On 08/13/2012 10:57 PM, Huacai Chen wrote:

Hi, David,

Seems like you are the original author of code in
arch/mips/cavium-octeon/dma-octeon.c. Could you please tell me why we
need mb() in alloc_coherent(), map_page(), map_sg()? It seems like
because of cache coherency (CPU write some data, then map the page for
a device, if without mb(), then device may read wrong data.) but I'm
not sure.

That is essentially correct.

The DMA API requires certain memory barrier semantics.  These are 
achieved with the mb() in the OCTEON code.

On Tue, Aug 14, 2012 at 1:54 AM, Konrad Rzeszutek Wilk
[off-list ref]  wrote:

quoted

quoted

+static void *loongson_dma_alloc_coherent(struct device *dev, size_t size,
+                             dma_addr_t *dma_handle, gfp_t gfp, struct dma_attrs *attrs)

I know nothing about Loongson, so I cannot comment on what is required 
for it.

David Daney