Skip to contentSkip to conversation

Fix journal flush problem [56/70]

GFS2 Steven Whitehouse 2 participants 3 messages Nov 30, 2006 – Dec 4, 2006 GFS2 file system

lore
Share & tools
lore
#0 lore raw

From b004157ab5b374a498a5874cda68c389219d23e7 Mon Sep 17 00:00:00 2001
From: Steven Whitehouse <[email protected]>
Date: Thu, 23 Nov 2006 10:51:34 -0500
Subject: [PATCH] [GFS2] Fix journal flush problem

This fixes a bug which resulted in poor performance due to flushing the journal too often. The code path in question was via the inode_go_sync() function in glops.c. The solution is not to flush the journal immediately when inodes are ejected from memory, but batch up the work for glockd to deal with later on. This means that glocks may now live on beyond the end of the lifetime of their inodes (but not very much longer in the normal case).

Also fixed in this patch is a bug (which was hidden by the bug mentioned above) in calculation of the number of free journal blocks.

The gfs2_logd process has been altered to be more responsive to the journal filling up. We now wake it up when the number of uncommitted journal blocks has reached the threshold level rather than trying to flush directly at the end of each transaction. This again means doing fewer, but larger, log flushes in general.

---
 fs/gfs2/daemon.c    |    7 +++-
 fs/gfs2/glock.c     |   17 +--------
 fs/gfs2/glock.h     |    1 -
 fs/gfs2/glops.c     |   93 +++++++++++++--------------------------------------
 fs/gfs2/log.c       |   17 +++++----
 fs/gfs2/meta_io.c   |    3 ++
 fs/gfs2/ops_super.c |    7 ++--
 7 files changed, 46 insertions(+), 99 deletions(-)
fs/gfs2/daemon.c +5 -2
index cab1f68..683cb5b 100644
@@ -112,6 +112,7 @@ int gfs2_logd(void *data)
 	struct gfs2_sbd *sdp = data;
 	struct gfs2_holder ji_gh;
 	unsigned long t;
+	int need_flush;
 
 	while (!kthread_should_stop()) {
 		/* Advance the log tail */
@@ -120,8 +121,10 @@ int gfs2_logd(void *data)
 		    gfs2_tune_get(sdp, gt_log_flush_secs) * HZ;
 
 		gfs2_ail1_empty(sdp, DIO_ALL);
-
-		if (time_after_eq(jiffies, t)) {
+		gfs2_log_lock(sdp);
+		need_flush = sdp->sd_log_num_buf > gfs2_tune_get(sdp, gt_incore_log_blocks);
+		gfs2_log_unlock(sdp);
+		if (need_flush || time_after_eq(jiffies, t)) {
 			gfs2_log_flush(sdp, NULL);
 			sdp->sd_log_flush_time = jiffies;
 		}
fs/gfs2/glock.c +1 -16
index b8ba4d5..3c2ff81 100644
@@ -785,21 +785,6 @@ out:
 		gfs2_holder_put(new_gh);
 }
 
-void gfs2_glock_inode_squish(struct inode *inode)
-{
-	struct gfs2_holder gh;
-	struct gfs2_glock *gl = GFS2_I(inode)->i_gl;
-	gfs2_holder_init(gl, LM_ST_UNLOCKED, 0, &gh);
-	set_bit(HIF_DEMOTE, &gh.gh_iflags);
-	spin_lock(&gl->gl_spin);
-	gfs2_assert(inode->i_sb->s_fs_info, list_empty(&gl->gl_holders));
-	list_add_tail(&gh.gh_list, &gl->gl_waiters2);
-	run_queue(gl);
-	spin_unlock(&gl->gl_spin);
-	wait_for_completion(&gh.gh_wait);
-	gfs2_holder_uninit(&gh);
-}
-
 /**
  * state_change - record that the glock is now in a different state
  * @gl: the glock
@@ -1920,7 +1905,7 @@ out:
 
 static void scan_glock(struct gfs2_glock *gl)
 {
-	if (gl->gl_ops == &gfs2_inode_glops)
+	if (gl->gl_ops == &gfs2_inode_glops && gl->gl_object)
 		return;
 
 	if (gfs2_glmutex_trylock(gl)) {
fs/gfs2/glock.h +0 -1
index a331bf8..fb39108 100644
@@ -106,7 +106,6 @@ void gfs2_glock_dq_uninit_m(unsigned int
 void gfs2_glock_prefetch_num(struct gfs2_sbd *sdp, u64 number,
 			     const struct gfs2_glock_operations *glops,
 			     unsigned int state, int flags);
-void gfs2_glock_inode_squish(struct inode *inode);
 
