Re: [RFC PATCH v2 4/9] add generic dmabounce support

[Albert Herranz <albert_herranz@yahoo.es>]

Russell King - ARM Linux wrote:
> On Sun, Feb 28, 2010 at 03:07:57PM +0100, Albert Herranz wrote:
>> This patch makes part of the ARM dmabounce code available to other
>> architectures as a generic API.
> 
> There is already a generic dma bounce implementation - it's called
> swiotlb - lib/swiotlb.c.  We should eventually switch the ARM
> dmabounce stuff over to that instead of keeping dmabounce around.
> 
> The only problem I forsee is that on ARM, we have devices which can
> only address the least significant N bits of RAM, but there may be
> an offset on the base address of RAM on the bus which isn't included
> in these N bits.
> 
> Even more fun is where we have a DMA controller which can address
> N bits of RAM, except for bit M which must be zero... (where N > M).
> 

In the Wii we have several limitations:
- it is a NOT_COHERENT_CACHE platform
- write accesses to coherent memory from the main processor must be done always in 32-bit chunks
- some devices can only reliably perform DMA to/from a specific region of memory (the second block of RAM, 64MB at 0x10000000, called MEM2)

So if swiotlb is the way to go I can try to adapt it to take into account these cases:
- it should allocate the io_tlb_start and io_tlb_overflow_buffer areas from MEM2 (add allocation/freeing hooks for these areas?)
- it should copy data from coherent memory in 32-bit chunks (add copy to/from coherent hooks?)
- it should decide to bounce buffers sitting in MEM1 (add a dma_needs_bounce() hook?)

Thanks,
Albert