(GLF_LOCK, &gl->gl_flags);
|
|
|
|
run_queue(gl);
|
|
|
|
spin_unlock(&gl->gl_spin);
|
|
|
|
}
|
|
|
|
|
|
|
|
return gh->gh_error;
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline struct gfs2_holder *
|
2007-03-23 23:05:15 +01:00
|
|
|
find_holder_by_owner(struct list_head *head, pid_t pid)
|
2006-01-16 17:50:04 +01:00
|
|
|
{
|
|
|
|
struct gfs2_holder *gh;
|
|
|
|
|
|
|
|
list_for_each_entry(gh, head, gh_list) {
|
2007-03-23 23:05:15 +01:00
|
|
|
if (gh->gh_owner_pid == pid)
|
2006-01-16 17:50:04 +01:00
|
|
|
return gh;
|
|
|
|
}
|
|
|
|
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
2007-03-16 11:26:37 +01:00
|
|
|
static void print_dbg(struct glock_iter *gi, const char *fmt, ...)
|
|
|
|
{
|
|
|
|
va_list args;
|
|
|
|
|
|
|
|
va_start(args, fmt);
|
|
|
|
if (gi) {
|
|
|
|
vsprintf(gi->string, fmt, args);
|
|
|
|
seq_printf(gi->seq, gi->string);
|
|
|
|
}
|
|
|
|
else
|
|
|
|
vprintk(fmt, args);
|
|
|
|
va_end(args);
|
|
|
|
}
|
|
|
|
|
2006-01-16 17:50:04 +01:00
|
|
|
/**
|
|
|
|
* add_to_queue - Add a holder to the wait queue (but look for recursion)
|
|
|
|
* @gh: the holder structure to add
|
|
|
|
*
|
|
|
|
*/
|
|
|
|
|
|
|
|
static void add_to_queue(struct gfs2_holder *gh)
|
|
|
|
{
|
|
|
|
struct gfs2_glock *gl = gh->gh_gl;
|
|
|
|
struct gfs2_holder *existing;
|
|
|
|
|
2007-03-23 23:05:15 +01:00
|
|
|
BUG_ON(!gh->gh_owner_pid);
|
2007-01-17 16:33:23 +01:00
|
|
|
if (test_and_set_bit(HIF_WAIT, &gh->gh_iflags))
|
|
|
|
BUG();
|
2006-04-20 22:57:23 +02:00
|
|
|
|
2007-09-14 06:35:27 +02:00
|
|
|
if (!(gh->gh_flags & GL_FLOCK)) {
|
|
|
|
existing = find_holder_by_owner(&gl->gl_holders,
|
|
|
|
gh->gh_owner_pid);
|
|
|
|
if (existing) {
|
|
|
|
print_symbol(KERN_WARNING "original: %s\n",
|
|
|
|
existing->gh_ip);
|
|
|
|
printk(KERN_INFO "pid : %d\n", existing->gh_owner_pid);
|
|
|
|
printk(KERN_INFO "lock type : %d lock state : %d\n",
|
|
|
|
existing->gh_gl->gl_name.ln_type,
|
|
|
|
existing->gh_gl->gl_state);
|
|
|
|
print_symbol(KERN_WARNING "new: %s\n", gh->gh_ip);
|
|
|
|
printk(KERN_INFO "pid : %d\n", gh->gh_owner_pid);
|
|
|
|
printk(KERN_INFO "lock type : %d lock state : %d\n",
|
|
|
|
gl->gl_name.ln_type, gl->gl_state);
|
|
|
|
BUG();
|
|
|
|
}
|
|
|
|
|
|
|
|
existing = find_holder_by_owner(&gl->gl_waiters3,
|
|
|
|
gh->gh_owner_pid);
|
|
|
|
if (existing) {
|
|
|
|
print_symbol(KERN_WARNING "original: %s\n",
|
|
|
|
existing->gh_ip);
|
|
|
|
print_symbol(KERN_WARNING "new: %s\n", gh->gh_ip);
|
|
|
|
BUG();
|
|
|
|
}
|
2006-01-16 17:50:04 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
if (gh->gh_flags & LM_FLAG_PRIORITY)
|
|
|
|
list_add(&gh->gh_list, &gl->gl_waiters3);
|
|
|
|
else
|
2006-09-25 15:26:04 +02:00
|
|
|
list_add_tail(&gh->gh_list, &gl->gl_waiters3);
|
2006-01-16 17:50:04 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* gfs2_glock_nq - enqueue a struct gfs2_holder onto a glock (acquire a glock)
|
|
|
|
* @gh: the holder structure
|
|
|
|
*
|
|
|
|
* if (gh->gh_flags & GL_ASYNC), this never returns an error
|
|
|
|
*
|
|
|
|
* Returns: 0, GLR_TRYFAILED, or errno on failure
|
|
|
|
*/
|
|
|
|
|
|
|
|
int gfs2_glock_nq(struct gfs2_holder *gh)
|
|
|
|
{
|
|
|
|
struct gfs2_glock *gl = gh->gh_gl;
|
|
|
|
struct gfs2_sbd *sdp = gl->gl_sbd;
|
|
|
|
int error = 0;
|
|
|
|
|
2006-05-18 22:25:27 +02:00
|
|
|
restart:
|
2006-01-16 17:50:04 +01:00
|
|
|
if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags))) {
|
|
|
|
set_bit(HIF_ABORTED, &gh->gh_iflags);
|
|
|
|
return -EIO;
|
|
|
|
}
|
|
|
|
|
|
|
|
set_bit(HIF_PROMOTE, &gh->gh_iflags);
|
|
|
|
|
|
|
|
spin_lock(&gl->gl_spin);
|
|
|
|
add_to_queue(gh);
|
|
|
|
run_queue(gl);
|
|
|
|
spin_unlock(&gl->gl_spin);
|
|
|
|
|
|
|
|
if (!(gh->gh_flags & GL_ASYNC)) {
|
|
|
|
error = glock_wait_internal(gh);
|
|
|
|
if (error == GLR_CANCELED) {
|
2006-04-20 22:57:23 +02:00
|
|
|
msleep(100);
|
2006-01-16 17:50:04 +01:00
|
|
|
goto restart;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
return error;
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* gfs2_glock_poll - poll to see if an async request has been completed
|
|
|
|
* @gh: the holder
|
|
|
|
*
|
|
|
|
* Returns: 1 if the request is ready to be gfs2_glock_wait()ed on
|
|
|
|
*/
|
|
|
|
|
|
|
|
int gfs2_glock_poll(struct gfs2_holder *gh)
|
|
|
|
{
|
|
|
|
struct gfs2_glock *gl = gh->gh_gl;
|
|
|
|
int ready = 0;
|
|
|
|
|
|
|
|
spin_lock(&gl->gl_spin);
|
|
|
|
|
|
|
|
if (test_bit(HIF_HOLDER, &gh->gh_iflags))
|
|
|
|
ready = 1;
|
|
|
|
else if (list_empty(&gh->gh_list)) {
|
|
|
|
if (gh->gh_error == GLR_CANCELED) {
|
|
|
|
spin_unlock(&gl->gl_spin);
|
2006-04-20 22:57:23 +02:00
|
|
|
msleep(100);
|
2006-01-16 17:50:04 +01:00
|
|
|
if (gfs2_glock_nq(gh))
|
|
|
|
return 1;
|
|
|
|
return 0;
|
|
|
|
} else
|
|
|
|
ready = 1;
|
|
|
|
}
|
|
|
|
|
|
|
|
spin_unlock(&gl->gl_spin);
|
|
|
|
|
|
|
|
return ready;
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* gfs2_glock_wait - wait for a lock acquisition that ended in a GLR_ASYNC
|
|
|
|
* @gh: the holder structure
|
|
|
|
*
|
|
|
|
* Returns: 0, GLR_TRYFAILED, or errno on failure
|
|
|
|
*/
|
|
|
|
|
|
|
|
int gfs2_glock_wait(struct gfs2_holder *gh)
|
|
|
|
{
|
|
|
|
int error;
|
|
|
|
|
|
|
|
error = glock_wait_internal(gh);
|
|
|
|
if (error == GLR_CANCELED) {
|
2006-04-20 22:57:23 +02:00
|
|
|
msleep(100);
|
2006-01-16 17:50:04 +01:00
|
|
|
gh->gh_flags &= ~GL_ASYNC;
|
|
|
|
error = gfs2_glock_nq(gh);
|
|
|
|
}
|
|
|
|
|
|
|
|
return error;
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* gfs2_glock_dq - dequeue a struct gfs2_holder from a glock (release a glock)
|
|
|
|
* @gh: the glock holder
|
|
|
|
*
|
|
|
|
*/
|
|
|
|
|
|
|
|
void gfs2_glock_dq(struct gfs2_holder *gh)
|
|
|
|
{
|
|
|
|
struct gfs2_glock *gl = gh->gh_gl;
|
2006-08-30 15:30:00 +02:00
|
|
|
const struct gfs2_glock_operations *glops = gl->gl_ops;
|
[GFS2] delay glock demote for a minimum hold time
When a lot of IO, with some distributed mmap IO, is run on a GFS2 filesystem in
a cluster, it will deadlock. The reason is that do_no_page() will repeatedly
call gfs2_sharewrite_nopage(), because each node keeps giving up the glock
too early, and is forced to call unmap_mapping_range(). This bumps the
mapping->truncate_count sequence count, forcing do_no_page() to retry. This
patch institutes a minimum glock hold time a tenth a second. This insures
that even in heavy contention cases, the node has enough time to get some
useful work done before it gives up the glock.
A second issue is that when gfs2_glock_dq() is called from within a page fault
to demote a lock, and the associated page needs to be written out, it will
try to acqire a lock on it, but it has already been locked at a higher level.
