[PATCH 0/3] cpu: idle state framework for offline CPUs.

[Peter Zijlstra]

On Mon, 2009-08-17 at 11:54 +0530, Dipankar Sarma wrote:
> On Sun, Aug 16, 2009 at 11:53:22PM +0200, Peter Zijlstra wrote:
> > On Mon, 2009-08-17 at 01:14 +0530, Balbir Singh wrote:
> > > Agreed, I've tried to come with a little ASCII art to depict your
> > > scenairos graphically
> > > 
> > > 
> > >         +--------+ don't need (offline)
> > >         |  OS    +----------->+------------+
> > >         +--+-----+            | hypervisor +-----> Reuse CPU
> > >            |                  |            |       for something
> > >            |                  |            |       else
> > >            |                  |            |   (visible to users)
> > >            |                  |            |    as resource changed
> > >            |                  +----------- +
> > >            V (needed, but can cede)
> > >        +------------+
> > >        | hypervisor | Don't reuse CPU
> > >        |            |  (CPU ceded)
> > >        |            | give back to OS
> > >        +------------+ when needed.
> > >                         (Not visible to
> > >                         users as so resource
> > >                         binding changed)
> > 
> > I still don't get it... _why_ should this be exposed in the guest
> > kernel? Why not let the hypervisor manage a guest's offline cpus in a
> > way it sees fit?
> 
> For most parts, we do. The guest kernel doesn't manage the offline
> CPU state. That is typically done by the hypervisor. However, offline
> operation as defined now always result in a VM resize in some hypervisor
> systems (like pseries) - it would be convenient to have a non-resize
> offline operation which lets the guest cede the cpu to hypervisor
> with the hint that the VM shouldn't be resized and the guest needs the guarantee
> to get the cpu back any time. The hypervisor can do whatever it wants
> with the ceded CPU including putting it in a low power state, but
> not change the physical cpu shares of the VM. The pseries hypervisor,
> for example, clearly distinguishes between the two - "rtas-stop-self" call
> to resize VM vs. H_CEDE hypercall with a hint. What I am suggesting
> is that we allow this with an extension to existing interfaces because it 
> makes sense to allow sort of "hibernation" of the cpus without changing any
> configuration of the VMs.

>From my POV the thing you call cede is the only sane thing to do for a
guest. Let the hypervisor management interface deal with resizing guests
if and when that's needed.

Thing is, you don't want a guest to be able to influence the amount of
cpu shares attributed to it. You want that in explicit control of
whomever manages the hypervisor.