Re: [PATCH] mm/mmu_notifier: Remove non_block_start/end() from notifier invocation

[David Woodhouse]

On Wed, 2026-08-20 at 15:30 +0200, Sebastian Andrzej Siewior wrote:
> So if it is *just* the annoying splat on RT, I could try to teach
> might_sleep() no to be upset here.

It *wasn't* just the splat on RT... but I think it can be reduced to
that. My full set of reasons were:

 • RT turns rwlock_t/spinlock_t into sleeping locks.
 • We believed a waiting KVM_REQ (IPI + wait for acks) would sleep;
   it actually spins with preemption disabled¹.
 • synchronize_srcu_expedited() in the invalidation path really does
   sleep on a workqueue completion, even on !RT.

If you deal with the first, the second turns out not to have been
needed, and that just leaves the third which is the SRCU atomic thing
(qv).

Aside from those selfish technical reasons, the check is actually
*wrong* because "must not sleep" isn't the real requirement here, as
discussed. And when I went looking, not only did I have to hack the
kernel to even make it trigger at all, I also couldn't find any example
of it finding anything that *wasn't* a false positive. The only
concrete thing it *did* achieve was making us strip a necessary
KVM_REQUEST_WAIT to appease it, which opened up a use-after-free.

So if you fix the might_sleep() and Paul comes through with the atomic
SRCU thing, I no longer care *much*, but still think the whole thing
deserves to die.

Going back to SRCU atomic (citing your mail in reverse order):
> In
>    https://lore.kernel.org/all/a94b75e8b9b1d2191f365c0f4275fd5082332df5.camel@xxxxxxxxxxxxx
> there was a suggestion for "srcu: Add an ATOMIC reader flavor and a spinning
> synchronize_srcu_atomic()". This one had preempt_disable() in
> srcu_read_lock_atomic(). This is bad as not only forbids to acquire any
> sleeping locks (such as spinlock_t) but it also does not allow any
> scheduling of any kind within this section.

That was kind of the point :) 

My srcu_read_lock_atomic/synchronize_srcu_atomic() were born of the
"need" to avoid scheduling in OOM (RT or no), but actually I think they
survive even when that requirement goes away. Even with my
try_synchronize_srcu() optimisation, we saw higher tail latencies when
deferring to the workqueue, and Sean was very much in favour of
eliminating those.

So I think we do want synchronize_srcu_atomic() if Paul can find a way
to give us that.

To your point about preempt_disable() in srcu_read_lock_atomic() on the
read side... I don't know if we *need* it to be like that. When I threw
it together, I was thinking of it as a raw_spinlock, in the "must not
sleep" mindset.

In fact, as long as the splats are dealt with, it would probably be OK
for it to be equivalent to a non-raw spinlock: disable preemption on
non-RT, but not on RT. I think that's perfectly OK for us from the KVM
point of view. But does heavily depend on how Paul wants to implement
it, of course.


¹ I wonder if we might want to change that on RT?

Attachment: smime.p7s
Description: S/MIME cryptographic signature