This patch puts makes gfs2_glock_dq() use the work queue as well, to avoid this
issue. This is the same patch as Steve Whitehouse originally proposed to fix
this issue, execpt that gfs2_glock_dq() now grabs a reference to the glock
before it queues up the work on it.
Signed-off-by: Benjamin E. Marzinski <bmarzins@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
2007-08-23 20:19:05 +02:00
|
|
|
unsigned delay = 0;
|
2006-01-16 17:50:04 +01:00
|
|
|
|
|
|
|
if (gh->gh_flags & GL_NOCACHE)
|
[GFS2] delay glock demote for a minimum hold time
When a lot of IO, with some distributed mmap IO, is run on a GFS2 filesystem in
a cluster, it will deadlock. The reason is that do_no_page() will repeatedly
call gfs2_sharewrite_nopage(), because each node keeps giving up the glock
too early, and is forced to call unmap_mapping_range(). This bumps the
mapping->truncate_count sequence count, forcing do_no_page() to retry. This
patch institutes a minimum glock hold time a tenth a second. This insures
that even in heavy contention cases, the node has enough time to get some
useful work done before it gives up the glock.
A second issue is that when gfs2_glock_dq() is called from within a page fault
to demote a lock, and the associated page needs to be written out, it will
try to acqire a lock on it, but it has already been locked at a higher level.
This patch puts makes gfs2_glock_dq() use the work queue as well, to avoid this
issue. This is the same patch as Steve Whitehouse originally proposed to fix
this issue, execpt that gfs2_glock_dq() now grabs a reference to the glock
before it queues up the work on it.
Signed-off-by: Benjamin E. Marzinski <bmarzins@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
2007-08-23 20:19:05 +02:00
|
|
|
handle_callback(gl, LM_ST_UNLOCKED, 0, 0);
|
2006-01-16 17:50:04 +01:00
|
|
|
|
|
|
|
gfs2_glmutex_lock(gl);
|
|
|
|
|
|
|
|
spin_lock(&gl->gl_spin);
|
|
|
|
list_del_init(&gh->gh_list);
|
|
|
|
|
|
|
|
if (list_empty(&gl->gl_holders)) {
|
|
|
|
spin_unlock(&gl->gl_spin);
|
|
|
|
|
|
|
|
if (glops->go_unlock)
|
|
|
|
glops->go_unlock(gh);
|
|
|
|
|
|
|
|
spin_lock(&gl->gl_spin);
|
2007-03-16 10:40:31 +01:00
|
|
|
gl->gl_stamp = jiffies;
|
2006-01-16 17:50:04 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
clear_bit(GLF_LOCK, &gl->gl_flags);
|
|
|
|
spin_unlock(&gl->gl_spin);
|
[GFS2] delay glock demote for a minimum hold time
When a lot of IO, with some distributed mmap IO, is run on a GFS2 filesystem in
a cluster, it will deadlock. The reason is that do_no_page() will repeatedly
call gfs2_sharewrite_nopage(), because each node keeps giving up the glock
too early, and is forced to call unmap_mapping_range(). This bumps the
mapping->truncate_count sequence count, forcing do_no_page() to retry. This
patch institutes a minimum glock hold time a tenth a second. This insures
that even in heavy contention cases, the node has enough time to get some
useful work done before it gives up the glock.
A second issue is that when gfs2_glock_dq() is called from within a page fault
to demote a lock, and the associated page needs to be written out, it will
try to acqire a lock on it, but it has already been locked at a higher level.
This patch puts makes gfs2_glock_dq() use the work queue as well, to avoid this
issue. This is the same patch as Steve Whitehouse originally proposed to fix
this issue, execpt that gfs2_glock_dq() now grabs a reference to the glock
before it queues up the work on it.
Signed-off-by: Benjamin E. Marzinski <bmarzins@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
2007-08-23 20:19:05 +02:00
|
|
|
|
|
|
|
gfs2_glock_hold(gl);
|
|
|
|
if (test_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
|
|
|
|
!test_bit(GLF_DEMOTE, &gl->gl_flags))
|
|
|
|
delay = gl->gl_ops->go_min_hold_time;
|
|
|
|
if (queue_delayed_work(glock_workqueue, &gl->gl_work, delay) == 0)
|
|
|
|
gfs2_glock_put(gl);
|
2006-01-16 17:50:04 +01:00
|
|
|
}
|
|
|
|
|
2007-06-11 09:22:32 +02:00
|
|
|
void gfs2_glock_dq_wait(struct gfs2_holder *gh)
|
|
|
|
{
|
|
|
|
struct gfs2_glock *gl = gh->gh_gl;
|
|
|
|
gfs2_glock_dq(gh);
|
|
|
|
wait_on_demote(gl);
|
|
|
|
}
|
|
|
|
|
2006-01-16 17:50:04 +01:00
|
|
|
/**
|
|
|
|
* gfs2_glock_dq_uninit - dequeue a holder from a glock and initialize it
|
|
|
|
* @gh: the holder structure
|
|
|
|
*
|
|
|
|
*/
|
|
|
|
|
|
|
|
void gfs2_glock_dq_uninit(struct gfs2_holder *gh)
|
|
|
|
{
|
|
|
|
gfs2_glock_dq(gh);
|
|
|
|
gfs2_holder_uninit(gh);
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* gfs2_glock_nq_num - acquire a glock based on lock number
|
|
|
|
* @sdp: the filesystem
|
|
|
|
* @number: the lock number
|
|
|
|
* @glops: the glock operations for the type of glock
|
|
|
|
* @state: the state to acquire the glock in
|
|
|
|
* @flags: modifier flags for the aquisition
|
|
|
|
* @gh: the struct gfs2_holder
|
|
|
|
*
|
|
|
|
* Returns: errno
|
|
|
|
*/
|
|
|
|
|
2006-09-04 18:49:07 +02:00
|
|
|
int gfs2_glock_nq_num(struct gfs2_sbd *sdp, u64 number,
|
2006-08-30 15:30:00 +02:00
|
|
|
const struct gfs2_glock_operations *glops,
|
|
|
|
unsigned int state, int flags, struct gfs2_holder *gh)
|
2006-01-16 17:50:04 +01:00
|
|
|
{
|
|
|
|
struct gfs2_glock *gl;
|
|
|
|
int error;
|
|
|
|
|
|
|
|
error = gfs2_glock_get(sdp, number, glops, CREATE, &gl);
|
|
|
|
if (!error) {
|
|
|
|
error = gfs2_glock_nq_init(gl, state, flags, gh);
|
|
|
|
gfs2_glock_put(gl);
|
|
|
|
}
|
|
|
|
|
|
|
|
return error;
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* glock_compare - Compare two struct gfs2_glock structures for sorting
|
|
|
|
* @arg_a: the first structure
|
|
|
|
* @arg_b: the second structure
|
|
|
|
*
|
|
|
|
*/
|
|
|
|
|
|
|
|
static int glock_compare(const void *arg_a, const void *arg_b)
|
|
|
|
{
|
2006-09-09 23:07:05 +02:00
|
|
|
const struct gfs2_holder *gh_a = *(const struct gfs2_holder **)arg_a;
|
|
|
|
const struct gfs2_holder *gh_b = *(const struct gfs2_holder **)arg_b;
|
|
|
|
const struct lm_lockname *a = &gh_a->gh_gl->gl_name;
|
|
|
|
const struct lm_lockname *b = &gh_b->gh_gl->gl_name;
|
2006-01-16 17:50:04 +01:00
|
|
|
|
|
|
|
if (a->ln_number > b->ln_number)
|
2006-09-09 23:07:05 +02:00
|
|
|
return 1;
|
|
|
|
if (a->ln_number < b->ln_number)
|
|
|
|
return -1;
|
2007-01-22 18:10:39 +01:00
|
|
|
BUG_ON(gh_a->gh_gl->gl_ops->go_type == gh_b->gh_gl->gl_ops->go_type);
|
2006-09-09 23:07:05 +02:00
|
|
|
return 0;
|
2006-01-16 17:50:04 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* nq_m_sync - synchonously acquire more than one glock in deadlock free order
|
|
|
|
* @num_gh: the number of structures
|
|
|
|
* @ghs: an array of struct gfs2_holder structures
|
|
|
|
*
|
|
|
|
* Returns: 0 on success (all glocks acquired),
|
|
|
|
* errno on failure (no glocks acquired)
|
|
|
|
*/
|
|
|
|
|
|
|
|
static int nq_m_sync(unsigned int num_gh, struct gfs2_holder *ghs,
|
|
|
|
struct gfs2_holder **p)
|
|
|
|
{
|
|
|
|
unsigned int x;
|
|
|
|
int error = 0;
|
|
|
|
|
|
|
|
for (x = 0; x < num_gh; x++)
|
|
|
|
p[x] = &ghs[x];
|
|
|
|
|
|
|
|
sort(p, num_gh, sizeof(struct gfs2_holder *), glock_compare, NULL);
|
|
|
|
|
|
|
|
for (x = 0; x < num_gh; x++) {
|
|
|
|
p[x]->gh_flags &= ~(LM_FLAG_TRY | GL_ASYNC);
|
|
|
|
|
|
|
|
error = gfs2_glock_nq(p[x]);
|
|
|
|
if (error) {
|
|
|
|
while (x--)
|
|
|
|
gfs2_glock_dq(p[x]);
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
return error;
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* gfs2_glock_nq_m - acquire multiple glocks
|
|
|
|
* @num_gh: the number of structures
|
|
|
|