 /**
  * gfs2_glock_nq_init - intialize a holder and enqueue it on a glock
fs/gfs2/glops.c +24 -69

119 lines not on this page · raw message

fs/gfs2/log.c +10 -7
index 0cace3d..6456fc3 100644
@@ -261,6 +261,12 @@ static void ail2_empty(struct gfs2_sbd *
  * @sdp: The GFS2 superblock
  * @blks: The number of blocks to reserve
  *
+ * Note that we never give out the last 6 blocks of the journal. Thats
+ * due to the fact that there is are a small number of header blocks
+ * associated with each log flush. The exact number can't be known until
+ * flush time, so we ensure that we have just enough free blocks at all
+ * times to avoid running out during a log flush.
+ *
  * Returns: errno
  */
 
@@ -274,7 +280,7 @@ int gfs2_log_reserve(struct gfs2_sbd *sd
 
 	mutex_lock(&sdp->sd_log_reserve_mutex);
 	gfs2_log_lock(sdp);
-	while(sdp->sd_log_blks_free <= blks) {
+	while(sdp->sd_log_blks_free <= (blks + 6)) {
 		gfs2_log_unlock(sdp);
 		gfs2_ail1_empty(sdp, 0);
 		gfs2_log_flush(sdp, NULL);
@@ -643,12 +649,9 @@ void gfs2_log_commit(struct gfs2_sbd *sd
 	up_read(&sdp->sd_log_flush_lock);
 
 	gfs2_log_lock(sdp);
-	if (sdp->sd_log_num_buf > gfs2_tune_get(sdp, gt_incore_log_blocks)) {
-		gfs2_log_unlock(sdp);
-		gfs2_log_flush(sdp, NULL);
-	} else {
-		gfs2_log_unlock(sdp);
-	}
+	if (sdp->sd_log_num_buf > gfs2_tune_get(sdp, gt_incore_log_blocks))
+		wake_up_process(sdp->sd_logd_process);
+	gfs2_log_unlock(sdp);
 }
 
 /**
fs/gfs2/meta_io.c +3 -0
index 3912d6a..939a09f 100644
@@ -472,6 +472,9 @@ int gfs2_meta_indirect_buffer(struct gfs
 	struct buffer_head *bh = NULL, **bh_slot = ip->i_cache + height;
 	int in_cache = 0;
 
+	BUG_ON(!gl);
+	BUG_ON(!sdp);
+
 	spin_lock(&ip->i_spin);
 	if (*bh_slot && (*bh_slot)->b_blocknr == num) {
 		bh = *bh_slot;
fs/gfs2/ops_super.c +3 -4
index 8635175..7685b46 100644
@@ -157,7 +157,8 @@ static void gfs2_write_super(struct supe
 static int gfs2_sync_fs(struct super_block *sb, int wait)
 {
 	sb->s_dirt = 0;
-	gfs2_log_flush(sb->s_fs_info, NULL);
+	if (wait)
+		gfs2_log_flush(sb->s_fs_info, NULL);
 	return 0;
 }
 
@@ -293,8 +294,6 @@ static void gfs2_clear_inode(struct inod
 	 */
 	if (inode->i_private) {
 		struct gfs2_inode *ip = GFS2_I(inode);
-		gfs2_glock_inode_squish(inode);
-		gfs2_assert(inode->i_sb->s_fs_info, ip->i_gl->gl_state == LM_ST_UNLOCKED);
 		ip->i_gl->gl_object = NULL;
 		gfs2_glock_schedule_for_reclaim(ip->i_gl);
 		gfs2_glock_put(ip->i_gl);
@@ -395,7 +394,7 @@ static void gfs2_delete_inode(struct ino
 	if (!inode->i_private)
 		goto out;
 
-	error = gfs2_glock_nq_init(ip->i_gl, LM_ST_EXCLUSIVE, LM_FLAG_TRY_1CB | GL_NOCACHE, &gh);
+	error = gfs2_glock_nq_init(ip->i_gl, LM_ST_EXCLUSIVE, LM_FLAG_TRY_1CB, &gh);
 	if (unlikely(error)) {
 		gfs2_glock_dq_uninit(&ip->i_iopen_gh);
 		goto out;
Signature · 1 line
1.4.1
#1 lore raw
Quoted 11 lines · Steven Whitehouse
Quoted 1 line

From b004157ab5b374a498a5874cda68c389219d23e7 Mon Sep 17 00:00:00 2001

From: Steven Whitehouse <[email protected]>
Date: Thu, 23 Nov 2006 10:51:34 -0500
Subject: [PATCH] [GFS2] Fix journal flush problem

This fixes a bug which resulted in poor performance due to flushing the journal too often. The code path in question was via the inode_go_sync() function in glops.c. The solution is not to flush the journal immediately when inodes are ejected from memory, but batch up the work for glockd to deal with later on. This means that glocks may now live on beyond the end of the lifetime of their inodes (but not very much longer in the normal case).

