Re: [PATCH RFC Linux] x86/xen: allow using x2APIC under Xen without CONFIG_XEN enabled

Re: [PATCH RFC Linux] x86/xen: allow using x2APIC under Xen without CONFIG_XEN enabled

[Boris Ostrovsky]

On 08/20/2015 08:04 AM, Jan Beulich wrote:
> While commit 4cca6ea04d31c claims to not have any functional effect on
> Xen, this isn't the case: Before that change, kernels built without
> CONFIG_XEN_PVHVM (a dependency which meanwhile became just CONFIG_XEN)
> were able to run in x2APIC mode just fine. Restore that behavior.
>
> This, however, still doesn't fix the case where CONFIG_HYPERVISOR_GUEST
> is not enabled, but I suppose this may be regarded as intentional.
>
> Signed-off-by: Jan Beulich<jbeulich@suse.com>
> ---
> The patch is RFC solely because the way the issue gets fixed doesn't
> look very neat, but I couldn't figure out a better way.

(+ x86 maintainers)

Can we provide something like xen_stub.c (that will have its own 
x86_hyper ops, probably only x2apic_available and maybe detect) which is 
built when !CONFIG_XEN?

Otherwise I see little reason to keep x2apic_available op and we should 
revert the portion of 4cca6ea04d31c that introduced it.

-boris

> ---
>   arch/x86/kernel/cpu/hypervisor.c |    4 ++++
>   1 file changed, 4 insertions(+)
>
> --- 4.2-rc7/arch/x86/kernel/cpu/hypervisor.c
> +++ 4.2-rc7-x86-xen-x2apic-available/arch/x86/kernel/cpu/hypervisor.c
> @@ -81,6 +81,10 @@ void __init init_hypervisor_platform(voi
>   
>   bool __init hypervisor_x2apic_available(void)
>   {
> +#ifndef CONFIG_XEN
> +	if (xen_x2apic_para_available())
> +		return true;
> +#endif
>   	return x86_hyper                   &&
>   	       x86_hyper->x2apic_available &&
>   	       x86_hyper->x2apic_available();
>
>
>

Re: [PATCH RFC Linux] x86/xen: allow using x2APIC under Xen without CONFIG_XEN enabled

[Jan Beulich]

>>> On 31.08.15 at 21:19, <boris.ostrovsky@oracle.com> wrote:
> On 08/20/2015 08:04 AM, Jan Beulich wrote:
>> While commit 4cca6ea04d31c claims to not have any functional effect on
>> Xen, this isn't the case: Before that change, kernels built without
>> CONFIG_XEN_PVHVM (a dependency which meanwhile became just CONFIG_XEN)
>> were able to run in x2APIC mode just fine. Restore that behavior.
>>
>> This, however, still doesn't fix the case where CONFIG_HYPERVISOR_GUEST
>> is not enabled, but I suppose this may be regarded as intentional.
>>
>> Signed-off-by: Jan Beulich<jbeulich@suse.com>
>> ---
>> The patch is RFC solely because the way the issue gets fixed doesn't
>> look very neat, but I couldn't figure out a better way.
> 
> (+ x86 maintainers)
> 
> Can we provide something like xen_stub.c (that will have its own 
> x86_hyper ops, probably only x2apic_available and maybe detect) which is 
> built when !CONFIG_XEN?
> 
> Otherwise I see little reason to keep x2apic_available op and we should 
> revert the portion of 4cca6ea04d31c that introduced it.

Actually I've considered that last option too, but it seemed no neater
than the presented patch (yet imo better than introducing a stub file).

Jan