Re: [PATCH net] macmace: Set platform device coherent_dma_mask

[Michael Schmitz]

Hi Finn,

Am 11.05.2018 um 17:28 schrieb Finn Thain:
> On Fri, 11 May 2018, Michael Schmitz wrote:
> 
>>
>> I'm afraid using platform_device_register() (which you already use for 
>> the SCC devices) is the only option handling this on a per-device basis 
>> without touching platform core code, while at the same time keeping the 
>> DMA mask setup out of device drivers
> 
> I don't think that will fly. If you call platform_device_register() and 
> follow that with a dma mask assignment, you could race with the bus 
> matching and driver probe, and we are back to the same WARNING message.

I wasn't proposing to follow platform_device_register() with the mask
assignment, but rather to use the same strategy from the Coldfire FEC
patch (f61e64310b75): set pdev.dev.coherent_dma_mask and
pdev.dev.dma_mask _before_ calling platform_device_register().

> If you want to use platform_device_register(), you'd have to implement 
> arch_setup_pdev_archdata() and use that to set up the dma mask.

If you want to avoid the overhead of defining the macmace platform
device and using platform_device_register() ... seeing as you would not
be defining just the DMA mask and nothing else, that's probably the
least effort for the MACE and SONIC drivers.

>> (I can see Geert's point there - device driver code might be shared 
>> across implementations of the device on platforms with different DMA 
>> mask requirements,, something the driver can't be expected to know 
>> about).
> 
> As I said, these drivers might be expected to be portable between Macs and 
> early PowerMacs, but the same dma mask would apply AFAIK.
> 
> If a platform driver isn't expected to be portable, I think either method 
> is reasonable: arch_setup_pdev_archdata() or the method in the patch.
> 
> Anyway, there is this in arch/powerpc/kernel/setup-common.c:
> 
> void arch_setup_pdev_archdata(struct platform_device *pdev)
> {
>         pdev->archdata.dma_mask = DMA_BIT_MASK(32);
>         pdev->dev.dma_mask = &pdev->archdata.dma_mask;
> 	...

You would have to be careful not to overwrite a pdev->dev.dma_mask and
pdev->dev.dma_coherent_mask that might have been set in a platform
device passed via platform_device_register here. Coldfire is the only
m68k platform currently using that, but there might be others in future.

> I'm inclined to propose something similar for m68k. That should fix the 
> problem, since arch_setup_pdev_archdata() is already in the call chain:
> 
> 	platform_device_register_simple()
> 		platform_device_register_resndata()
> 			platform_device_register_full()
> 				platform_device_alloc()
> 					arch_setup_pdev_archdata()
> 
> Thoughts? Will this have nasty side effects for m68k platforms that use 
> smaller dma masks?

These can still set a smaller mask in pdev->dev directly and use
platform_device_register() instead. But I don't think there are smaller
DMA masks used by m68k drivers that use the platform device mechanism at
present. I've only looked at arch/m68k though.

Cheers,

	Michael