Re: [Xen-devel] [RFC] scf: SCF device tree and configuration documentation

[Julien Grall <julien.grall@xxxxxxx>]

On 04/05/17 16:50, Andrii Anisov wrote:
> Julien,
Hi Andrii,

> > What I would like to understand is what are the information that the
> > hypervisors as to know for sharing co-processor? So far I have:
> >     - MMIOs
> >     - Interrupts
> >
> > Anything else?
> IOMMU bindings.
> This knowledge enough to get the physical coprocessor shared.
>
> In order to spawn a virtual coprocessor (vcoproc) for some domain you
> have to provide additional configuration information:
>     - Which physical coprocessor this vcoproc should represent to a
> domain ( a SoC could have several physical coprocs shared through the
> framework)
>     - IRQ(s) (provided that no IRQ remapping is implemented in XEN)
> could be omitted (or used for verification only)
>     - IOMEM ranges correspondence between this vcoproc instance and a
> physical coprocessor.
>
> The latest point in the configuration is the most complex.
> Let me explain a use case we faced now:
>     - a GPU has two different firmwares implementing OpenGL and OpenCL
>     - we need both GL and CL in the same domain working simultaneously
> (actually concurrently, but the concurrency should be transparent for
> domain, GPU drivers and firmwares)
> In current case we are lucky, the GPU has a single mmio range.
> We can implement such system using SCF: spawn two vcoprocs for a domain.
> Those vcoprocs will have own mmio range within the domain.
> In a hypervisor those mmio ranges would be served by the same handler,
> but must be associated with the own vcoproc context.
I have CCed Ian and Wei to comment on the difficult to describe a such 
interface in libxl. They may have insights how to do this properly.
@Ian @Wei: Andrii is suggesting to use Device-Tree for describing 
virtual co-processor as it seems it would be very difficult to do the 
same with the configuration file. See the suggested binding in [1].
> In case a coprocessor has several mmio ranges things are getting worse.
>
> In a device tree configuration concept I explicitly link vcoproc to
> pcoproc and keep mmio ranges correspondency with names.
> I'm not sure how to keep this coincidence in a simple way.
Cheers,

[1] https://www.mail-archive.com/xen-devel@xxxxxxxxxxxxx/msg106924.html


--
Julien Grall

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
https://lists.xen.org/xen-devel