Re: [RFC PATCH] iommu/arm-smmu-v3: Allow nested attach for PCI bridges without vDEVICE

[Jason Gunthorpe]

On Mon, Aug 17, 2026 at 02:26:46PM +0000, Dmitry Malkin wrote:
> 
> > This kind of situation is very challenging to setup in the VM, and at
> > this point you may as well include the switch to the VM too.
> 
> Passing through the PCI switch would not help because the device without
> a vDEVICE is the host Root Port. Neither the Root Port nor the switch
> should be exposed to the VM.

Oh that's probably because of the unfixed grouping bugs in the
kernel. The root port should not be part of the group with correctly
set ACS flags for switch P2P.

> A multi-device IOMMU group is an intentional effect of using
> 'pci=config_acs', and I would prefer to restore support for this
> configuration upstream.

It doesn't work. We tried to make it work, there are many problems
unfortunately. Don't use multi-device groups in the hypervisor.

This patch is pure hacky, there is no good justification for the
kernel to ignore certain devices in the group inside an iommu driver.

Jason