This seems like multiple changes in one patch. The inode flush handling is changing quite significantly. The log flushing is also being changed.

Quoted 34 lines

Also fixed in this patch is a bug (which was hidden by the bug mentioned above) in calculation of the number of free journal blocks.

The gfs2_logd process has been altered to be more responsive to the journal filling up. We now wake it up when the number of uncommitted journal blocks has reached the threshold level rather than trying to flush directly at the end of each transaction. This again means doing fewer, but larger, log flushes in general.

---
 fs/gfs2/daemon.c    |    7 +++-
 fs/gfs2/glock.c     |   17 +--------
 fs/gfs2/glock.h     |    1 -
 fs/gfs2/glops.c     |   93 +++++++++++++--------------------------------------
 fs/gfs2/log.c       |   17 +++++----
 fs/gfs2/meta_io.c   |    3 ++
 fs/gfs2/ops_super.c |    7 ++--
 7 files changed, 46 insertions(+), 99 deletions(-)
fs/gfs2/daemon.c +1 -0
index cab1f68..683cb5b 100644
@@ -112,6 +112,7 @@ int gfs2_logd(void *data)
 	struct gfs2_sbd *sdp = data;
 	struct gfs2_holder ji_gh;
 	unsigned long t;
+	int need_flush;
 
 	while (!kthread_should_stop()) {
 		/* Advance the log tail */

fs/gfs2/daemon.c int gfs2_logd(void *data) · line 126

Quoted 8 lines
fs/gfs2/daemon.c +3 -2
@@ -120,8 +121,10 @@ int gfs2_logd(void *data)
 		    gfs2_tune_get(sdp, gt_log_flush_secs) * HZ;
 
 		gfs2_ail1_empty(sdp, DIO_ALL);
-
-		if (time_after_eq(jiffies, t)) {
+		gfs2_log_lock(sdp);
+		need_flush = sdp->sd_log_num_buf > gfs2_tune_get(sdp, gt_incore_log_blocks);
+		gfs2_log_unlock(sdp);

Do we really need to lock the log just to get the log_num_buf? Seems like a serialization we don't need?

So why does this loop have a sleep timeout and a flush interval? Shouldn't the schedual timeout be the same as the flush interval?

Quoted 265 lines
fs/gfs2/daemon.c +1 -0
+		if (need_flush || time_after_eq(jiffies, t)) {
 			gfs2_log_flush(sdp, NULL);
 			sdp->sd_log_flush_time = jiffies;
 		}
fs/gfs2/glock.c +1 -16
index b8ba4d5..3c2ff81 100644
@@ -785,21 +785,6 @@ out:
 		gfs2_holder_put(new_gh);
 }
 
-void gfs2_glock_inode_squish(struct inode *inode)
-{
-	struct gfs2_holder gh;
-	struct gfs2_glock *gl = GFS2_I(inode)->i_gl;
-	gfs2_holder_init(gl, LM_ST_UNLOCKED, 0, &gh);
-	set_bit(HIF_DEMOTE, &gh.gh_iflags);
-	spin_lock(&gl->gl_spin);
-	gfs2_assert(inode->i_sb->s_fs_info, list_empty(&gl->gl_holders));
-	list_add_tail(&gh.gh_list, &gl->gl_waiters2);
-	run_queue(gl);
-	spin_unlock(&gl->gl_spin);
-	wait_for_completion(&gh.gh_wait);
-	gfs2_holder_uninit(&gh);
-}
-
 /**
  * state_change - record that the glock is now in a different state
  * @gl: the glock
@@ -1920,7 +1905,7 @@ out:
 
 static void scan_glock(struct gfs2_glock *gl)
 {
-	if (gl->gl_ops == &gfs2_inode_glops)
+	if (gl->gl_ops == &gfs2_inode_glops && gl->gl_object)
 		return;
 
 	if (gfs2_glmutex_trylock(gl)) {
fs/gfs2/glock.h +0 -1
index a331bf8..fb39108 100644
@@ -106,7 +106,6 @@ void gfs2_glock_dq_uninit_m(unsigned int
 void gfs2_glock_prefetch_num(struct gfs2_sbd *sdp, u64 number,
 			     const struct gfs2_glock_operations *glops,
 			     unsigned int state, int flags);
-void gfs2_glock_inode_squish(struct inode *inode);
 