* @ghs: an array of struct gfs2_holder structures
|
|
|
|
*
|
|
|
|
*
|
|
|
|
* Returns: 0 on success (all glocks acquired),
|
|
|
|
* errno on failure (no glocks acquired)
|
|
|
|
*/
|
|
|
|
|
|
|
|
int gfs2_glock_nq_m(unsigned int num_gh, struct gfs2_holder *ghs)
|
|
|
|
{
|
2007-06-19 16:38:17 +02:00
|
|
|
struct gfs2_holder *tmp[4];
|
|
|
|
struct gfs2_holder **pph = tmp;
|
2006-01-16 17:50:04 +01:00
|
|
|
int error = 0;
|
|
|
|
|
2007-06-19 16:38:17 +02:00
|
|
|
switch(num_gh) {
|
|
|
|
case 0:
|
2006-01-16 17:50:04 +01:00
|
|
|
return 0;
|
2007-06-19 16:38:17 +02:00
|
|
|
case 1:
|
2006-01-16 17:50:04 +01:00
|
|
|
ghs->gh_flags &= ~(LM_FLAG_TRY | GL_ASYNC);
|
|
|
|
return gfs2_glock_nq(ghs);
|
2007-06-19 16:38:17 +02:00
|
|
|
default:
|
|
|
|
if (num_gh <= 4)
|
2006-01-16 17:50:04 +01:00
|
|
|
break;
|
2007-06-19 16:38:17 +02:00
|
|
|
pph = kmalloc(num_gh * sizeof(struct gfs2_holder *), GFP_NOFS);
|
|
|
|
if (!pph)
|
|
|
|
return -ENOMEM;
|
2006-01-16 17:50:04 +01:00
|
|
|
}
|
|
|
|
|
2007-06-19 16:38:17 +02:00
|
|
|
error = nq_m_sync(num_gh, ghs, pph);
|
2006-01-16 17:50:04 +01:00
|
|
|
|
2007-06-19 16:38:17 +02:00
|
|
|
if (pph != tmp)
|
|
|
|
kfree(pph);
|
2006-01-16 17:50:04 +01:00
|
|
|
|
|
|
|
return error;
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* gfs2_glock_dq_m - release multiple glocks
|
|
|
|
* @num_gh: the number of structures
|
|
|
|
* @ghs: an array of struct gfs2_holder structures
|
|
|
|
*
|
|
|
|
*/
|
|
|
|
|
|
|
|
void gfs2_glock_dq_m(unsigned int num_gh, struct gfs2_holder *ghs)
|
|
|
|
{
|
|
|
|
unsigned int x;
|
|
|
|
|
|
|
|
for (x = 0; x < num_gh; x++)
|
|
|
|
gfs2_glock_dq(&ghs[x]);
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* gfs2_glock_dq_uninit_m - release multiple glocks
|
|
|
|
* @num_gh: the number of structures
|
|
|
|
* @ghs: an array of struct gfs2_holder structures
|
|
|
|
*
|
|
|
|
*/
|
|
|
|
|
|
|
|
void gfs2_glock_dq_uninit_m(unsigned int num_gh, struct gfs2_holder *ghs)
|
|
|
|
{
|
|
|
|
unsigned int x;
|
|
|
|
|
|
|
|
for (x = 0; x < num_gh; x++)
|
|
|
|
gfs2_glock_dq_uninit(&ghs[x]);
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* gfs2_lvb_hold - attach a LVB from a glock
|
|
|
|
* @gl: The glock in question
|
|
|
|
*
|
|
|
|
*/
|
|
|
|
|
|
|
|
int gfs2_lvb_hold(struct gfs2_glock *gl)
|
|
|
|
{
|
|
|
|
int error;
|
|
|
|
|
|
|
|
gfs2_glmutex_lock(gl);
|
|
|
|
|
|
|
|
if (!atomic_read(&gl->gl_lvb_count)) {
|
|
|
|
error = gfs2_lm_hold_lvb(gl->gl_sbd, gl->gl_lock, &gl->gl_lvb);
|
|
|
|
if (error) {
|
|
|
|
gfs2_glmutex_unlock(gl);
|
|
|
|
return error;
|
|
|
|
}
|
|
|
|
gfs2_glock_hold(gl);
|
|
|
|
}
|
|
|
|
atomic_inc(&gl->gl_lvb_count);
|
|
|
|
|
|
|
|
gfs2_glmutex_unlock(gl);
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* gfs2_lvb_unhold - detach a LVB from a glock
|
|
|
|
* @gl: The glock in question
|
|
|
|
*
|
|
|
|
*/
|
|
|
|
|
|
|
|
void gfs2_lvb_unhold(struct gfs2_glock *gl)
|
|
|
|
{
|
|
|
|
gfs2_glock_hold(gl);
|
|
|
|
gfs2_glmutex_lock(gl);
|
|
|
|
|
|
|
|
gfs2_assert(gl->gl_sbd, atomic_read(&gl->gl_lvb_count) > 0);
|
|
|
|
if (atomic_dec_and_test(&gl->gl_lvb_count)) {
|
|
|
|
gfs2_lm_unhold_lvb(gl->gl_sbd, gl->gl_lock, gl->gl_lvb);
|
|
|
|
gl->gl_lvb = NULL;
|
|
|
|
gfs2_glock_put(gl);
|
|
|
|
}
|
|
|
|
|
|
|
|
gfs2_glmutex_unlock(gl);
|
|
|
|
gfs2_glock_put(gl);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void blocking_cb(struct gfs2_sbd *sdp, struct lm_lockname *name,
|
|
|
|
unsigned int state)
|
|
|
|
{
|
|
|
|
struct gfs2_glock *gl;
|
[GFS2] delay glock demote for a minimum hold time
When a lot of IO, with some distributed mmap IO, is run on a GFS2 filesystem in
a cluster, it will deadlock. The reason is that do_no_page() will repeatedly
call gfs2_sharewrite_nopage(), because each node keeps giving up the glock
too early, and is forced to call unmap_mapping_range(). This bumps the
mapping->truncate_count sequence count, forcing do_no_page() to retry. This
patch institutes a minimum glock hold time a tenth a second. This insures
that even in heavy contention cases, the node has enough time to get some
useful work done before it gives up the glock.
A second issue is that when gfs2_glock_dq() is called from within a page fault
to demote a lock, and the associated page needs to be written out, it will
try to acqire a lock on it, but it has already been locked at a higher level.
This patch puts makes gfs2_glock_dq() use the work queue as well, to avoid this
issue. This is the same patch as Steve Whitehouse originally proposed to fix
this issue, execpt that gfs2_glock_dq() now grabs a reference to the glock
before it queues up the work on it.
Signed-off-by: Benjamin E. Marzinski <bmarzins@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
2007-08-23 20:19:05 +02:00
|
|
|
unsigned long delay = 0;
|
|
|
|
unsigned long holdtime;
|
|
|
|
unsigned long now = jiffies;
|
2006-01-16 17:50:04 +01:00
|
|
|
|
|
|
|
gl = gfs2_glock_find(sdp, name);
|
|
|
|
if (!gl)
|
|
|
|
return;
|
|
|
|
|
[GFS2] delay glock demote for a minimum hold time
When a lot of IO, with some distributed mmap IO, is run on a GFS2 filesystem in
a cluster, it will deadlock. The reason is that do_no_page() will repeatedly
call gfs2_sharewrite_nopage(), because each node keeps giving up the glock
too early, and is forced to call unmap_mapping_range(). This bumps the
mapping->truncate_count sequence count, forcing do_no_page() to retry. This
patch institutes a minimum glock hold time a tenth a second. This insures
that even in heavy contention cases, the node has enough time to get some
useful work done before it gives up the glock.
A second issue is that when gfs2_glock_dq() is called from within a page fault
to demote a lock, and the associated page needs to be written out, it will
try to acqire a lock on it, but it has already been locked at a higher level.
This patch puts makes gfs2_glock_dq() use the work queue as well, to avoid this
issue. This is the same patch as Steve Whitehouse originally proposed to fix
this issue, execpt that gfs2_glock_dq() now grabs a reference to the glock
before it queues up the work on it.
Signed-off-by: Benjamin E. Marzinski <bmarzins@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
2007-08-23 20:19:05 +02:00
|
|
|
holdtime = gl->gl_tchange + gl->gl_ops->go_min_hold_time;
|
|
|
|
if (time_before(now, holdtime))
|
|
|
|
delay = holdtime - now;
|
2006-01-16 17:50:04 +01:00
|
|
|
|
[GFS2] delay glock demote for a minimum hold time
When a lot of IO, with some distributed mmap IO, is run on a GFS2 filesystem in
a cluster, it will deadlock. The reason is that do_no_page() will repeatedly
call gfs2_sharewrite_nopage(), because each node keeps giving up the glock
too early, and is forced to call unmap_mapping_range(). This bumps the
mapping->truncate_count sequence count, forcing do_no_page() to retry. This
patch institutes a minimum glock hold time a tenth a second. This insures
that even in heavy contention cases, the node has enough time to get some
useful work done before it gives up the glock.
A second issue is that when gfs2_glock_dq() is called from within a page fault
to demote a lock, and the associated page needs to be written out, it will
try to acqire a lock on it, but it has already been locked at a higher level.
This patch puts makes gfs2_glock_dq() use the work queue as well, to avoid this
issue. This is the same patch as Steve Whitehouse originally proposed to fix
this issue, execpt that gfs2_glock_dq() now grabs a reference to the glock
before it queues up the work on it.
