Re: [PATCH 1/3] x86/mm/pat: Disable preemption around __flush_tlb_all()

[Peter Zijlstra]

On Wed, Oct 17, 2018 at 01:17:18PM +0200, Thomas Gleixner wrote:
> On Wed, 17 Oct 2018, Peter Zijlstra wrote:
> > On Wed, Oct 17, 2018 at 09:54:38AM +0000, David Laight wrote:
> > > >  	 * We should perform an IPI and flush all tlbs,
> > > >  	 * but that can deadlock->flush only current cpu:
> > > >  	 */
> > > > +	preempt_disable();
> > > >  	__flush_tlb_all();
> > > > +	preempt_enable();
> > > 
> > > Can it make any sense to flush the tlb with preemption enabled?
> > > Surely preemption must be disabled over something else as well?
> > 
> > This code is fishy anyway, for only doing that local invalidate.
> > 
> > Ideally we'd never ever merge anything that only does local invalidates,
> > on a global address space, that's just broken.
> 
> A little bit late to lament about that.

For this, yes :/ But for future stuff we should really not allow such
things anymore.

> So should we just replace it with cpa_flush_all() ?

The comment there suggests that will deadlock, supposedly because the
kernel_map_page() call can happen with IRQs disabled or such.

I've not deeply looked at this.