Re: swiotlb_alloc_coherent: allocated memory is out of range fordevice

From: FUJITA Tomonori
Date: Wed Oct 22 2008 - 22:37:17 EST

On Wed, 22 Oct 2008 15:32:52 +0200
Takashi Iwai <tiwai@xxxxxxx> wrote:

> > About the bug that you hit, I suspect that dma_map_coherent() in
> > asm-x86/dma-mapping.h doesn't set gfp flags correctly.
> >
> > dma_map_coherent() calls swiotlb_alloc_coherent with the flags GFP_DMA
> > set? parport driver set dev->coherent_dma_mask properly?
> The parport driver itself passes always GFP_KERNEL. So I added
> GFP_DMA in my initial patch as a workaround.

Hmm, have you actually tried your patch?

dma_alloc_coherent clears the gfp zone flags that the callers pass. So
even if a driver passes GFP_DMA, swiotlb_alloc_coherent doesn't get
GFP_DMA. So I'm not sure how your patch fixes the parport problem.

The current dma_alloc_coherent (asm-x86/dma-mapping.h) handles the gfp
flags in the exact same way as the old dma_alloc_coherent
(pci-dma.c). Neither sets GFP_DMA even if coherent_dma_mask is
24bits. The old code is fine because of the GFP_DMA retry
mechanism. But if coherent_dma_mask is 24bits, there is no point to go
into the GFP_DMA retry mechanism. We should use GFP_DMA in the first

How about the following patch?

diff --git a/include/asm-x86/dma-mapping.h b/include/asm-x86/dma-mapping.h
index 219c33d..05fcec5 100644
--- a/include/asm-x86/dma-mapping.h
+++ b/include/asm-x86/dma-mapping.h
@@ -255,9 +255,11 @@ static inline unsigned long dma_alloc_coherent_mask(struct device *dev,

static inline gfp_t dma_alloc_coherent_gfp_flags(struct device *dev, gfp_t gfp)
-#ifdef CONFIG_X86_64
unsigned long dma_mask = dma_alloc_coherent_mask(dev, gfp);

+ if (dma_mask <= DMA_24BIT_MASK)
+ gfp |= GFP_DMA;
+#ifdef CONFIG_X86_64
if (dma_mask <= DMA_32BIT_MASK && !(gfp & GFP_DMA))
gfp |= GFP_DMA32;
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at
Please read the FAQ at