On Wed, Sep 16, 2009 at 05:32:51PM +0200, Peter Zijlstra wrote:

On Wed, 2009-09-16 at 20:58 +0530, Dipankar Sarma wrote:

quoted

On Tue, Sep 15, 2009 at 02:11:41PM +0200, Peter Zijlstra wrote:

quoted

On Tue, 2009-09-15 at 17:36 +0530, Gautham R Shenoy wrote:

quoted

This patchset contains the offline state driver implemented for
pSeries. For pSeries, we define three available_hotplug_states. They are:

        online: The processor is online.

        offline: This is the the default behaviour when the cpu is offlined

        inactive: This cedes the vCPU to the hypervisor with a cede latency

Any feedback on the patchset will be immensely valuable.

I still think its a layering violation... its the hypervisor manager
that should be bothered in what state an off-lined cpu is in. 

The problem is that all hypervisor managers cannot figure out what sort
of latency guest OS can tolerate under the situation. They wouldn't know
from what context guest OS has ceded the vcpu. It has to have
some sort of hint, which is what the guest OS provides.

I'm missing something here, hot-unplug is a slow path and should not
ever be latency critical..?

You aren't, I did :)

No, for this specific case, latency isn't an issue. The issue is -
how do we cede unused vcpus to hypervisor for better energy management ?
Yes, it can be done by a hypervisor manager telling the kernel to
offline and make a bunch of vcpus "inactive". It does have to choose
offline (release vcpu) vs. inactive (cede but guranteed if needed).
The problem is that long ago we exported a lot of hotplug stuff to
userspace through the sysfs interface and we cannot do something
inside the kernel without keeping the sysfs stuff consistent.
This seems like a sane way to do that without undoing all the
virtual cpu hotplug infrastructure in different supporting archs.

Thanks
Dipankar