On Fri, Jun 19, 2020 at 08:00:41PM -0700, Cong Wang wrote: > On Fri, Jun 19, 2020 at 6:14 PM Roman Gushchin <g...@fb.com> wrote: > > > > On Sat, Jun 20, 2020 at 09:00:40AM +0800, Zefan Li wrote: > > > I think so, though I'm not familiar with the bfp cgroup code. > > > > > > > If so, we might wanna fix it in a different way, > > > > just checking if (!(css->flags & CSS_NO_REF)) in cgroup_bpf_put() > > > > like in cgroup_put(). It feels more reliable to me. > > > > > > > > > > Yeah I also have this idea in my mind. > > > > I wonder if the following patch will fix the issue? > > Interesting, AFAIU, this refcnt is for bpf programs attached > to the cgroup. By this suggestion, do you mean the root > cgroup does not need to refcnt the bpf programs attached > to it? This seems odd, as I don't see how root is different > from others in terms of bpf programs which can be attached > and detached in the same way. > > I certainly understand the root cgroup is never gone, but this > does not mean the bpf programs attached to it too. > > What am I missing?
It's different because the root cgroup can't be deleted. All this reference counting is required to automatically detach bpf programs from a _deleted_ cgroup (look at cgroup_bpf_offline()). It's required because a cgroup can be in dying state for a long time being pinned by a pagecache page, for example. Only a user can detach a bpf program from an existing cgroup. Thanks!
