Re: [RFC PATCH] of: platform: Skip mapping of interrupts in of_device_alloc()

On Thu, Dec 9, 2021 at 5:35 AM Lad, Prabhakar
[off-list ref] wrote:

Hi Rob and Marc,

On Thu, Dec 9, 2021 at 10:33 AM Marc Zyngier [off-list ref] wrote:

quoted

On Thu, 09 Dec 2021 10:00:44 +0000,
"Lad, Prabhakar" [off-list ref] wrote:

quoted

quoted

The root of the issue is that all the resource allocation is done
upfront, way before we even have a driver that could potentially
deal with this device. This is a potential waste of resource, and
it triggers the issue you noticed.

If you delay the resource allocation until there is an actual
match with a driver, you could have a per-driver flag telling you
whether the IRQ allocation should be performed before the probe()
function is called.

As suggested by Rob, if we switch the drivers to use
platform_get_resource(pdev, IORESOURCE_IRQ, n) call with
platform_get_irq() this code should go away and with this switch the
resource allocation will happen demand. Is this approach OK?

If you get rid of of_irq_to_resource_table() altogether, then yes,
this has a fighting chance to work.

Yes, switching to platform_get_irq() will eventually cause
of_irq_to_resource_table() to go away.

On second thought, instead of touching all the drivers, if we update
platform_get_resource/platform_get_resource_byname to internally call
platform_get_irq() internally if it's a IORESOURCE_IRQ resource. Does
that sound good or should I just get on changing all the drivers to
use platform_get_irq() instead?

Except that platform_get_irq() already internally calls
platform_get_resource()... I think changing the drivers is the right
way. Happy to do some if you want to divide it up.

Using coccigrep, I think I've found all the places using
platform_device.resource directly. A large swath are Sparc drivers
which don't matter. The few that do matter I've prepared patches for
here[1]. Most of what I found were DT based drivers that copy
resources to a child platform device. That case will not work with
platform_get_irq() callers either unless the child device has it's DT
node set to the parent node which is the change I made.

Rob

[1] git://git.kernel.org/pub/scm/linux/kernel/git/robh/linux.git for-kernelci