diff options
author | Jan Kara <jack@suse.cz> | 2009-12-03 13:39:28 +0100 |
---|---|---|
committer | Jan Kara <jack@suse.cz> | 2009-12-14 21:40:04 +0100 |
commit | 2c948b3f86e5f0327e2e57858600af6e6f0ae29a (patch) | |
tree | a40d5fd7a5757d26fc41829e5620b4685bac44e2 /fs/udf/file.c | |
parent | e971b0b9e0dd50d9ceecb67a6a6ab80a80906033 (diff) |
udf: Avoid IO in udf_clear_inode
It is not very good to do IO in udf_clear_inode. First, VFS does not really
expect inode to become dirty there and thus we have to write it ourselves,
second, memory reclaim gets blocked waiting for IO when it does not really
expect it, third, the IO pattern (e.g. on umount) resulting from writes in
udf_clear_inode is bad and it slows down writing a lot.
The reason why UDF needed to do IO in udf_clear_inode is that UDF standard
mandates extent length to exactly match inode size. But when we allocate
extents to a file or directory, we don't really know what exactly the final
file size will be and thus temporarily set it to block boundary and later
truncate it to exact length in udf_clear_inode. Now, this is changed to
truncate to final file size in udf_release_file for regular files. For
directories and symlinks, we do the truncation at the moment when learn
what the final file size will be.
Signed-off-by: Jan Kara <jack@suse.cz>
Diffstat (limited to 'fs/udf/file.c')
-rw-r--r-- | fs/udf/file.c | 1 |
1 files changed, 1 insertions, 0 deletions
diff --git a/fs/udf/file.c b/fs/udf/file.c index b80cbd78833..f311d509b6a 100644 --- a/fs/udf/file.c +++ b/fs/udf/file.c @@ -196,6 +196,7 @@ static int udf_release_file(struct inode *inode, struct file *filp) mutex_lock(&inode->i_mutex); lock_kernel(); udf_discard_prealloc(inode); + udf_truncate_tail_extent(inode); unlock_kernel(); mutex_unlock(&inode->i_mutex); } |