pv_ops dom0 kernel and VT cpu extensions

[Gerry Reno <greno@verizon.net>]

I started this discussion over on xen-users and I'm bringing it over to 
xen-devel.

Summarizing....



I compiled a pv_ops dom0 kernel from Jeremy's latest git tree and 
upgraded Xen to 3.4.1 following one of Boris' blog entries.  From the 
Xen perspective everything seems to work as expected.  But, I want to be 
able to run both Xen and KVM existing guests in a single cloud on a farm 
of VT-enabled machines.  So after booting into the new pv_ops dom0 
kernel as a test I tried starting one of my KVM guests.  It ran 
extremely slow so I knew immediately that it wasnt' using the VT 
acceleration.  I then looked at /proc/cpuinfo and saw that the VT cpu 
flag (svm on this machine) was not there.  The machine has a VT 
processor and hardware virtualization is enabled in the BIOS and I 
rebooted to check with a regular kernel and it of course the guest runs 
with VT acceleration there. 

My question is how can I get pv_ops dom0 kernel or Xen 3.4.1 to pass the 
cpu VT flags through so that existing KVM guests will run with cpu VT 
acceleration at full speed?

----------
Fajar A. Nugraha wrote:
 > On Tue, Dec 29, 2009 at 8:11 AM, Gerry Reno <greno@verizon.net> wrote:
 > 
 >> My question is how can I
 >> get pv_ops dom0 kernel or Xen 3.4.1 to pass the cpu VT flags through 
so that
 >> existing KVM guests will run with cpu VT acceleration at full speed?
 >>    
 >
 > AFAIK you can't. The same reason why you can't get Virtualbox and KVM
 > to use VT together.
 > Only one virtualization technology can use VT at the same time.
 >
 >  
As physical boxes gain more and more processing capability it makes no 
sense to restrict a physical machine to only a single hypervisor.

Libvirt will support nested VM's that pass through the VT capabilities 
to the next level.  So it makes sense that Xen should be able to do the 
same thing and pass through the cpu VT capabilities to other 
hypervisors.  Is there some law of the universe that prevents this?

-Gerry

----------

Chris wrote:
 > Following that logic we will need to start working on a vmmm (virtual 
machine monitor monitor) to handle multiple vmm's :)

.....What you say about monitor monitor is not quite what is needed but 
somewhat along those lines.

We need some type of small hypervisor-monitor / scheduler-kernel that 
would exist in Ring 0 and mediate between hypervisors.  Then all 
hypervisors/kernels could be in Ring 1.  Domains in Ring 2.  All apps in 
Ring 3.   Something like this.

-Gerry

----------


So with libvirt being able to pass the cpu VT flags between nested 
levels can Xen do this as well and pass the cpu VT flags so that a KVM 
guest would be able to run with VT acceleration at full speed?  How soon 
could this capability be added to pv_ops dom0 kernel or Xen?

-Gerry