On Thu, Oct 27, 2022 at 11:26:50AM -0700, Alexander Potapenko wrote:
> On Thu, Oct 27, 2022 at 1:05 AM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> >
> > On Wed, Oct 26, 2022 at 11:38:53AM -0700, Alexander Potapenko wrote:
> > > A bigger issue from the NMI perspective is probably
> > > having __msan_poison_alloca() inserted in every non-noinstr kernel
> > > function, because that hook may acquire the stackdepot lock.
> >
> > *urgghhh* that's broken, that must not be. There is a *TON* of NMI
> > functions that are non-noinstr.
>
> __msan_poison_alloca() is guarded by kmsan_in_runtime(), which is
> currently implemented as:
>
> static __always_inline bool kmsan_in_runtime(void)
> {
> if ((hardirq_count() >> HARDIRQ_SHIFT) > 1)
> return true;
> return kmsan_get_context()->kmsan_in_runtime;
> }
>
> I think the easiest way to fix the NMI situation would be adding "if
> in_nmi() return true"?
It might help to look through these threads:
https://lore.kernel.org/lkml/20220916135953.1320601-1-keescook@xxxxxxxxxxxx/
https://lore.kernel.org/all/20220919201648.2250764-1-keescook@xxxxxxxxxxxx/
I wandered around attempting to deal with in_nmi(), etc. And in
the end just drop the attempt to cover it. It's worth noting that
copy_from_user_nmi() exists on 1 architecture and has exactly 1
call-site...
--
Kees Cook
