Re: [PATCH v4 08/11] PCI: liveupdate: Require preserved devices are in immutable singleton IOMMU groups

[Samiullah Khawaja <skhawaja@google.com>]

On Thu, Apr 23, 2026 at 04:09:01PM -0700, David Matlack wrote:
>On Thu, Apr 23, 2026 at 3:53 PM Jason Gunthorpe <jgg@nvidia.com> wrote:
>>
>> On Thu, Apr 23, 2026 at 03:10:55PM -0700, David Matlack wrote:
>> > On Thu, Apr 23, 2026 at 2:23 PM David Matlack <dmatlack@google.com> wrote:
>> > >
>> > > Restrict support for preserving PCI devices across Live Update to
>> > > devices in immutable singleton IOMMU groups. A device's group is
>> > > considered immutable if all bridges upstream from the device up to the
>> > > root port have the required ACS features enabled.
>> > >
>> > > Since ACS flags are inherited across a Live Update for preserved devices
>> > > and all the way up to the root port, the preserved device should be in a
>> > > singleton IOMMU group after kexec in the new kernel.
>> > >
>> > > This change should still permit all the current use-cases for PCI device
>> > > preservation across Live Update, since it is intended to be used in
>> > > Cloud enviroments which should have the required ACS features enabled
>> > > for virtualization purposes.
>> > >
>> > > If a device is part of a multi-device IOMMU group, preserving it will
>> > > now fail with an error. This restriction may be lifted in the future if
>> > > support for preserving multi-device groups is desired.
>> > >
>> > > Signed-off-by: David Matlack <dmatlack@google.com>
>> >
>> > Jason, do you think requiring singleton iommu groups is still
>> > necessary/useful now that this series preserves ACS flags on preserved
>> > devices and upstream bridges?
>>
>> I have forgotten why we introduced that? There are alot of funky
>> things about iommu groups that might be important upon restoration..
>
>You had originally suggested it in this thread:
>
>  https://lore.kernel.org/kvm/20260301192236.GQ5933@nvidia.com/
>
>> Like if you preserve one group member but not the other what do you ?
>
>Yeah I imagine there could be some tricky cases there...
>
>I wonder if PCI core is the right layer to enforce this. Maybe this
>fits better into Sami's IOMMU core series since that is where all
>those tricky cases will be (I imagine?).

+1

Also I think this should probably be checked by iommufd and invoked
through vfio cdev. Basically when vfio cdev calls into iommufd to
preserve IOMMU specific aspects of device (PASID table etc), iommufd can
check this and return error.
>
>> Even if you have ACS flags there are cases where groups are still
>> aliasing DMA..
>
>Hm, if a DMA alias can be created after boot time enumeration even
>with the REQ_ACS_FLAGS check, then
>pci_device_group_immutable_singleton() is not really immutable.
>
>
>
>> Frankly, multi-device iommu groups don't even work fully last time we
>> tried to use them in a VMM. So I think I would not expect them to ever
>> intersect with live update. Blocking something tricky you can't test
>> does seem like a reasonable thing.