Re: [PATCH] acpi: fix generic acpi_register_gsi() handling of shared interrupts

[Mark Salter <msalter@redhat.com>]

On Tue, 2015-04-28 at 18:08 +0100, Marc Zyngier wrote:
> On 28/04/15 17:51, Mark Salter wrote:
> > If there are devices sharing an interrupt, acpi_register_gsi() could
> > be called multiple times for the same gsi. Currently, it just maps
> > the gsi without checking for a previous mapping. This patch adds a
> > check for an existing mapping and uses that if found.
> > 
> > Cc: Hanjun Guo <hanjun.guo@linaro.org>
> > Cc: Marc Zyngier <marc.zyngier@arm.com>
> > Cc: Lorenzo Pieralisi <lorenzo.pieralisi@arm.com>
> > Cc: Will Deacon <will.deacon@arm.com>
> > Signed-off-by: Mark Salter <msalter@redhat.com>
> > ---
> >  drivers/acpi/gsi.c | 7 +++++++
> >  1 file changed, 7 insertions(+)
> > 
> > diff --git a/drivers/acpi/gsi.c b/drivers/acpi/gsi.c
> > index 38208f2..5698c4f 100644
> > --- a/drivers/acpi/gsi.c
> > +++ b/drivers/acpi/gsi.c
> > @@ -76,6 +76,13 @@ int acpi_register_gsi(struct device *dev, u32 gsi, int trigger,
> >  	unsigned int irq_type = acpi_gsi_get_irq_type(trigger, polarity);
> >  
> >  	/*
> > +	 * The gsi may be shared with other devices, so check for a
> > +	 * previous mapping and use that if found.
> > +	 */
> > +	if (acpi_gsi_to_irq(gsi, &irq) > 0)
> > +		return irq;
> 
> I'm afraid that's completely racy, and doesn't prevent much (two CPUs
> can execute this at the same time and still end up executing the rest of
> the code).
> 
> Furthermore, irq_create_mapping is *designed* to handle that situation,

You're right. irq_create_mapping does that check already. But if it
is truly racy, irq_create_mapping has the same problem.

> and returns the same virtual IRQ for a given hwirq (and it has the
> proper locking...).
> 
> So maybe it would be worth explaining the actual problem you're seeing,
> because the code as it stands should handle the problem you seem to
> describe.

Sorry for the noise. I was experimenting with some other code
in acpi_register_gsi which called __irq_domain_alloc_irqs
directly which bypassed the check in irq_create_mapping. And
that turns out to be a silly thing to do anyway, so nevermind.