Re: [RFC PATCH v1 06/38] iommufd: Add and option to request for bar mapping with IORESOURCE_EXCLUSIVE

[Xu Yilun]

On Tue, Jul 29, 2025 at 11:29:17AM -0300, Jason Gunthorpe wrote:
> On Tue, Jul 29, 2025 at 01:58:54PM +0530, Aneesh Kumar K.V wrote:
> > Jason Gunthorpe <jgg@xxxxxxxx> writes:
> > 
> > > On Mon, Jul 28, 2025 at 07:21:43PM +0530, Aneesh Kumar K.V (Arm) wrote:
> > >> Signed-off-by: Aneesh Kumar K.V (Arm) <aneesh.kumar@xxxxxxxxxx>
> > >
> > > Why would we need this?
> > >
> > > I can sort of understand why Intel would need it due to their issues
> > > with MCE, but ARM shouldn't care either way, should it?
> > >
> > > But also why is it an iommufd option? That doesn't seem right..
> > >
> > > Jason
> > 
> > This is based on our previous discussion https://lore.kernel.org/all/20250606120919.GH19710@xxxxxxxxxx
> 
> I suggested a global option, this is a per-device option, and that
> especially seems wrong for iommufd. If it is per-device that is vfio,

I think this should be per-device. The original purpose of this
pci_region_request_*() is to prevent further mmap/read/write against
a vfio_cdev FD which would be used for private assignment. You shouldn't
prevent all other devices from working with userspace APPs (e.g. DPDK)
if there is one private assignment in system.

> if it is global then vfio can pick it up during the early phases of
> opening the device.
> 
> > IIUC, we intend to request the resource in exclusive mode for secure
> > guests—regardless of whether the platform is Intel or ARM. Could you
> > help clarify the MCE issue observed on Intel platforms in this context?
> 
> As I understand it Intel MCEs if the non-secure side ever reads from
> secure'd address space. So there is alot of emphasis there to ensure

Yeah, Intel TDX doesn't have a lower access control table for CC. So if
host reads, the TLP sends and MCE happens.

Thanks,
Yilun

> there are no CPU mappings.
> 
> Jason