Re: [Qemu-devel] [RFC v1 3/3] intel_iommu: add scalable-mode option to make scalable mode work

[Yi Sun]

On 19-02-15 13:39:05, Peter Xu wrote:
> On Fri, Feb 15, 2019 at 01:22:34PM +0800, Yi Sun wrote:
> 
> [...]
> 
> > > > +    /* TODO: read cap/ecap from host to decide which cap to be 
> > > > exposed. */
> > > > +    if (s->scalable_mode) {
> > > > +        if (!s->caching_mode) {
> > > > +            error_report("Need to set caching-mode for scalable mode");
> > > 
> > > Could I ask why?
> > > 
> > My intention is to make guest explicitly send to make sure SLT shadowing
> > correctly.
> > 
> > For this point, I also have question. Why does legacy mode not check CM?
> > If CM is not set, may the DMA remapping be wrong because SLT cannot
> > match guest's latest change?
> 
> Because CM is currently only required for device assignment.  For
> example, if you only have an emulated device like virtio-net-pci in
> the guest, then you don't need the CM capability to let it work with
> the vIOMMU.  That's because we'll walk the 2nd level IOMMU page table
> only when the virtio-net-pci device does DMA, and QEMU can easily do
> that (QEMU is emulating the DMA of the virtio-net-pci device, and QEMU
> has the knowledge of the guest 2nd level IOMMU page tables).  Assigned
> devices are special because the host hardware knows nothing about
> guest 2nd level page tables, so QEMU needs to shadow them before DMA
> starts.
> 
Thanks for the explanation! As CM is must for device assignment, I may
remove the check in current codes. In fact, CM is optional for SM.

> That's also why I listed "device assignment" as a special case in the
> test device matrix, because that's the only case where we can torture
> the IOMMU page shadowing code a bit.
> 
> For the scalable mode I would suppose you will still allow it to work
> without caching mode.  The example is the same as above - when there's
> only emulated devices in the guest.
> 
Yes, I verified it by using virtio-net-pci without device assignment.
The netperf simple case runs well.

> Regards,
> 
> -- 
> Peter Xu