Re: [PATCH] genirq: fix trigger flags check for shared irqs

From: Brian Starkey
Date: Thu Jan 28 2016 - 06:10:51 EST

On Thu, Jan 28, 2016 at 11:29:21AM +0100, Thomas Gleixner wrote:

On Thu, 28 Jan 2016, Brian Starkey wrote:
On Tue, Jan 26, 2016 at 09:45:32PM +0100, Thomas Gleixner wrote:
> On Tue, 26 Jan 2016, Brian Starkey wrote:
> > For shared interrupts, if one requester passes in any IRQF_TRIGGER_*
> > flags whilst another doesn't, __setup_irq() can erroneously fail.
> >
> > The no-flags case should be treated as "already configured", so change
> > __setup_irq() to only check that the flags match if any have been
> > provided.
> What happens if that "already configured", i.e. the default setting, is
> conflicting with the newly requested interrupt?
> I rather prefer the failure than the resulting silent wreckage.

Yes, I agree that would be best avoided. It seems to me that this case
is actually handled a bit lower down:

} else if (new->flags & IRQF_TRIGGER_MASK) {
unsigned int nmsk = new->flags & IRQF_TRIGGER_MASK;
unsigned int omsk = irq_settings_get_trigger_mask(desc);

if (nmsk != omsk)
/* hope the handler works with current trigger mode
pr_warning("irq %d uses trigger mode %u; requested
irq, nmsk, omsk);

Perhaps that should be louder/fatal?

Perhaps. So what's the actual problem case you are trying to solve?

I've got a few devices on the same interrupt line. One driver does
something along these lines:

res = platform_get_resource(dev, IORESOURCE_IRQ, 0);
flags = (res->flags & IRQF_TRIGGER_MASK) | IRQF_SHARED;
request_irq(res->start, handler, flags, "name", dev);

This seems pretty reasonable. The problem is since 4a43d686fe33:
of/irq: Pass trigger type in IRQ resource flags[1]
the trigger type information from device-tree is in res->flags.

So when the other drivers don't pass in any flags, they fail the check
in __setup_irq().

Changing the former driver to remove the flags doesn't seem right, and
adding flags to the latter would imply adding flags to _every_ driver,
which is an awful lot to change - and I'm not sure it would be possible
and/or effective in all cases.