Signed-off-by: Benjamin E. Marzinski <bmarzins@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
2007-08-23 20:19:05 +02:00
|
|
|
handle_callback(gl, state, 1, delay);
|
|
|
|
if (queue_delayed_work(glock_workqueue, &gl->gl_work, delay) == 0)
|
|
|
|
gfs2_glock_put(gl);
|
2006-01-16 17:50:04 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* gfs2_glock_cb - Callback used by locking module
|
2006-09-07 21:50:20 +02:00
|
|
|
* @sdp: Pointer to the superblock
|
2006-01-16 17:50:04 +01:00
|
|
|
* @type: Type of callback
|
|
|
|
* @data: Type dependent data pointer
|
|
|
|
*
|
|
|
|
* Called by the locking module when it wants to tell us something.
|
|
|
|
* Either we need to drop a lock, one of our ASYNC requests completed, or
|
|
|
|
* a journal from another client needs to be recovered.
|
|
|
|
*/
|
|
|
|
|
2006-09-08 16:17:58 +02:00
|
|
|
void gfs2_glock_cb(void *cb_data, unsigned int type, void *data)
|
2006-01-16 17:50:04 +01:00
|
|
|
{
|
2006-09-08 16:17:58 +02:00
|
|
|
struct gfs2_sbd *sdp = cb_data;
|
2006-01-16 17:50:04 +01:00
|
|
|
|
|
|
|
switch (type) {
|
|
|
|
case LM_CB_NEED_E:
|
2006-04-27 17:25:45 +02:00
|
|
|
blocking_cb(sdp, data, LM_ST_UNLOCKED);
|
2006-01-16 17:50:04 +01:00
|
|
|
return;
|
|
|
|
|
|
|
|
case LM_CB_NEED_D:
|
2006-04-27 17:25:45 +02:00
|
|
|
blocking_cb(sdp, data, LM_ST_DEFERRED);
|
2006-01-16 17:50:04 +01:00
|
|
|
return;
|
|
|
|
|
|
|
|
case LM_CB_NEED_S:
|
2006-04-27 17:25:45 +02:00
|
|
|
blocking_cb(sdp, data, LM_ST_SHARED);
|
2006-01-16 17:50:04 +01:00
|
|
|
return;
|
|
|
|
|
|
|
|
case LM_CB_ASYNC: {
|
2006-04-27 17:25:45 +02:00
|
|
|
struct lm_async_cb *async = data;
|
2006-01-16 17:50:04 +01:00
|
|
|
struct gfs2_glock *gl;
|
|
|
|
|
2007-01-29 12:51:45 +01:00
|
|
|
down_read(&gfs2_umount_flush_sem);
|
2006-01-16 17:50:04 +01:00
|
|
|
gl = gfs2_glock_find(sdp, &async->lc_name);
|
|
|
|
if (gfs2_assert_warn(sdp, gl))
|
|
|
|
return;
|
|
|
|
if (!gfs2_assert_warn(sdp, gl->gl_req_bh))
|
|
|
|
gl->gl_req_bh(gl, async->lc_ret);
|
[GFS2] delay glock demote for a minimum hold time
When a lot of IO, with some distributed mmap IO, is run on a GFS2 filesystem in
a cluster, it will deadlock. The reason is that do_no_page() will repeatedly
call gfs2_sharewrite_nopage(), because each node keeps giving up the glock
too early, and is forced to call unmap_mapping_range(). This bumps the
mapping->truncate_count sequence count, forcing do_no_page() to retry. This
patch institutes a minimum glock hold time a tenth a second. This insures
that even in heavy contention cases, the node has enough time to get some
useful work done before it gives up the glock.
A second issue is that when gfs2_glock_dq() is called from within a page fault
to demote a lock, and the associated page needs to be written out, it will
try to acqire a lock on it, but it has already been locked at a higher level.
This patch puts makes gfs2_glock_dq() use the work queue as well, to avoid this
issue. This is the same patch as Steve Whitehouse originally proposed to fix
this issue, execpt that gfs2_glock_dq() now grabs a reference to the glock
before it queues up the work on it.
Signed-off-by: Benjamin E. Marzinski <bmarzins@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
2007-08-23 20:19:05 +02:00
|
|
|
if (queue_delayed_work(glock_workqueue, &gl->gl_work, 0) == 0)
|
|
|
|
gfs2_glock_put(gl);
|
2007-01-29 12:51:45 +01:00
|
|
|
up_read(&gfs2_umount_flush_sem);
|
2006-01-16 17:50:04 +01:00
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
case LM_CB_NEED_RECOVERY:
|
|
|
|
gfs2_jdesc_make_dirty(sdp, *(unsigned int *)data);
|
|
|
|
if (sdp->sd_recoverd_process)
|
|
|
|
wake_up_process(sdp->sd_recoverd_process);
|
|
|
|
return;
|
|
|
|
|
|
|
|
case LM_CB_DROPLOCKS:
|
|
|
|
gfs2_gl_hash_clear(sdp, NO_WAIT);
|
|
|
|
gfs2_quota_scan(sdp);
|
|
|
|
return;
|
|
|
|
|
|
|
|
default:
|
|
|
|
gfs2_assert_warn(sdp, 0);
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* demote_ok - Check to see if it's ok to unlock a glock
|
|
|
|
* @gl: the glock
|
|
|
|
*
|
|
|
|
* Returns: 1 if it's ok
|
|
|
|
*/
|
|
|
|
|
|
|
|
static int demote_ok(struct gfs2_glock *gl)
|
|
|
|
{
|
2006-08-30 15:30:00 +02:00
|
|
|
const struct gfs2_glock_operations *glops = gl->gl_ops;
|
2006-01-16 17:50:04 +01:00
|
|
|
int demote = 1;
|
|
|
|
|
|
|
|
if (test_bit(GLF_STICKY, &gl->gl_flags))
|
|
|
|
demote = 0;
|
|
|
|
else if (glops->go_demote_ok)
|
|
|
|
demote = glops->go_demote_ok(gl);
|
|
|
|
|
|
|
|
return demote;
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* gfs2_glock_schedule_for_reclaim - Add a glock to the reclaim list
|
|
|
|
* @gl: the glock
|
|
|
|
*
|
|
|
|
*/
|
|
|
|
|
|
|
|
void gfs2_glock_schedule_for_reclaim(struct gfs2_glock *gl)
|
|
|
|
{
|
|
|
|
struct gfs2_sbd *sdp = gl->gl_sbd;
|
|
|
|
|
|
|
|
spin_lock(&sdp->sd_reclaim_lock);
|
|
|
|
if (list_empty(&gl->gl_reclaim)) {
|
|
|
|
gfs2_glock_hold(gl);
|
|
|
|
list_add(&gl->gl_reclaim, &sdp->sd_reclaim_list);
|
|
|
|
atomic_inc(&sdp->sd_reclaim_count);
|
|
|
|
}
|
|
|
|
spin_unlock(&sdp->sd_reclaim_lock);
|
|
|
|
|
|
|
|
wake_up(&sdp->sd_reclaim_wq);
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* gfs2_reclaim_glock - process the next glock on the filesystem's reclaim list
|
|
|
|
* @sdp: the filesystem
|
|
|
|
*
|
|
|
|
* Called from gfs2_glockd() glock reclaim daemon, or when promoting a
|
|
|
|
* different glock and we notice that there are a lot of glocks in the
|
|
|
|
* reclaim list.
|
|
|
|
*
|
|
|
|
*/
|
|
|
|
|
|
|
|
void gfs2_reclaim_glock(struct gfs2_sbd *sdp)
|
|
|
|
{
|
|
|
|
struct gfs2_glock *gl;
|
|
|
|
|
|
|
|
spin_lock(&sdp->sd_reclaim_lock);
|
|
|
|
if (list_empty(&sdp->sd_reclaim_list)) {
|
|
|
|
spin_unlock(&sdp->sd_reclaim_lock);
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
gl = list_entry(sdp->sd_reclaim_list.next,
|
|
|
|
struct gfs2_glock, gl_reclaim);
|
|
|
|
list_del_init(&gl->gl_reclaim);
|
|
|
|
spin_unlock(&sdp->sd_reclaim_lock);
|
|
|
|
|
|
|
|
atomic_dec(&sdp->sd_reclaim_count);
|
|
|
|
atomic_inc(&sdp->sd_reclaimed);
|
|
|
|
|
|
|
|
if (gfs2_glmutex_trylock(gl)) {
|
2007-01-22 19:09:04 +01:00
|
|
|
if (list_empty(&gl->gl_holders) &&
|
2006-09-04 15:49:55 +02:00
|
|
|
gl->gl_state != LM_ST_UNLOCKED && demote_ok(gl))
|
[GFS2] delay glock demote for a minimum hold time
When a lot of IO, with some distributed mmap IO, is run on a GFS2 filesystem in
a cluster, it will deadlock. The reason is that do_no_page() will repeatedly
call gfs2_sharewrite_nopage(), because each node keeps giving up the glock
too early, and is forced to call unmap_mapping_range(). This bumps the
mapping->truncate_count sequence count, forcing do_no_page() to retry. This
patch institutes a minimum glock hold time a tenth a second. This insures
that even in heavy contention cases, the node has enough time to get some
useful work done before it gives up the glock.
A second issue is that when gfs2_glock_dq() is called from within a page fault
to demote a lock, and the associated page needs to be written out, it will
try to acqire a lock on it, but it has already been locked at a higher level.
This patch puts makes gfs2_glock_dq() use the work queue as well, to avoid this
issue. This is the same patch as Steve Whitehouse originally proposed to fix
this issue, execpt that gfs2_glock_dq() now grabs a reference to the glock
before it queues up the work on it.
