On Mon, Dec 19, 2011 at 10:56:40PM +0000, David Woodhouse wrote: > On Tue, 2011-12-20 at 09:31 +1100, David Gibson wrote: > > When we're running paravirtualized under pHyp, it's impossible to > > merge multiple PEs into one domain per se. We could fake it rather > > nastily by replicating all map/unmaps across mutiple PEs. When > > running bare metal, we could do so a bit more nicely by assigning > > multiple PEs the same TCE pointer, but we have no mechanism to do so > > at present. > > VT-d does share the page tables, as you could on bare metal. But it's an > implementation detail — there's nothing *fundamentally* wrong with > having to do the map/unmap for each PE, is there? It's only at VM setup > time, so it doesn't really matter if it's slow. > > Surely that's the only way you're going to present the guest with the > illusion of having no IOMMU; so that DMA to any given guest physical > address "just works". > > On the other hand, perhaps you don't want to do that at all. Perhaps > you're better off presenting a virtualised IOMMU to the guest and > *insisting* that it fully uses it in order to do any DMA at all?
Not only do we want to, we more or less *have* to. Existing kernels, which are used to being paravirt under phyp expect and need a paravirt iommu. DMA without iommu setup just doesn't happen. And the map/unmap hypercalls are frequently a hot path, so slow does matter. The other problem is that each domain's IOVA window is often fairly small, a limitation that would get even worse if we try to put too many devices in there. -- David Gibson | I'll have my music baroque, and my code david AT gibson.dropbear.id.au | minimalist, thank you. NOT _the_ _other_ | _way_ _around_! http://www.ozlabs.org/~dgibson