Re: [patch] mm, debug: allow suppressing panic on CONFIG_DEBUG_VM checks

[Justin Forbes]

On Wed, May 24, 2023 at 3:42 AM David Hildenbrand <david@xxxxxxxxxx> wrote:
>
> On 23.05.23 03:47, Linus Torvalds wrote:
> > On Mon, May 22, 2023 at 5:52 PM David Rientjes <rientjes@xxxxxxxxxx> wrote:
> >>
> >> Right now kernel.panic_on_warn can either be 0 or 1.  We can keep the
> >> lowest bit to be "panic on all warnings" and then bit-1 as "panic on debug
> >> VM warnings."  When CONFIG_DEBUG_VM is enabled, set the new bit by
> >> default so there's no behavior change.
> >
> > So right now CONFIG_DEBUG_VM being off means that there's nothing at
> > all - not just no output, but also no code generation.
> >
> > I don't think CONFIG_DEBUG_VM in itself should enable that bit-1 behavior.
> >
> > That may be what *you* as a VM person wants, but VM people are not
> > exactly the common case.
> >
> > So I think we've got several cases:
> >
> >   (a) the "don't even build it" case (CONFIG_DEBUG_VM being off)
> >
> >   (b) the "build it, and it is a WARN_ON_ONCE()" case
> >
> >   (c) the *normal* "panic_on_warn=1" case, which by default would panic
> > on all warnings, including any warnings from CONFIG_DEBUG_VM
> >
> >   (d) the "VM person" case, which might not panic on normal warnings,
> > but would panic on the VM warnings.
> >
> > and I think the use-cases are for different classes of kernel use:
> >
> >   (a) is for people who disable debugging code until they feel it is
> > needed (which I think covers a lot of kernel developers - I certainly
> > personally tend to not build with debug support unless I'm chasing
> > some issue down)
> >
> >   (b) would probably be most distros - enable the warning so that the
> > distro can report it, but try not to kill the machine of random people
> >
> >   (c) would be most cloud use cases, presumably together with reboot-on-panic
> >
> >   (d) would be people who are actual VM developers, and basically want
> > the *current* behavior of VM_BUG_ON() with a machine that stops
> >
> > and I think (d) is the smallest set of cases of all, but is the one
> > you're personally interested in.
>
> Just as a side note, I stumbled yesterday over [1], which apparently
> disables CONFIG_DEBUG_VM on !debug Fedora builds.
>
> The commit description does not contain a rational ( it's empty :) ),
> and I don't know if this is just a temporary change.
>
> I'll CC Justin, maybe Fedora also would like to keep building with
> CONFIG_DEBUG_VM, but default to WARN_ON_ONCE() instead.
>
>
> [1]
> https://gitlab.com/cki-project/kernel-ark/-/commit/ade780e10ae1fdcb575ab100bf02d61eb12dd406

Do not read too much into this right now.  The RHEL performance folks
did a comparison of the RHEL config vs the Fedora config for 6.3 and
found Fedora was considerably slower in a couple of tests. We are
re-running those tests with some DEBUG configs turned off to see which
is the culprit. FWIW, CONFIG_DEBUG_VM made very little difference. As
we have not found the specific cause yet though, final configs have
not been restored.
As for my prefeernce, WARN_ON_ONCE() behavior would be much preferred.

Justin

> --
> Thanks,
>
> David / dhildenb
>