Hey,

On Fri, Jun 29, 2012 at 10:20:11AM +0800, Li Zefan wrote:

quoted

Can you please elaborate a bit?  I'm not really following?  Where does
the last root->d_count-- come from?

From my limited knowledge about vfs internal, seems the parent's refcnt won't go down
to 0 before its children. When mkdir, the parent's refcnt will be incremented, and
after rmdir, dput(subdir) will drop subdir's refcnt and then drop parent's.

So when dropping the subdir's refcnt and leading the superblock to be killed, the root's
dentry is still > 0.

Heh, yeah, I thought you found who was holding out on the refcnt. :)

quoted

quoted

I use this script to reproduce the bug:

mount -t cgroup -o cpu xxx /mnt
mkdir /mnt/sub
sleep 100 < /mnt/sub &
kill $!
wait $!
rmdir /mnt/sub
umount /mnt

Unfortunately, this doesn't reproduce the bug here either. :(

I can reproduce the bug reliably.. Try s/cpu/perf or s/cpu/net_cls, which have fewer
cgroup files?

Hmm... weird.  Maybe some debug config option I have is preventing the
race from occurring as reliably?  Can you please attach your .config?

Thanks.

-- 
tejun