Re: [RFC PATCH 00/16] x86/split_lock: Enable #AC exception for split locked accesses

[Peter Zijlstra <peterz@infradead.org>]

On Thu, Jun 21, 2018 at 03:42:03PM -0700, Fenghua Yu wrote:
> On Thu, Jun 21, 2018 at 03:18:19PM -0700, Dave Hansen wrote:
> > On 06/21/2018 03:08 PM, Peter Zijlstra wrote:
> > > On Thu, Jun 21, 2018 at 03:00:03PM -0700, Fenghua Yu wrote:
> > >> The control knob for kernel is to choose re-executing the faulting
> > >> instruction (default) or kernel panic. Kernel panic may be useful in hard
> > >> real time which has less tolerant to bad performance.
> > > The kernel should never trigger this, ever. Any #AC is a plain bug and
> > > should be fixed. So unconditional WARN+disable is suffient.
> > 
> > It sounds like we need to start with that functionality and then if
> > Fenghua wants to do more, we have add-on patches that can be evaluated
> > by themselves.
> 
> For split lock in kernel, by default, I just do warning, disable feature
> in #AC handler, continue to execute the faulting instruction, and re-enable
> #AC for split lock after a delay. This is the default behavior in the
> current patches.
> 
> By "WAR+disable", you mean I just do this default behavior and remove the
> control knob?

No, remove all that broken delay crap, remove the knob. Basically remove
everything.

The handler really should be:

do_handle_ac_trap()
{
	if (user_mode(regs)) {
		send_signal(SIGBUS);
		return;
	}

	WARN_ON(1);
	disable_ac_on_all_cpus();
}

done. No broken delay code, no nothing.