Re: [PATCH v11 net-next 1/9] mfd: ocelot: add helper to get regmap from a resource

[Vladimir Oltean <vladimir.oltean@xxxxxxx>]

On Tue, Jun 28, 2022 at 10:25:18AM -0700, Colin Foster wrote:
> > struct regmap *
> > ocelot_regmap_from_resource(struct platform_device *pdev,
> > 			    unsigned int index,
> > 			    const struct regmap_config *config)
> > {
> > 	struct regmap *map;
> > 
> > 	map = ocelot_regmap_from_resource_optional(pdev, index, config);
> > 	return map ? : ERR_PTR(-ENOENT);
> > }
> > 
> > I hope I didn't get something wrong, this is all code written within the
> > email client, so it is obviously not compiled/tested....
> 
> Yep - I definitely get the point. And thanks for the review.
> 
> The other (bigger?) issue is around how this MFD can be loaded as a
> module. Right now it is pretty straightforward to say
> #if IS_ENABLED(CONFIG_MFD_OCELOT). Theres a level of nuance if
> CONFIG_MFD_OCELOT=m while the child devices are compiled in
> (CONFIG_PINCTRL_MICROCHIP_SGPIO=y for example). It still feels like this
> code belongs somewhere in platform / resource / device / mfd...?
> 
> It might be perfectly valid to have multiple SGPIO controllers - one
> local and one remote / SPI. But without the CONFIG_MFD_OCELOT module
> loaded, I don't think the SGPIO module would work.
> 
> This patch set deals with the issue by setting MFD_OCELOT to a boolean -
> but in the long run I think a module makes sense. I admittedly haven't
> spent enough time researching (bashing my head against the wall) this,
> but this seems like a good opportunity to at least express that I'm
> expecting to have to deal with this issue soon. I met with Alexandre at
> ELC this past week, and he said Arnd (both added to CC) might be a good
> resource - but again I'd like to do a little more searching before
> throwing it over the wall.

Well, that's quite a different ball game. It sounds like what you want
is to have a dynamic list of regmap providers for a device, and when a
device probes, to not depend on all the modules of all the possible
providers being inserted at that particular time. Which makes sense,
and I agree it's something that should be handled by the kernel core if
you don't want the sgpio code to directly call symbols exported by the
ocelot mfd core.

I searched for 5 minutes or so, and I noticed regmap_attach_dev() and
dev_get_regmap(), maybe those could be of help? I'm not completely sure
about the mechanics of it all, but what I'm searching for is something
through which the mfd core could attach a regmap to a platform device
through devres before calling platform_device_add(), such that the
driver can use it (or manually fall back to an MMIO regmap) via
dev_get_regmap().