Re: [PATCH] xfs: don't invalidate whole file on DAX read/write

[Dave Chinner <david@fromorbit.com>]

On Wed, Aug 03, 2016 at 05:34:37PM +0200, Jan Kara wrote:
> On Wed 03-08-16 09:40:26, Dave Chinner wrote:
> > From: Dave Chinner <dchinner@redhat.com>
> > 
> > When we do DAX IO, we try to invalidate the entire page cache held
> > on the file. This is incorrect as it will trash the entire mapping
> > tree that now tracks dirty state in exceptional entries in the radix
> > tree slots.
> > 
> > What we are trying to do is remove cached pages (e.g from reads
> > into holes) that sit in the radix tree over the range we are about
> > to write to. Hence we should just limit the invalidation to the
> > range we are about to overwrite.
> 
> The patch looks good. Just one comment below.
> 
> > 
> > Reported-by: Jan Kara <jack@suse.cz>
> > Signed-off-by: Dave Chinner <dchinner@redhat.com>
> > ---
> >  fs/xfs/xfs_file.c | 13 ++++++++++++-
> >  1 file changed, 12 insertions(+), 1 deletion(-)
> > 
> > diff --git a/fs/xfs/xfs_file.c b/fs/xfs/xfs_file.c
> > index ed95e5b..e612a02 100644
> > --- a/fs/xfs/xfs_file.c
> > +++ b/fs/xfs/xfs_file.c
> > @@ -741,9 +741,20 @@ xfs_file_dax_write(
> >  	 * page is inserted into the pagecache when we have to serve a write
> >  	 * fault on a hole.  It should never be dirtied and can simply be
> >  	 * dropped from the pagecache once we get real data for the page.
> > +	 *
> > +	 * XXX: This is racy against mmap, and there's nothing we can do about
> > +	 * it. dax_do_io() should really do this invalidation internally as
> > +	 * it will know if we've allocated over a holei for this specific IO and
> > +	 * if so it needs to update the mapping tree and invalidate existing
> > +	 * PTEs over the newly allocated range. Remove this invalidation when
> > +	 * dax_do_io() is fixed up.
> 
> And would it be OK for XFS if dax_do_io() actually invalidated page cache /
> PTEs under just XFS_IOLOCK_SHARED? Because currently you seem to be careful
> to call invalidate_inode_pages2() only when holding the lock exclusively
> and then demote it to a shared one when calling dax_do_io().

That really only exists to prevent multiple IOs trying to do
invalidation at once. In the direct IO code, we don't want multiple
page cache flushers running at once - one is enough - so we
serialise on that state knowing that once the invalidation is done
the remaining EXCL lock waiters will pass straight through.

For DAX, I don't think that's a problem - the invalidation is
ranged, and it's unlikely there will be overlaps, and mapping/pte
invalidation is done under fine grained locks so we don't have to
worry about races there, either. So it seems fine to me to do this
under a SHARED lock. It will still serialise against truncate and
other extent manipulation operations, and that's mainly what we care
about here.

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com

_______________________________________________
xfs mailing list
xfs@oss.sgi.com
http://oss.sgi.com/mailman/listinfo/xfs