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"?
Currently that will render kmsan_unpoison_memory() useless in NMI
context, but I think we don't need a check for kmsan_in_runtime()
there, because unpoisoning is self-contained and normally does not
recurse (guess we can tolerate a pr_err() on the rare assertion
violation path?)
> What's worse, it seems to do a memory allocation as well, and that's out
> the window with PREEMPT_RT where you can't do even GFP_ATOMIC from
> regular IRQ context.
Yes, there's a lazy call to alloc_pages() in lib/stackdepot.c that is
done when we run out of storage space.
It would be a pity to ignore everything that is happening inside
regular IRQs (by making kmsan_in_runtime() bail out on in_irq()) - I
think we occasionally see errors from there, e.g.
https://syzkaller.appspot.com/bug?id=233563e79a8e00f86412eb3d2fb4eb1f425e70c3
We could make stackdepot avoid allocating memory in IRQs/NMIs and hope
that calls to __msan_poison_alloca() from regular contexts keep up
with draining the storage from interrupts.
Another option would be to preallocate a very big chunk of memory for
stackdepot and never do allocations again.
These tricks won't however save us from acquiring depot_lock from
lib/stackdepot.c every time we want to create a new origin.
But that should not be a problem by itself, because we always do
kmsan_enter_runtime() before accessing the stack depot - i.e. it won't
be taken recursively.
Given that PREEMPT_RT is not the default at the moment, shall we make
KMSAN incompatible with it for the time being?
> That function is wholly unacceptable to be added to every kernel
> function.
>
--
Alexander Potapenko
Software Engineer
Google Germany GmbH
Erika-Mann-Straße, 33
80636 München
Geschäftsführer: Paul Manicle, Liana Sebastian
Registergericht und -nummer: Hamburg, HRB 86891
Sitz der Gesellschaft: Hamburg
