Re: [PATCH 5/8] xen/pvh: Prevent PVH guests from using PIC, RTC and IOAPIC

[Roger Pau Monné]

On Wed, Oct 26, 2016 at 10:50:21AM -0400, Boris Ostrovsky wrote:
> On 10/26/2016 06:42 AM, Roger Pau Monné wrote:
> > On Fri, Oct 14, 2016 at 02:05:15PM -0400, Boris Ostrovsky wrote:
> >> Make sure they don't use these devices since they are not emulated
> >> for unprivileged PVH guest.
> >>
> >> Also don't initialize hypercall page for them in init_hvm_pv_info()
> >> since this has already been done.
> >>
> >> Signed-off-by: Boris Ostrovsky <boris.ostrovsky@xxxxxxxxxx>
> >> ---
> >>  arch/x86/xen/enlighten.c | 24 +++++++++++++++++-------
> >>  1 file changed, 17 insertions(+), 7 deletions(-)
> >>
> >> diff --git a/arch/x86/xen/enlighten.c b/arch/x86/xen/enlighten.c
> >> index d38d568..6c1a330 100644
> >> --- a/arch/x86/xen/enlighten.c
> >> +++ b/arch/x86/xen/enlighten.c
> >> @@ -1803,10 +1803,21 @@ static void __init init_hvm_pv_info(void)
> >>  	minor = eax & 0xffff;
> >>  	printk(KERN_INFO "Xen version %d.%d.\n", major, minor);
> >>  
> >> -	cpuid(base + 2, &pages, &msr, &ecx, &edx);
> >> +	xen_domain_type = XEN_HVM_DOMAIN;
> >>  
> >> -	pfn = __pa(hypercall_page);
> >> -	wrmsr_safe(msr, (u32)pfn, (u32)(pfn >> 32));
> >> +	/* PVH set up hypercall page earlier in xen_prepare_pvh() */
> >> +	if (xen_pvh_domain()) {
> >> +		pv_info.name = "Xen PVH";
> >> +#ifdef CONFIG_ACPI
> >> +		/* No PIC or IOAPIC */
> > Shouldn't this be fetched from the MADT ACPI table if ACPI is available 
> > (rsdp_paddr != 0 in start_info)?
> 
> 
> At this point we haven't parsed ACPI yet (with or without rsdp_paddr,
> which we don't set anyway for domU) so we don't know whether we have PIC
> or IOAPIC.

I guess I'm missing something, but if we are providing ACPI tables to a PVH
DomU rsdp_paddr should be set, or else we are failing to comply with our own
start info specification.
 
> Having said that, I will probably remove this ("acpi_irq_model =
> ACPI_IRQ_MODEL_PLATFORM", together with the comment) since I am working
> on adding SCI support via an event channel and ACPI_IRQ_MODEL_PIC model,
> which is default, seems to work OK.

Hm, right, that might be an option. Do you know how hardware-reduced ACPI 
implementations (which IIRC also don't have a SCI interrupt) deliver events 
to the OS? Or it's simply not possible in that case?

Roger.