Re: [PATCH v2 0/2] Detect stalls on guest vCPUS

[Sebastian Ene <sebastianene@google.com>]

On Sat, Apr 23, 2022 at 10:36:56AM +0100, Marc Zyngier wrote:
> On Sat, 23 Apr 2022 10:02:24 +0100,
> Sebastian Ene <sebastianene@google.com> wrote:
> > 
> > On Sat, Apr 23, 2022 at 08:51:16AM +0200, Greg Kroah-Hartman wrote:
> > > On Fri, Apr 22, 2022 at 02:19:48PM +0000, Sebastian Ene wrote:
> > > > This adds a mechanism to detect stalls on the guest vCPUS by creating a
> > > > per CPU hrtimer which periodically 'pets' the host backend driver.
> > > > 
> > > > This device driver acts as a soft lockup detector by relying on the host
> > > > backend driver to measure the elapesed time between subsequent 'pet' events.
> > > > If the elapsed time doesn't match an expected value, the backend driver
> > > > decides that the guest vCPU is locked and resets the guest. The host
> > > > backend driver takes into account the time that the guest is not
> > > > running. The communication with the backend driver is done through MMIO
> > > > and the register layout of the virtual watchdog is described as part of
> > > > the backend driver changes.
> > > > 
> > > > The host backend driver is implemented as part of:
> > > > https://chromium-review.googlesource.com/c/chromiumos/platform/crosvm/+/3548817
> > > > 
> > > > Changelog v2:
> > > >  - move the driver to misc as this does not cope with watchdog core
> > > >    subsystem
> > 
> > Hello Greg,
> > 
> > > 
> > > Wait, why does it not cope with it?  That's not documented anywhere in
> > > your patch that adds the driver.  In fact, most of the text here needs
> > > to be in the changelog for the driver submission, not thrown away in the
> > > 00/XX email that will never end up in the kernel tree.
> > > 
> > > thanks,
> > > 
> > > greg k-h
> > 
> > From the previous feedback that I received on this patch it seems that
> > watchdog core is not intended to be used for this type of driver. This
> > watchdog device tracks the elapsed time on a per-cpu basis,
> > since KVM schedules vCPUs independently. Watchdog core is not intended
> > to detect CPU stalls and the drivers don't have a notion of CPU.

Hello Marc,

> 
> I must say that I don't really get the objection against the watchdog
> approach. OK, there is no userspace aspect to this.  But we already
> use watchdogs for more than just userspace (reboot is one of the major
> use cases).
> 
> There already are per-CPU watchdog in the tree: see how the
> fsl-ls208xa platform has one SP805 per CPU (8 of them in total). As
> far as I can tell, there was no objection to this. So what is special
> about this one?

I think the difference is in the fact that this driver expects hrtimers
which are CPU binded to execute the periodic watchdog 'pet'. We would
require a strong thread affinity setting if we rely on userspace to
do this 'pet' operation.

Thanks,
Sebastian

> 
> Thanks,
> 
> 	M.
> 
> -- 
> Without deviation from the norm, progress is not possible.