在 2021/12/03 18:27, Michal Koutný 写道:

On Fri, Dec 03, 2021 at 03:50:01PM +0800, "yukuai (C)" [off-list ref] wrote:

quoted

blkg_destroy() is protected by the queue_lock，so I think queue_lock can
protect such concurrent scenario.

blkg_destroy() is not as destroying :-) as actual free, you should
synchronize against (the queue_lock ensures this for
pd_free_fn=throtl_pd_free but you may still trip on blkcg after
blkcg_css_free()).

Hi, Michal

I was thinking that if there are active blkgs, holding queue_lock will 
ensure blkcg won't be freed. However, if there are no active blkgs in
the first place, it seems right rcu_read_lock() can prevent this
iteration concurrent with css_release->css_release_work_fn->
css_free_rwork_fn.

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

Thanks,
Kuai

[Actually, I think you should see a warning in your situation if you
enable CONFIG_PROVE_RCU.]

HTH,
Michal