 /**
  * gfs2_glock_nq_init - intialize a holder and enqueue it on a glock
fs/gfs2/glops.c +24 -69
index 60561ca..b068d10 100644
@@ -107,70 +107,6 @@ static void gfs2_pte_inval(struct gfs2_g
 }
 
 /**
- * gfs2_page_inval - Invalidate all pages associated with a glock
- * @gl: the glock
- *
- */
-
-static void gfs2_page_inval(struct gfs2_glock *gl)
-{
-	struct gfs2_inode *ip;
-	struct inode *inode;
-
-	ip = gl->gl_object;
-	inode = &ip->i_inode;
-	if (!ip || !S_ISREG(inode->i_mode))
-		return;
-
-	truncate_inode_pages(inode->i_mapping, 0);
-	gfs2_assert_withdraw(GFS2_SB(&ip->i_inode), !inode->i_mapping->nrpages);
-	clear_bit(GIF_PAGED, &ip->i_flags);
-}
-
-/**
- * gfs2_page_wait - Wait for writeback of data
- * @gl: the glock
- *
- * Syncs data (not metadata) for a regular file.
- * No-op for all other types.
- */
-
-static void gfs2_page_wait(struct gfs2_glock *gl)
-{
-	struct gfs2_inode *ip = gl->gl_object;
-	struct inode *inode = &ip->i_inode;
-	struct address_space *mapping = inode->i_mapping;
-	int error;
-
-	if (!S_ISREG(inode->i_mode))
-		return;
-
-	error = filemap_fdatawait(mapping);
-
-	/* Put back any errors cleared by filemap_fdatawait()
-	   so they can be caught by someone who can pass them
-	   up to user space. */
-
-	if (error == -ENOSPC)
-		set_bit(AS_ENOSPC, &mapping->flags);
-	else if (error)
-		set_bit(AS_EIO, &mapping->flags);
-
-}
-
-static void gfs2_page_writeback(struct gfs2_glock *gl)
-{
-	struct gfs2_inode *ip = gl->gl_object;
-	struct inode *inode = &ip->i_inode;
-	struct address_space *mapping = inode->i_mapping;
-
-	if (!S_ISREG(inode->i_mode))
-		return;
-
-	filemap_fdatawrite(mapping);
-}
-
-/**
  * meta_go_sync - sync out the metadata for this glock
  * @gl: the glock
  *
@@ -264,11 +200,24 @@ static void inode_go_drop_th(struct gfs2
 
 static void inode_go_sync(struct gfs2_glock *gl)
 {
+	struct gfs2_inode *ip = gl->gl_object;
+
+	if (ip && !S_ISREG(ip->i_inode.i_mode))
+		ip = NULL;
+
 	if (test_bit(GLF_DIRTY, &gl->gl_flags)) {
-		gfs2_page_writeback(gl);
 		gfs2_log_flush(gl->gl_sbd, gl);
+		if (ip)
+			filemap_fdatawrite(ip->i_inode.i_mapping);
 		gfs2_meta_sync(gl);
-		gfs2_page_wait(gl);
+		if (ip) {
+			struct address_space *mapping = ip->i_inode.i_mapping;
+			int error = filemap_fdatawait(mapping);
+			if (error == -ENOSPC)
+				set_bit(AS_ENOSPC, &mapping->flags);
+			else if (error)
+				set_bit(AS_EIO, &mapping->flags);
+		}
 		clear_bit(GLF_DIRTY, &gl->gl_flags);
 		gfs2_ail_empty_gl(gl);
 	}
@@ -283,14 +232,20 @@ static void inode_go_sync(struct gfs2_gl
 
 static void inode_go_inval(struct gfs2_glock *gl, int flags)
 {
+	struct gfs2_inode *ip = gl->gl_object;
 	int meta = (flags & DIO_METADATA);
 
 	if (meta) {
-		struct gfs2_inode *ip = gl->gl_object;
 		gfs2_meta_inval(gl);
-		set_bit(GIF_INVALID, &ip->i_flags);
+		if (ip)
+			set_bit(GIF_INVALID, &ip->i_flags);
+	}
+
+	if (ip && S_ISREG(ip->i_inode.i_mode)) {
+		truncate_inode_pages(ip->i_inode.i_mapping, 0);
+		gfs2_assert_withdraw(GFS2_SB(&ip->i_inode), !ip->i_inode.i_mapping->nrpages);
+		clear_bit(GIF_PAGED, &ip->i_flags);
 	}
-	gfs2_page_inval(gl);
 }
 
