[PATCH V4] notifier/panic: Introduce panic_notifier_filter

[bhe at redhat.com]

On 03/07/22 at 10:11am, Guilherme G. Piccoli wrote:
> On 07/03/2022 00:42, bhe at redhat.com wrote:
> > [...]
> >> Let me know your thoughts Petr / Baoquan - it would add slightly more
> >> code / complexity, but in my opinion the payback is very good.
> >> Cheers,
> > 
> > The ideal situation is each panic notifier has an order or index to
> > indicate its priority. Wondering how to make it. What I think of is
> > copying initcall. We have several priorities, at the same priority,
> > execution sequence is not important. Not sure if I get your point.
> > 
> > ~~~~~~~
> > #define core_initcall(fn)               __define_initcall(fn, 1)
> > #define core_initcall_sync(fn)          __define_initcall(fn, 1s)
> > ......
> > #define late_initcall(fn)               __define_initcall(fn, 7)
> > #define late_initcall_sync(fn)          __define_initcall(fn, 7s)
> > 
> 
> Hi Baoquan, thanks for you consideration! In fact, the notifiers
> infrastructure already have a mechanism of ordering, my idea is to make
> use of that. It's not that different from the initcall system...

Ah, sorry, I even didn't notice that. That's awesome if we can make use
of that. While I still have concerns:

1) about those we have decided to take out from panic notifier list and
put before kdump, e.g the Hypver-V notifier, how will we do with it? Are
we going to handle them as we have discussed?

2) Combing and settling priority for all existing panic notifier looks
great, even though it will take some effort. How about the later newly
added one? How can we guarantee that those new notifiers are getting
appropriate priority to mark their order? Sometime we even don't know
a new panic notifier is added since code change may be made in any
component or driver.

> 
> For instance, the code in the notifier register function checks for the
> priority field in the notifier block:
> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/kernel/notifier.c#n31
> .
> 
> For example, the Xen panic notifier is one of the few blocks that make
> use of that, currently:
> 
> static struct notifier_block xen_panic_block = {
> 	.notifier_call = xen_panic_event,
> 	.priority = INT_MIN
> };
> 
> (see
> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/arch/x86/xen/enlighten.c#n286)
> 
> In this case, xen_panic_event() will be the latest one to run. My idea
> is to make use of that in *all* panic notifiers, having a table/list on
> panic_notifier.h (defines or enum, I'll think about that when writing
> this part) so all notifiers are documented and the ordering is clear and
> enforced.
> 
> Makes sense to you?
> Cheers,
> 
> 
> Guilherme
>