Re: [RFC 13/14] irq_domain: Remove 'new' irq_domain in favour of the ppc one

On Thu, Jan 12, 2012 at 5:31 PM, Rob Herring [off-list ref] wrote:

Adding lakml...

On 01/11/2012 03:27 PM, Grant Likely wrote:

quoted

On Wed, Jan 11, 2012 at 2:15 PM, Rob Herring [off-list ref] wro=

te:

quoted

quoted

Grant,

On 01/11/2012 02:22 PM, Grant Likely wrote:

quoted

This patch removes the simplistic implementation of irq_domains and en=

ables

quoted

quoted

quoted

the powerpc infrastructure for all irq_domain users. =A0The powerpc
infrastructure includes support for complex mappings between Linux and
hardware irq numbers, and can manage allocation of irq_descs.

This patch also converts the few users of irq_domain_add()/irq_domain_=

del()

quoted

quoted

quoted

to call irq_domain_add_legacy() instead.

So what is the non-legacy way? Legacy implies we don't want to do it
that way. I guess until we remove all non-DT platforms with GIC we are
stuck with legacy. That seems like it could be a ways out until we get
there.

Non-legacy is letting the irq_domain manage the irq_desc allocations.
Some of the controllers will be easy to convert, some will be more
difficult. =A0The primary thing that really blocks getting away from the
legacy method is anything that expects hardcoded #defined irq numbers.
=A0The goal is to convert all users over to the linear revmap method.

So I gave this a spin on highbank. I ran into a couple problems.

I had to revert "irqdesc: Consolidate irq reservation logic" which is in
your branch, but not this series. irq_alloc_desc_from was returning -EEXI=

ST.

Hmmm... I thought I sorted that out.  Thanks for letting me know.

The GIC code did not work which I think is specific to using gic_of_init
which makes irq_start =3D -1. With that it still doesn't work. It dies in
gic_set_type... I've found one problem which I'll reply inline to, but I
think this is a dead end path anyway.

Haha, I'm not surprised.  That last patch was only compile tested on
platforms using the gic.  I'm not surprised that I flubbed it.

You have removed the irq_alloc_descs call from the GIC which is a step
backwards. Several of the ARM DT enabled platforms are at the point they
can fully support dynamic virq base for each irqchip. I changed the
domain from legacy to linear and got things working.
The issue with

I hadn't actually intended to remove the irq_alloc_descs in this
patch.  That was a leftover hunk from when I was playing with going
straight to irq_domain_add_linear().  For this specific patch, I'll
put the alloc back in and test it that way.  A follow-on patch can do
a proper conversion to the linear revmap.

linear is for SPARSE_IRQ. The default behavior on ARM for SPARSE_IRQ is
all nr_irqs are allocated at boot time before any controller is
initialized. The only platform with a GIC and requiring SPARSE_IRQ is
shmobile, but it is also the only one that calls irq_alloc_desc
functions for it's interrupts. So I think we are okay there. The problem
occurs when enabling SPARSE_IRQ for a non-DT platform with a GIC and
with irqchips that don't call irq_alloc_desc for their irqs. IMHO, this
should be an okay trade-off. There's no advantage to enabling SPARSE_IRQ
on ARM for platforms that don't require it. All the platforms with a GIC
have active work to convert to DT (except shmobile which I think is
okay), so it's a temporary issue.

Actually, I believe Thomas' long term goal is to always enable
SPARSE_IRQ and remove the option entirely, so it should still be
properly resolved. I'll take a look next week if I don't get to it
tomorrow.  I need to resurrect my vexpress qemu test environment so I
can test the permutations.

g.

--=20
Grant Likely, B.Sc., P.Eng.
Secret Lab Technologies Ltd.