On Mon, Feb 18, 2013 at 04:42:55PM +0100, Stefan Hajnoczi wrote:
> On Wed, Feb 13, 2013 at 02:22:09PM +0100, Kevin Wolf wrote:
> > diff --git a/block/qcow2.c b/block/qcow2.c
> > index 57552aa..2819336 100644
> > --- a/block/qcow2.c
> > +++ b/block/qcow2.c
> > @@ -774,11 +774,33 @@ static void coroutine_fn process_l2meta(void *opaque)
> > m->sleeping = false;
> > }
> >
> > +again:
> > qemu_co_mutex_lock(&s->lock);
> >
> > ret = qcow2_alloc_cluster_link_l2(bs, m);
> > if (ret < 0) {
> > - /* FIXME */
> > + /*
> > + * This is a nasty situation: We have already completed the
> > allocation
> > + * write request and returned success, so just failing it isn't
> > + * possible. We need to make sure to return an error during the
> > next
> > + * flush.
> > + *
> > + * However, we still can't drop the l2meta because we want I/O
> > errors
> > + * to be recoverable e.g. after the block device has been grown or
> > the
> > + * network connection restored. Sleep until the next flush comes
> > and
> > + * then retry.
> > + */
>
> A failed flush is live migrated by hw/virtio-blk.c but what happens when
> we fail during drain?
That's a very good questions. Looks like things become rather hairy...
This would be a case where we really need a VMState for block drivers
(which is in fact how the whole rerror/werror handling would have been
implemented best).
Juan, any chance to introduce such a thing without breaking everything?
Is there something like optional top-level sections?
Kevin
