[U-Boot] [PATCH 00/12] Define ARCH_DMA_MINALIGN for all architectures

[Anton Staaf]

On Wed, Oct 12, 2011 at 8:47 PM, Kumar Gala <galak at kernel.crashing.org> wrote:
>
> On Oct 12, 2011, at 4:01 PM, Anton Staaf wrote:
>
>> ARCH_DMA_MINALIGN is a new define borrowed from the Linux kernel.  It is
>> used to define the minimum alignement requirement for a DMA buffer.  This
>> series of patches ensures that it will always be defined in the arch specific
>> asm/cache.h header and that asm/cache.h is included in common.h.  I chose to
>> add this new define instead of reusing CONFIG_SYS_CACHELINE_SIZE for two
>> reasons.  First, CONFIG_SYS_CACHELINE_SIZE is not defined for all architectures
>> and thus to use it I would have to first define it correctly for all used
>> architectures.  I wasn't quite up to that task right now.  Defining
>> ARCH_DMA_MINALIGN is easier because it has slightly different semantics that
>> allow it to be the maximum cacheline size known for an architecture.  And
>> secondly, CONFIG_SYS_CACHELINE_SIZE may not always be the correct value to use
>> to align DMA buffers.  In particular, on cache coherent architectures you can
>> define ARCH_DMA_MINALIGN as 1 and be safe (unless there are other restrictions
>> such as PCI alignment requirements).
>>
>> This patch set will allow me to resubmit my ALLOC_CACHE_ALIGN_BUFFER patch set
>> modified to use this new define in place of CONFIG_SYS_CACHELINE_SIZE.  I will
>> repost those patches based on this set shortly.
>
> I haven't looked at this patchset you reference, but what is causing the need for buffer alignment?
>
> I question if on PPC we really need this to be at a cache line granularity.

Briefly, if your cache architecture is not coherent with respect to
DMA transactions
then you need to flush and invalidate the cache before and after read
and write DMA
operations.

To safely do this you need the region that you wish to invalidate
and/or flush to
be aligned with the cache line at both ends.  If it is not then you
run the risk of
corrupting variables that are in the unaligned memory at either end of
the buffer.

So, in short, if the PowerPC cache architecture is coherent with
respect to memory
access done by other bus masters then you can set this value to 1.  If
it is not,
then you must set it to at least the cache line size.

Thanks,
    Anton

> - k