Re: [GEDI] [PATCH 07/17] gluster: Drop useless has_zero_init callback

[Niels de Vos <ndevos@redhat.com>]

On Mon, Feb 17, 2020 at 06:03:40AM -0600, Eric Blake wrote:
> On 2/17/20 2:06 AM, Niels de Vos wrote:
> > On Fri, Jan 31, 2020 at 11:44:26AM -0600, Eric Blake wrote:
> > > block.c already defaults to 0 if we don't provide a callback; there's
> > > no need to write a callback that always fails.
> > > 
> > > Signed-off-by: Eric Blake <eblake@redhat.com>
> > 
> > Reviewed-by: Niels de Vos <ndevos@redhat.com>
> > 
> 
> Per your other message,
> 
> On 2/17/20 2:16 AM, Niels de Vos wrote:
> > On Fri, Jan 31, 2020 at 11:44:31AM -0600, Eric Blake wrote:
> >> Since gluster already copies file-posix for lseek usage in block
> >> status, it also makes sense to copy it for learning if the image
> >> currently reads as all zeroes.
> >>
> 
> >> +static int qemu_gluster_known_zeroes(BlockDriverState *bs)
> >> +{
> >> +    /*
> >> +     * GlusterFS volume could be backed by a block device, with no way
> >
> > Actually, Gluster dropped support for volumes backed by block devices
> > (LVM) a few releases back. Nobody could be found that used it, and it
> > could not be combined with other Gluster features. All contents on a
> > Gluster volume is now always backed by 'normal' files on a filesystem.
> 
> That's useful to know.  Thanks!
> 
> >
> > Creation or truncation should behave just as on a file on a local
> > filesystem. So maybe qemu_gluster_known_zeroes is not needed at all?
> 
> Which version of gluster first required a regular filesystem backing for all
> gluster files?  Does qemu support older versions (in which case, what is the
> correct version-probing invocation to return 0 prior to that point, and 1
> after), or do all versions supported by qemu already guarantee zero
> initialization on creation or widening truncation by virtue of POSIX file
> semantics (in which case, patch 7 should instead switch to using
> .bdrv_has_zero_init_1 for both functions)?  Per configure, we probe for
> glusterfs_xlator_opt from gluster 4, which implies the code still tries to
> be portable to even older gluster, but I'm not sure if this squares with
> qemu-doc.texi which mentions our minimum distro policy (for example, now
> that qemu requires python 3 consistent with our distro policy, that rules
> out several older systems where older gluster was likely to be present).

The block device feature (storage/bd xlator) got deprecated in Gluster
5.0, and was removed with Gluster 6.0. Fedora 29 is the last version
that contained the bd.so xlator (glusterfs-server 5.0, deprecated).

All currently maintained and available Gluster releases should have
glusterfs_xlator_opt (introduced with glusterfs-3.5 in 2014). However, I
am not sure what versions are provided with different distributions. The
expectation is that at least Gluster 5 is provided, as older releases
will not get any updates anymore. See
https://www.gluster.org/release-schedule/ for a more detailed timeline.

Unfortunately there is no reasonable way to probe for the type of
backend (block or filesystem) that is used. So, a runtime check to be on
the extreme safe side to fallback on block device backends is not an
option.

HTH,
Niels