Re: [Qemu-devel] active block commit bug?

[Fam Zheng <famz@redhat.com>]

On Wed, 06/04 19:54, Eric Blake wrote:
> On 06/04/2014 06:12 PM, Jeff Cody wrote:
> 
> >> Huh? I thought that an active commit was not supposed to complete
> >> automatically, but that the job would remain around until I either
> >> 'block-job-cancel' or 'block-job-complete' it.  That is, I should have
> >> gotten a BLOCK_JOB_READY event and still see the job when I query for
> >> it.  Where am I going wrong, or did I uncover a bug in active commit?
> >>
> > 
> > I tried repeating your findings, but I couldn't, until I noticed that
> > 'f' was just a 0-length raw image in your test.
> > 
> > The snapshot file will be the same size, 0.  So when we go to perform
> > the active commit, we short-circuit at the beginning, since we are
> > committing a zero-length image:
> 
> That explains it.
> 
> >  
> >      s->common.len = bdrv_getlength(bs);
> >      if (s->common.len <= 0) {
> >          block_job_completed(&s->common, s->common.len);
> >          return;
> >      }
> >      ^^^^^
> >      we exit early here, with a completed message, since there is
> >      nothing to do.
> > 
> > If 'g' had increased to non-zero size, then you would have received a
> > BLOCK_JOB_READY instead.
> 
> Sounds like we have an off-by-one condition if empty files behave
> differently from other files.  We ought to fix that bug (not that your
> normal guest will ever have a 0-length backing file, but this was what I
> was trying to use for libvirt's probing of whether active commit is
> supported)
> 

Yes, agreed, this special case is only going to make management confused. I
will send a patch to fix this.

Eric, is this a good way to probe the active commit? I was expecting full
instrospection of QMP could do it, but I don't know about the status of that
piece of work. Amos, any ideas?

Fam