Re: [PATCH v2 2/2] blk-cgroup: Optimize blkcg_rstat_flush()

[Tejun Heo]

On Wed, Jun 01, 2022 at 05:25:53PM -0400, Waiman Long wrote:
> I think the best way to protect against blkg destruction is to get a percpu
> reference when put into lockless list and put it back when removed.
> 
> BTW, when I ran a test that continuously create and destroy containers, the
> total number of blkcg's kept on increasing. There are some freeing of
> blkcg's but no freeing of blkg's at all. Maybe we have a similar dying
> blkcg's problem here. I will take a further look at that when I have time.

They get pinned by per-cgroup writebacks which gets pinned by lingering page
cache and other remaining accounted memory areas, so I think they can hang
around if there's no memory pressure. But, yeah, it'd be great to verify
that they actually go away under memory pressure.

Thanks.

-- 
tejun