We only recently have the min version of libvirt high enough for us to use > v1.3.4. So I guess its pluasable.
+1 Dan's comment on live-migration permissions, many users have access to it, although that is not the default.
+1 on Dan's comment around the data leak being one of the worst possible failure modes here.
I guess the patch is tricker for when min_libvirt is < v1.3.4.
Does this not also affect pinned CPU cores as well? Because we might pick a different set of CPUs on the desitnation hypervisor (train onwards)? With all the speclative execution stuff, that is also a possible data leak. Certainly leads to performance oddness.
Do we have an understanding of what backends use this operation mode? I remember discussing this with Cinder around multi-attach time frame, and it sounded like very few backends (if any upstream?) actually use these host based connections.
We only recently have the min version of libvirt high enough for us to use > v1.3.4. So I guess its pluasable.
+1 Dan's comment on live-migration permissions, many users have access to it, although that is not the default.
+1 on Dan's comment around the data leak being one of the worst possible failure modes here.
I guess the patch is tricker for when min_libvirt is < v1.3.4.
Does this not also affect pinned CPU cores as well? Because we might pick a different set of CPUs on the desitnation hypervisor (train onwards)? With all the speclative execution stuff, that is also a possible data leak. Certainly leads to performance oddness.
Do we have an understanding of what backends use this operation mode? I remember discussing this with Cinder around multi-attach time frame, and it sounded like very few backends (if any upstream?) actually use these host based connections.