Re: [v1][PATCH 1/1] KVM: PPC: disable preemption when using hard_irq_disable()

[Benjamin Herrenschmidt <benh@kernel.crashing.org>]

On Thu, 2013-07-11 at 15:07 +0200, Alexander Graf wrote:
> Ok, let me quickly explain the problem.
> 
> We are leaving host context, switching slowly into guest context.
> During that transition we call get_paca() indirectly (apparently by
> another call to hard_disable() which sounds bogus, but that's another
> story).
> 
> get_paca() warns when we're preemptible. We're only not preemptible
> when either preempt is disabled or irqs are disabled. Irqs are
> disabled, but arch_irqs_disabled() doesn't know, because it only
> checks for soft disabled IRQs.
> 
> So we can fix this either by setting IRQs as soft disabled as well or
> by disabling preemption until we enter the guest for real. Any
> preferences?

Well, if you hard disable first (ie, direct transition from full enabled
to hard disabled), you know you have nothing lazy pending in
irq_pending, then it's ok to mess around with local_paca->soft_enabled
to make it "look disabled".

IE. Call hard_irq_disable(), then only turn local_paca->soft_enabled
back on late in the process, some time before the final rfi(d).

That works as long as you had a transition from full enabled to full
disabled and don't hard re-enable in the process. IE, You are certain
that there is nothing pending in irq_happened.

HOWEVER !

If you do that, you *ALSO* need to clear irq_happened. You must *NEVER*
leave PACA_IRQ_HARD_DIS set in irq_happened if you are soft-enabled, and
since the above means that you *will* be seen as soft enabled on the way
out of the guest, you can kaboom.

BTW. I'm fine with a patch that does:

#define hard_irq_disable()	do {			\
	u8 _was_enabled = get_paca()->soft_enabled;	\
	__hard_irq_disable();				\
-	get_paca()->soft_enabled = 0;			\
+	local_paca->soft_enabled = 0;			\

In fact we should probably do it anyway.

Cheers,
Ben.