Re: [PATCH] intel-iommu: Obey coherent_dma_mask for alloc_coherent on passthrough

[Alex Williamson <alex.williamson@hp.com>]

On Tue, 2009-11-10 at 01:01 +0000, David Woodhouse wrote:
> On Tue, 2009-11-10 at 00:46 +0000, David Woodhouse wrote:
> > 
> >         if (iommu_no_mapping(hwdev))
> > -               return paddr;
> > +               return paddr > dma_mask - size ? 0 : paddr; 
> 
> Hm, that's still wrong. If your mask is 0xffffffff and you map
> 0xfffff000 + 0x1000, that should be allowed, right? How about:
> 
> 	return paddr <= dma_mask - size + 1 ? paddr : 0;

Ok, I was going to suggest that there was probably a -1 missing from my
patch.  I'd also be surprised if you could actually get a paddr + size -
1 from a single allocation that would wrap.  Would a device even be able
to handle that for DMA?

> And is this strictly necessary as part of the patch we're discussing?

Yeah, we could probably do without it.  As for the commit comment, it's
rather scathing, but I can live with it.  Thanks,

Alex