Re: [Qemu-devel] [PATCH V7 4/6] qcow2: cancel the modification on fail in qcow2_snapshot_create()

[Wenchao Xia]

于 2013/12/23 10:57, Wenchao Xia 写道:
> 于 2013/12/20 22:20, Stefan Hajnoczi 写道:
> > On Thu, Dec 05, 2013 at 08:02:50PM +0800, Wenchao Xia wrote:
> > > +restore_refcount:
> > > +    if (qcow2_update_snapshot_refcount(bs, s->l1_table_offset,
> > > s->l1_size, -1)
> > > +        < 0 && errp) {
> > > +        /* Nothing can be done now, need image check later */
> > > +        error_setg(&err, "%s\nqcow2: Error in restoring refcount in
> > > snapshot",
> > > +                   error_get_pretty(*errp));
> > > +        error_free(*errp);
> > > +        *errp = NULL;
> > > +        error_propagate(errp, err);
> > > +    }
> >
> > We get here if writing the new snapshot list failed.  If
> > qcow2_update_snapshot_refcount(..., -1) also fails I think we should
> > skip qcow2_free_clusters() below.  We don't know the exact state of the
> > disk image anymore - better to leak clusters than to have a dangling
> > reference.
>
>    Make sense, dangling point should be avoid in any case, will fix,
> Thanks for reviewing!
>
> > > +dealloc_cluster:
> > > +    qcow2_free_clusters(bs, sn->l1_table_offset,
> > > +                        sn->l1_size * sizeof(uint64_t),
> > > +                        QCOW2_DISCARD_ALWAYS);
> > > +
> > >   fail:
> > >       g_free(sn->id_str);
> > >       g_free(sn->name);
> > > --
> > > 1.7.1
Hi, Stefan
  I have reconsidered the roll back process, there is many case we
should take care, so it is better to summarize a general rule to do such
cancel operations. I suggest: do a series of roll back operations,
when one fail, skip following roll back operation. For snapshot create,
the create action is:
allocate new L1 -> refcount+1 -> allocate sn_list -> update header
The mirrored rollback action can be:
deallocate L1 <- refcount-1 <- deallocate sn_list <- restore header

When fail happens in rollback action, simply stop following ones.
If you agree, I'd like to reorganize the patch as above.