[PATCH v2 00/26] genirq: fix use of irq_find_mapping outside of legal RCU context

[Thomas Gleixner]

On Wed, 3 Sep 2014, Jason Cooper wrote:

> On Wed, Aug 27, 2014 at 10:33:44AM +0100, Marc Zyngier wrote:
> > Hi Thomas,
> > 
> > On Tue, Aug 26 2014 at 10:34:51 pm BST, Thomas Gleixner <tglx at linutronix.de> wrote:
> > > On Tue, 26 Aug 2014, Marc Zyngier wrote:
> > >
> > >> A number of irqchip drivers are directly calling irq_find_mapping,
> > >> which may use a rcu_read_lock call when walking the radix tree.
> > >> 
> > >> Turns out that if you hit that point with CONFIG_PROVE_RCU enabled,
> > >> the kernel will shout at you, as using RCU in this context may be
> > >> illegal (specially if coming from the idle state, where RCU would be
> > >> in a quiescent state).
> > >> 
> > >> A possible fix would be to wrap calls to irq_find_mapping into a
> > >> RCU_NONIDLE macro, but that really looks ugly.
> > >> 
> > >> This patch series introduce another generic IRQ entry point
> > >> (handle_domain_irq), which has the exact same behaviour as handle_IRQ
> > >> (as defined on arm, arm64 and openrisc), except that it also takes a
> > >> irq_domain pointer. This allows the logical IRQ lookup to be done
> > >> inside the irq_{enter,exit} section, which contains a
> > >> rcu_irq_{enter,exit}, making it safe.
> > >
> > > Looks good. Should this be routed to the genirq tree?
> > 
> > I'm happy for you to take this series, provided the architecture
> > maintainers agree on it (I'm still to hear from the openrisc guys, and
> > their mailing-list seems to positively hate my guts).
> 
> I think everyone's had a chance to look over it by now.  Thomas, shall I
> take the series?

Yes please.