Re: [virtio] [OASIS Issue Tracker] Created: (VIRTIO-35) race condition with multi-dword config accesses

["Michael S. Tsirkin" <mst@redhat.com>]

On Tue, Sep 24, 2013 at 08:18:20PM +0930, Rusty Russell wrote:
> OASIS Issues Tracker <workgroup_mailer@lists.oasis-open.org> writes:
> > race condition with multi-dword config accesses
> > -----------------------------------------------
> >
> >                  Key: VIRTIO-35
> >                  URL: http://tools.oasis-open.org/issues/browse/VIRTIO-35
> >              Project: OASIS Virtual I/O Device (VIRTIO) TC
> >           Issue Type: Bug
> >             Reporter: Michael Tsirkin
> >
> >
> > on many architectures, accesses larger than 32 bit can not be atomic.
> > Thus access to a device config field of >4 bytes is inherently racy
> > in case field can change.
> >
> > For example, virtio-blk has
> >                 u64 capacity;
> > The following race can trigger:
> >   driver reads low 32 bit
> >   both low and high 32 bit change
> >   driver reads high 32 bit
> >
> > as a result, capacity observed is composed of
> > old low bits and new high bits which does not
> > make sense.
> >
> > For legacy devices, spec allowed byte by byte access,
> > making the race even more common.
> 
> Yes, in theory this is a problem with virtio_blk, which could change
> capacity (the spec is silent on this, but the Linux driver will handle
> it by re-reading capacity, exposing this race).
> 
> I don't think qemu supports changing a disk underneath a device anyway?

QEMU supports resizing I think.

> And in practice, I'm not sure it would ever change fast enough to
> trigger this race...

Well it's easy to make guest very slow :)

> So, should we try to fix it?
> Rusty.

Not really sure but I thought I'd put it on the table.

-- 
MST