Re: [Xen-devel] [xen-unstable test] 12876: tolerable FAIL - PUSHED

[Stefano Stabellini <stefano.stabellini@xxxxxxxxxxxxx>]

On Tue, 15 May 2012, Jan Beulich wrote:
> >>> On 15.05.12 at 12:27, Stefano Stabellini 
> >>> <stefano.stabellini@xxxxxxxxxxxxx>
> wrote:
> > On Tue, 15 May 2012, Ian Campbell wrote:
> >> >  After having fixed
> >> > the gntdev driver in our kernels and the pvops-centric shortcomings in
> >> > both qemu-s, the qdisk backend still looks somewhat unreliable in
> >> > testing that Olaf has performed. We haven't narrowed it so far, but
> >> > a resulting question of course is whether using that backend (and/or
> >> > qemu-upstream) by default for any guests is a good idea.
> >> 
> >> CCing Stefano who made the patch to have PV guests use this guy. Please
> >> do share details when you have them.
> > 
> > I would prefer precise bug reports, and possibly patches, to "somewhat
> > unreliable" :-)
> 
> Of course. But we barely got past all the basic issues...
> 
> > Please note that the userspace disk backend is basically the same in
> > upstream QEMU and qemu-xen-traditional,
> 
> I understand that, ...
> 
> > so switching back to the old
> > QEMU for pv guests wouldn't improve anything.
> 
> ... and I didn't mean to suggest that. I was rather trying to hint
> towards continuing to use blkback as default backend.

blkback is still the default backend for physical partitions and LVM
volumes, but without direct_IO support in loop.c is unsafe for files.
I wouldn't want to run my VM on a disk that is basically stored in RAM.
Also we don't really have choice when it comes to QCOW and QCOW2 images.

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel