[PATCH] ext3: reduce allocate-with-reservation lock latencies

Currently in ext3 block reservation code, the global filesystem reservation
tree lock (rsv_block) is hold during the process of searching for a space
to make a new reservation window, including while scaning the block bitmap
to verify if the avalible window has a free block.  Holding the lock during
bitmap scan is unnecessary and could possibly cause scalability issue and
latency issues.

This patch tries to address this by dropping the lock before scan the
bitmap.  Before that we need to reserve the open window in case someone
else is targetting at the same window.  Question was should we reserve the
whole free reservable space or just the window size we need.  Reserve the
whole free reservable space will possibly force other threads which
intended to do block allocation nearby move to another block group(cause
bad layout).  In this patch, we just reserve the desired size before drop
the lock and scan the block bitmap.  This patch fixed a ext3 reservation
latency issue seen on a cvs check out test.  Patch is tested with many fsx,
tiobench, dbench and untar a kernel test.

Signed-Off-By: Mingming Cao <cmm@us.ibm.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>

1 files changed, 4 insertions, 0 deletions

diff --git a/fs/ext3/file.c b/fs/ext3/file.c
index 5ad8cf0292df..98e78345ead9 100644
--- a/fs/ext3/file.c
+++ b/fs/ext3/file.c
@@ -36,7 +36,11 @@ static int ext3_release_file (struct inode * inode, struct file * filp)
 	/* if we are the last writer on the inode, drop the block reservation */
 	if ((filp->f_mode & FMODE_WRITE) &&
 			(atomic_read(&inode->i_writecount) == 1))
+	{
+		down(&EXT3_I(inode)->truncate_sem);
 		ext3_discard_reservation(inode);
+		up(&EXT3_I(inode)->truncate_sem);
+	}
 	if (is_dx(inode) && filp->private_data)
 		ext3_htree_free_dir_info(filp->private_data);