On the shutdown question from the Sashiko review of this patch: The gap is real, but it predates this patch. The driver in drm-misc-next has no .shutdown callback at all, so a reboot or kexec with a job in flight already leaves the cores running today. The callback added here takes the devfreq device down and puts the clock back to its boot rate; it does not stop, drain or reset the cores.
Quiescing the cores on shutdown (stopping the schedulers and resetting the cores through rocket_core_reset()) is part of the same teardown work as the use-after-free under "Not done" in the cover letter, where a core is unbound while clients submit: both need the driver to stop its schedulers and cores while files are still open. I would rather fix both in one follow-up, once Tomeu has said which way he wants the teardown to go. Igor
