On Tue, 21 Sep 2010 16:43:12 -0500
Timur Tabi [off-list ref] wrote:

On Tue, Sep 21, 2010 at 4:34 PM, Scott Wood [off-list ref] wrote:

quoted

This doesn't control allocation (it probably should with
dma_alloc_coherent, though I don't see it in the code), it controls
whether swiotlb will create a bounce buffer -- defeating the point of
using DMA to accelerate a memcpy.

But it would do that only for the 'dev' used in the dma_set_mask()
call.  That dev is only used here:

	chan->desc_pool = dma_pool_create("fsl_dma_engine_desc_pool",
					  chan->dev,
					  sizeof(struct fsl_desc_sw),
					  __alignof__(struct fsl_desc_sw), 0);

Since we don't DMA the descriptors themselves, I just don't see how
this patch does anything.

Look in dmaengine.c, there are calls to dma_map_single() and
dma_map_page(), using what I assume is that same device pointer --
unless there's confusion between the channel and the controller.

-Scott