 /**
fs/gfs2/log.c +10 -7
index 0cace3d..6456fc3 100644
@@ -261,6 +261,12 @@ static void ail2_empty(struct gfs2_sbd *
  * @sdp: The GFS2 superblock
  * @blks: The number of blocks to reserve
  *
+ * Note that we never give out the last 6 blocks of the journal. Thats
+ * due to the fact that there is are a small number of header blocks
+ * associated with each log flush. The exact number can't be known until
+ * flush time, so we ensure that we have just enough free blocks at all
+ * times to avoid running out during a log flush.
+ *
  * Returns: errno
  */
 
@@ -274,7 +280,7 @@ int gfs2_log_reserve(struct gfs2_sbd *sd
 
 	mutex_lock(&sdp->sd_log_reserve_mutex);
 	gfs2_log_lock(sdp);
-	while(sdp->sd_log_blks_free <= blks) {
+	while(sdp->sd_log_blks_free <= (blks + 6)) {
 		gfs2_log_unlock(sdp);
 		gfs2_ail1_empty(sdp, 0);
 		gfs2_log_flush(sdp, NULL);
@@ -643,12 +649,9 @@ void gfs2_log_commit(struct gfs2_sbd *sd
 	up_read(&sdp->sd_log_flush_lock);
 
 	gfs2_log_lock(sdp);
-	if (sdp->sd_log_num_buf > gfs2_tune_get(sdp, gt_incore_log_blocks)) {
-		gfs2_log_unlock(sdp);
-		gfs2_log_flush(sdp, NULL);
-	} else {
-		gfs2_log_unlock(sdp);
-	}
+	if (sdp->sd_log_num_buf > gfs2_tune_get(sdp, gt_incore_log_blocks))
+		wake_up_process(sdp->sd_logd_process);
+	gfs2_log_unlock(sdp);
 }
 
 /**
fs/gfs2/meta_io.c +3 -0
index 3912d6a..939a09f 100644
@@ -472,6 +472,9 @@ int gfs2_meta_indirect_buffer(struct gfs
 	struct buffer_head *bh = NULL, **bh_slot = ip->i_cache + height;
 	int in_cache = 0;
 
+	BUG_ON(!gl);
+	BUG_ON(!sdp);
+
 	spin_lock(&ip->i_spin);
 	if (*bh_slot && (*bh_slot)->b_blocknr == num) {
 		bh = *bh_slot;
fs/gfs2/ops_super.c +3 -4
index 8635175..7685b46 100644
@@ -157,7 +157,8 @@ static void gfs2_write_super(struct supe
 static int gfs2_sync_fs(struct super_block *sb, int wait)
 {
 	sb->s_dirt = 0;
-	gfs2_log_flush(sb->s_fs_info, NULL);
+	if (wait)
+		gfs2_log_flush(sb->s_fs_info, NULL);
 	return 0;
 }
 
@@ -293,8 +294,6 @@ static void gfs2_clear_inode(struct inod
 	 */
 	if (inode->i_private) {
 		struct gfs2_inode *ip = GFS2_I(inode);
-		gfs2_glock_inode_squish(inode);
-		gfs2_assert(inode->i_sb->s_fs_info, ip->i_gl->gl_state == LM_ST_UNLOCKED);
 		ip->i_gl->gl_object = NULL;
 		gfs2_glock_schedule_for_reclaim(ip->i_gl);
 		gfs2_glock_put(ip->i_gl);
@@ -395,7 +394,7 @@ static void gfs2_delete_inode(struct ino
 	if (!inode->i_private)
 		goto out;
 
-	error = gfs2_glock_nq_init(ip->i_gl, LM_ST_EXCLUSIVE, LM_FLAG_TRY_1CB | GL_NOCACHE, &gh);
+	error = gfs2_glock_nq_init(ip->i_gl, LM_ST_EXCLUSIVE, LM_FLAG_TRY_1CB, &gh);
 	if (unlikely(error)) {
 		gfs2_glock_dq_uninit(&ip->i_iopen_gh);
 		goto out;
Signature · 8 lines
Russell Cattelan <[email protected]>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 189 bytes
Desc: This is a digitally signed message part
URL: <http://listman.redhat.com/archives/cluster-devel/attachments/20061201/b047adaa/attachment.sig>
#2 lore raw

Hi,

Quoted 13 lines · Russell Cattelan
Quoted 11 lines · Steven Whitehouse
Quoted 1 line

From b004157ab5b374a498a5874cda68c389219d23e7 Mon Sep 17 00:00:00 2001

From: Steven Whitehouse <[email protected]>
Date: Thu, 23 Nov 2006 10:51:34 -0500
Subject: [PATCH] [GFS2] Fix journal flush problem

This fixes a bug which resulted in poor performance due to flushing the journal too often. The code path in question was via the inode_go_sync() function in glops.c. The solution is not to flush the journal immediately when inodes are ejected from memory, but batch up the work for glockd to deal with later on. This means that glocks may now live on beyond the end of the lifetime of their inodes (but not very much longer in the normal case).

