Re: [PATCH][RFC] asm-generic:remove calling flush_write_buffers() in dma_sync_*_for_cpu

[Arnd Bergmann <arnd@arndb.de>]

On Tuesday 07 July 2009, Ming Lei wrote:
> > ARM has two (normal, and dma bounce), and in the long run we need to do
> 
> OK, Can we use dma-mapping-common.h on ARM?

It should work in principle. It may be a good idea to also move to
the generic swiotlb instead of the traditional dma bounce at the
same time.

Note that dma-mapping-common.h is only needed if you want to support
two or more different DMA implementations in a single kernel, which
I'm not sure is needed for ARM.

> > cache handling on unmap as well as map due to CPU speculative fetches.
> 
> IMHO, it seems we can fix this problem now.
> 
> For DMA_TO_DEVICE transfer, clean cache in dma map, but does nothing in
> dma unmap;
> 
> For  DMA_FROM_DEVICE, we may do nothing in dma map, but invaliate cache
> in dma unmap.

A number of other architectures do this already. You also need to
have dma_sync_*_for_cpu and dma_sync_*_for_device, where the *_for_device
operation needs to do the same flushing as dma_map_* and *_for_cpu
does the same as dma_unmap_*.

Note that actually you need to do writeback+invalidate in DMA_TO_DEVICE
and at least an invalidate in DMA_FROM_DEVICE during dma_map_*.
For the unmap, I don't think you ever need to invalidate the cache.
If you invalidate only at unmap time for DMA_FROM_DEVICE, a dirty
cache line might be accidentally flushed to the buffer after
the device has written to it.

	Arnd <><