Signed-off-by: Benjamin E. Marzinski <bmarzins@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
2007-08-23 20:19:05 +02:00
|
|
|
handle_callback(gl, LM_ST_UNLOCKED, 0, 0);
|
2006-01-16 17:50:04 +01:00
|
|
|
gfs2_glmutex_unlock(gl);
|
|
|
|
}
|
|
|
|
|
|
|
|
gfs2_glock_put(gl);
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* examine_bucket - Call a function for glock in a hash bucket
|
|
|
|
* @examiner: the function
|
|
|
|
* @sdp: the filesystem
|
|
|
|
* @bucket: the bucket
|
|
|
|
*
|
|
|
|
* Returns: 1 if the bucket has entries
|
|
|
|
*/
|
|
|
|
|
|
|
|
static int examine_bucket(glock_examiner examiner, struct gfs2_sbd *sdp,
|
2006-09-08 19:35:56 +02:00
|
|
|
unsigned int hash)
|
2006-01-16 17:50:04 +01:00
|
|
|
{
|
2006-09-12 03:40:30 +02:00
|
|
|
struct gfs2_glock *gl, *prev = NULL;
|
|
|
|
int has_entries = 0;
|
2006-09-12 16:10:01 +02:00
|
|
|
struct hlist_head *head = &gl_hash_table[hash].hb_list;
|
2006-01-16 17:50:04 +01:00
|
|
|
|
2006-09-12 03:40:30 +02:00
|
|
|
read_lock(gl_lock_addr(hash));
|
2006-09-12 16:10:01 +02:00
|
|
|
/* Can't use hlist_for_each_entry - don't want prefetch here */
|
|
|
|
if (hlist_empty(head))
|
2006-09-12 03:40:30 +02:00
|
|
|
goto out;
|
2006-09-12 16:10:01 +02:00
|
|
|
gl = list_entry(head->first, struct gfs2_glock, gl_list);
|
|
|
|
while(1) {
|
2007-08-01 14:57:10 +02:00
|
|
|
if (!sdp || gl->gl_sbd == sdp) {
|
2006-01-16 17:50:04 +01:00
|
|
|
gfs2_glock_hold(gl);
|
2006-09-12 03:40:30 +02:00
|
|
|
read_unlock(gl_lock_addr(hash));
|
|
|
|
if (prev)
|
|
|
|
gfs2_glock_put(prev);
|
|
|
|
prev = gl;
|
|
|
|
examiner(gl);
|
2006-09-14 19:57:38 +02:00
|
|
|
has_entries = 1;
|
2006-09-12 03:40:30 +02:00
|
|
|
read_lock(gl_lock_addr(hash));
|
2006-01-16 17:50:04 +01:00
|
|
|
}
|
2006-09-12 16:10:01 +02:00
|
|
|
if (gl->gl_list.next == NULL)
|
|
|
|
break;
|
2006-09-12 03:40:30 +02:00
|
|
|
gl = list_entry(gl->gl_list.next, struct gfs2_glock, gl_list);
|
2006-01-16 17:50:04 +01:00
|
|
|
}
|
2006-09-12 03:40:30 +02:00
|
|
|
out:
|
|
|
|
read_unlock(gl_lock_addr(hash));
|
|
|
|
if (prev)
|
|
|
|
gfs2_glock_put(prev);
|
2007-08-01 14:57:10 +02:00
|
|
|
cond_resched();
|
2006-09-12 03:40:30 +02:00
|
|
|
return has_entries;
|
2006-01-16 17:50:04 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* scan_glock - look at a glock and see if we can reclaim it
|
|
|
|
* @gl: the glock to look at
|
|
|
|
*
|
|
|
|
*/
|
|
|
|
|
|
|
|
static void scan_glock(struct gfs2_glock *gl)
|
|
|
|
{
|
2006-11-23 16:51:34 +01:00
|
|
|
if (gl->gl_ops == &gfs2_inode_glops && gl->gl_object)
|
2006-09-12 03:40:30 +02:00
|
|
|
return;
|
2006-08-24 23:03:05 +02:00
|
|
|
|
2006-01-16 17:50:04 +01:00
|
|
|
if (gfs2_glmutex_trylock(gl)) {
|
2007-01-22 19:09:04 +01:00
|
|
|
if (list_empty(&gl->gl_holders) &&
|
2006-09-12 03:40:30 +02:00
|
|
|
gl->gl_state != LM_ST_UNLOCKED && demote_ok(gl))
|
2006-01-16 17:50:04 +01:00
|
|
|
goto out_schedule;
|
|
|
|
gfs2_glmutex_unlock(gl);
|
|
|
|
}
|
|
|
|
return;
|
|
|
|
|
2006-07-05 19:16:19 +02:00
|
|
|
out_schedule:
|
2006-01-16 17:50:04 +01:00
|
|
|
gfs2_glmutex_unlock(gl);
|
|
|
|
gfs2_glock_schedule_for_reclaim(gl);
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* clear_glock - look at a glock and see if we can free it from glock cache
|
|
|
|
* @gl: the glock to look at
|
|
|
|
*
|
|
|
|
*/
|
|
|
|
|
|
|
|
static void clear_glock(struct gfs2_glock *gl)
|
|
|
|
{
|
|
|
|
struct gfs2_sbd *sdp = gl->gl_sbd;
|
|
|
|
int released;
|
|
|
|
|
|
|
|
spin_lock(&sdp->sd_reclaim_lock);
|
|
|
|
if (!list_empty(&gl->gl_reclaim)) {
|
|
|
|
list_del_init(&gl->gl_reclaim);
|
|
|
|
atomic_dec(&sdp->sd_reclaim_count);
|
2006-04-20 22:57:23 +02:00
|
|
|
spin_unlock(&sdp->sd_reclaim_lock);
|
2006-01-16 17:50:04 +01:00
|
|
|
released = gfs2_glock_put(gl);
|
|
|
|
gfs2_assert(sdp, !released);
|
2006-04-20 22:57:23 +02:00
|
|
|
} else {
|
|
|
|
spin_unlock(&sdp->sd_reclaim_lock);
|
2006-01-16 17:50:04 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
if (gfs2_glmutex_trylock(gl)) {
|
2007-01-23 19:20:41 +01:00
|
|
|
if (list_empty(&gl->gl_holders) &&
|
2006-01-16 17:50:04 +01:00
|
|
|
gl->gl_state != LM_ST_UNLOCKED)
|
[GFS2] delay glock demote for a minimum hold time
When a lot of IO, with some distributed mmap IO, is run on a GFS2 filesystem in
a cluster, it will deadlock. The reason is that do_no_page() will repeatedly
call gfs2_sharewrite_nopage(), because each node keeps giving up the glock
too early, and is forced to call unmap_mapping_range(). This bumps the
mapping->truncate_count sequence count, forcing do_no_page() to retry. This
patch institutes a minimum glock hold time a tenth a second. This insures
that even in heavy contention cases, the node has enough time to get some
useful work done before it gives up the glock.
A second issue is that when gfs2_glock_dq() is called from within a page fault
to demote a lock, and the associated page needs to be written out, it will
try to acqire a lock on it, but it has already been locked at a higher level.
This patch puts makes gfs2_glock_dq() use the work queue as well, to avoid this
issue. This is the same patch as Steve Whitehouse originally proposed to fix
this issue, execpt that gfs2_glock_dq() now grabs a reference to the glock
before it queues up the work on it.
Signed-off-by: Benjamin E. Marzinski <bmarzins@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
2007-08-23 20:19:05 +02:00
|
|
|
handle_callback(gl, LM_ST_UNLOCKED, 0, 0);
|
2006-01-16 17:50:04 +01:00
|
|
|
gfs2_glmutex_unlock(gl);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* gfs2_gl_hash_clear - Empty out the glock hash table
|
|
|
|
* @sdp: the filesystem
|
|
|
|
* @wait: wait until it's all gone
|
|
|
|
*
|
|
|
|
* Called when unmounting the filesystem, or when inter-node lock manager
|
|
|
|
* requests DROPLOCKS because it is running out of capacity.
