Re: [PATCH] blk-cgroup: Flush stats before releasing blkcg_gq

[Ming Lei <ming.lei@redhat.com>]

On Wed, May 24, 2023 at 12:10:55AM -0400, Waiman Long wrote:
> 
> On 5/23/23 22:37, Ming Lei wrote:
> > Hi Yosry,
> > 
> > On Tue, May 23, 2023 at 07:06:38PM -0700, Yosry Ahmed wrote:
> > > Hi Ming,
> > > 
> > > On Tue, May 23, 2023 at 6:21 PM Ming Lei <ming.lei@redhat.com> wrote:
> > > > As noted by Michal, the blkg_iostat_set's in the lockless list
> > > > hold reference to blkg's to protect against their removal. Those
> > > > blkg's hold reference to blkcg. When a cgroup is being destroyed,
> > > > cgroup_rstat_flush() is only called at css_release_work_fn() which
> > > > is called when the blkcg reference count reaches 0. This circular
> > > > dependency will prevent blkcg and some blkgs from being freed after
> > > > they are made offline.
> > > I am not at all familiar with blkcg, but does calling
> > > cgroup_rstat_flush() in offline_css() fix the problem?
> > Except for offline, this list needs to be flushed after the associated disk
> > is deleted.
> > 
> > > or can items be
> > > added to the lockless list(s) after the blkcg is offlined?
> > Yeah.
> > 
> > percpu_ref_*get(&blkg->refcnt) still can succeed after the percpu refcnt
> > is killed in blkg_destroy() which is called from both offline css and
> > removing disk.
> 
> As suggested by Tejun, we can use percpu_ref_tryget(&blkg->refcnt) to make
> sure that we can only take a reference when the blkg is online. I think it
> is a bit safer to take a percpu refcnt to avoid use after free. My other

blkg_release() does guarantee that no new stat associated with this
blkg can be added any more, so what is the use-after-free?

> concern about your patch is that the per cpu list iterations will be done
> multiple times when a blkcg is destroyed if many blkgs are attached to the

Yeah, but is it one big deal? The percpu list should be flushed just in one
of blkg's release handler.

> blkcg. I still prefer to do it once in blkcg_destroy_blkgs(). I am going to

Problem is that new stat still may be added to this list after
percpu_ref_kill() returns.

To be honest, I really hate to grab/put blkg refcnt for adding/consuming
state, and this way is too fragile.

Thanks,
Ming