delalloc makes data=writeback safer

delalloc makes data=writeback safer

[Dmitry Monakhov]

Everybody know that writeback mode for ext3 is potentially result in
stale data in case of unclean umount. The same is true for ext4 with
nodelalloc.
But as far as i can see things is getting better in case of delalloc.
In case of delalloc we have following calltrace:
[writeback thread]
ext4_da_writepages
  journal_start()
  ext4_map_blocks() -> real block allocation
  mpage_da_submit_io() -> submit_bh() -> submit_bio()
  journal_stop()
So journal will be closed only after data was issued. 
And if journal->j_dev == journal->j_fs_dev (which is usually the case)
data will be serialized with metadata. So it looks like it is impassible
to see stale data even after unclean umount.

Do I miss something? May be this was already discussed before, 
if so please post a link to previous discussion.

Re: delalloc makes data=writeback safer

[Jan Kara]

> Everybody know that writeback mode for ext3 is potentially result in
> stale data in case of unclean umount. The same is true for ext4 with
> nodelalloc.
> But as far as i can see things is getting better in case of delalloc.
> In case of delalloc we have following calltrace:
> [writeback thread]
> ext4_da_writepages
>   journal_start()
>   ext4_map_blocks() -> real block allocation
>   mpage_da_submit_io() -> submit_bh() -> submit_bio()
>   journal_stop()
> So journal will be closed only after data was issued. 
> And if journal->j_dev == journal->j_fs_dev (which is usually the case)
> data will be serialized with metadata. So it looks like it is impassible
> to see stale data even after unclean umount.
> 
> Do I miss something? May be this was already discussed before, 
> if so please post a link to previous discussion.
  Well, we turn off delayed allocation when a filesystem gets close to
ENOSPC so it's not completely reliable. But besides this, your observation
looks correct to me.

								Honza
-- 
Jan Kara <jack@suse.cz>
SuSE CR Labs