Re: [PATCH v5 4/7] drivers: dma-coherent: Introduce default DMA pool

From: Robin Murphy
Date: Tue Jun 27 2017 - 10:36:41 EST


On 26/06/17 10:42, Christoph Hellwig wrote:
> On Tue, Jun 20, 2017 at 03:24:21PM +0100, Robin Murphy wrote:
>> True, but the case here is where we need a special piece of coherent
>> memory for *all* devices, and it was more complicated *not* to reuse the
>> existing infrastructure. This would already be achievable by specifying
>> a separate rmem carveout per device, but the shared pool just makes life
>> easier, and mirrors the functionality dma-contiguous already supports.
>
> Ð'm really worried about the code in dma-coherent.c - the original
> version clearly intends to have a coherent pool per device, declared
> in the driver. Then Marek added the reserved_mem interface, and
> now we get another variant of it. Conceptually the per-device
> and global pool are very different, and to me it seems like the
> reserved mem should be a different interface.

Per-device reserved mem is still a private per-device pool though, it's
just discovered and declared by common firmware code rather than in some
device-specific way by driver code - once it's assigned there's no
distinction. The global/per-device issue is essentially entirely
orthogonal, and has the dubious pleasure of being a massive conceptual
difference yet a much smaller implementation difference.

>>> If you're allocating out of the global allocator the memory should
>>> come from the normal dma_ops ->alloc allocator - and also take
>>> the attrs into account (e.g. for DMA_ATTR_NON_CONSISTENT or
>>> DMA_ATTR_NO_KERNEL_MAPPING requests you don't need coherent memory)
>>
>> The context here is noMMU but with caches - the problem being that the
>> normal allocator will give back kernel memory, and there's no way to
>> make that coherent with devices short of not enabling the caches in the
>> first place, which is obviously undesirable. The trick is that RAM is
>> aliased (in hardware) at two addresses, one of which makes CPU accesses
>> non-cacheable, so by only ever accessing the RAM set aside for the
>> coherent DMA pool using the non-cacheable alias (represented by the
>> dma_pfn_offset) we can achieve DMA coherency.
>
> Yes, and I think this is something we already have to deal with
> for example on mips. A simple genalloc allocator from your pool
> in the normal dma_ops implementation should do the work just fine.

I admit I'm almost in agreement, were it not for the fact that
dma-contiguous already supports all four combinations of both per-device
and global pools, and both reserved mem and direct declarations from
arch/platform code, all through the same interface to boot, and nobody's
complaining about that. The only real difference for dma-coherent seems
to be the way it's baked into the existing API.

If it is just a matter of interfaces, I'd have no objection to exporting
a separate e.g. dma_alloc_from_global_coherent() or somesuch as a
conceptually separate interface to dma_coherent_default_memory, which
the arch code can then call from ->alloc in the same manner they
currently call dma_alloc_from_contiguous(). That seems like a reasonable
way to keep the per-device and global pools conceptually distinct
without needlessly duplicating implementations. In fact, I'm now
wondering if the regular arm/arm64 atomic pools couldn't also make use
of such a thing as well...

Robin.