Re: [PATCH v18 12/14] net/tap: support local migration with virtio-net

["Michael S. Tsirkin" <mst@redhat.com>]

On Tue, Jul 14, 2026 at 09:43:52AM +0300, Vladimir Sementsov-Ogievskiy wrote:
> On 13.07.26 22:36, Michael S. Tsirkin wrote:
> > On Fri, Jul 10, 2026 at 10:18:01PM +0300, Vladimir Sementsov-Ogievskiy wrote:
> > > Support transferring of TAP state (including open fd).
> > > 
> > > Add new property "local-migration-supported", which defines
> > > whether local-migration is actually supported for this TAP device.
> > > Starting from 11.2 QEMU Machine Types it's enabled by default.
> > > 
> > > Note, that local-migration (including migrating opened FDs
> > > through migration channel, which must be UNIX socket), is
> > > enabled by global "local" migration parameters. But individual
> > > devices may have additional options to enable/disable it
> > > per device.
> > > 
> > > The tricky thing, is that we need to know, should we do
> > > call open/connect in TAP initialization code. That means,
> > > that on incoming migration we want to know status of
> > > "local" migration parameter at time of creating TAP device.
> > > So, currently, for incoming local migration only TAP device
> > > created through QMP is supported, as you need to set migration
> > > parameter first (through QMP as well).
> > > 
> > > So the full picture is:
> > > 
> > > On source, to start outgoing "local" migration you need:
> > > 
> > >   - migration parameter "local" set to true
> > >   - "local-migration-supported" TAP option set to true (the
> > >     default, starting from 11.2 QEMU Machine Types)
> > > 
> > > If at least one of these options is not set, TAP backend
> > > doesn't participate in migration.
> > > 
> > > On target, things are more difficult:
> > > 
> > > Same, you need both "local" and "local-migration-supported"
> > > be set. And same, if one of them is not set, TAP backend
> > > is initialized as usual, and doesn't accept any incoming
> > > state.
> > > 
> > > Additionally, if you are going to set "local", it must be
> > > set before creating the TAP device. If TAP device created
> > > with "local" unset, it initializes as usual. If you enable
> > > "local" after it and start incoming migration, it will fail
> > > in .pre_load handler of TAP backend.
> > > 
> > > Moreover, there are interface restrictions: if you create TAP
> > > device when QEMU is in INCOMING state, and both "local"
> > > and "local-migration-supported" set, most of TAP options are
> > > not allowed, and script/downscript are required to be explicitly
> > > unset (set to "" or "no").
> > > 
> > > Signed-off-by: Vladimir Sementsov-Ogievskiy<vsementsov@yandex-team.ru>
> > 
> > I am still confused why is this per device. I just do not
> > see where is it useful to set it for some devices
> > but not others. So it looks like a machine property
> > "incoming local migration" to me?
> > 
> > Sorry if I am being dense.
> 
> 1. We need at least "per device class": when local-migration-support is
> implemented for some new device, we still need to migrate from/to older
> version without this support.
> 
> Assume we have "local" migration parameter and support for TAP in 11.2.
> Then, in 12.0 support for vhost-user-blk is implemented. If we have only
> global "local" parameter, the migration will be broken between 11.2 and
> 12.0,
> with local=true, as in 11.2 it means only TAP local migration, but
> in 12.0 it means both TAP and vhost-user-blk local migration.
> 
> Of course, now it's a bit redundant, as we have only TAP, and may define,
> that TAP is always supported (as support is introduced together with
> "local"). But there are another considerations:

If this is your way to version the behaviour, that is fine, but it should
be an internal property without ABI guarantees, starting with "x-".

> 
> 2. Symmetry: if have a possiblity to disable feature only for vhost-user-blk,
> seems good to have same possibility for TAP.
> 
> 3. Disabling fd-migration per device may be useful to switch this specific
> device to another backend.

Unclear what good is it, and the drawback is that it blows up the size
of our testing/support matrix.

> -- 
> Best regards,
> Vladimir