Re: [RFC XEN PATCH v3 2/3] x86/pvh: Add (un)map_pirq and setup_gsi for PVH dom0

[Roger Pau Monné <roger.pau@xxxxxxxxxx>]

On Thu, Dec 14, 2023 at 10:58:24AM +0100, Jan Beulich wrote:
> On 14.12.2023 10:55, Roger Pau Monné wrote:
> > On Thu, Dec 14, 2023 at 08:55:45AM +0000, Chen, Jiqian wrote:
> >> On 2023/12/13 15:03, Jan Beulich wrote:
> >>> On 13.12.2023 03:47, Chen, Jiqian wrote:
> >>>> On 2023/12/12 17:30, Jan Beulich wrote:
> >>>>> On 12.12.2023 07:49, Chen, Jiqian wrote:
> >>>>>> On 2023/12/11 23:31, Roger Pau Monné wrote:
> >>>>>>> On Mon, Dec 11, 2023 at 12:40:08AM +0800, Jiqian Chen wrote:
> >>>>>>>> --- a/xen/arch/x86/hvm/hypercall.c
> >>>>>>>> +++ b/xen/arch/x86/hvm/hypercall.c
> >>>>>>>> @@ -72,8 +72,11 @@ long hvm_physdev_op(int cmd, 
> >>>>>>>> XEN_GUEST_HANDLE_PARAM(void) arg)
> >>>>>>>>  
> >>>>>>>>      switch ( cmd )
> >>>>>>>>      {
> >>>>>>>> +    case PHYSDEVOP_setup_gsi:
> >>>>>>>
> >>>>>>> I think given the new approach on the Linux side patches, where
> >>>>>>> pciback will configure the interrupt, there's no need to expose
> >>>>>>> setup_gsi anymore?
> >>>>>> The latest patch(the second patch of v3 on kernel side) does setup_gsi 
> >>>>>> and map_pirq for passthrough device in pciback, so we need this and 
> >>>>>> below.
> >>>>>>
> >>>>>>>
> >>>>>>>>      case PHYSDEVOP_map_pirq:
> >>>>>>>>      case PHYSDEVOP_unmap_pirq:
> >>>>>>>> +        if ( is_hardware_domain(currd) )
> >>>>>>>> +            break;
> >>>>>>>
> >>>>>>> Also Jan already pointed this out in v2: this hypercall needs to be
> >>>>>>> limited so a PVH dom0 cannot execute it against itself.  IOW: refuse
> >>>>>>> the hypercall if DOMID_SELF or the passed domid matches the current
> >>>>>>> domain domid.
> >>>>>> Yes, I remember Jan's suggestion, but since the latest patch(the 
> >>>>>> second patch of v3 on kernel side) has change the implementation, it 
> >>>>>> does setup_gsi and map_pirq for dom0 itself, so I didn't add the 
> >>>>>> DOMID_SELF check.
> >>>>>
> >>>>> And why exactly would it do specifically the map_pirq? (Even the 
> >>>>> setup_gsi
> >>>>> looks questionable to me, but there might be reasons there.)
> >>>> Map_pirq is to solve the check failure problem. (pci_add_dm_done-> 
> >>>> xc_domain_irq_permission-> XEN_DOMCTL_irq_permission-> 
> >>>> pirq_access_permitted->domain_pirq_to_irq->return irq is 0)
> >>>> Setup_gsi is because the gsi is never be unmasked, so the gsi is never 
> >>>> be registered( vioapic_hwdom_map_gsi-> mp_register_gsi is never be 
> >>>> called).
> >>>
> >>> And it was previously made pretty clear by Roger, I think, that doing a 
> >>> "map"
> >>> just for the purpose of granting permission is, well, at best a temporary
> >>> workaround in the early development phase. If there's presently no 
> >>> hypercall
> >>> to _only_ grant permission to IRQ, we need to add one.
> >> Could you please describe it in detail? Do you mean to add a new hypercall 
> >> to grant irq access for dom0 or domU?
> >> It seems XEN_DOMCTL_irq_permission is the hypercall to grant irq access 
> >> from dom0 to domU(see XEN_DOMCTL_irq_permission-> irq_permit_access). 
> >> There is no need to add hypercall to grant irq access.
> >> We failed here (XEN_DOMCTL_irq_permission-> 
> >> pirq_access_permitted->domain_pirq_to_irq->return irq is 0) is because the 
> >> PVH dom0 didn't use PIRQ, so we can't get irq from pirq if "current" is 
> >> PVH dom0.
> > 
> > One way to bodge this would be to detect whether the caller of
> > XEN_DOMCTL_irq_permission is a PV or an HVM domain, and in case of HVM
> > assume the pirq field is a GSI.  I'm unsure however how that will work
> > with non-x86 architectures.
> > 
> > It would  be better to introduce a new XEN_DOMCTL_gsi_permission, or
> > maybe XEN_DOMCTL_intr_permission that can take a struct we can use to
> > accommodate GSIs and other arch specific interrupt identifiers.
> 
> How would you see MSI being handled then?

I wasn't really accounting for MSI here, as MSI is not handled by
XEN_DOMCTL_irq_permission now either.  My plan long term was to
introduce a new hypercall (part of dm_ops possibly) in order to be
able to bind MSI directly without having to 'map' it first.

Roger.