Re: [PATCH] dma-pool: use single atomic pool for both DMA zones

[Robin Murphy <robin.murphy@arm.com>]

On 2020-07-08 16:36, Christoph Hellwig wrote:
> On Wed, Jul 08, 2020 at 12:35:34PM +0200, Nicolas Saenz Julienne wrote:
>>> Which allows me to switch between ACPI/DT on the machine. In DT mode it
>>> works fine now,
>>
>> Nice, would that count as a Tested-by from you?
>>
>>> but with ACPI I continue to have failures unless I
>>> disable CMA via cma=0 on the kernel command line.
>>
>> Yes, I see why, in atomic_pool_expand() memory is allocated from CMA without
>> checking its correctness. That calls for a separate fix. I'll try to think of
>> something.
> 
> I think we need a dma_coherent_ok for the allocations from the
> pool and then fall back to the next better one to get started.  And
> yes, CMA is a bit of a mess, that generally needs better checks.

Yeah, another thought that came to mind later is that iommu-dma can use 
pages from any pool regardless of the device's DMA mask, so we could 
stand to be a lot less restrictive in that case too.

Perhaps it is better to just bite the bullet, keep the straightforward 
one-pool-per-zone setup, and implement the dma_coherent_ok() type 
fallback logic. More complexity for dma_alloc_from_pool(), but 
everything else stays nice and simple - lose the assumption that 
dev_to_pool() can work for this and and just let callers pass an 
allocation mask directly, and have dma_free_from_pool() simply check all 
available pools.

Robin.