Re: [PATCH] lockdep: Have assert functions test for actual interrupts disabled

From: Steven Rostedt
Date: Thu Sep 06 2018 - 10:17:07 EST


On Thu, 6 Sep 2018 15:52:58 +0200
Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:

> On Wed, Sep 05, 2018 at 11:20:31AM -0400, Steven Rostedt wrote:
> >
> > Peter, you OK with this patch? I'm currently triggering a bug (in rc2)
> > where this patch is telling me that lockdep is getting it wrong. It
> > would be good to have this upstream such that we know if it is really a
> > bug in the code itself, or if lockdep didn't keep up properly.
>
> I thought we had something for that under CONFIG_DEBUG_LOCKDEP.
>
> /me checks and finds:
>
> kernel/locking/lockdep.c:check_flags()
>
> Doesn't that work?

Perhaps it does, but DEBUG_LOCKDEP wasn't set. Thus, when a use case
like this happens it will confuse developers because all they see is:

------------[ cut here ]------------
IRQs not enabled as expected
[...]


And there's no reason to assume that lockdep is broken.

Commits like ebf3adbad012b ("timers/nohz: Use lockdep to assert IRQs
are disabled/enabled") which replace WARN_ON_ONCE(!irqs_disabled())
with lockdep_assert_irqs_disabled(), were done for performance reasons
(which I agree with). My patch doesn't affect performance as it only
does the "irqs_disabled()" check when the WARN_ONCE() actually
triggers. And gives useful information.

If I haven't worked on lockdep in the past, I would have been spending
a lot more time trying to figure out why interrupts were disabled here
and never looking into the fact that the report was wrong.

I still think checking if IRQS are really disabled or not when lockdep
thinks it is (or not) is valuable and doesn't cause any other problems.

-- Steve