Re: workqueues and percpu (was: [PATCH] dm: remake of the verity target)

[Mandeep Singh Baines <msb@chromium.org>]

Tejun Heo (tj@kernel.org) wrote:
> Hello,
> 
> On Fri, Mar 09, 2012 at 01:15:12PM -0800, Mandeep Singh Baines wrote:
> > In this case, I need the per-cpu data for the duration of calculating
> > a cryptographics hash on a 4K page of data. That's a long time to disable
> > pre-emption.
> 
> How long are we talking about?  Tens of microsecs, tens of millisecs?
> 

It depends on the H/W. We are running on Atom w/ PREEMPT_DESKTOP so
we aren't pre-emptible even now. But I think there is interest from other
folks in embedded to use the work. So it potentially be millisecs on a
slower few hundred MHZ embedded processor.

> > I could fix the bug temporarily by adding get/put for the per_cpu data
> > but would that be acceptable? I'm not sure what the OK limit is for how
> > long one can disable preemption. An alternative fix would be not allow
> > CONFIG_VERITY when CONFIG_HOTPLUG_CPU. Once workqueues are fixed, I could
> > remove that restriction.
> 
> I think the right thing to do for now is to add cpu hotplug notifier
> and do flush_work_sync() on the work item.  We can later move that
> logic into workqueue and remove it from crypto.
> 

That seems like the correct solution. I will implement that.

Thanks,
Mandeep

> Thanks.
> 
> -- 
> tejun