Re: [PATCH] x86/CPUID: unconditionally set XEN_HVM_CPUID_IOMMU_MAPPINGS

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

On Tue, Jan 19, 2021 at 11:16:06AM +0000, Andrew Cooper wrote:
> On 19/01/2021 11:14, Roger Pau Monné wrote:
> > On Mon, Jan 18, 2021 at 05:48:37PM +0000, Andrew Cooper wrote:
> >> On 18/01/2021 17:10, Roger Pau Monné wrote:
> >>> On Mon, Jan 18, 2021 at 05:04:19PM +0100, Jan Beulich wrote:
> >>>> On 18.01.2021 16:54, Roger Pau Monné wrote:
> >>>>> On Mon, Jan 18, 2021 at 12:05:12PM +0100, Jan Beulich wrote:
> >>>>>> On 15.01.2021 16:01, Roger Pau Monne wrote:
> >>>>>>> --- a/xen/arch/x86/traps.c
> >>>>>>> +++ b/xen/arch/x86/traps.c
> >>>>>>> @@ -1049,11 +1049,10 @@ void cpuid_hypervisor_leaves(const struct 
> >>>>>>> vcpu *v, uint32_t leaf,
> >>>>>>>              res->a |= XEN_HVM_CPUID_X2APIC_VIRT;
> >>>>>>>  
> >>>>>>>          /*
> >>>>>>> -         * Indicate that memory mapped from other domains (either 
> >>>>>>> grants or
> >>>>>>> -         * foreign pages) has valid IOMMU entries.
> >>>>>>> +         * Unconditionally set the flag to indicate this version of 
> >>>>>>> Xen has
> >>>>>>> +         * been fixed to create IOMMU mappings for grant/foreign 
> >>>>>>> maps.
> >>>>>>>           */
> >>>>>>> -        if ( is_iommu_enabled(d) )
> >>>>>>> -            res->a |= XEN_HVM_CPUID_IOMMU_MAPPINGS;
> >>>>>>> +        res->a |= XEN_HVM_CPUID_IOMMU_MAPPINGS;
> >>>>>> ... try to clarify the "Unconditionally" here?
> >>>>> I guess Unconditionally doesn't make much sense, so would be better to
> >>>>> start the sentence with 'Set ...' instead?
> >>>> Hmm, this would further move us away from the goal of the comment
> >>>> making sufficiently clear that a conditional shouldn't be (re-)
> >>>> introduced here, I would think. Since I can't seem to think of a
> >>>> good way to express this more briefly than in the description,
> >>>> and if maybe you can't either, perhaps there's no choice then to
> >>>> leave it as is, hoping that people would look at the commit before
> >>>> proposing a further change here.
> >>> /*
> >>>  * Unconditionally set the flag to indicate this version of Xen has
> >>>  * been fixed to create IOMMU mappings for grant/foreign maps.
> >>>  *
> >>>  * NB: this flag shouldn't be made conditional on IOMMU presence, as
> >>>  * it could force guests to resort to using bounce buffers when using
> >>>  * grant/foreign maps with devices.
> >>>  */
> >>>
> >>> Would be better? (albeit too verbose maybe).
> >> The comment should be rather more direct.
> >>
> >> 1) Xen 4.10 and older was broken WRT grant maps requesting a DMA
> >> mapping, and forgot to honour the guest's request.
> >> 2) 4.11 (and presumably backports) fixed the bug, so the map hypercall
> >> actually did what the guest asked.
> >> 3) To work around the bug, guests must bounce buffer all DMA, because it
> >> doesn't know whether the DMA is originating from an emulated or a real
> >> device.
> >> 4) This flag tells guests it is safe not to bounce-buffer all DMA to
> >> work around the bug.
> > /*
> >  * Old versions of Xen are broken when creating grant/foreign maps,
> >  * and will never create IOMMU entries for such mappings. This was
> >  * fixed in later versions of Xen, but guests wanting to work on
> >  * unpatched versions will need to use a bounce buffer in order to
> >  * avoid sending grant/foreign maps to devices. Whether such bounce
> >  * buffer mechanism is not needed is indicated by the presence of the
> >  * following CPUID flag.
> >  */
> >
> > Does that seem better?
> 
> Better, but the key point is that all DMA, emulated or real, needs
> bounce buffering because the guest kernel doesn't know the difference. 
> *This* is why the flag needs to be always present, because otherwise a
> guest will bounce buffer DMA for emulated devices.

I think this is clear from the text as I explicitly say 'devices'
(not emulated or passed through devices).

/*
 * Old versions of Xen are broken when creating grant/foreign maps,
 * and will never create IOMMU entries for such mappings. This was
 * fixed in later versions of Xen, but guests wanting to work on
 * unpatched versions will need to use a bounce buffer in order to
 * avoid sending grant/foreign maps to devices. Whether such bounce
 * buffer mechanism is not needed is indicated by the presence of the
 * following CPUID flag. Not exposing the flag will force guests to
 * fallback to bounce buffering when sending grant/foreign maps to any
 * device.
 */

If not I don't mind just using the list that you provided if Jan
agrees.

Thanks, Roger.