x2apic mode vs. interrupt remapping

x2apic mode vs. interrupt remapping

[Jan Beulich]

Hi,

while the Linux kernel enters x2apic mode only when it was able to
initialize interrupt remapping (leaving aside a KVM special case), Xen
treats the two features as independent ones. We're having a report
where interrupt delivery doesn't work in x2apic mode (i.e. no-x2apic
makes the system boot successfully), and hence I'm wondering why
Xen would use this mode on a wider range of machines than Linux.

Thanks, Jan

RE: x2apic mode vs. interrupt remapping

[Zhang, Xiantao]

According to the #Cset20168, x2apic is only enabled when all VT-d engines support IR and EIM support, so the policy should be same with Linux's.
Xiantao

Jan Beulich wrote:
> Hi,
> 
> while the Linux kernel enters x2apic mode only when it was able to
> initialize interrupt remapping (leaving aside a KVM special case), Xen
> treats the two features as independent ones. We're having a report
> where interrupt delivery doesn't work in x2apic mode (i.e. no-x2apic
> makes the system boot successfully), and hence I'm wondering why
> Xen would use this mode on a wider range of machines than Linux.
> 
> Thanks, Jan
> 
> 
> _______________________________________________
> Xen-devel mailing list
> Xen-devel@lists.xensource.com
> http://lists.xensource.com/xen-devel

RE: x2apic mode vs. interrupt remapping

[Jan Beulich]

>>> "Zhang, Xiantao" <xiantao.zhang@intel.com> 03.03.10 14:32 >>>
>According to the #Cset20168, x2apic is only enabled when all VT-d
>engines support IR and EIM support, so the policy should be same
>with Linux's.

I would say that c/s has the intended behavior you describe, but since
iommu_setup() runs after enable_x2apic(), the call to
iommu_supports_eim() still returns true (iommu_enabled, iommu_qinval,
and iommu_inremap all have an initial value of 1, and with no drhd
present the function returns 1.

Jan

RE: x2apic mode vs. interrupt remapping

[Zhang, Xiantao]

When iommu_supports_eim is called in enable_x2apic, drhd is already parsed in acpi_boot_init, so iommu_supports_eim should return hardware's real capability. That is to say, if eim is no supported on platform, x2apic shouldn't be enabled at all. 
Xiantao


Jan Beulich wrote:
>>>> "Zhang, Xiantao" <xiantao.zhang@intel.com> 03.03.10 14:32 >>>
>> According to the #Cset20168, x2apic is only enabled when all VT-d
>> engines support IR and EIM support, so the policy should be same
>> with Linux's.
> 
> I would say that c/s has the intended behavior you describe, but since
> iommu_setup() runs after enable_x2apic(), the call to
> iommu_supports_eim() still returns true (iommu_enabled, iommu_qinval,
> and iommu_inremap all have an initial value of 1, and with no drhd
> present the function returns 1.
> 
> Jan

RE: x2apic mode vs. interrupt remapping

[Jan Beulich]

>>> "Zhang, Xiantao" <xiantao.zhang@intel.com> 03.03.10 15:09 >>>
>When iommu_supports_eim is called in enable_x2apic, drhd is already
>parsed in acpi_boot_init,

Provided there is any - in the given case there is none. Perhaps the
function should also return 0 if the list of drhd-s is empty?

>so iommu_supports_eim should return hardware's real capability. That
>is to say, if eim is no supported on platform, x2apic shouldn't be
>enabled at all. 

Jan

RE: x2apic mode vs. interrupt remapping

[Zhang, Xiantao]

Jan Beulich wrote:
>>>> "Zhang, Xiantao" <xiantao.zhang@intel.com> 03.03.10 15:09 >>>
>> When iommu_supports_eim is called in enable_x2apic, drhd is already
>> parsed in acpi_boot_init,
> 
> Provided there is any - in the given case there is none. Perhaps the
> function should also return 0 if the list of drhd-s is empty?

Yeah, if the drhd list is empty, it should return 0 instead of the default 1.  The condition check maybe need enhancement. 
Xiantao


>> so iommu_supports_eim should return hardware's real capability. That
>> is to say, if eim is no supported on platform, x2apic shouldn't be
>> enabled at all.
> 
> Jan

RE: x2apic mode vs. interrupt remapping

[Cui, Dexuan]

Zhang, Xiantao wrote:
> Jan Beulich wrote:
>>>>> "Zhang, Xiantao" <xiantao.zhang@intel.com> 03.03.10 15:09 >>>
>>> When iommu_supports_eim is called in enable_x2apic, drhd is already
>>> parsed in acpi_boot_init,
>> 
>> Provided there is any - in the given case there is none. Perhaps the
>> function should also return 0 if the list of drhd-s is empty?
> 
> Yeah, if the drhd list is empty, it should return 0 instead of the
> default 1.  The condition check maybe need enhancement. Xiantao

Oh, yes, I missed that!
Thank you for pointing this out. 

I'll make a patch for this soon.

-- Dexuan

>>> so iommu_supports_eim should return hardware's real capability. That
>>> is to say, if eim is no supported on platform, x2apic shouldn't be
>>> enabled at all.
>> 
>> Jan