Hello, Li.

On Mon, Nov 19, 2012 at 05:02:42PM +0800, Li Zefan wrote:

On 2012/11/13 11:01, Tejun Heo wrote:

quoted

struct cgroup is made RCU-safe by synchronize_rcu() in cgroup_diput().

but synchronize_rcu() is called before ss->destroy().

rcu_read_lock();
for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
	print_cfs_rq(m, cpu, cfs_rq);
	-> call cgroup_path(task_group->css.cgroup);
rcu_read_unlock();

With this patch, if the above code race with cgroup_diput(), we might
end up accessing a cgroup which has been freed.

Ah, okay.  So, the problem here is that sched is using ->css_free() as
a de-registration point rather than freeing and may end up walking it
after ->css_free() is complete inside RCU period.

I think the correct solution is using ->css_offline() for that.  It's
ugly to require double RCU grace periods.

quoted

diff --git a/kernel/cgroup.c b/kernel/cgroup.c
index 278752e..a91e7ad 100644
--- a/kernel/cgroup.c
+++ b/kernel/cgroup.c

@@ -893,7 +893,7 @@ static void cgroup_diput(struct dentry *dentry, struct inode *inode)
 
 		simple_xattrs_free(&cgrp->xattrs);
 
-		kfree_rcu(cgrp, rcu_head);
+		kfree(cgrp);

This was also added to prevent a race in group scheduling code, and I think the race still
exists.

Care to point out which one?  I don't think the double-RCU workaround
is a good idea.  We really should sort it out by following object
lifecycle rules consistently.

Thanks.

-- 
tejun