Re: [Qemu-devel] Question: SRIOV support over Win Hyper-V VM running in QEMU process on Linux host

[Knut Omang <knut.omang@oracle.com>]

On Thu, 2018-07-26 at 20:11 +0300, Michael S. Tsirkin wrote:
> On Thu, Jul 26, 2018 at 10:42:52AM -0600, Alex Williamson wrote:
> > On Thu, 26 Jul 2018 19:14:45 +0300
> > "Michael S. Tsirkin" <mst@redhat.com> wrote:
> > 
> > > On Thu, Jul 26, 2018 at 06:51:13PM +0300, Marcel Apfelbaum wrote:
> > > > Hi
> > > > 
> > > > On 07/26/2018 05:52 PM, Stefan Hajnoczi wrote:  
> > > > > On Thu, Jul 12, 2018 at 07:33:14AM +0000, Elijah Shakkour wrote:  
> > > > > > Hey,
> > > > > > 
> > > > > > Our team is adding a NIC functional  emulation to QEMU.
> > > > > > One of the features we are adding to this NIC is SRIOV.
> > > > > > 
> > > > > > Here is the error message I get when checking SRIOV support of  our
> > > > > > emulated NIC on Win2016 server (the hyper-v VM).
> > > > > > "
> > > > > > SR-IOV cannot be used on this system as the PCI Express hardware
> > > > > > does not support Access Control Services (ACS) at any root port.  
> > > > 
> > > > QEMU's emulated PCI Express Root Ports do not support ACS yet, however I
> > > > am
> > > > not sure ACS is a prerequisite
> > > > for SR-IOV.
> > 
> > ACS is certainly not a prerequisite for SR-IOV.
> >  
> > > Looks like windows blocks dev assignment in nested VMs without it.
> > > Thinking about it, doesn't vfio do the same by default? I think vfio has
> > > a flag to override this though.
> > 
> > IOMMU grouping in Linux takes isolation via ACS and device specific
> > mechanisms into account, limiting the granularity with which a
> > userspace driver can claim ownership of devices,
> 
> In that you must assign all devices behind this port as a group?

Yes, that's why I needed to add PCIe config space ACS support in ioh3420:
to be able to use VFIO to pass through individual VFs to different L2 VMs with
"plain" kernels on the L1 host which owns the SR/IOV device.

Knut

> > but it doesn't
> > actually prevent enabling SR-IOV on the endpoint.  It just makes it
> > less useful if your intention is to use SR-IOV for device assignment.
> > The only overrides for this are out-of-tree.  Thanks,
> >
> > 
> > Alex
> 
>