x86: restore the old swiotlb alloc_coherent behavior

This restores the old swiotlb alloc_coherent behavior (before the
alloc_coherent rewrite):

  http://lkml.org/lkml/2008/8/12/200

The old alloc_coherent avoids GFP_DMA allocation first and if the
allocated address is not fit for the device's coherent_dma_mask, then
dma_alloc_coherent does GFP_DMA allocation. If it fails,
alloc_coherent calls swiotlb_alloc_coherent (in short, we rarely used
swiotlb_alloc_coherent).

After the alloc_coherent rewrite, dma_alloc_coherent
(include/asm-x86/dma-mapping.h) directly calls swiotlb_alloc_coherent.
It means that we possibly can't handle a device having dma_masks >
24bit < 32bits since swiotlb_alloc_coherent doesn't have the above
GFP_DMA retry mechanism.

This patch fixes x86's swiotlb alloc_coherent to use the GFP_DMA retry
mechanism, which dma_generic_alloc_coherent() provides now
(pci-nommu.c and GART IOMMU driver also use
dma_generic_alloc_coherent).
Signed-off-by: FUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

arch/x86/kernel/pci-swiotlb_64.c

@@ -18,9 +18,21 @@ swiotlb_map_single_phys(struct device *hwdev, phys_addr_t paddr, size_t size,
 	return swiotlb_map_single(hwdev, phys_to_virt(paddr), size, direction);
 }
 
+static void *x86_swiotlb_alloc_coherent(struct device *hwdev, size_t size,
+					dma_addr_t *dma_handle, gfp_t flags)
+{
+	void *vaddr;
+
+	vaddr = dma_generic_alloc_coherent(hwdev, size, dma_handle, flags);
+	if (vaddr)
+		return vaddr;
+
+	return swiotlb_alloc_coherent(hwdev, size, dma_handle, flags);
+}
+
 struct dma_mapping_ops swiotlb_dma_ops = {
 	.mapping_error = swiotlb_dma_mapping_error,
-	.alloc_coherent = swiotlb_alloc_coherent,
+	.alloc_coherent = x86_swiotlb_alloc_coherent,
 	.free_coherent = swiotlb_free_coherent,
 	.map_single = swiotlb_map_single_phys,
 	.unmap_single = swiotlb_unmap_single,