Re: [PATCH 3/4] xfs/259: try to force loop device block size

["Darrick J. Wong" <djwong@kernel.org>]

On Sun, Jun 01, 2025 at 10:07:00PM -0700, Christoph Hellwig wrote:
> On Wed, May 28, 2025 at 03:22:26PM -0700, Darrick J. Wong wrote:
> > Welll... the only reason I patched the loop driver to turn ovn directio
> > by default is because writeback throttling for loop devices keeps
> > getting turned on and off randomly.  At this point I have NFI if
> > throttling is actually the desired behavior or not.  It makes fstests
> > crawl really slowly.
> > 
> > On one hand it seems bogus that a loopbacked filesystem with enough
> > dirty pages to trip the thresholds then gets throttled doing writeback
> > to the pagecache of the loop file, but OTOH it /is/ more dirty
> > pagecache.  Ultimately I think non-directio loop devices are stupid
> > especially when there are filesystems on top of them, but I bet there's
> > some user that would break if we suddenly started requiring directio
> > alignments.
> > 
> > Maybe RWF_DONTCACHE will solve this whenever it stabilizes.
> 
> Well, I'm all for using direct I/O loop devices by default.  But having
> non-standard kernel hacks for that is pretty silly.  Can we just make
> xfstests use direct I/O by default so that everyone uses the same
> configuration?

I guess we could just modify _create_loop_device to set directio from
creation and fall back to pagecache io if need be, instead of the weird
"create it then try to change the mode" dance that we do now.  Does that
sound better?

--D