Re: [PATCH v5 02/18] iommu: Implement IOMMU Live update FLB callbacks

[Samiullah Khawaja]

On Mon, Oct 12, 2026 at 05:25:42PM -0700, Nicolin Chen wrote:
> On Sat, Oct 10, 2026 at 03:18:11AM +0000, Samiullah Khawaja wrote:
> > On Tue, Oct 06, 2026 at 04:31:01PM -0700, Nicolin Chen wrote:
> > > On Mon, Sep 21, 2026 at 12:48:18AM +0000, Samiullah Khawaja wrote:
> > > > +struct iommu_flb_obj {
> > > > +	struct mutex lock;
> > > > +	struct iommu_flb_ser *ser;
> > > > +
> > > > +	struct iommu_hw_array_ser *curr_iommu_array;
> > > > +	struct iommu_domain_array_ser *curr_domain_array;
> > > > +	struct iommu_device_array_ser *curr_device_array;
> > > > +};
> > >
> > > IIUIC, there should be one pair of obj + ser in the entire system:
> > >  - old kernel has one outgoing obj + ser
> > >  - new kernel has one incoming obj + ser
> > > right?
> > >
> > > If so, things in iommu_flb_obj (except ser) are all transient, and
> > > there is no need to preserve them across the two kernels.
> >
> > The contents of iommu_flb_obj are not preserved. Only the contents of
> > iommu_flb_ser are preserved. Other "curr_" structs in iommu_flb_obj are
> > only there for quick access to add new iommus, domains and devices.
> > >
> > > It also feels redundant to have this iommu_flb_obj structure. Why
> > > not link liveupdate_flb_op_args directly to the ser? Then, things
> > > in iommu_flb_obj could be global?
> >
> > The liveupdate_flb_op_args is linked directly to the ser, but it only
> > contains the physical address for the next kernel. Basically LUO
> > provides following mechanism to handle FLBs:
> >
> > - data: ser structure for the next kernel.
> > - obj: Live object that can be used in the current or next kernel for
> >   easy access or staging.
> >
> > For example here, iommu_flb_obj has pointers to the end of the array
> > linked list of each type.
> >
> > Also note LUO keeps separate incoming and outgoing data and obj for each
> > FLB. In the new kernel both can exist at the same time, as the incoming
> > one stays until finish and preserving for the next live update creates
> > the outgoing one. So keeping these in the obj avoids managing two sets
> > of globals and their lifetime in the iommu code.
>
> Could incoming and outgoing objects exist in the same time?
>
> Correct me if I am wrong: I expect the lifecycle of an incoming
> object will end once liveupdate completes, and the outgoing one
> can only exist after that?

LUO can have incoming and outgoing at the same time. The incoming FLB
stays until all the sessions from the previous kernel that use it are
finished, and the outgoing FLB is created when the first FD that uses it
is preserved in the current kernel. So if a new preservation starts
before all the incoming sessions are finished, both exist.

Generally, it should be rare, as after the live update the userspace
orchestrator (like systemd) would probably finish all the sessions. At
the FD level it is mutually exclusive, but at the FLB level it is
possible.
> Nicolin

Sami