Re: [XEN PATCH v12 2/7] x86/pvh: Allow (un)map_pirq when dom0 is PVH

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

On Fri, Aug 02, 2024 at 02:37:24AM +0000, Chen, Jiqian wrote:
> On 2024/7/31 21:03, Roger Pau Monné wrote:
> > On Wed, Jul 31, 2024 at 01:39:40PM +0200, Jan Beulich wrote:
> >> On 31.07.2024 13:29, Roger Pau Monné wrote:
> >>> On Wed, Jul 31, 2024 at 11:55:35AM +0200, Jan Beulich wrote:
> >>>> On 31.07.2024 11:37, Roger Pau Monné wrote:
> >>>>> On Wed, Jul 31, 2024 at 11:02:01AM +0200, Jan Beulich wrote:
> >>>>>> On 31.07.2024 10:51, Roger Pau Monné wrote:
> >>>>>>> I agree with (a), but I don't think enabling PVH dom0 usage of the
> >>>>>>> hypercalls should be gated on this.  As said a PV dom0 is already
> >>>>>>> capable of issuing PHYSDEVOP_{,un}map_pirq operations against a PVH
> >>>>>>> domU.
> >>>>>>
> >>>>>> Okay, I can accept that as an intermediate position. We ought to deny
> >>>>>> such requests at some point though for PVH domains, the latest in the
> >>>>>> course of making vPCI work there.
> >>>>>
> >>>>> Hm, once physdev_map_pirq() works as intended against PVH domains, I
> >>>>> don't see why we would prevent the usage of PHYSDEVOP_{,un}map_pirq
> >>>>> against such domains.
> >>>>
> >>>> Well. If it can be made work as intended, then I certainly agree. 
> >>>> However,
> >>>> without even the concept of pIRQ in PVH I'm having a hard time seeing how
> >>>> it can be made work. Iirc you were advocating for us to not introduce 
> >>>> pIRQ
> >>>> into PVH.
> >>>
> >>> From what I'm seeing here the intention is to expose
> >>> PHYSDEVOP_{,un}map_pirq to PVH dom0, so there must be some notion of
> >>> pIRQs or akin in a PVH dom0?  Even if only for passthrough needs.
> >>
> >> Only in so far as it is an abstract, handle-like value pertaining solely
> >> to the target domain.
> >>
> >>>> Maybe you're thinking of re-using the sub-ops, requiring PVH domains to
> >>>> pass in GSIs?
> >>>
> >>> I think that was one my proposals, to either introduce a new
> >>> hypercall that takes a GSI, or to modify the PHYSDEVOP_{,un}map_pirq
> >>> in an ABI compatible way so that semantically the field could be a GSI
> >>> rather than a pIRQ.  We however would also need a way to reference an
> >>> MSI entry.
> >>
> >> Of course.
> >>
> >>> My main concern is not with pIRQs by itself, pIRQs are just an
> >>> abstract way to reference interrupts, my concern and what I wanted to
> >>> avoid on PVH is being able to route pIRQs over event channels.  IOW:
> >>> have interrupts from physical devices delivered over event channels.
> >>
> >> Oh, I might have slightly misunderstood your intentions then.
> > 
> > My intention would be to not even use pIRQs at all, in order to avoid
> > the temptation of the guest itself managing interrupts using
> > hypercalls, hence I would have preferred that abstract interface to be
> > something else.
> > 
> > Maybe we could even expose the Xen IRQ space directly, and just use
> > that as interrupt handles, but since I'm not the one doing the work
> > I'm not sure it's fair to ask for something that would require more
> > changes internally to Xen.
> > 
> >>>> I think I suggested something along these lines also to
> >>>> Jiqian, yet with the now intended exposure to !has_pirq() domains I'm
> >>>> not sure this could be made work reliably.
> >>>
> >>> I'm afraid I've been lacking behind on reviewing those series.
> >>>
> >>>> Which reminds me of another question I had: What meaning does the pirq
> >>>> field have right now, if Dom0 would issue the request against a PVH DomU?
> >>>> What meaning will it have for a !has_pirq() HVM domain?
> >>>
> >>> The pirq field could be a way to reference an interrupt.  It doesn't
> >>> need to be exposed to the PVH domU at all, but it's a way for the
> >>> device model to identify which interrupt should be mapped to which
> >>> domain.
> >>
> >> Since pIRQ-s are per-domain, _that_ kind of association won't be
> >> helped. But yes, as per above it could serve as an abstract handle-
> >> like value.
> > 
> > I would be fine with doing the interrupt bindings based on IRQs
> > instead of pIRQs, but I'm afraid that would require more changes to
> > hypercalls and Xen internals.
> > 
> > At some point I need to work on a new interface to do passthrough, so
> > that we can remove the usage of domctls from QEMU.  That might be a
> > good opportunity to switch from using pIRQs.
> 
> Thanks for your input, but I may be a bit behind you with my knowledge and 
> can't fully understand the discussion.
> How should I modify this question later?
> Should I add a new hypercall specifically for passthrough?
> Or if it is to prevent the (un)map from being used for PVH guests, can I just 
> add a new function to check if the subject domain is a PVH type? Like 
> is_pvh_domain().

I think that would be part of a new interface, as said before I don't
think it would be fair to force you to do all this work.  I won't
oppose with the approach to attempt to re-use the existing interfaces
as much as possible.

I think this patch needs to be adjusted to drop the change to
xen/arch/x86/physdev.c, as just allowing PHYSDEVOP_{,un}map_pirq
without any change to do_physdev_op() should result in the correct
behavior.

Thanks, Roger.