Re: [PATCH v2] mm/kfence: add reboot notifier to disable KFENCE on shutdown

[Breno Leitao <leitao@debian.org>]

Hello Chris,

On Tue, Jan 13, 2026 at 06:02:27AM -0800, Chris Mason wrote:
> On Thu, 27 Nov 2025 06:51:54 -0800 Breno Leitao <leitao@debian.org> wrote:
> > @@ -820,6 +821,25 @@ static struct notifier_block kfence_check_canary_notifier = {
> >  static struct delayed_work kfence_timer;
> >
> >  #ifdef CONFIG_KFENCE_STATIC_KEYS
> > +static int kfence_reboot_callback(struct notifier_block *nb,
> > +				  unsigned long action, void *data)
> > +{
> > +	/*
> > +	 * Disable kfence to avoid static keys IPI synchronization during
> > +	 * late shutdown/kexec
> > +	 */
> > +	WRITE_ONCE(kfence_enabled, false);
> > +	/* Cancel any pending timer work */
> > +	cancel_delayed_work_sync(&kfence_timer);
>                    ^^^^^^^^^^^^^^^
> 
> Can cancel_delayed_work_sync() deadlock here?
> 
> If toggle_allocation_gate() is currently executing and blocked inside
> wait_event_idle() (waiting for kfence_allocation_gate > 0), then
> cancel_delayed_work_sync() will block forever waiting for the work to
> complete.
> 
> The wait_event_idle() condition depends only on allocations occurring
> to increment kfence_allocation_gate - setting kfence_enabled to false
> does not wake up this wait. During shutdown when allocations may have
> stopped, the work item could remain blocked indefinitely, causing the
> reboot notifier to hang.
> 
> The call chain is:
>   kfence_reboot_callback()
>     -> cancel_delayed_work_sync(&kfence_timer)
>        -> __flush_work()
>           -> wait_for_completion(&barr.done)
>              // waits forever because...
> 
>   toggle_allocation_gate() [currently running]
>     -> wait_event_idle(allocation_wait, kfence_allocation_gate > 0)
>        // never wakes up if no allocations happen

This is spot on, I think this is a real case if the following happen:


1) toggle_allocation_gate() passed beyond kfence_enabled and is waiting
   for kfence_allocation_gate to be > 0.
   a) kfence_allocation_gate is increased on allocation time

2) There is no more kernel allocation, thus, kfence_allocation_gate is
   not incremented

3) cancel_delayed_work_sync() is for kfence_allocation_gate > 0, but
   given there is no more allocation, this will never happen.

> Would it be safer to use cancel_delayed_work() (non-sync) here.

In this case toggle_allocation_gate() task will continue to be idle,
waiting for to be kfence_allocation_gate > 0 forever, but it will not
block the notifiers, unless we wake them up.

Is this a problem?

Maybe a more robust solution would include:

1) s/cancel_delayed_work_sync()/cancel_delayed_work().
  - This would unblock the notifier

or/and some of the followings

2) Return from wait_event_idle() if kfence_enabled got disabled.
  - Remove the waiters once kfence got disabled
  - Cons: kfence_allocation_gate will continue to be negative

3) Wake up everyone in the allocation_wait() list
  - This might not be necessary if we got 2, since they will wake
    themselves once kfence_enabled got to 0
  - Cons: kfence_allocation_gate will continue to be negative

4) bump kfence_allocation_gate > 1 on the notifier
  - Avoid kfence allocation completely after it got disabled.
  - Cons: it is unclear if we I cant set kfence_allocation_gate = 1 from
    the notifier.


Thanks for the report, 
--breno