[RFC PATCH 3/7] iommu/arm-smmu-v3: Introduce smmu option USE_SHARED_IRQS for Silicon errata

[Will Deacon]

On Tue, Apr 11, 2017 at 05:38:21PM +0100, Robin Murphy wrote:
> On 11/04/17 17:21, Will Deacon wrote:
> > On Tue, Apr 11, 2017 at 04:54:26PM +0100, Robin Murphy wrote:
> >> On 11/04/17 15:42, linucherian at gmail.com wrote:
> >>> From: Geetha <gakula at cavium.com>
> >>>
> >>> Cavium 99xx SMMU implementation doesn't not support unique irq lines for
> >>> gerror, eventq and cmdq-sync. USE_SHARED_IRQS option enables to use single
> >>> irq line for all three interrupts.
> >>
> >> AFAICS, there's nothing actually wrong with using shared wired IRQs -
> >> the architecture spec doesn't appear to say anything about it. I think
> >> it might suffice to simply add IRQF_SHARED if we can see the SMMU
> >> doesn't support MSIs anyway - it doesn't really seem like something we
> >> need to treat as a specific quirk.
> > 
> > No, this is not permitted by the spec. See 3.18.2 ("Interrupt sources"),
> > where it's clear that each source asserts a *unique* wired interrupt.
> 
> Perhaps I'm reading it too generously; it does indeed specify that the
> *implementation* has to provide a unique output for each source, but
> other than suggesting a particular mode of operation based on that I
> don't see anything actually forbidding the *integration* from then just
> munging those lines together externally, as integrators so often like to
> do. That's the case I had in mind.

Sure, but then there wouldn't be any point in the architecture mandating
a unique source, would there? What next, OR all the address lines together
too?

Will