Re: [PATCH 1/2] dma-mapping: support setting memory uncached in place

[Christoph Hellwig]

On Thu, Feb 20, 2020 at 05:21:35PM +0000, Robin Murphy wrote:
>> @@ -196,10 +192,15 @@ void *dma_direct_alloc_pages(struct device *dev, size_t size,
>>     	memset(ret, 0, size);
>>   -	if (IS_ENABLED(CONFIG_ARCH_HAS_UNCACHED_SEGMENT) &&
>> -	    dma_alloc_need_uncached(dev, attrs)) {
>> +	if (dma_alloc_need_uncached(dev, attrs)) {
>>   		arch_dma_prep_coherent(page, size);
>> -		ret = uncached_kernel_address(ret);
>> +
>> +		if (IS_ENABLED(CONFIG_ARCH_HAS_DMA_SET_UNCACHED)) {
>> +			if (!arch_dma_set_uncached(ret, size))
>> +				goto out_free_pages;
>> +		} else if (IS_ENABLED(CONFIG_ARCH_HAS_UNCACHED_SEGMENT)) {
>> +			ret = uncached_kernel_address(ret);
>
> Hmm, would we actually need to keep ARCH_HAS_UNCACHED_SEGMENT? If 
> arch_dma_set_uncached() returned void*/ERR_PTR instead, then it could work 
> for both cases (with arch_dma_clear_uncached() being a no-op for segments).

Yes, I think so.  I was a little worried about what to do with
cached_kernel_address() in that scheme, but it turns out with the recent
round of dma-direct cleanup that is actually entirely unused now.