Quoted 4 lines

This seems like multiple changes in one patch. The inode flush handling is changing quite significantly. The log flushing is also being changed.

They are rather closely linked, so it seems to make sense to change both at once. See further comments below.

Quoted 36 lines
Quoted 36 lines

Also fixed in this patch is a bug (which was hidden by the bug mentioned above) in calculation of the number of free journal blocks.

The gfs2_logd process has been altered to be more responsive to the journal filling up. We now wake it up when the number of uncommitted journal blocks has reached the threshold level rather than trying to flush directly at the end of each transaction. This again means doing fewer, but larger, log flushes in general.

---
 fs/gfs2/daemon.c    |    7 +++-
 fs/gfs2/glock.c     |   17 +--------
 fs/gfs2/glock.h     |    1 -
 fs/gfs2/glops.c     |   93 +++++++++++++--------------------------------------
 fs/gfs2/log.c       |   17 +++++----
 fs/gfs2/meta_io.c   |    3 ++
 fs/gfs2/ops_super.c |    7 ++--
 7 files changed, 46 insertions(+), 99 deletions(-)
fs/gfs2/daemon.c +1 -0
index cab1f68..683cb5b 100644
@@ -112,6 +112,7 @@ int gfs2_logd(void *data)
 	struct gfs2_sbd *sdp = data;
 	struct gfs2_holder ji_gh;
 	unsigned long t;
+	int need_flush;
 
 	while (!kthread_should_stop()) {
 		/* Advance the log tail */
@@ -120,8 +121,10 @@ int gfs2_logd(void *data)
 		    gfs2_tune_get(sdp, gt_log_flush_secs) * HZ;
 
Quoted 9 lines

fs/gfs2/daemon.c int gfs2_logd(void *data) · line 126

Quoted 6 lines
fs/gfs2/daemon.c +3 -2
 		gfs2_ail1_empty(sdp, DIO_ALL);
-
-		if (time_after_eq(jiffies, t)) {
+		gfs2_log_lock(sdp);
+		need_flush = sdp->sd_log_num_buf > gfs2_tune_get(sdp, gt_incore_log_blocks);
+		gfs2_log_unlock(sdp);

Do we really need to lock the log just to get the log_num_buf? Seems like a serialization we don't need?

Yes we do need a lock, and bearing in mind that its only a spinlock I don't see that its going to be that much of a problem. Compared with the length of time it takes to flush the journal, it must surely be a completely minimal overhead.

Quoted 3 lines

So why does this loop have a sleep timeout and a flush interval? Shouldn't the schedual timeout be the same as the flush interval?

No it shouldn't. There are two things happening in this loop. The first is running gfs2_ail1_empty() which happens on a much more frequent basis than the log flushing. The log flushing runs every few seconds, or when its triggered by the journal getting close to being full:

Quoted 4 lines

fs/gfs2/daemon.c int gfs2_logd(void *data) · line 130

Quoted 4 lines
fs/gfs2/daemon.c +1 -0
+		if (need_flush || time_after_eq(jiffies, t)) {
 			gfs2_log_flush(sdp, NULL);
 			sdp->sd_log_flush_time = jiffies;
 		}

Steve.

Quoted 261 lines
Quoted 261 lines
fs/gfs2/glock.c +1 -16
index b8ba4d5..3c2ff81 100644
@@ -785,21 +785,6 @@ out:
 		gfs2_holder_put(new_gh);
 }
 
-void gfs2_glock_inode_squish(struct inode *inode)
-{
-	struct gfs2_holder gh;
-	struct gfs2_glock *gl = GFS2_I(inode)->i_gl;
-	gfs2_holder_init(gl, LM_ST_UNLOCKED, 0, &gh);
-	set_bit(HIF_DEMOTE, &gh.gh_iflags);
-	spin_lock(&gl->gl_spin);
-	gfs2_assert(inode->i_sb->s_fs_info, list_empty(&gl->gl_holders));
-	list_add_tail(&gh.gh_list, &gl->gl_waiters2);
-	run_queue(gl);
-	spin_unlock(&gl->gl_spin);
-	wait_for_completion(&gh.gh_wait);
-	gfs2_holder_uninit(&gh);
-}
-
 /**
  * state_change - record that the glock is now in a different state
  * @gl: the glock
@@ -1920,7 +1905,7 @@ out:
 
