Re: [RFC PATCH 2/2] dma-mapping: Clean up dma_get_required_mask() hooks

[Christoph Hellwig <hch@lst.de>]

On Tue, Jul 10, 2018 at 01:29:20PM +0100, Robin Murphy wrote:
>> What I've done is to:
>>
>>   1) provide the get_required_mask unconditionally in struct dma_map_ops
>>   2) default to what is the current dma_get_required_mask implementation
>>      if nothing else is specified.
>
> Yeah, there's already 17 pointers in dma_map_ops of which about half are 
> optional, so these awkward #ifdefs to save one more probably aren't worth 
> the inconsistency they bring. It feels like this guy mostly goes 
> hand-in-hand with dma_supported, so ack to giving it the same look and 
> feel.

This whole area needs a major refactoring - we currentl have three
different APIs to deal with addressability: dma_get_required_mask,
dma_capable/dma_set_mask and dma_capable from dma-direct.h, and there
is plenty of unexplainable mismatches between them.

Sorting this out has been on my TODO list, but I think it can only
effectively be done once the direct mapping implementations are
reasonably consolidated.

>> What I still had on my todo list but not done yet:
>>
>>   3) go through all instances and check if the current default
>>      makes sense, at it based on direct addressability.  For most
>>      iommu instances it seems like we should just return a 64-bit mask.
>
> That's reasonable, although in many cases we should know the effective 
> IOMMU input address size which would be even neater.

Sure.  Maybe I just need to steps 1 and 2 and let maintainers fill
in.

>>   4) figure out how to take the dma offsets into account for it
>
> AFAICS it might boil down to simply:
>
> 	mask = roundup_pow_of_two(phys_to_dma(dev, PFN_PHYS(max_pfn))) - 1;

That looks way to sensible.  Which reminds me that I need to research
the history behind the low_totalram/high_totalram magic in
dma_get_required_mask.