Re: [PATCH v4 00/18] iommu: Add live update state preservation

[Samiullah Khawaja]

On Wed, Aug 26, 2026 at 09:12:53AM +0100, David Woodhouse wrote:
> On Sat, 2026-08-08 at 02:27 +0000, Samiullah Khawaja wrote:
> > This patch series introduces a mechanism for IOMMU state preservation
> > across live update, including the Intel VT-d driver support
> > implementation.
>
> Nice work; thanks. Do you have any timing information? Every
> millisecond spent on work here is perceived as steal time by the guests
> which are paused while it happens.
>
> So very quickly after the "yay! it works!" stage, we get to the "how do
> we do it in <100ms overall steal time" stage...

I don't have timing information for this series yet, but I completely
agree that every millisecond spent here is perceived as steal time by
the guests which are paused while it happens. And this is why we do
preservation of the IOMMU state early during LUO preserve() when the
guest is not paused. And later LUO freeze(), during shutdown, does the
minimum.

But we (Google) did some analysis internally on the time for which the
guest is paused, and the major time spent in such a kexec liveupdate is
in bringing back the rest of the system including important unpreserved
devices, other kernel drivers, userspace (VMM) and other major
components. Note the Orphaned VMs to reduce the VM pause time:

https://lore.kernel.org/all/afEwWZksU0Fw61oT@plex/

Thanks,
Sami