xen-hotplug.patch

xen-hotplug.patch

[Jan Beulich]

Could someone point out the relevance of that patch (in patches/linux-2.6.16.13/). For my eyes it looks just wrong, as
it reverts the general effort of not touching per-CPU data of offline CPUs.

Thanks, Jan

Re: xen-hotplug.patch

[Keir Fraser]

On 2 Aug 2006, at 15:02, Jan Beulich wrote:

> Could someone point out the relevance of that patch (in 
> patches/linux-2.6.16.13/). For my eyes it looks just wrong, as
> it reverts the general effort of not touching per-CPU data of offline 
> CPUs.
>
> Thanks, Jan

top (at least old versions) would crash if the number of online 
processors changed while they ran. Hence this patch. Not clear it's 
really a good fix, but it does okay as a workaround.

   -- Keir

Re: xen-hotplug.patch

[Jan Beulich]

>top (at least old versions) would crash if the number of online 
>processors changed while they ran. Hence this patch. Not clear it's 
>really a good fix, but it does okay as a workaround.

That must be a problem with CPU hotplug on real hardware then, too.

Jan

Re: xen-hotplug.patch

[Emmanuel Ackaouy]

On Wed, Aug 02, 2006 at 03:45:27PM +0100, Jan Beulich wrote:
> >top (at least old versions) would crash if the number of online 
> >processors changed while they ran. Hence this patch. Not clear it's 
> >really a good fix, but it does okay as a workaround.
> 
> That must be a problem with CPU hotplug on real hardware then, too.

Yeah.

In practice the problem is less visible on real hardware because
CPU hotplugging is not be used as aggressively in that scenario
as it might when multiple SMP virtual guests compete for a fixed
number of physical CPUs on Xen.

Re: xen-hotplug.patch

[Emmanuel Ackaouy]

On Wed, Aug 02, 2006 at 03:02:51PM +0100, Jan Beulich wrote:
> Could someone point out the relevance of that patch (in patches/linux-2.6.16.13/). For my eyes it looks just wrong, as
> it reverts the general effort of not touching per-CPU data of offline CPUs.

Yeah it's an argueable hack but it does make things nicer when
you hotplug CPUs in and out of guests.

Here's the changeset description:

------------------------------------------------------------
# HG changeset patch
# User ack@kneesa.uk.xensource.com
# Node ID c6da0c5b03ffec5ce4130aeba90a49f218772622
# Parent  dec7f819293c11f88281aa0fe408a10f8c1345ce
Patch linux so that /proc/stat shows all -- not just online -- CPUs.
This way, tools such as TOP don't freak out when CPUs are hotplugged.
------------------------------------------------------------

Without this, top crashes when you hotplug CPUs out and it
just doesn't notice new CPUs coming online.

I'm all for a better fix if we can come up with one. Aside
from changing tons of libs and apps to deal with the reality
of CPUs coming online and going offline dynamically, I'm not
sure what we can do though.