On Tue, Jan 02, 2018 at 04:43:15PM +0000, Vladimir Murzin wrote:

On 29/12/17 08:18, Christoph Hellwig wrote:

quoted

If we got back an allocation that wasn't inside the support coherent mask,
retry the allocation using GFP_DMA.

Based on the x86 code.

Signed-off-by: Christoph Hellwig <hch@lst.de>
---
 lib/dma-direct.c | 25 ++++++++++++++++++++++++-
 1 file changed, 24 insertions(+), 1 deletion(-)

diff --git a/lib/dma-direct.c b/lib/dma-direct.c
index ab81de3ac1d3..f8467cb3d89a 100644
--- a/lib/dma-direct.c
+++ b/lib/dma-direct.c

@@ -28,6 +28,11 @@ check_addr(struct device *dev, dma_addr_t dma_addr, size_t size,
 	return true;
 }
 
+static bool dma_coherent_ok(struct device *dev, phys_addr_t phys, size_t size)
+{
+	return phys_to_dma(dev, phys) + size <= dev->coherent_dma_mask;

Shouldn't it be: phys_to_dma(dev, phys) + size - 1 <= dev->coherent_dma_mask ?

Yes, I think it should.  The existing code was blindly copy and pasted
from x86.

quoted

+	if (page && !dma_coherent_ok(dev, page_to_phys(page), size)) {
+		__free_pages(page, page_order);
+		page = NULL;
+
+		if (dev->coherent_dma_mask < DMA_BIT_MASK(32) &&
+		    !(gfp & GFP_DMA)) {
+			gfp = (gfp & ~GFP_DMA32) | GFP_DMA;
+			goto again;

Shouldn't we limit number of attempts?

We only retty once anyway, due to the !GFP_DMA check first and then
ORing in GFP_DMA.