ARM: 2.6.3[45] PCI regression (IXP4xx and PXA?)

[Russell King - ARM Linux]

On Fri, Aug 13, 2010 at 03:23:53PM +0900, FUJITA Tomonori wrote:
> On Wed, 11 Aug 2010 08:25:32 +0100
> Russell King - ARM Linux <linux at arm.linux.org.uk> wrote:
> > It doesn't break dmabounce.
> > 
> > What it breaks is the fact that a PCI device which can do 32-bit DMA is
> > connected to a PCI bus which can only access the first 64MB of memory
> > through the host bridge, but the system has more than 64MB available.
> > 
> > Allowing a 32-bit DMA mask means that dmabounce can't detect that memory
> > above 64MB needs to be bounced to memory below the 64MB boundary.
> 
> But dmabounce doesn't look at dev->coherent_dma_mask.
> 
> The change breaks __dma_alloc_buffer()? If we set dev->coherent_dma_mask
> to DMA_BIT_MASK(32) for ixp4xx's pci devices, __dma_alloc_buffer()
> doesn't use GFP_DMA.

With an incorrect coherent_dma_mask, dma_alloc_coherent() will return
memory outside of the 64MB window.  This means that when dmabounce comes
to allocate the replacement buffer, it gets a buffer which won't be
accessible to the DMA controller - which is a condition it doesn't check
for.

Ergo, it breaks dmabounce.