|
|
|
|
*/
|
|
|
|
|
|
|
|
void gfs2_gl_hash_clear(struct gfs2_sbd *sdp, int wait)
|
|
|
|
{
|
|
|
|
unsigned long t;
|
|
|
|
unsigned int x;
|
|
|
|
int cont;
|
|
|
|
|
|
|
|
t = jiffies;
|
|
|
|
|
|
|
|
for (;;) {
|
|
|
|
cont = 0;
|
2006-09-12 03:40:30 +02:00
|
|
|
for (x = 0; x < GFS2_GL_HASH_SIZE; x++) {
|
2006-09-25 15:26:04 +02:00
|
|
|
if (examine_bucket(clear_glock, sdp, x))
|
2006-01-16 17:50:04 +01:00
|
|
|
cont = 1;
|
2006-09-12 03:40:30 +02:00
|
|
|
}
|
2006-01-16 17:50:04 +01:00
|
|
|
|
|
|
|
if (!wait || !cont)
|
|
|
|
break;
|
|
|
|
|
|
|
|
if (time_after_eq(jiffies,
|
|
|
|
t + gfs2_tune_get(sdp, gt_stall_secs) * HZ)) {
|
|
|
|
fs_warn(sdp, "Unmount seems to be stalled. "
|
|
|
|
"Dumping lock state...\n");
|
|
|
|
gfs2_dump_lockstate(sdp);
|
|
|
|
t = jiffies;
|
|
|
|
}
|
|
|
|
|
2007-01-29 12:51:45 +01:00
|
|
|
down_write(&gfs2_umount_flush_sem);
|
2006-01-16 17:50:04 +01:00
|
|
|
invalidate_inodes(sdp->sd_vfs);
|
2007-01-29 12:51:45 +01:00
|
|
|
up_write(&gfs2_umount_flush_sem);
|
2006-05-05 22:59:11 +02:00
|
|
|
msleep(10);
|
2006-01-16 17:50:04 +01:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Diagnostic routines to help debug distributed deadlock
|
|
|
|
*/
|
|
|
|
|
2007-03-23 23:05:15 +01:00
|
|
|
static void gfs2_print_symbol(struct glock_iter *gi, const char *fmt,
|
|
|
|
unsigned long address)
|
|
|
|
{
|
2007-04-17 18:37:11 +02:00
|
|
|
char buffer[KSYM_SYMBOL_LEN];
|
2007-03-23 23:05:15 +01:00
|
|
|
|
2007-04-17 18:37:11 +02:00
|
|
|
sprint_symbol(buffer, address);
|
|
|
|
print_dbg(gi, fmt, buffer);
|
2007-03-23 23:05:15 +01:00
|
|
|
}
|
|
|
|
|
2006-01-16 17:50:04 +01:00
|
|
|
/**
|
|
|
|
* dump_holder - print information about a glock holder
|
|
|
|
* @str: a string naming the type of holder
|
|
|
|
* @gh: the glock holder
|
|
|
|
*
|
|
|
|
* Returns: 0 on success, -ENOBUFS when we run out of space
|
|
|
|
*/
|
|
|
|
|
2007-03-16 11:26:37 +01:00
|
|
|
static int dump_holder(struct glock_iter *gi, char *str,
|
|
|
|
struct gfs2_holder *gh)
|
2006-01-16 17:50:04 +01:00
|
|
|
{
|
|
|
|
unsigned int x;
|
2007-03-23 23:05:15 +01:00
|
|
|
struct task_struct *gh_owner;
|
2006-01-16 17:50:04 +01:00
|
|
|
|
2007-03-16 11:26:37 +01:00
|
|
|
print_dbg(gi, " %s\n", str);
|
2007-03-23 23:05:15 +01:00
|
|
|
if (gh->gh_owner_pid) {
|
|
|
|
print_dbg(gi, " owner = %ld ", (long)gh->gh_owner_pid);
|
|
|
|
gh_owner = find_task_by_pid(gh->gh_owner_pid);
|
|
|
|
if (gh_owner)
|
|
|
|
print_dbg(gi, "(%s)\n", gh_owner->comm);
|
|
|
|
else
|
|
|
|
print_dbg(gi, "(ended)\n");
|
|
|
|
} else
|
|
|
|
print_dbg(gi, " owner = -1\n");
|
2007-03-16 11:26:37 +01:00
|
|
|
print_dbg(gi, " gh_state = %u\n", gh->gh_state);
|
|
|
|
print_dbg(gi, " gh_flags =");
|
2006-01-16 17:50:04 +01:00
|
|
|
for (x = 0; x < 32; x++)
|
|
|
|
if (gh->gh_flags & (1 << x))
|
2007-03-16 11:26:37 +01:00
|
|
|
print_dbg(gi, " %u", x);
|
|
|
|
print_dbg(gi, " \n");
|
|
|
|
print_dbg(gi, " error = %d\n", gh->gh_error);
|
|
|
|
print_dbg(gi, " gh_iflags =");
|
2006-01-16 17:50:04 +01:00
|
|
|
for (x = 0; x < 32; x++)
|
|
|
|
if (test_bit(x, &gh->gh_iflags))
|
2007-03-16 11:26:37 +01:00
|
|
|
print_dbg(gi, " %u", x);
|
|
|
|
print_dbg(gi, " \n");
|
2007-03-23 23:05:15 +01:00
|
|
|
gfs2_print_symbol(gi, " initialized at: %s\n", gh->gh_ip);
|
2006-01-16 17:50:04 +01:00
|
|
|
|
2007-03-16 11:26:37 +01:00
|
|
|
return 0;
|
2006-01-16 17:50:04 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* dump_inode - print information about an inode
|
|
|
|
* @ip: the inode
|
|
|
|
*
|
|
|
|
* Returns: 0 on success, -ENOBUFS when we run out of space
|
|
|
|
*/
|
|
|
|
|
2007-03-16 11:26:37 +01:00
|
|
|
static int dump_inode(struct glock_iter *gi, struct gfs2_inode *ip)
|
2006-01-16 17:50:04 +01:00
|
|
|
{
|
|
|
|
unsigned int x;
|
|
|
|
|
2007-03-16 11:26:37 +01:00
|
|
|
print_dbg(gi, " Inode:\n");
|
|
|
|
print_dbg(gi, " num = %llu/%llu\n",
|
2007-05-15 16:37:50 +02:00
|
|
|
(unsigned long long)ip->i_no_formal_ino,
|
|
|
|
(unsigned long long)ip->i_no_addr);
|
2007-03-16 11:26:37 +01:00
|
|
|
print_dbg(gi, " type = %u\n", IF2DT(ip->i_inode.i_mode));
|
|
|
|
print_dbg(gi, " i_flags =");
|
2006-01-16 17:50:04 +01:00
|
|
|
for (x = 0; x < 32; x++)
|
|
|
|
if (test_bit(x, &ip->i_flags))
|
2007-03-16 11:26:37 +01:00
|
|
|
print_dbg(gi, " %u", x);
|
|
|
|
print_dbg(gi, " \n");
|
|
|
|
return 0;
|
2006-01-16 17:50:04 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* dump_glock - print information about a glock
|
|
|
|
* @gl: the glock
|
|
|
|
* @count: where we are in the buffer
|
|
|
|
*
|
|
|
|
* Returns: 0 on success, -ENOBUFS when we run out of space
|
|
|
|
*/
|
|
|
|
|
2007-03-16 11:26:37 +01:00
|
|
|
static int dump_glock(struct glock_iter *gi, struct gfs2_glock *gl)
|
2006-01-16 17:50:04 +01:00
|
|
|
{
|
|
|
|
struct gfs2_holder *gh;
|
|
|
|
unsigned int x;
|
|
|
|
int error = -ENOBUFS;
|
2007-03-23 23:05:15 +01:00
|
|
|
struct task_struct *gl_owner;
|
2006-01-16 17:50:04 +01:00
|
|
|
|
|
|
|
spin_lock(&gl->gl_spin);
|
|
|
|
|
2007-08-21 16:57:29 +02:00
|
|
|
print_dbg(gi, "Glock 0x%p (%u, 0x%llx)\n", gl, gl->gl_name.ln_type,
|
2007-03-16 11:26:37 +01:00
|
|
|
(unsigned long long)gl->gl_name.ln_number);
|
|
|
|
print_dbg(gi, " gl_flags =");
|
2006-09-07 20:40:21 +02:00
|
|
|
for (x = 0; x < 32; x++) {
|
2006-01-16 17:50:04 +01:00
|
|
|
if (test_bit(x, &gl->gl_flags))
|
2007-03-16 11:26:37 +01:00
|
|
|
print_dbg(gi, " %u", x);
|
|
|
|
}
|
2007-03-23 23:05:15 +01:00
|
|
|
if (!test_bit(GLF_LOCK, &gl->gl_flags))
|
|
|
|
print_dbg(gi, " (unlocked)");
|
2007-03-16 11:26:37 +01:00
|
|
|
print_dbg(gi, " \n");
|
|
|
|
print_dbg(gi, " gl_ref = %d\n", atomic_read(&gl->gl_ref));
|
|
|
|
print_dbg(gi, " gl_state = %u\n", gl->gl_state);
|
2007-03-23 23:05:15 +01:00
|
|
|
if (gl->gl_owner_pid) {
|
|
|
|
gl_owner = find_task_by_pid(gl->gl_owner_pid);
|
|
|
|
if (gl_owner)
|
|
|
|
print_dbg(gi, " gl_owner = pid %d (%s)\n",
|
|
|
|
gl->gl_owner_pid, gl_owner->comm);
|
|
|
|
else
|
|
|
|
print_dbg(gi, " gl_owner = %d (ended)\n",
|
|
|
|
gl->gl_owner_pid);
|
|
|
|
} else
|
|
|
|
print_dbg(gi, " gl_owner = -1\n");
|
2007-03-16 11:26:37 +01:00
|
|
|
print_dbg(gi, " gl_ip = %lu\n", gl->gl_ip);
|
|
|
|
print_dbg(gi, " req_gh = %s\n", (gl->gl_req_gh) ? "yes" : "no");
|
|
|
|
print_dbg(gi, " req_bh = %s\n", (gl->gl_req_bh) ? "yes" : "no");
|
|
|
|
print_dbg(gi, " lvb_count = %d\n", atomic_read(&gl->gl_lvb_count));
|
|
|
|
print_dbg(gi, " object = %s\n", (gl->gl_object) ? "yes" : "no");
|
|
|
|
print_dbg(gi, " le = %s\n",
|
2006-01-16 17:50:04 +01:00
|
|
|
(list_empty(&gl->gl_le.le_list)) ? "no" : "yes");
|
2007-03-16 11:26:37 +01:00
|
|
|
print_dbg(gi, " reclaim = %s\n",
|
|
|
|
(list_empty(&gl->gl_reclaim)) ? "no" : "yes");
|
2006-01-16 17:50:04 +01:00
|
|
|
if (gl->gl_aspace)
|
2007-03-16 11:26:37 +01:00
|
|
|
print_dbg(gi, " aspace = 0x%p nrpages = %lu\n", gl->gl_aspace,
|
|
|
|
gl->gl_aspace->i_mapping->nrpages);
|
2006-01-16 17:50:04 +01:00
|
|
|
else
|
2007-03-16 11:26:37 +01:00
|
|
|
print_dbg(gi, " aspace = no\n");
|
|
|
|
print_dbg(gi, " ail = %d\n", atomic_read(&gl->gl_ail_count));
|
2006-01-16 17:50:04 +01:00
|
|
|
if (gl->gl_req_gh) {
|
2007-03-16 11:26:37 +01:00
|
|
|
