Re: [PATCH] intel-iommu: Fix alloc_coherent for pass-through devices

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

On Thu, 2009-10-22 at 23:47 +0900, David Woodhouse wrote:
> On Thu, 2009-10-22 at 06:24 -0600, Alex Williamson wrote:
> > The coherent_dma_mask is independent of the dma_mask and can be set
> > separately by the device.  The default for any device that doesn't
> > specify one is 32bits.  iommu_should_identity_map() only checks the
> > dma_mask, not the coherent_dma_mask. 
> 
> Are you telling me that this particular device supports only a 32-bit
> coherent DMA mask, but that it _does_ support a 64-bit DMA mask for
> non-coherent DMA? On x86?

Yes, yes.  AFAIK, this is not that exceptional.

> >  BTW, we skip RMRR setup when doing hardware pass-through, but I can't
> > find where they get reloaded if we then end up removing the device
> > from the si_domain.  Is this another issue?
> 
> Maybe, theoretically. In practice, the whole RMRR thing is just broken
> by design anyway. We have to quiesce the offending devices before we
> turn on the IOMMU, because BIOSes tend to leave things out of the RMRR
> table... and then crash in SMM mode when their DMA goes AWOL. 

Hmm, we've had a lot of trouble getting our RMRRs to reflect shared
memory regions correctly, so I'm reluctant to just drop them like that.

Another issue, iommu_should_identity_map() only dumps devices if their
dma_mask is 32bit or less, but being greater than 32bits does not imply
64bit DMA support.  If the device exports a DMA mask that's less than
the physical address width of the processor, you might be in trouble
(for example, a 40bit dma_mask on a platform that supports 44bits for
physical memory).  Maybe dropping swiotlb out of the picture isn't such
a clean solution?  Thanks,

Alex