Re: [PATCH] docs: fusa: Add requirements for Device Passthrough

[Julien Grall <julien@xxxxxxx>]

Hi Bertrand,

On 09/10/2024 07:30, Bertrand Marquis wrote:
> Hi Stefano,
>
> > On 9 Oct 2024, at 00:46, Stefano Stabellini <sstabellini@xxxxxxxxxx> wrote:
> >
> > On Tue, 8 Oct 2024, Oleksandr Tyshchenko wrote:
> > > > > On 7 Oct 2024, at 20:55, Oleksandr Tyshchenko <olekstysh@xxxxxxxxx> wrote:
> > > > > +
> > > > > +Rationale:
> > > > > +
> > > > > +Comments:
> > > > > +
> > > > > +Covers:
> > > > > + - `XenProd~device_passthrough~1`
> > > > > +
> > > > > +Assign PCI device to domain (with IOMMU)
> > > > > +----------------------------------------
> > > > > +
> > > > > +`XenSwdgn~passthrough_assign_pci_device_with_iommu~1`
> > > > > +
> > > > > +Description:
> > > > > +Xen shall assign a specified PCI device (always implied as DMA-capable)
> > > > > to
> > > > > +a domain during its creation using passthrough (partial) device tree on
> > > > > Arm64
> > > > > +and Hyperlaunch device tree on AMD-x86. The physical device to be
> > > > > assigned is
> > > > > +protected by the IOMMU.
> > > > This is a very long and complex requirement.
> > > > I would suggest to split it in 3:
> > > > - generic: Xen shall support assign PCI devices to domains.
> > > > - arm64 one: Xen shall assign PCI devices based on device tree (explain how
> > > > this is configured in dts)
> > > > - amd: xxxx based on hyperlaunch
> > > I agree, will split, but ...
> > >
> > > > - Xen shall use the IOMMU to enforce DMA operations done by a PCI device
> > > > assigned to a domain to be restricted to the memory of the given domain.
> > >
> > > ... does this need to be a separate 4th requirement here (and for the similar
> > > requirement for the platform device down the document) or this sentence is
> > > meant to be added to all resulting generic/arm64/amd requirements?
> > This is not specific to PCI, though? The generic requirement is "Xen
> > shall use the IOMMU to enforce DMA operations done by a DMA-capable
> > device assigned to a domain to be restricted to the memory of the given
> > domain".
> >
> > I think it is also OK to both have a PCI-specific and a
> > non-PCI-specific requirement for that, I just wanted to mention that it
> > doesn't look like something to PCI-specific.
> I think we should completely abstract out the DMA engine problem as it
> is not specific to a PCI or not device but to any DMA capable stuff.
+1

> Maybe the same also goes for interrupts and how those are assigned or
> forwarded or not to a domain (attached or not to a device).
I am not sure we can abstract this one because:
 * For platform devices, we should be able to support SPIs and MSI(-x) 
(only the former is so far).
 * For PCI devices, at least on Arm, AFAIR, we only intend to support 
MSI(-x).
Cheers,

--
Julien Grall