max vcpus in dom0

max vcpus in dom0

[Mukesh Rathor]

Hi,

Looking at a bug where dom0 is crashing coming up with more than 32
vcpus, the problem happens trying to initialize 32nd vcpu. I see
the shared info is limited to 32vcpus, implying we'd have a hard limit
of 32 vcpus in dom0, correct? 

For now, I can just fix our dom0 to not go above 32, and look at ways
of fixing it if there's a strong case for more than 32 vcpus in dom0.

thanks,
Mukesh

Re: max vcpus in dom0

[Jan Beulich]

>>> Mukesh Rathor  04/27/10 4:20 AM >>>
>Looking at a bug where dom0 is crashing coming up with more than 32
>vcpus, the problem happens trying to initialize 32nd vcpu. I see
>the shared info is limited to 32vcpus, implying we'd have a hard limit
>of 32 vcpus in dom0, correct? 

That's the legacy limit, yes. Any attempt to go beyond 32 CPUs requires
the kernel to use the per-vCPU-info-placement hypercall prior to bringing
up those vCPU-s.

Jan

Re: max vcpus in dom0

[Pasi Kärkkäinen]

On Mon, Apr 26, 2010 at 07:17:33PM -0700, Mukesh Rathor wrote:
> Hi,
> 
> Looking at a bug where dom0 is crashing coming up with more than 32
> vcpus, the problem happens trying to initialize 32nd vcpu. I see
> the shared info is limited to 32vcpus, implying we'd have a hard limit
> of 32 vcpus in dom0, correct? 
> 

'shared info' in Xen hypervisor or in the dom0 kernel?

> For now, I can just fix our dom0 to not go above 32, and look at ways
> of fixing it if there's a strong case for more than 32 vcpus in dom0.
> 

Which dom0 kernel did you test with? 

-- Pasi

Re: max vcpus in dom0

[Jeremy Fitzhardinge]

On 04/27/2010 01:59 AM, Pasi Kärkkäinen wrote:
>> Looking at a bug where dom0 is crashing coming up with more than 32
>> vcpus, the problem happens trying to initialize 32nd vcpu. I see
>> the shared info is limited to 32vcpus, implying we'd have a hard limit
>> of 32 vcpus in dom0, correct? 
>>
>>     
> 'shared info' in Xen hypervisor or in the dom0 kernel?
>   

Both ;) It's shared.

    J

Re: max vcpus in dom0

[Pasi Kärkkäinen]

On Tue, Apr 27, 2010 at 11:38:30AM -0700, Jeremy Fitzhardinge wrote:
> On 04/27/2010 01:59 AM, Pasi Kärkkäinen wrote:
> >> Looking at a bug where dom0 is crashing coming up with more than 32
> >> vcpus, the problem happens trying to initialize 32nd vcpu. I see
> >> the shared info is limited to 32vcpus, implying we'd have a hard limit
> >> of 32 vcpus in dom0, correct? 
> >>
> >>     
> > 'shared info' in Xen hypervisor or in the dom0 kernel?
> >   
> 
> Both ;) It's shared.
> 

Oh, shared like *that* :) Thanks.

-- Pasi

Re: max vcpus in dom0

[Mukesh Rathor]

On Tue, 27 Apr 2010 09:23:19 +0100
"Jan Beulich" <jbeulich@novell.com> wrote:

> >>> Mukesh Rathor  04/27/10 4:20 AM >>>
> >Looking at a bug where dom0 is crashing coming up with more than 32
> >vcpus, the problem happens trying to initialize 32nd vcpu. I see
> >the shared info is limited to 32vcpus, implying we'd have a hard
> >limit of 32 vcpus in dom0, correct? 
> 
> That's the legacy limit, yes. Any attempt to go beyond 32 CPUs
> requires the kernel to use the per-vCPU-info-placement hypercall
> prior to bringing up those vCPU-s.
> 
> Jan

Hi Jan,

Yes, I found your patch searching xen-devel implementing
VCPUOP_register_vcpu_info after I posted my msg. However, I don't find
any kernel using this other than pv-ops kernel. So I assume this is
work in progress. If you already have a patch for 2.6.18-xxx, please let
me know, if not, I'll work on it.


thanks,
Mukesh

Re: max vcpus in dom0

[Jan Beulich]

>>> Mukesh Rathor  04/28/10 4:35 AM >>>
>Yes, I found your patch searching xen-devel implementing
>VCPUOP_register_vcpu_info after I posted my msg. However, I don't find
>any kernel using this other than pv-ops kernel. So I assume this is
>work in progress. If you already have a patch for 2.6.18-xxx, please let
>me know, if not, I'll work on it.

While we use it in our newer forward ported trees (2.6.32
onwards), the patch isn't easily backported to 2.6.18, and
hence I refrained from submitting anything to xen-devel.

Jan

Re: Re: max vcpus in dom0

[Jeremy Fitzhardinge]

On 04/27/2010 07:34 PM, Mukesh Rathor wrote:
> Yes, I found your patch searching xen-devel implementing
> VCPUOP_register_vcpu_info after I posted my msg. However, I don't find
> any kernel using this other than pv-ops kernel. So I assume this is
> work in progress. If you already have a patch for 2.6.18-xxx, please let
> me know, if not, I'll work on it.
>   

What's the use-case for a dom0 with a large number of vcpus?

    J