Re: Preserving TDX module during kexec with live update

On Thu, Jul 16, 2026 at 9:55 AM Pasha Tatashin
<[email protected]> wrote:
>
> On 07-15 18:28, Sagi Shahar wrote:
> > Hi,
> >
> > I started looking into preserving the TDX module across live updates
> > and I'm not sure what the best approach is. I tried looking into using
> > flb as David M suggested but I don't think this approach works for the
> > TDX module.
> >
> > Right now the way that TDX supports kexec is by disabling the TDX
> > module on kexec and then reinitilizing it on the next boot [1]. This
> > approach clears the TDX module state so any TDX VM from the source
> > kernel won't be able to resume running on the destination kernel.
> >
> > In order to support live update while keeping the VMs alive we need to
> > skip the TDH_SYS_DISABLE call and instead preserve pages used by the
> > TDX module such as the TDMR pages from the source kernel.
>
> Agreed.
>
> >
> > From what I understand, using flb will tie the TDX module save/restore
> > to the lifecycle of the VMs depending on it. This means the module
> > will be restored too late during kernel boot, after the VMs are
> > restored using the live update IOCTLs.
>
> That is not correct. FLBs specifically can be restored very early in
> boot. We do that for HugeTLB preservation, for example.
>
> liveupdate_flb_get_incoming() does not require userspace to resume or
> restore VMs. You can call it super early in boot to see if there is FLB
> data present and, if so, change the way the module is initialized during
> boot.
>
Thanks Pasha, I also got a few pointers from Samiullah and think I
should be able to make it work. I'm going to be on vacation for the
next couple of weeks so I'm going to miss the next meeting but
hopefully I can start working on it once I'm back.
> >
> > I'm thinking we should preserve the TDX module during the shutdown
> > stage of the source kernel. Perhaps we can preserve the pages used by
> > the TDX module as we allocate them on the source kernel and pass
> > pointers and additional metadata during KHO. Then on the destination
> > we need to pipe this information to where the TDX module gets
> > initialized during the KVM module initialization.
>
> Please don't invent another state machine; instead, use the existing
> one.
>
> Pasha
>
> >
> > I haven't found a device that acts in a similar way so I'm not sure
> > how complicated this will be or if there's a better approach.
> >
> > I would love to hear your thoughts and if you have any pointers on
> > where this code can be implemented.
> >
> > Thanks,
> > Sagi
> >
> > [1] https://lore.kernel.org/all/[email protected]/