 static void scan_glock(struct gfs2_glock *gl)
 {
-	if (gl->gl_ops == &gfs2_inode_glops)
+	if (gl->gl_ops == &gfs2_inode_glops && gl->gl_object)
 		return;
 
 	if (gfs2_glmutex_trylock(gl)) {
fs/gfs2/glock.h +0 -1
index a331bf8..fb39108 100644
@@ -106,7 +106,6 @@ void gfs2_glock_dq_uninit_m(unsigned int
 void gfs2_glock_prefetch_num(struct gfs2_sbd *sdp, u64 number,
 			     const struct gfs2_glock_operations *glops,
 			     unsigned int state, int flags);
-void gfs2_glock_inode_squish(struct inode *inode);
 
 /**
  * gfs2_glock_nq_init - intialize a holder and enqueue it on a glock
fs/gfs2/glops.c +24 -69
index 60561ca..b068d10 100644
@@ -107,70 +107,6 @@ static void gfs2_pte_inval(struct gfs2_g
 }
 
 /**
- * gfs2_page_inval - Invalidate all pages associated with a glock
- * @gl: the glock
- *
- */
-
-static void gfs2_page_inval(struct gfs2_glock *gl)
-{
-	struct gfs2_inode *ip;
-	struct inode *inode;
-
-	ip = gl->gl_object;
-	inode = &ip->i_inode;
-	if (!ip || !S_ISREG(inode->i_mode))
-		return;
-
-	truncate_inode_pages(inode->i_mapping, 0);
-	gfs2_assert_withdraw(GFS2_SB(&ip->i_inode), !inode->i_mapping->nrpages);
-	clear_bit(GIF_PAGED, &ip->i_flags);
-}
-
-/**
- * gfs2_page_wait - Wait for writeback of data
- * @gl: the glock
- *
- * Syncs data (not metadata) for a regular file.
- * No-op for all other types.
- */
-
-static void gfs2_page_wait(struct gfs2_glock *gl)
-{
-	struct gfs2_inode *ip = gl->gl_object;
-	struct inode *inode = &ip->i_inode;
-	struct address_space *mapping = inode->i_mapping;
-	int error;
-
-	if (!S_ISREG(inode->i_mode))
-		return;
-
-	error = filemap_fdatawait(mapping);
-
-	/* Put back any errors cleared by filemap_fdatawait()
-	   so they can be caught by someone who can pass them
-	   up to user space. */
-
-	if (error == -ENOSPC)
-		set_bit(AS_ENOSPC, &mapping->flags);
-	else if (error)
-		set_bit(AS_EIO, &mapping->flags);
-
-}
-
-static void gfs2_page_writeback(struct gfs2_glock *gl)
-{
-	struct gfs2_inode *ip = gl->gl_object;
-	struct inode *inode = &ip->i_inode;
-	struct address_space *mapping = inode->i_mapping;
-
-	if (!S_ISREG(inode->i_mode))
-		return;
-
-	filemap_fdatawrite(mapping);
-}
-
-/**
  * meta_go_sync - sync out the metadata for this glock
  * @gl: the glock
  *
@@ -264,11 +200,24 @@ static void inode_go_drop_th(struct gfs2
 
 static void inode_go_sync(struct gfs2_glock *gl)
 {
+	struct gfs2_inode *ip = gl->gl_object;
+
+	if (ip && !S_ISREG(ip->i_inode.i_mode))
+		ip = NULL;
+
 	if (test_bit(GLF_DIRTY, &gl->gl_flags)) {
-		gfs2_page_writeback(gl);
 		gfs2_log_flush(gl->gl_sbd, gl);
+		if (ip)
+			filemap_fdatawrite(ip->i_inode.i_mapping);
 		gfs2_meta_sync(gl);
-		gfs2_page_wait(gl);
+		if (ip) {
+			struct address_space *mapping = ip->i_inode.i_mapping;
+			int error = filemap_fdatawait(mapping);
+			if (error == -ENOSPC)
+				set_bit(AS_ENOSPC, &mapping->flags);
+			else if (error)
+				set_bit(AS_EIO, &mapping->flags);
+		}
 		clear_bit(GLF_DIRTY, &gl->gl_flags);
 		gfs2_ail_empty_gl(gl);
 	}
@@ -283,14 +232,20 @@ static void inode_go_sync(struct gfs2_gl
 
 static void inode_go_inval(struct gfs2_glock *gl, int flags)
 {
+	struct gfs2_inode *ip = gl->gl_object;
 	int meta = (flags & DIO_METADATA);
 
 	if (meta) {
-		struct gfs2_inode *ip = gl->gl_object;
 		gfs2_meta_inval(gl);
-		set_bit(GIF_INVALID, &ip->i_flags);
+		if (ip)
+			set_bit(GIF_INVALID, &ip->i_flags);
+	}
+
+	if (ip && S_ISREG(ip->i_inode.i_mode)) {
+		truncate_inode_pages(ip->i_inode.i_mapping, 0);
+		gfs2_assert_withdraw(GFS2_SB(&ip->i_inode), !ip->i_inode.i_mapping->nrpages);
+		clear_bit(GIF_PAGED, &ip->i_flags);
 	}
-	gfs2_page_inval(gl);
 }
 