error = dump_holder(gi, "Request", gl->gl_req_gh);
|
2006-01-16 17:50:04 +01:00
|
|
|
if (error)
|
|
|
|
goto out;
|
|
|
|
}
|
|
|
|
list_for_each_entry(gh, &gl->gl_holders, gh_list) {
|
2007-03-16 11:26:37 +01:00
|
|
|
error = dump_holder(gi, "Holder", gh);
|
2006-01-16 17:50:04 +01:00
|
|
|
if (error)
|
|
|
|
goto out;
|
|
|
|
}
|
|
|
|
list_for_each_entry(gh, &gl->gl_waiters1, gh_list) {
|
2007-03-16 11:26:37 +01:00
|
|
|
error = dump_holder(gi, "Waiter1", gh);
|
2006-01-16 17:50:04 +01:00
|
|
|
if (error)
|
|
|
|
goto out;
|
|
|
|
}
|
|
|
|
list_for_each_entry(gh, &gl->gl_waiters3, gh_list) {
|
2007-03-16 11:26:37 +01:00
|
|
|
error = dump_holder(gi, "Waiter3", gh);
|
2006-01-16 17:50:04 +01:00
|
|
|
if (error)
|
|
|
|
goto out;
|
|
|
|
}
|
2007-03-16 10:40:31 +01:00
|
|
|
if (test_bit(GLF_DEMOTE, &gl->gl_flags)) {
|
|
|
|
print_dbg(gi, " Demotion req to state %u (%llu uS ago)\n",
|
2007-05-14 19:42:18 +02:00
|
|
|
gl->gl_demote_state, (unsigned long long)
|
|
|
|
(jiffies - gl->gl_demote_time)*(1000000/HZ));
|
2007-03-16 10:40:31 +01:00
|
|
|
}
|
2006-02-27 23:23:27 +01:00
|
|
|
if (gl->gl_ops == &gfs2_inode_glops && gl->gl_object) {
|
2006-01-16 17:50:04 +01:00
|
|
|
if (!test_bit(GLF_LOCK, &gl->gl_flags) &&
|
2007-03-16 11:26:37 +01:00
|
|
|
list_empty(&gl->gl_holders)) {
|
|
|
|
error = dump_inode(gi, gl->gl_object);
|
2006-01-16 17:50:04 +01:00
|
|
|
if (error)
|
|
|
|
goto out;
|
|
|
|
} else {
|
|
|
|
error = -ENOBUFS;
|
2007-03-16 11:26:37 +01:00
|
|
|
print_dbg(gi, " Inode: busy\n");
|
2006-01-16 17:50:04 +01:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
error = 0;
|
|
|
|
|
2006-09-04 18:04:26 +02:00
|
|
|
out:
|
2006-01-16 17:50:04 +01:00
|
|
|
spin_unlock(&gl->gl_spin);
|
|
|
|
return error;
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* gfs2_dump_lockstate - print out the current lockstate
|
|
|
|
* @sdp: the filesystem
|
|
|
|
* @ub: the buffer to copy the information into
|
|
|
|
*
|
|
|
|
* If @ub is NULL, dump the lockstate to the console.
|
|
|
|
*
|
|
|
|
*/
|
|
|
|
|
2006-04-28 16:59:12 +02:00
|
|
|
static int gfs2_dump_lockstate(struct gfs2_sbd *sdp)
|
2006-01-16 17:50:04 +01:00
|
|
|
{
|
|
|
|
struct gfs2_glock *gl;
|
2006-09-12 16:10:01 +02:00
|
|
|
struct hlist_node *h;
|
2006-01-16 17:50:04 +01:00
|
|
|
unsigned int x;
|
|
|
|
int error = 0;
|
|
|
|
|
|
|
|
for (x = 0; x < GFS2_GL_HASH_SIZE; x++) {
|
|
|
|
|
2006-09-09 22:59:11 +02:00
|
|
|
read_lock(gl_lock_addr(x));
|
2006-01-16 17:50:04 +01:00
|
|
|
|
2006-09-12 16:10:01 +02:00
|
|
|
hlist_for_each_entry(gl, h, &gl_hash_table[x].hb_list, gl_list) {
|
2006-09-07 20:40:21 +02:00
|
|
|
if (gl->gl_sbd != sdp)
|
|
|
|
continue;
|
2006-01-16 17:50:04 +01:00
|
|
|
|
2007-03-16 11:26:37 +01:00
|
|
|
error = dump_glock(NULL, gl);
|
2006-01-16 17:50:04 +01:00
|
|
|
if (error)
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
2006-09-09 22:59:11 +02:00
|
|
|
read_unlock(gl_lock_addr(x));
|
2006-01-16 17:50:04 +01:00
|
|
|
|
|
|
|
if (error)
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
return error;
|
|
|
|
}
|
|
|
|
|
2007-08-01 14:57:10 +02:00
|
|
|
/**
|
|
|
|
* gfs2_scand - Look for cached glocks and inodes to toss from memory
|
|
|
|
* @sdp: Pointer to GFS2 superblock
|
|
|
|
*
|
|
|
|
* One of these daemons runs, finding candidates to add to sd_reclaim_list.
|
|
|
|
* See gfs2_glockd()
|
|
|
|
*/
|
|
|
|
|
|
|
|
static int gfs2_scand(void *data)
|
|
|
|
{
|
|
|
|
unsigned x;
|
|
|
|
unsigned delay;
|
|
|
|
|
|
|
|
while (!kthread_should_stop()) {
|
|
|
|
for (x = 0; x < GFS2_GL_HASH_SIZE; x++)
|
|
|
|
examine_bucket(scan_glock, NULL, x);
|
|
|
|
if (freezing(current))
|
|
|
|
refrigerator();
|
|
|
|
delay = scand_secs;
|
|
|
|
if (delay < 1)
|
|
|
|
delay = 1;
|
|
|
|
schedule_timeout_interruptible(delay * HZ);
|
|
|
|
}
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
|
2006-09-07 20:40:21 +02:00
|
|
|
int __init gfs2_glock_init(void)
|
|
|
|
{
|
|
|
|
unsigned i;
|
|
|
|
for(i = 0; i < GFS2_GL_HASH_SIZE; i++) {
|
2006-09-12 16:10:01 +02:00
|
|
|
INIT_HLIST_HEAD(&gl_hash_table[i].hb_list);
|
2006-09-07 20:40:21 +02:00
|
|
|
}
|
2006-09-09 22:59:11 +02:00
|
|
|
#ifdef GL_HASH_LOCK_SZ
|
|
|
|
for(i = 0; i < GL_HASH_LOCK_SZ; i++) {
|
|
|
|
rwlock_init(&gl_hash_locks[i]);
|
|
|
|
}
|
|
|
|
#endif
|
2007-08-01 14:57:10 +02:00
|
|
|
|
|
|
|
scand_process = kthread_run(gfs2_scand, NULL, "gfs2_scand");
|
|
|
|
if (IS_ERR(scand_process))
|
|
|
|
return PTR_ERR(scand_process);
|
|
|
|
|
[GFS2] delay glock demote for a minimum hold time
When a lot of IO, with some distributed mmap IO, is run on a GFS2 filesystem in
a cluster, it will deadlock. The reason is that do_no_page() will repeatedly
call gfs2_sharewrite_nopage(), because each node keeps giving up the glock
too early, and is forced to call unmap_mapping_range(). This bumps the
mapping->truncate_count sequence count, forcing do_no_page() to retry. This
patch institutes a minimum glock hold time a tenth a second. This insures
that even in heavy contention cases, the node has enough time to get some
useful work done before it gives up the glock.
A second issue is that when gfs2_glock_dq() is called from within a page fault
to demote a lock, and the associated page needs to be written out, it will
try to acqire a lock on it, but it has already been locked at a higher level.
This patch puts makes gfs2_glock_dq() use the work queue as well, to avoid this
issue. This is the same patch as Steve Whitehouse originally proposed to fix
this issue, execpt that gfs2_glock_dq() now grabs a reference to the glock
before it queues up the work on it.
Signed-off-by: Benjamin E. Marzinski <bmarzins@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
2007-08-23 20:19:05 +02:00
|
|
|
glock_workqueue = create_workqueue("glock_workqueue");
|
|
|
|
if (IS_ERR(glock_workqueue)) {
|
|
|
|
kthread_stop(scand_process);
|
|
|
|
return PTR_ERR(glock_workqueue);
|
|
|
|
}
|
|
|
|
|
2006-09-07 20:40:21 +02:00
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2007-08-01 14:57:10 +02:00
|
|
|
void gfs2_glock_exit(void)
|
|
|
|
{
|
[GFS2] delay glock demote for a minimum hold time
When a lot of IO, with some distributed mmap IO, is run on a GFS2 filesystem in
a cluster, it will deadlock. The reason is that do_no_page() will repeatedly
call gfs2_sharewrite_nopage(), because each node keeps giving up the glock
too early, and is forced to call unmap_mapping_range(). This bumps the
mapping->truncate_count sequence count, forcing do_no_page() to retry. This
patch institutes a minimum glock hold time a tenth a second. This insures
that even in heavy contention cases, the node has enough time to get some
useful work done before it gives up the glock.
A second issue is that when gfs2_glock_dq() is called from within a page fault
to demote a lock, and the associated page needs to be written out, it will
try to acqire a lock on it, but it has already been locked at a higher level.
This patch puts makes gfs2_glock_dq() use the work queue as well, to avoid this
issue. This is the same patch as Steve Whitehouse originally proposed to fix
this issue, execpt that gfs2_glock_dq() now grabs a reference to the glock
before it queues up the work on it.
