On Wed, Feb 27, 2013 at 04:00:28PM +0100, Benoît Canet wrote:
> > > - if ((refcount == 1) != ((l2_entry & QCOW_OFLAG_COPIED)
> > > != 0)) {
> > > + if (!s->has_dedup &&
> > > + (refcount == 1) != ((l2_entry & QCOW_OFLAG_COPIED)
> > > != 0)) {
> > > + fprintf(stderr, "ERROR OFLAG_COPIED: offset=%"
> > > + PRIx64 " refcount=%d\n", l2_entry, refcount);
> > > + res->corruptions++;
> > > + }
> >
> > Why is this warning suppressed when dedup is enabled? The meaning of
> > QCOW_OFLAG_COPIED is that refcount == 1. If this invariant is violated
> > then something is wrong.
>
> When deduplication is done refcount will be bigger than one and
> QCOW_OFLAG_COPIED will be cleared.
>
> Then if enough logical clustere pointing to the same physical cluster are
> rewritten with something else the refcount will goes down back to one.
>
> But this time QCOW_OFLAG_COPIED can be set back so this equality won't be
> true.
When the refcount decreases to 1 again we need to set QCOW_OFLAG_COPIED
again. qcow2-snapshot.c:qcow2_snapshot_delete() does this with:
/* must update the copied flag on the current cluster offsets */
ret = qcow2_update_snapshot_refcount(bs, s->l1_table_offset, s->l1_size, 0);
Is dedup not restoring QCOW_OFLAG_COPIED?
Stefan
