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. > 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) 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. 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).