Signed-off-by: Benjamin E. Marzinski <bmarzins@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
2007-08-23 20:19:05 +02:00
|
|
|
destroy_workqueue(glock_workqueue);
|
2007-08-01 14:57:10 +02:00
|
|
|
kthread_stop(scand_process);
|
|
|
|
}
|
|
|
|
|
|
|
|
module_param(scand_secs, uint, S_IRUGO|S_IWUSR);
|
|
|
|
MODULE_PARM_DESC(scand_secs, "The number of seconds between scand runs");
|
|
|
|
|
2007-03-16 11:26:37 +01:00
|
|
|
static int gfs2_glock_iter_next(struct glock_iter *gi)
|
|
|
|
{
|
2007-07-24 14:53:36 +02:00
|
|
|
struct gfs2_glock *gl;
|
|
|
|
|
2007-08-21 16:57:29 +02:00
|
|
|
restart:
|
2007-04-17 18:37:11 +02:00
|
|
|
read_lock(gl_lock_addr(gi->hash));
|
2007-07-24 14:53:36 +02:00
|
|
|
gl = gi->gl;
|
|
|
|
if (gl) {
|
2007-08-21 16:57:29 +02:00
|
|
|
gi->gl = hlist_entry(gl->gl_list.next,
|
|
|
|
struct gfs2_glock, gl_list);
|
2007-03-16 11:26:37 +01:00
|
|
|
if (gi->gl)
|
2007-07-24 14:53:36 +02:00
|
|
|
gfs2_glock_hold(gi->gl);
|
2007-03-16 11:26:37 +01:00
|
|
|
}
|
2007-04-17 18:37:11 +02:00
|
|
|
read_unlock(gl_lock_addr(gi->hash));
|
2007-07-24 14:53:36 +02:00
|
|
|
if (gl)
|
|
|
|
gfs2_glock_put(gl);
|
2007-08-21 16:57:29 +02:00
|
|
|
if (gl && gi->gl == NULL)
|
2007-07-24 14:53:36 +02:00
|
|
|
gi->hash++;
|
2007-08-21 16:57:29 +02:00
|
|
|
while(gi->gl == NULL) {
|
2007-07-24 14:53:36 +02:00
|
|
|
if (gi->hash >= GFS2_GL_HASH_SIZE)
|
|
|
|
return 1;
|
|
|
|
read_lock(gl_lock_addr(gi->hash));
|
|
|
|
gi->gl = hlist_entry(gl_hash_table[gi->hash].hb_list.first,
|
|
|
|
struct gfs2_glock, gl_list);
|
|
|
|
if (gi->gl)
|
|
|
|
gfs2_glock_hold(gi->gl);
|
|
|
|
read_unlock(gl_lock_addr(gi->hash));
|
2007-08-21 16:57:29 +02:00
|
|
|
gi->hash++;
|
2007-07-24 14:53:36 +02:00
|
|
|
}
|
2007-08-21 16:57:29 +02:00
|
|
|
|
|
|
|
if (gi->sdp != gi->gl->gl_sbd)
|
|
|
|
goto restart;
|
|
|
|
|
2007-03-16 11:26:37 +01:00
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void gfs2_glock_iter_free(struct glock_iter *gi)
|
|
|
|
{
|
2007-07-24 14:53:36 +02:00
|
|
|
if (gi->gl)
|
|
|
|
gfs2_glock_put(gi->gl);
|
2007-03-16 11:26:37 +01:00
|
|
|
kfree(gi);
|
|
|
|
}
|
|
|
|
|
|
|
|
static struct glock_iter *gfs2_glock_iter_init(struct gfs2_sbd *sdp)
|
|
|
|
{
|
|
|
|
struct glock_iter *gi;
|
|
|
|
|
|
|
|
gi = kmalloc(sizeof (*gi), GFP_KERNEL);
|
|
|
|
if (!gi)
|
|
|
|
return NULL;
|
|
|
|
|
|
|
|
gi->sdp = sdp;
|
|
|
|
gi->hash = 0;
|
|
|
|
gi->seq = NULL;
|
2007-08-21 16:57:29 +02:00
|
|
|
gi->gl = NULL;
|
2007-03-16 11:26:37 +01:00
|
|
|
memset(gi->string, 0, sizeof(gi->string));
|
|
|
|
|
2007-08-21 16:57:29 +02:00
|
|
|
if (gfs2_glock_iter_next(gi)) {
|
2007-03-16 11:26:37 +01:00
|
|
|
gfs2_glock_iter_free(gi);
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
|
|
|
return gi;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void *gfs2_glock_seq_start(struct seq_file *file, loff_t *pos)
|
|
|
|
{
|
|
|
|
struct glock_iter *gi;
|
|
|
|
loff_t n = *pos;
|
|
|
|
|
|
|
|
gi = gfs2_glock_iter_init(file->private);
|
|
|
|
if (!gi)
|
|
|
|
return NULL;
|
|
|
|
|
2007-07-24 14:53:36 +02:00
|
|
|
while(n--) {
|
2007-03-16 11:26:37 +01:00
|
|
|
if (gfs2_glock_iter_next(gi)) {
|
|
|
|
gfs2_glock_iter_free(gi);
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
return gi;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void *gfs2_glock_seq_next(struct seq_file *file, void *iter_ptr,
|
|
|
|
loff_t *pos)
|
|
|
|
{
|
|
|
|
struct glock_iter *gi = iter_ptr;
|
|
|
|
|
|
|
|
(*pos)++;
|
|
|
|
|
|
|
|
if (gfs2_glock_iter_next(gi)) {
|
|
|
|
gfs2_glock_iter_free(gi);
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
|
|
|
return gi;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void gfs2_glock_seq_stop(struct seq_file *file, void *iter_ptr)
|
|
|
|
{
|
2007-07-24 14:53:36 +02:00
|
|
|
struct glock_iter *gi = iter_ptr;
|
|
|
|
if (gi)
|
|
|
|
gfs2_glock_iter_free(gi);
|
2007-03-16 11:26:37 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
static int gfs2_glock_seq_show(struct seq_file *file, void *iter_ptr)
|
|
|
|
{
|
|
|
|
struct glock_iter *gi = iter_ptr;
|
|
|
|
|
|
|
|
gi->seq = file;
|
|
|
|
dump_glock(gi, gi->gl);
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2007-07-31 12:31:11 +02:00
|
|
|
static const struct seq_operations gfs2_glock_seq_ops = {
|
2007-03-16 11:26:37 +01:00
|
|
|
.start = gfs2_glock_seq_start,
|
|
|
|
.next = gfs2_glock_seq_next,
|
|
|
|
.stop = gfs2_glock_seq_stop,
|
|
|
|
.show = gfs2_glock_seq_show,
|
|
|
|
};
|
|
|
|
|
|
|
|
static int gfs2_debugfs_open(struct inode *inode, struct file *file)
|
|
|
|
{
|
|
|
|
struct seq_file *seq;
|
|
|
|
int ret;
|
|
|
|
|
|
|
|
ret = seq_open(file, &gfs2_glock_seq_ops);
|
|
|
|
if (ret)
|
|
|
|
return ret;
|
|
|
|
|
|
|
|
seq = file->private_data;
|
|
|
|
seq->private = inode->i_private;
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
static const struct file_operations gfs2_debug_fops = {
|
|
|
|
.owner = THIS_MODULE,
|
|
|
|
.open = gfs2_debugfs_open,
|
|
|
|
.read = seq_read,
|
|
|
|
.llseek = seq_lseek,
|
|
|
|
.release = seq_release
|
|
|
|
};
|
|
|
|
|
|
|
|
int gfs2_create_debugfs_file(struct gfs2_sbd *sdp)
|
|
|
|
{
|
2007-04-18 18:41:11 +02:00
|
|
|
sdp->debugfs_dir = debugfs_create_dir(sdp->sd_table_name, gfs2_root);
|
|
|
|
if (!sdp->debugfs_dir)
|
|
|
|
return -ENOMEM;
|
|
|
|
sdp->debugfs_dentry_glocks = debugfs_create_file("glocks",
|
|
|
|
S_IFREG | S_IRUGO,
|
|
|
|
sdp->debugfs_dir, sdp,
|
|
|
|
&gfs2_debug_fops);
|
|
|
|
if (!sdp->debugfs_dentry_glocks)
|
2007-03-16 11:26:37 +01:00
|
|
|
return -ENOMEM;
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
void gfs2_delete_debugfs_file(struct gfs2_sbd *sdp)
|
|
|
|
{
|
2007-04-18 18:41:11 +02:00
|
|
|
if (sdp && sdp->debugfs_dir) {
|
|
|
|
if (sdp->debugfs_dentry_glocks) {
|
|
|
|
debugfs_remove(sdp->debugfs_dentry_glocks);
|
|
|
|
sdp->debugfs_dentry_glocks = NULL;
|
|
|
|
}
|
|
|
|
debugfs_remove(sdp->debugfs_dir);
|
|
|
|
sdp->debugfs_dir = NULL;
|
|
|
|
}
|
2007-03-16 11:26:37 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
int gfs2_register_debugfs(void)
|
|
|
|
{
|
|
|
|
gfs2_root = debugfs_create_dir("gfs2", NULL);
|
|
|
|
return gfs2_root ? 0 : -ENOMEM;
|
|
|
|
}
|
|
|
|
|
|
|
|
void gfs2_unregister_debugfs(void)
|
|
|
|
{
|
|
|
|
debugfs_remove(gfs2_root);
|
2007-04-18 18:41:11 +02:00
|
|
|
gfs2_root = NULL;
|
2007-03-16 11:26:37 +01:00
|
|
|
}
|