Re: [PATCH vfs/for-next v2] cgroup: fix top cgroup refcnt leak

[David Howells <dhowells@redhat.com>]

Andrei Vagin <avagin@gmail.com> wrote:

> Here is a reproducer for this problem:
> 
> [root@fc24 ~]# cat fs-vs-cg 
> set -m
> d=$(mktemp -d /tmp/cg.XXXXXX)
> for i in `seq 2`; do 
> 	mkdir $d/a$i
> 	mount -t cgroup -o none,name=xxxxy xxx $d/a$i
> done
> mkdir -p $d/a1/test
> for i in `seq 2`; do 
> 	umount $d/a$i
> done
> d=$(mktemp -d /tmp/cg.XXXXXX)
> for i in `seq 2`; do 
> 	mkdir $d/a$i
> 	mount -t cgroup -o none,name=xxxxy xxx $d/a$i
> done
> rmdir $d/a1/test
> for i in `seq 2`; do 
> 	umount $d/a$i
> done
> 
> You need to execute this script twice and it will stuck on a second
> attmept.

This causes:

	percpu ref (css_release) <= 0 (0) after switching to atomic

without your patch and:

	percpu ref (css_release) <= 0 (-2) after switching to atomic

with your patch, both followed by:

	WARNING: CPU: 0 PID: 0 at lib/percpu-refcount.c:155 percpu_ref_switch_to_atomic_rcu+0x90/0x1a0

on Al's for-next branch.

David