 /**
fs/gfs2/log.c +10 -7
index 0cace3d..6456fc3 100644
@@ -261,6 +261,12 @@ static void ail2_empty(struct gfs2_sbd *
  * @sdp: The GFS2 superblock
  * @blks: The number of blocks to reserve
  *
+ * Note that we never give out the last 6 blocks of the journal. Thats
+ * due to the fact that there is are a small number of header blocks
+ * associated with each log flush. The exact number can't be known until
+ * flush time, so we ensure that we have just enough free blocks at all
+ * times to avoid running out during a log flush.
+ *
  * Returns: errno
  */
 
@@ -274,7 +280,7 @@ int gfs2_log_reserve(struct gfs2_sbd *sd
 
 	mutex_lock(&sdp->sd_log_reserve_mutex);
 	gfs2_log_lock(sdp);
-	while(sdp->sd_log_blks_free <= blks) {
+	while(sdp->sd_log_blks_free <= (blks + 6)) {
 		gfs2_log_unlock(sdp);
 		gfs2_ail1_empty(sdp, 0);
 		gfs2_log_flush(sdp, NULL);
@@ -643,12 +649,9 @@ void gfs2_log_commit(struct gfs2_sbd *sd
 	up_read(&sdp->sd_log_flush_lock);
 
 	gfs2_log_lock(sdp);
-	if (sdp->sd_log_num_buf > gfs2_tune_get(sdp, gt_incore_log_blocks)) {
-		gfs2_log_unlock(sdp);
-		gfs2_log_flush(sdp, NULL);
-	} else {
-		gfs2_log_unlock(sdp);
-	}
+	if (sdp->sd_log_num_buf > gfs2_tune_get(sdp, gt_incore_log_blocks))
+		wake_up_process(sdp->sd_logd_process);
+	gfs2_log_unlock(sdp);
 }
 
 /**
fs/gfs2/meta_io.c +3 -0
index 3912d6a..939a09f 100644
@@ -472,6 +472,9 @@ int gfs2_meta_indirect_buffer(struct gfs
 	struct buffer_head *bh = NULL, **bh_slot = ip->i_cache + height;
 	int in_cache = 0;
 
+	BUG_ON(!gl);
+	BUG_ON(!sdp);
+
 	spin_lock(&ip->i_spin);
 	if (*bh_slot && (*bh_slot)->b_blocknr == num) {
 		bh = *bh_slot;
fs/gfs2/ops_super.c +3 -4
index 8635175..7685b46 100644
@@ -157,7 +157,8 @@ static void gfs2_write_super(struct supe
 static int gfs2_sync_fs(struct super_block *sb, int wait)
 {
 	sb->s_dirt = 0;
-	gfs2_log_flush(sb->s_fs_info, NULL);
+	if (wait)
+		gfs2_log_flush(sb->s_fs_info, NULL);
 	return 0;
 }
 
@@ -293,8 +294,6 @@ static void gfs2_clear_inode(struct inod
 	 */
 	if (inode->i_private) {
 		struct gfs2_inode *ip = GFS2_I(inode);
-		gfs2_glock_inode_squish(inode);
-		gfs2_assert(inode->i_sb->s_fs_info, ip->i_gl->gl_state == LM_ST_UNLOCKED);
 		ip->i_gl->gl_object = NULL;
 		gfs2_glock_schedule_for_reclaim(ip->i_gl);
 		gfs2_glock_put(ip->i_gl);
@@ -395,7 +394,7 @@ static void gfs2_delete_inode(struct ino
 	if (!inode->i_private)
 		goto out;
 
-	error = gfs2_glock_nq_init(ip->i_gl, LM_ST_EXCLUSIVE, LM_FLAG_TRY_1CB | GL_NOCACHE, &gh);
+	error = gfs2_glock_nq_init(ip->i_gl, LM_ST_EXCLUSIVE, LM_FLAG_TRY_1CB, &gh);
 	if (unlikely(error)) {
 		gfs2_glock_dq_uninit(&ip->i_iopen_gh);
 		goto out;