Re: [Xen-devel] Wrong cpupool handling

[Juergen Gross <jgross@xxxxxxxx>]

On 11/12/2014 10:53 AM, Dietmar Hahn wrote:
> Am Dienstag 11 November 2014, 15:21:01 schrieb Juergen Gross:
> > Hi again,
> >
> > On 11/11/2014 01:18 PM, Dietmar Hahn wrote:
> > > Hi list,
> > >
> > > When creating a cpupool, starting and destroying a guest within this pool,
> > > then removing this pool doesn't work because of EBUSY.
> > >
> > > It seems the cause of this behavior is the commit
> > > bac6334b51d9bcfe57ecf4a4cb5288348fcf044a.
> > >
> > > In domain_kill() the function sched_move_domain() gets called changing the
> > > d->cpupool pointer to the new cpupool without incrementing/decrementing the
> > > counters "n_dom" of the new/old cpupool.
> > >
> > > This leads to decrementing the wrong  cpupool0->n_dom counter when
> > > cpupool_rm_domain() gets called at the end and my own cpupool can't be
> > > destroyed because n_dom = 1!
> > >
> > > I don't have a fast patch because I'am not enough familiar with the code
> > > this time but I think it should be fixed for 4.5.
> > Please discard previous patch, try this one.
> Yes this patch works.
Thanks. Can I add your "tested-by:"?

> But I think in general a better solution would be to have the changing of the
> cpupool pointer in sched_move_domain() together with the increment/decrement
> of the counters but I see the locking problem.
The scheduler should never change cpupool owned data. The cpupool
pointer is domain data, so changing this is okay.


Juergen


_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel