Re: [RFC PATCH v7 03/14] x86/domain: Defer domain iommu initialization.

[Teddy Astie <teddy.astie@xxxxxxxxxx>]

Le 19/08/2026 à 17:10, Jan Beulich a écrit :
> On 20.11.2025 12:09, Teddy Astie wrote:
> > For the IOMMU redesign, the iommu context pagetable is defined once during
> > initialization. When reusing P2M pagetable, we want to ensure that this
> > pagetable is properly initialized.
> >
> > Signed-off-by Teddy Astie <teddy.astie@xxxxxxxxxx>
>  From what is said, I for one cannot deduce why the move is (a) necessary and
> (b) correct / safe to do.
>
> Jan
>
> > --- a/xen/arch/x86/domain.c
> > +++ b/xen/arch/x86/domain.c
> > @@ -927,9 +927,6 @@ int arch_domain_create(struct domain *d,
> >       if ( (rc = init_domain_irq_mapping(d)) != 0 )
> >           goto fail;
> >   
> > -    if ( (rc = iommu_domain_init(d, config->iommu_opts)) != 0 )
> > -        goto fail;
> > -
> >       psr_domain_init(d);
> >   
> >       if ( is_hvm_domain(d) )
> > @@ -948,6 +945,9 @@ int arch_domain_create(struct domain *d,
> >       else
> >           ASSERT_UNREACHABLE(); /* Not HVM and not PV? */
> >   
> > +    if ( (rc = iommu_domain_init(d, config->iommu_opts)) != 0 )
> > +        goto fail;
> > +
> >       if ( (rc = tsc_set_info(d, XEN_CPUID_TSC_MODE_DEFAULT, 0, 0, 0)) != 0 )
> >       {
> >           ASSERT_UNREACHABLE();
(a)

For now in Xen, iommu_domain_init() doesn't create or reuse any 
pagetable (for hap_pt_share) *yet*, this is deferred to 
per-implementation code that is called much later.
The end idea is to move some of the pagetable initialization/tracking to 
iommu_domain_init() (then to iommu_context_init()), but in the 
hap_pt_share case, the P2M needs to be initialized first (e.g through 
paging_enable).
(b)

There is no dependency of iommu_domain_init() to be called before 
hvm_domain_initialise()/pv_domain_initialise(). In particular, we're not 
adding new p2m entries to the guest that would require IOMMU mappings.
So I don't see any issue moving the call later.

AIA, it's mostly preparatory patches (to avoid making the next ones more 
complex than they already are); it's not necessarily useful on its own.
Teddy

Attachment: OpenPGP_signature.asc
Description: OpenPGP digital signature