Yeah, as far as I understand, this looks like something RPU wasn't really designed to handle, and I'm not deep enough in the internals to say for sure how tricky it would be to add.
My instinct is this sits more in feature request territory than a bug, so I'd suggest filing it on https://feedback.vates.tech with something like "RPU: graceful shutdown/restart for PCI-pinned VMs" and your actual topology in the description. The product team will understand the constraint much better with a concrete Ceph setup to look at.
Someone here probably knows if there's a cleaner workaround than manually draining and shutting down the pinned VMs before each host update, but that's what I'd fall back on in the meantime.
