Re: [Qemu-devel] [PATCH qom-next 04/59] pc: Add CPU as /machine/cpu[n]

[Igor Mammedov]

On 06/08/2012 03:04 PM, Andreas Färber wrote:
> Am 08.06.2012 14:52, schrieb Jan Kiszka:
> > On 2012-06-08 14:47, Igor Mammedov wrote:
> > > ----- Original Message -----
> > > > From: "Jan Kiszka" <address@hidden>
> > > > To: "Andreas Färber" <address@hidden>
> > > > Cc: "Igor Mammedov" <address@hidden>, "Anthony Liguori" <address@hidden>, 
> > > > address@hidden, "Igor
> > > > Mammedov" <address@hidden>, "Paolo Bonzini" <address@hidden>
> > > > Sent: Friday, June 8, 2012 2:36:53 PM
> > > > Subject: Re: [Qemu-devel] [PATCH qom-next 04/59] pc: Add CPU as /machine/cpu[n]
> > > >
> > > > On 2012-06-08 14:34, Andreas Färber wrote:
> > > > > Am 08.06.2012 14:05, schrieb Igor Mammedov:
> > > > > > On Fri, Jun 08, 2012 at 11:11:11AM +0200, Andreas Färber wrote:
> > > > > > > Another factor that is making this slightly difficult is that
> > > > > > > there are
> > > > > > > three APIC subclasses. Currently they all have an instance_size
> > > > > > > of
> > > > > > > sizeof(APICCommonState) so it could be created in-place if it
> > > > > > > actually
> > > > > > > is a part (child<>) of the CPU wrt hot-plug. Creating objects
> > > > > > > with
> > > > > > > object_new() in QOM instance_init is forbidden.
> > > > > > Any particular reason why object_new() in intifn is not
> > > > > > acceptable?
> > > > >
> > > > > It allocates memory, which may fail. The initfn must not fail, the
> > > > > realizefn may return an Error object.
> > > >
> > > > Since when do we fail gracefully on OOM again?
> > > Maybe Andreas means that we cannot report error to caller?
> > > If it's a case then lets pass error to object_new() and fail gracefully
> > > or simply abort on OOM.
> >
> > QEMU's policy on OOM is abort (that's what glib already does for us
> > theses days).
>
> Nah, that's not the whole truth.
Could you elaborate more on subj?
I've looked at different initfns, many of them call object_property_add() which 
may cause
OOM as well. So if object_property_add() is permitted then why not object_new()?

> (More on that when I've finished fixing my series.)
>
> Andreas

--
-----
 Igor