On 19/07/2019 18:05, Christoph Hellwig wrote:
On Fri, Jul 19, 2019 at 06:00:25PM +1000, Alexey Kardashevskiy wrote:
But shouldn't we force usage of the direct ops in that case as the
IOMMU is not neededed at all?
We do, for mappings, but not unmappings and syncing.
Well, I mean as in literally not setting a dma_ops so that the
dma_direct code is used without the indirection through the iommu ops.
This is not only more obvious, but also faster as you avoid the
indirect call (although that probably doesn't matter much if you
are bounce buffering anyway).
I was not precise. We cannot avoid IOMMU in the guest for passed through
devices.
Also isn't that support non-upstream so far?
How is this relevant? I expect the existing "swiotlb=force" just work.
I though the whole secure VM support was still not upstream.
It is getting there and I still do not see why "swiotlb=force" should
not work if chosed in the cmdline.
--
Alexey
