This bug is used to track libvirt support on postcopy preemption. There's one new capability called "postcopy-preempt" supported for QEMU migration to speedup postcopy page request latencies. To enable it, one can simply enable the capability before starting the migration for postcopy. It does not need other special parameters. The only limitation is it only works with protocol that supports multiple channels (similar to multi-fd in this case).
Peter is there any downside of this feature or can we just enable it automatically without an explicit request from a user? I guess we could enabled it whenever postcopy and multi-fd are enabled... And should it be enabled on both sides of migration or just one of them?
(In reply to Jiri Denemark from comment #2) > Peter is there any downside of this feature or can we just enable it > automatically without an explicit request from a user? I guess we could > enabled it whenever postcopy and multi-fd are enabled... Postcopy and multifd are not yet compatible, afaict, so the new preempt mode is not compatible with multifd as well. However yeah if any VM that can enable multifd should also be able to enable preempt when with postcopy, and it should always benefit postcopy, iiuc. Currently one downside of preempt is we haven't gone though enough test, e.g., one known issue is multi channels can cause wrong ordered connections, similar to the multifd bug here: https://bugzilla.redhat.com/show_bug.cgi?id=2137740 I'll need to further look into it, but hopefully shouldn't affect the long term plan of having it by default. > > And should it be enabled on both sides of migration or just one of them? Both sides. Thanks!
Jiri, any updates for this one? Thanks,