Re: [PATCH v4 02/21] iommufd: Add iommufd_sw_map_msi()

[Andrew Jones]

On Wed, Aug 26, 2026 at 08:11:58PM +0200, Andrew Jones wrote:
> On Wed, Aug 26, 2026 at 01:32:34PM -0300, Jason Gunthorpe wrote:
> > On Wed, Aug 26, 2026 at 10:23:20AM +0200, Andrew Jones wrote:
> > > RISC-V uses the same basic model as x86 (an IOMMU-owned interrupt
> > > remapping IRQ domain). x86 makes that domain per-IOMMU because its IRTE
> > > resources belong to the IOMMU instance. RISC-V makes it per-device because
> > > its callbacks operate on the device's attached IOMMU domain and preserve
> > > that device's existing MSI parent. The IRQ callbacks provide the lifecycle
> > > hooks, even though the mappings, MSI table, and device context all remain
> > > IOMMU state.
> > 
> > The irq domain isn't per device, it is per msi table, many devices can
> > share the same table, and the VM should have a single shared table
> > across all devices. That's the core uniquely RISCV wrinkle I think.
> > 
> > IDK how Thomas would like to model this, but you will get a better
> > result if it is very clearly explained and asked.
> > 
> > I would like to keep the iommu out of it, so I prefer if IRQ code
> > patches in its MSI table pointer broadly the same way AMD does.
> > Though AMD doesn't have the restriction that remapping only works with
> > a S2 domain :(
> >
> 
> The per-device irqdomain model allows multiple devices to share the same
> S2 IOMMU domain and MSI table. Unlike AMD, however, the RISC-V remapping
> state belongs to a replaceable S2 domain rather than the physical IOMMU.
> The IRQ hierarchy must remain stable across domain changes, so a stable
> per-device irqdomain which operates on the currently attached S2 domain
> may be the closest practical equivalent. I'll explore this further while
> reworking the irqbypass series.
>

I think it should be possible to follow AMD's model more closely. I'll
create one IOMMU-owned irqdomain and install it as each device's MSI
parent at probe time, provided the device currently uses the expected
IMSIC MSI domain, which will be the IOMMU irqdomain's parent. The
irqdomain will contain no per-device or MSI-table state and its callbacks
can recover the device and operate on its currently attached S2 domain.
This is better than the per-device irqdomain since sharing it avoids
redundant domains.

I'll give that a go for the next posting of the irqbypass series, based on
v5 of this host-remapping series. v5 itself will not introduce an IOMMU
irqdomain.

Thanks,
drew