On Fri, 1 May 2020, Julien Grall wrote:
> Hi Stefano,
>
> On 01/05/2020 02:31, Stefano Stabellini wrote:
> > On Wed, 15 Apr 2020, Julien Grall wrote:
> > > > diff --git a/xen/arch/arm/vgic-v3.c b/xen/arch/arm/vgic-v3.c
> > > > index 4e60ba15cc..4cf430f865 100644
> > > > --- a/xen/arch/arm/vgic-v3.c
> > > > +++ b/xen/arch/arm/vgic-v3.c
> > > > @@ -1677,13 +1677,25 @@ static int vgic_v3_domain_init(struct domain *d)
> > >
> > >
> > > I think you also want to modify vgic_v3_max_rdist_count().
> >
> > I don't think so: domUs even direct-mapped still only get 1 rdist
> > region. This patch is not changing the layout of the domU gic, it is
> > only finding a "hole" in the physical address space to make sure there
> > are no conflicts (or at least minimize the chance of conflicts.)
>
> How do you know the "hole" is big enough?
>
> >
> > > > * Domain 0 gets the hardware address.
> > > > * Guests get the virtual platform layout.
> > >
> > > This comment needs to be updated.
> >
> > Yep, I'll do
> >
> >
> > > > */
> > > > - if ( is_hardware_domain(d) )
> > > > + if ( is_domain_direct_mapped(d) )
> > > > {
> > > > unsigned int first_cpu = 0;
> > > > + unsigned int nr_rdist_regions;
> > > > d->arch.vgic.dbase = vgic_v3_hw.dbase;
> > > > - for ( i = 0; i < vgic_v3_hw.nr_rdist_regions; i++ )
> > > > + if ( is_hardware_domain(d) )
> > > > + {
> > > > + nr_rdist_regions = vgic_v3_hw.nr_rdist_regions;
> > > > + d->arch.vgic.intid_bits = vgic_v3_hw.intid_bits;
> > > > + }
> > > > + else
> > > > + {
> > > > + nr_rdist_regions = 1;
> > >
> > > What does promise your the rdist region will be big enough to cater all
> > > the
> > > re-distributors for your domain?
> >
> > Good point. I'll add an explicit check for that with at least a warning.
> > I don't think we want to return error because the configuration it is
> > still likely to work.
>
> No it is not going to work. Imagine you have have a guest with 3 vCPUs but the
> first re-distributor region can only cater 2 re-distributor. How is this going
> to be fine to continue?
>
> For dom0, we are re-using the same hole but possibly not all of them. Why
> can't we do that for domU?
I implemented what you suggested
