XENFEAT_supervisor_mode_kernel ???

XENFEAT_supervisor_mode_kernel ???

[Mukesh Rathor]

Hi all,

I'm trying ot figure out what that's all about? I see in 64 PV domU, it's 
being checked prior to returning from syscall.  On the xen side, I see it tied 
to supervisor_mode_kernel, but the only info I find is changelog comment which 
suggests it's purpose is to allow dom0 run in ring 0... So why the check in a 
domU kernel?


Thanks,
Mukesh
|     |
|     | <= Invisible ribbon campain against overuse of macros :).
|     |

Re: XENFEAT_supervisor_mode_kernel ???

[Keir Fraser]

On 20/11/08 22:42, "Mukesh Rathor" <mukesh.rathor@oracle.com> wrote:

> I'm trying ot figure out what that's all about? I see in 64 PV domU, it's
> being checked prior to returning from syscall.  On the xen side, I see it tied
> to supervisor_mode_kernel, but the only info I find is changelog comment which
> suggests it's purpose is to allow dom0 run in ring 0... So why the check in a
> domU kernel?

It's useful if the hypervisor runs the guest kernel in ring 0. Current Xen
doesn't do that, but in future we may run 64-bit PV guests in HVM context as
that may actually be faster.

 -- Keir

Re: XENFEAT_supervisor_mode_kernel ???

[Mukesh Rathor]

Oh, cool. I was looking at syscalls on PV x86_64 guests, and one of my 
thoughts was running in HVM mode but do all MMU stuff using hypercalls like 
PV. Other than having to change the hyp domain create portions, setting things 
up, is there a technical reason it can't be done today... thanks, Mukesh

Keir Fraser wrote:
> On 20/11/08 22:42, "Mukesh Rathor" <mukesh.rathor@oracle.com> wrote:
> 
>> I'm trying ot figure out what that's all about? I see in 64 PV domU, it's
>> being checked prior to returning from syscall.  On the xen side, I see it tied
>> to supervisor_mode_kernel, but the only info I find is changelog comment which
>> suggests it's purpose is to allow dom0 run in ring 0... So why the check in a
>> domU kernel?
> 
> It's useful if the hypervisor runs the guest kernel in ring 0. Current Xen
> doesn't do that, but in future we may run 64-bit PV guests in HVM context as
> that may actually be faster.
> 
>  -- Keir
> 

Re: XENFEAT_supervisor_mode_kernel ???

[Keir Fraser]

No, not really. The hypervisor changes could likely not be that large.

 -- Keir

On 20/11/08 23:03, "Mukesh Rathor" <mukesh.rathor@oracle.com> wrote:

> Oh, cool. I was looking at syscalls on PV x86_64 guests, and one of my
> thoughts was running in HVM mode but do all MMU stuff using hypercalls like
> PV. Other than having to change the hyp domain create portions, setting things
> up, is there a technical reason it can't be done today... thanks, Mukesh
> 
> Keir Fraser wrote:
>> On 20/11/08 22:42, "Mukesh Rathor" <mukesh.rathor@oracle.com> wrote:
>> 
>>> I'm trying ot figure out what that's all about? I see in 64 PV domU, it's
>>> being checked prior to returning from syscall.  On the xen side, I see it
>>> tied
>>> to supervisor_mode_kernel, but the only info I find is changelog comment
>>> which
>>> suggests it's purpose is to allow dom0 run in ring 0... So why the check in
>>> a
>>> domU kernel?
>> 
>> It's useful if the hypervisor runs the guest kernel in ring 0. Current Xen
>> doesn't do that, but in future we may run 64-bit PV guests in HVM context as
>> that may actually be faster.
>> 
>>  -- Keir
>> 

Re: XENFEAT_supervisor_mode_kernel ???

[Jeremy Fitzhardinge]

Mukesh Rathor wrote:
> Oh, cool. I was looking at syscalls on PV x86_64 guests, and one of my 
> thoughts was running in HVM mode but do all MMU stuff using hypercalls 
> like PV. Other than having to change the hyp domain create portions, 
> setting things up, is there a technical reason it can't be done 
> today... thanks, Mukesh

It's sort of on my list of things to look at.  But I'm still rather tied 
up in the dom0 work, so if you want to have a stab at it, go ahead!

    J

Re: XENFEAT_supervisor_mode_kernel ???

[Mukesh Rathor]

Yup, I started looking into it, but significant progress after 
holidays/vacation late January. Thanks, mukesh


Jeremy Fitzhardinge wrote:
> Mukesh Rathor wrote:
>> Oh, cool. I was looking at syscalls on PV x86_64 guests, and one of my 
>> thoughts was running in HVM mode but do all MMU stuff using hypercalls 
>> like PV. Other than having to change the hyp domain create portions, 
>> setting things up, is there a technical reason it can't be done 
>> today... thanks, Mukesh
> 
> It's sort of on my list of things to look at.  But I'm still rather tied 
> up in the dom0 work, so if you want to have a stab at it, go ahead!
> 
>    J