On Wed, Aug 18, 2010 at 09:27:55AM -0500, Mike Gao wrote:
> Hi, all,
>
> I sync latest XFS last month and build into our embedded system. When I
> test, I found below issue.
> It seems that the xfs_iomap failed when it did allocat with
> xfs_ilock_nowait().
> I debug this and found that ip->i_lock is locked and there are sometime no
> unlock coming between two lock trying.
> This issue will give wrong data in file on disk exactly on that offset.
Ah, it looks like we dropped the EAGAIN error from xfs_map_blocks()
on the floor.
Christoph, the 2.6.34 code did this:
error:
......
/*
* If it's delalloc and we have nowhere to put it,
* throw it away, unless the lower layers told
* us to try again.
*/
if (err != -EAGAIN) {
if (!unmapped)
xfs_aops_discard_page(page);
ClearPageUptodate(page);
}
and the caller redirtied the page on EAGAIN. The EAGAIN comes from
the BMAPI_TRYLOCK case, so it is a valid return indicating that we
would have blocked on a non-blocking IO.
The current code does not handle EAGAIN at all, just discards the
page. This looks incorrect - I think that EAGAIN should still cause
the page to be redirtied, not discarded. I'm not sure if I missed
something else, though - can you remember why you removed the EAGAIN
case?
Cheers,
Dave.
--
Dave Chinner
david@xxxxxxxxxxxxx
|