On Fri, Feb 06, 2026 at 04:50:15PM -0300, Fabiano Rosas wrote: > Peter Xu <[email protected]> writes: > > > On Fri, Feb 06, 2026 at 09:29:04AM -0300, Fabiano Rosas wrote: > >> There's some amount of rigidness caused by qdev requirements > >> unfortunately. > >> > >> I'm not sure if I ever posted it, but I wrote some code to move the > >> parameters into a MigrationOptions object so we could make > >> MigrationState not be a TYPE_DEVICE anymore. But then we'd end up with > >> something like -global migration-options by default, so it kinda killed > >> the idea. > > > > It's not strictly about TYPE_DEVICE, but reusing of qdev properties, right? > > At least the issue described by your comment was about offseting and it > > sounds like so. > > > > Absolutely, I was just rambling. > > Of course, TYPE_DEVICE brings us weirdness such as: > > (qemu) device_add migration,help > migration options: > announce-initial=<size> - (default: 50) > announce-max=<size> - (default: 550) > announce-rounds=<size> - (default: 5) > ... > > So it would be nice to drop this dependency.
Yep. I'll move my (below) series higher priority to respin, likely I can drop the QOBJECT_COMPAT idea that didn't attract much attention, but instead I can stick with exporting the qdev property helpers. It'll be enough for us to make migration object to get rid of TYPE_DEVICE. > > > I just want to double check with you that I think the problem you described > > will also present even after applying my other series: > > > > https://lore.kernel.org/r/[email protected] > > > > That series only removes the TYPE_DEVICE dependency, but not qdev > > properties. I think it's the qdev property trick that is relevant at least > > to the offset issue you mentioned so MigrationParameters cannot be > > g_new()ed. > > > > The major use case for this qdev reuse is: (1) help scripting, so as to use > > -global migration.XXX=YYY, (2) support migration in machine compat > > properties. IIUC (1) isn't a major thing we ask for (again, maybe I used > > the most of it.. but maybe only me; I'm not sure..), as long as anything > > can keep (2) working then we can consider. > > The properties are also useful in providing defaults for the migration > options and perhaps we could make use of the .get/.set methods in a more > convenient way in the migration code, such as implementing per-option > input validation instead of checking all options always. (although I > haven't thought about the overlap with QAPI) Yes good point. Said that, currently if we're still with qdev properties we can't easily inject those checks due to the hard-coded get()/set() for qdev properties, e.g. qdev_prop_uint8 will provide its static get()/set() hooks. We either need to teach qdev props to take check functions, or impl our own (then we'll need to provide the default_val and machine compat features). Maybe it's easier to do the former, and maybe some other qdev props can leverage too. Not an immediate concern. > > About -global, we should do better to separate the compat options from > the regular options. (and no, a blank line in options.c is not enough > separation!). I worry someday we'll break something in compat while > trying to change the regular options. Normally if we have something in machine compat properties, those shouldn't be used for normal users. Those, importantly, also shouldn't be present in QMP set parameters. That should IMHO be the line to identify a real compat option v.s. a real parameter, because we do not expect normaly user to use -global, only advanced, so one knows what one is doing and taking the risks. It actually makes sense to still be able to adjust some internal compat behaviors if an advanced user really wants; that might be handy for debugging in some cases to overwrite machine compat properties, even if extremely rare. > > Another point is that even after all the recent changes to options.c, we > are still left with a list of migration_properties that is optional to > use. Which means setting defaults is also optional. > > If we decide to retroactively add a default we'll suddenly have a new > option showing up in -global! Having to discover mid-way through the > implementation of a new migration feature that setting a default like > the other options do will also add property code to your option and now > you have one more source of SIGSEGV is not super fun either. > > I had the intention to somehow force every option to have an equivalent > in migration_properties so that every option would have a default > explicitly set, but seeing the recent work on fixing the StrOrNull > property, I don't think it's worth it to ask that people go through the > hassle of implementing a property (when the new option cannot use the > existing types). Right. We can discuss this case by case, normally I think we should always welcome the same parameter to be added into migration_properties[], most of them will be simple scalars so we shouldn't worry. We can make it optional if it needs extensive work on new types. -- Peter Xu
