On Sat, Dec 04, 2021 at 04:03:53PM +0800, "yukuai (C)" [off-list ref] wrote:

I was thinking that if there are active blkgs, holding queue_lock will
ensure blkcg won't be freed.

My take is that the function traverses the whole blkcg tree (from global
root) and nothing prevents concurrent blkcg_css_free() in a possibly
unrelated branch (or queue).

By the way, does spin_lock can guarantee this since it disables preempt
like what rcu_read_lock() does?

Yes (but don't quoRTe me on that :-).

(It even isn't issue with a non-preemptible kernel neither but the code
IMO should be generic to allow for different configs -- or as I
mentioned initially, make a comment why the tree traversal is not
affected by concurrent frees.)

Thanks,
Michal