Re: [PATCH 1/3] firmware: of: populate /firmware/ node during init

On 11/08/17 16:54, Arnd Bergmann wrote:

On Fri, Aug 11, 2017 at 5:05 PM, Rob Herring [off-list ref] wrote:

quoted

On Fri, Aug 11, 2017 at 9:37 AM, Arnd Bergmann [off-list ref] wrote:

quoted

On Fri, Aug 11, 2017 at 3:30 PM, Sudeep Holla [off-list ref] wrote:

quoted

Since "/firmware" does not have its own "compatible" property as it's
just collection of nodes representing firmware interface, it's sub-nodes
are not populated during system initialization.

Currently different firmware drivers search the /firmware/ node and
populate the sub-node devices selectively. Instead we can populate
the /firmware/ node during init to avoid more drivers continuing to
populate the devices selectively.

This patch adds initcall to achieve the same.

Hmm, I'm a bit skeptical whether representing anything under /firmware
as a platform device is a good idea. Having a more structured way to
probe those seems like a good idea, but maybe a different subsystem
would be more appropriate.

I do realize that a 'platform_device' has become a rather generic abstraction
for almost anything, but at some point we might want to draw the line
of what is a platform_device.

I guess the question how are they different? Most of what's under
drivers/firmware/ are platform drivers. I think they are mostly either
smc calls or mailbox interfaces. Would there be any advantage to
creating an smc bus or mailbox bus?

I guess one difference I see is between things that are purely software
based (smc, efi runtime, rtas,  ...) and those that talk to some
hardware other than the CPU running some firmware.

The first category seems like a good fit for /firmware in DT and
for /sys/firmware in sysfs, while the second category would be
represented elsewhere in both DT and sysfs.

drivers/base/firmware.c currently is extremely rudimentary but this
is where /sys/firmware objects hook into. How about extending
this with a firmware_device that gets populated from /firmware
in DT? Not using platform_device obviously means we lose
all of the automatic probing of reg/interrupts/... resources

While I agree conceptually, but with things like SCMI which provides
clock, power domain, dvfs and other providers which are generally
fit into the driver model and have other devices that are consumers
linked to it.

Also I noticed quite a few ACPI based static tables which are purely
firmware interface driven as platform device. As Rob pointed out we
already have quite a few DT based drivers too. I know all these points
doesn't prove anything against valid points you have made, but I was
just referring the reality out there.

One thing I observed the devices are getting "firmware:"(e.g.:
firmware:scmi) in the beginning of their name which is better than
explicitly adding in the driver.

-- 
Regards,
Sudeep