Re: [RFC] watchdog: pretimeout: add a notifier call chain governor

[Stefan Eichenberger <eichest@gmail.com>]

On Wed, Apr 14, 2021 at 04:50:38AM -0700, Guenter Roeck wrote:
> On 4/13/21 9:56 AM, Stefan Eichenberger wrote:
> > Hi Guenter,
> > 
> > Thanks a lot for the fast feedback!
> > 
> > On Tue, Apr 13, 2021 at 09:15:35AM -0700, Guenter Roeck wrote:
> >> On 4/13/21 8:45 AM, Stefan Eichenberger wrote:
> >>> Add a governor which calls a notifier call chain. This allows the
> >>> registration of custom functions to the watchdog pretimeout. This is
> >>> useful when writing out-of-tree kernel modules and it is more similar to
> >>> the register_reboot_notifier or register_oom_notifier API.
> >>>
> >>> Signed-off-by: Stefan Eichenberger <eichest@gmail.com>
> >>
> >> It seems questionable to implement such a notification as governor.
> >> A governor is configurable, and letting userspace override notifications
> >> seems very odd and unusual. Please provide a use case.
> > 
> > What I would like to do is to have an out-of-tree module which will
> > write a pattern into a reserved memory region when a watchdog occurs.
> > After a reboot the module then read this reserved memory region and can
> > differentiate between different reset reasons. Here the example kernel
> > module:
> > https://github.com/embear-engineering/sample-kernel-modules/blob/use-watchdog-register/reset-reason/reset-reason.c
> > 
> > Registering to the watchdog happens on line 180.
> > 
> > I think I could just implement a governor in the module but it is not
> > really flexible because then I would have to compile the kernel with
> > e.g. the noop governor and then switch to my governor after the module
> > loaded. Then I thought why not using a notifier chain similar to reboot,
> > panic, or oom.
> > 
> 
> I have two problems with this: First, we do not support in-kernel code purely
> for out-of-tree drivers. Second, the idea of using a configurable notifier
> (ie one that can be disabled by userspace action) seems odd. If anything,
> that would have to be an unconditional notifier. However, again, that would
> require a use case in the upstream kernel.

I see your point. Thanks for checking anyways.

Regards,
Stefan