Am 14.11.19 um 23:09 schrieb Rob Herring:
On Tue, Nov 12, 2019 at 4:47 AM Andreas Färber afaerber@suse.de wrote:
Finally, arch/arm seems unique in that it has the machine_desc mechanism that allows individual SoCs to force creating their soc_device early and using it as parent for further DT-created platform_devices. With arm64 we've lost that ability, and nios2 is not using it either. A bad side effect (with SUSE hat on) is that this parent design pattern does not allow to build such drivers as modules, which means that distro configs using arm's multi-platform, e.g., CONFIG_ARCH_MULTI_V7 will get unnecessary code creep as new platforms get added over time (beyond the basic clk, pinctrl, tty and maybe timer). Even if it were possible to call into a driver module that early, using it as parent would seem to imply that all the references by its children would not allow to unload the module, which I'd consider a flawed design for such an "optional" read-once driver. If we want the device hierarchy to have a soc root then most DT based platforms will have a /soc DT node anyway (although no device_type = "soc") that we probably could have a device registered for in common code rather than each SoC platform handling that differently? That might then make soc_register_device() not the creator of the device (if pre-existent) but the supplier of data to that core device, which should then allow to unload the data provider with just the sysfs data disappearing until re-inserted (speeding up the develop-and-test cycle on say not-so-resource-constrained platforms).
I for one would like to for this to be consistent. Either we always have an SoC device parent or never. I wouldn't decide based on having a DT node or not either.
Sure, if we can always be consistent, that might be best.
Where I was coming from here is that, if we're not supposed to use soc_device_to_device(), then we have no way to associate an of_node with the soc_device from the outside (and nobody was doing it today, as per my analysis). We'd either need a new helper of_soc_device_register() with additional argument for the node, or it would need to be done entirely in the infrastructure as I suggested, be it by looking for one hardcoded /soc node or for nodes with device_type = "soc".
Rob, in light of this discussion, should we start adding the latter property for new DTs such as my new Realtek SoCs, or was there a reason this has not been used consistently outside of powerpc and nios2? Intel/Altera appear to be the only two in arm64, with only three more in arm, none in mips.
(BTW my assumption was that we don't follow the booting-without-of.txt documented schema of soc<SOCname> so that we can share .dtsi across differently named SoCs, is that correct?)
Generally, we should always have MMIO devices under a bus node and perhaps more than one, but that doesn't always happen. I think building the drivers as modules is a good reason to do away with the parent device.
It would also allow getting rid of remaining of_platform_default_populate calls in arm platforms except for auxdata cases. Pretty much that's the ones you list below in arch/arm/.
The majority was indeed passing in NULL, so yeah, we might clean that up, if someone could come up with a plan of where/how to implement it.
Cheers, Andreas