On Thu, Jan 05, 2012 at 03:29:30PM +0100, Hiroshi Doyu wrote:

Ok, the above function order should be as below?

  page = alloc_page(GFP_KERNEL);
      cpu_addr = kmap(page);
      memset(cpu_addr, 0xa5, PAGE_SIZE);
  	dma_addr = dma_map_page(dev, page, 0, PAGE_SIZE, DMA_TO_DEVICE);
  	    < expect GPU does something >
  	dma_unmap_page(dev, dma_addr, PAGE_SIZE, DMA_TO_DEVICE);
      < CPU can access data here again>
      kunmap(cpu_addr);
  __free_page(page);

That should work, but it's not particularly nice to wrap kmap around
the DMA API.

Bear in mind that alloc_page(GFP_KERNEL) will always give you a lowmem
page.  So, you might as well do:

	page = alloc_page(GFP_KERNEL);
	cpu_addr = page_address(page);
	memset()
	dma_addr = dma_map_single(dev, cpu_addr, PAGE_SIZE, DMA_TO_DEVICE);
		< GPU access >
	dma_unmap_single(dev, dma_ddr, PAGE_SIZE, DMA_TO_DEVICE);
	< CPU access >
	__free_page(page);

and this doesn't raise any issues with kmap vs DMA API.