Re: [PATCH 00/39] irqdomain: Simplify interrupt handling

[Marc Zyngier]

On Thu, 20 May 2021 17:37:12 +0100,
Marc Zyngier <maz@xxxxxxxxxx> wrote:
> 
> Although most device drivers only deal with an interrupt number, the
> core IRQ code is mostly concerned with the irq_desc structure that
> describes the full interrupt context (hierarchy, handlers, state).
> 
> However, the low-level interrupt handling code that relies on the
> irqdomain abstraction has to perform an annoying dance to eventually
> get the core code to invoke interrupt handlers: the irqdomain code
> converts a low-level identifier to the unique Linux interrupt number,
> and the core code resolves this into an irq_desc pointer.
> 
> Each of these two lookups ends-up parsing a radix tree (although the
> irqdomain code can use a linear mapping for the smallest domains),
> which is obviously one too many. Wouldn't it be nice if the irqdomain
> would cache the irq_desc instead of forcing the core code to look it
> up on each and every interrupt? This is what this long series is all
> about.
> 
> There is roughly 3 parts here:
> 
> - a substantial amount of massaging for some architectures (nios, mips
>   and powerpc) to disentangle weird include constructs (asm/irq.h
>   including linux/irqdomain.h is pretty bad...) and simplify bits of
>   the irqdomain code
> 
> - some rework of the irqdomain code to allow the caching of a irq_data
>   pointer, unify the RCU behaviour and offer new APIs.
> 
> - Perform a bulk of conversions that turn constructs similar to
>   generic_handle_irq(irq_find_mapping(domain, hwirq)) into a simpler
>   call to generic_handle_domain_irq(domain, hwirq). Yes, this is a
>   mouthful.
> 
> I've kept most of the conversions per-subsystem/per-arch in order to
> keep the number of patches low (though it is debatable whether I have
> succeeded).
> 
> This ends up with a negative diffstat, so it can't be completely bad!
> Given the breadth of the changes, I do expect some breakage, although
> I've extensively compile-tested it and the kbuild robot has been
> invaluable in helping with the coverage.

Slight nudge in the direction of the cc'd arch maintainers. I'd like
to take the core of this series into -next (in practice, patches #1
through to #27). Any comment on the early, arch-specific patches would
be most welcome.

Thanks,

	M.

-- 
Without deviation from the norm, progress is not possible.