On Thu, Jan 09, 2014 at 06:59:07PM -0500, Steven Rostedt wrote:

On Thu, 9 Jan 2014 15:45:37 -0800
"Paul E. McKenney" [off-list ref] wrote:

quoted

quoted

 static void inode_free_security(struct inode *inode)
 {
 	struct inode_security_struct *isec = inode->i_security;

@@ -244,8 +252,7 @@ static void inode_free_security(struct i
 		list_del_init(&isec->list);
 	spin_unlock(&sbsec->isec_lock);

-	inode->i_security = NULL;
-	kmem_cache_free(sel_inode_cache, isec);
+	call_rcu(&isec->rcu, inode_free_rcu);

Does not clearing ->i_security mean that RCU readers can traverse
this pointer after the invocation of call_rcu()?  If so, this is
problematic.  (If something else already prevents readers from getting
here, no problem.)

This is called when we are about to free the inode. Look at
destroy_inode(). Basically, this is the same as doing:

	call_rcu(&isec->rcu, inode_free_rcu);
	call_rcu(&inode->i_rcu, i_callback);

Where i_callback() does the free of the inode.

If you can access inode->i_security, after a call_rcu, then you can
also access the inode itself that has just been freed.

Yes, technically, having two separate call_rcu(), the first grace
period can end before the second, but everything to remove the inode
from sight has already been set up before that first call_rcu() is
made. That means when the first call_rcu() is executed, the inode
should already be invisible to the readers.

Got it, should be fine then, sorry for the noise.

							Thanx, Paul