Re: [PATCH] kcsan: Treat runtime as NMI-like with interrupt tracing

[Marco Elver]

On Mon, 10 Aug 2020 at 22:18, Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote:
> Marco Elver <elver@xxxxxxxxxx> writes:
> > Since KCSAN instrumentation is everywhere, we need to treat the hooks
> > NMI-like for interrupt tracing. In order to present an as 'normal' as
> > possible context to the code called by KCSAN when reporting errors, we
> > need to update the IRQ-tracing state.
> >
> > Tested: Several runs through kcsan-test with different configuration
> > (PROVE_LOCKING on/off), as well as hours of syzbot testing with the
> > original config that caught the problem (without CONFIG_PARAVIRT=y,
> > which appears to cause IRQ state tracking inconsistencies even when
> > KCSAN remains off, see Link).
> >
> > Link: https://lkml.kernel.org/r/0000000000007d3b2d05ac1c303e@xxxxxxxxxx
> > Fixes: 248591f5d257 ("kcsan: Make KCSAN compatible with new IRQ state tracking")
> > Reported-by: syzbot+8db9e1ecde74e590a657@xxxxxxxxxxxxxxxxxxxxxxxxx
> > Co-developed-by: Peter Zijlstra (Intel) <peterz@xxxxxxxxxxxxx>
> > Signed-off-by: Marco Elver <elver@xxxxxxxxxx>
> > ---
> > Patch Note: This patch applies to latest mainline. While current
> > mainline suffers from the above problem, the configs required to hit the
> > issue are likely not enabled too often (of course with PROVE_LOCKING on;
> > we hit it on syzbot though). It'll probably be wise to queue this as
> > normal on -rcu, just in case something is still off, given the
> > non-trivial nature of the issue. (If it should instead go to mainline
> > right now as a fix, I'd like some more test time on syzbot.)
>
> I'd rather stick it into mainline before -rc1.
>
> Reviewed-by: Thomas Gleixner <tglx@xxxxxxxxxxxxx>

Thank you, sounds good.

FWIW I let it run on syzkaller over night once more, rebased against
Sunday's mainline, and found no DEBUG_LOCKDEP issues. (It still found
the known issue in irqentry_exit(), but is not specific to KCSAN:
https://lore.kernel.org/lkml/000000000000e3068105ac405407@xxxxxxxxxx/)

Thanks,
-- Marco