RE: [PATCH v2 2/2] x86: don't unmask disabled irqs when migrating them

[Ian Campbell <Ian.Campbell@eu.citrix.com>]

On Fri, 2011-05-06 at 14:24 +0100, Thomas Gleixner wrote:
> On Fri, 6 May 2011, Tian, Kevin wrote:
> > > From: Thomas Gleixner
> > > Sent: Friday, May 06, 2011 6:00 PM
> > > 
> > > On Fri, 6 May 2011, Tian, Kevin wrote:
> > > > x86: don't unmask disabled irqs when migrating them
> > > >
> > > > it doesn't make sense to mask/unmask a disabled irq when migrating it
> > > > from offlined cpu to another, because it's not expected to handle any
> > > > instance of it. Current mask/set_affinity/unmask steps may trigger
> > > > unexpected instance on disabled irq which then simply bug on when
> > > > there is no handler for it. One failing example is observed in Xen.
> > > > Xen pvops
> > > 
> > > So there is no handler, why the heck is there an irq action?
> > > 
> > > 	  if (!irq_has_action(irq) ....
> > > 	     	continue;
> > > 
> > > Should have caught an uninitialized interrupt. If Xen abuses interrupts that way,
> > > then it rightfully explodes. And we do not fix it by magic somewhere else.
> > 
> > sorry that my bad description here. there does be a dummy handler registered
> > on such irqs which simply throws out a BUG_ON when hit. I should just say such 
> > injection is not expected instead of no handler. :-)
> 
> So can please someone point me to that particular incarnation of
> nonsense and provide a reasonable explanation for this abuse?
> 
> What is the point of an interrupt, which is permanently disabled, has
> a handler with a BUG() inside and an irqaction assigned ?
> 
> What's the purpose of this? Why is the irqaction there in the first
> place? To be called by some other weird means than by the irq
> handling code?

The Xen PV spinlock code (arch/x86/xen/spinlock.c) allocates an IRQ
(per-cpu lock_kicker_irq). I think it is there purely in order to have
the associated underlying evtchn to use as the thing to poll (Xen has an
evtchn poll hypercall, see xen_poll_irq()) on the slow path and kick on
release. There is never any need to call a handler for that evtchn --
just notifying the evtchn is enough to wake the poller.

The irq is setup using request_irq(). Is there a different API to
register an IRQ without attaching a handler/action to it? (I can't think
why such a thing would exist).

I'm not really sure why these can't just be an evtchn without an
associated IRQ since it doesn't really have any interrupt-like
semantics. Perhaps just a general desire to keep event channels
abstracted into the core Xen event subsystem with IRQs as the public
facing API? Jeremy?

Ian.

> > > The only conditional which is interesting is the unmask path and that's a simple
> > > optimization and not a correctness problem.
> > > 
> > 
> > So what's your suggestion based on my updated information? Is there any
> > interface I may take to differentiate above exception with normal case? Basically
> > in Xen usage we want such irqs permanently disabled at the chip level. Or
> > could we only do mask/unmask for irqs which are unmasked atm if as you said
> > it's just an optimization step? :-)
> 
> No we can make the unmask conditional on !irqd_irq_disabled() because
> that's not violating any of the semantics. The interrupt would be
> masked anyway when it arrives and the handler code sees that it is
> lazy disabled. I mean real handler code, not the Xen abomination.
> 
> The only valid reason why I'd apply that patch is that it avoids a
> potential extra interrupt, but not to prevent screwed up handlers from
> exploding.
> 
> Thanks,
> 
> 	tglx