Hi Julien,
On Fri, Dec 24, 2021 at 02:30:50PM +0100, Julien Grall wrote:
> Hi,
>
> On 23/12/2021 20:06, Stefano Stabellini wrote:
> > On Wed, 22 Dec 2021, Stefano Stabellini wrote:
> > > # Future Ideas
> > >
> > > A great suggestion by Julien is to start supporting the dom0less partial
> > > device tree format in xl/libxl as well so that we can have a single
> > > "device_tree" option in dom.cfg instead of 4 (device_tree, iomem, irqs,
> > > dtdev).
> > >
> > > Even with that implemented, the user has still to provide a partial dtb,
> > > xen,reg and xen,path. I think this is a great step forward and we should
> > > do that, if nothing else to make it easier to switch between dom0less
> > > and normal domU configurations. But the number of options and
> > > information that the user has to provide is still similar to what we
> > > have today.
> >
> > I have just realized that if we start to parse the partial DTB in
> > xl/libxl the same way that we do for dom0less guests (parse "xen,path",
> > "xen,reg", and "interrupts", making dtdev, irqs and iomem optional)
> > actually we can achieve the goal below thanks to the combination:
> > "xen,path" + "xen,force-assign-without-iommu".
> >
> > In other words, with dom0less we already have a way to specify the link
> > to the host node even if the device is not a DMA master. We can do that
> > by specifying both xen,path and xen,force-assign-without-iommu for a
> > device.
> >
> > This is just FYI. I am not suggesting we should introduce dom0less-style
> > partial DTBs in order to get SCMI support in guests (although it would
> > be great to have). I think the best way forward for this series is one
> > of the combinations below, like a) + d), or a) + c)
>
> I strongly prefer a) + c) because a warning is easy to miss/ignore. At least
> with the extra property the user made an action to think about it and agree
> that this is the way do it.
>
> It is also easier to spot if we ask the user to provide the configuration
> file.
>
Let me share my thoughts about c), which is:
c) require force-assign-without-iommu="true" in dom.cfg
Adding this parameter to domain config means removing
xen,force-assign-without-iommu param from partial DTB. This will affect
dom0less configuration, which I can't test for now without extra effort.
What I suggest is to implement a) + d) in this patch series, which is:
a) extend dtdev to cover all devices, including non-DMA masters
d) or print a warning like:
"WARNING: device assignment safety for device XXX cannot be
verified. Please make sure XXX is not a DMA mastering device."
And introduce a) + c) with the next patch series where dom0less scmi
support will be done.
Maybe leave a comment in code that force-assign-without-iommu config parameter
should be implemened.
What do you think about this?
--
Best regards,
Oleksii.
