On Thu, 16 Jul 2015, Julien Grall wrote:

Hi Stefano,

On 16/07/2015 16:33, Stefano Stabellini wrote:

quoted

On Fri, 10 Jul 2015, Konrad Rzeszutek Wilk wrote:

quoted

On Thu, Jul 09, 2015 at 09:42:21PM +0100, Julien Grall wrote:

quoted

When Linux is using 64K page granularity, every page will be slipt in
multiple non-contiguous 4K MFN (page granularity of Xen).

But you don't care about that on the Linux layer I think?

As in, is there an SWIOTLB that does PFN to MFN and vice-versa
translation?

I thought that ARM guests are not exposed to the MFN<->PFN logic
and trying to figure that out to not screw up the DMA engine
on a PCIe device slurping up contingous MFNs which don't map
to contingous PFNs?

Dom0 is mapped 1:1, so pfn == mfn normally, however grant maps
unavoidably screw up the 1:1, so the swiotlb jumps in to save the day
when a foreign granted page is involved in a dma operation.

Regarding xen_biovec_phys_mergeable, we could check that all the pfn ==
mfn and return true in that case.

I mentioned it in the commit message. Although, we would have to loop on every
pfn which is slow on 64KB (16 times for every page). Given the biovec is
called often, I don't think we can do a such things.

We would have to run some benchmarks, but I think it would still be a
win. We should write an ad-hoc __pfn_to_mfn translation function that
operates on a range of pfns and simply checks whether an entry is
present in that range. It should be just as fast as __pfn_to_mfn. I
would definitely recommend it.