Re: Ext4 fiemap implementation

["Darrick J. Wong" <darrick.wong@oracle.com>]

On Thu, Jun 07, 2018 at 10:31:01AM +0200, Carlos Maiolino wrote:
> On Wed, Jun 06, 2018 at 07:40:18AM -0700, Darrick J. Wong wrote:
> > On Wed, Jun 06, 2018 at 03:13:39PM +0200, Carlos Maiolino wrote:
> > > Sigh..
> > > 
> > > On Mon, Jun 04, 2018 at 09:43:09AM -0700, Darrick J. Wong wrote:
> > > > On Sat, Jun 02, 2018 at 11:28:53PM -0400, Theodore Y. Ts'o wrote:
> > > > > On Fri, Jun 01, 2018 at 10:01:54AM -0500, Eric Sandeen wrote:
> > > > > > > Ted, is there any restriction why ext4_fiemap isn't using iomap_fiemap()? Or any
> > > > > > > reason why ext4 fiemap always returns the offset from the beginning of the
> > > > > > > extent? Would you oppose to have it updated to return the offset initially
> > > > > > > requested? Or maybe, change ext4_fiemap() to use iomap_fiemap()?
> > > > > 
> > > > > ext4_fiemap() predates iomap_fiemap().  In fact, it used to be that
> > > > > all of the file systems had their own fiemap() implementation.   
> > > > > 
> > > > > > > I read the fiemap documentation, but I didn't get a clear understanding if
> > > > > > > fiemap should be returning the beginning of the extent, the offset initially
> > > > > > > requested, or if it depends on FS implementation.
> > > > > > 
> > > > > > I think the fiemap docs[1] explicitly state that ext4's behavior is valid:
> > > > > > 
> > > > > > > Extents returned mirror
> > > > > > > those on disk - that is, the logical offset of the 1st returned extent
> > > > > > > may start before fm_start, and the range covered by the last returned
> > > > > > > extent may end after fm_length.
> > > > > 
> > > > > Actually, I read, "Extents returned mirror those on disk" as meaning
> > > > > that the ext4 behavior is *mandated* by the docs.  It would be
> > > > > interesting to see what XFS did before the iomap_fiemap() conversion.
> > > > > Or it could have been that the docs were inconsistent with what XFS
> > > > > was doing and then when when ext4_fiemap() was implemented, we
> > > > > followed the docs.  Some software archeology would be required to know
> > > > > for sure.
> > > > 
> > > > IIRC the pre-iomap xfs_vn_fiemap implementation only returned extent
> > > > data for the block range requested.  As far as I can tell, the current
> > > > xfs iomap implementation retains that behavior.
> > > > 
> > > > The fiemap spec says that "it is valid for an extents [sic] logical
> > > > offset to start before the request or its logical length to extend past
> > > > the request".  To my eyes, that means either behavior is acceptable.
> > > > 
> > > 
> > > Ok, thanks for the input everyone. I believe Eric's idea then is the one which
> > > makes more sense. If both behaviors are valid, to make fiemap() usage for
> > > fibmap, I think I'll need to get the extent returned by the filesystem and look
> > > for the block into the extent. Thanks a lot for the ideas.
> > 
> > Just to throw another monkey wrench into the machine, have you
> > considered using iomap_bmap() instead?  It's new for 4.18...
> 
> No, but thanks, I'll look into it.

You might also look at converting ext4 to use iomap_swapfile_activate
since that's new too.  iomap_bmap won't report blocks for unwritten
extents (because we shouldn't be pointing userspace at stale disk
blocks), which breaks swapfiles with unwritten extents.

--D

> 
> > 
> > --D
> > 
> > > 
> > > 
> > > -- 
> > > Carlos
> 
> -- 
> Carlos