On Tue, 2014-06-10 at 13:50 -0400, Benjamin LaHaise wrote: 

On Tue, Jun 10, 2014 at 05:47:28AM +0200, Mike Galbraith wrote:

quoted

On Mon, 2014-06-09 at 10:08 +0800, Lai Jiangshan wrote: 

quoted

Hi, rt-people

I don't think it is the correct direction.
Softirq (including local_bh_disable()) in RT kernel should be preemptible.

How about the below then?

I was sorely tempted to post a tiny variant that dropped taking ctx_lock
in free_ioctx_users() entirely, as someone diddling with no reference
didn't make sense.  Cc Ben, he would know.

That should be okay...  Let's ask Kent to chime in on whether this looks 
safe to him on the percpu ref front as well, since he's the one who wrote 
this code.

Looking at the gizzard of our in-tree user of kiocb_set_cancel_fn()
(gadget), cancel() leads to dequeue() methods, which take other sleeping
locks, so tiniest variant is not an option, patchlet stands.

-Mike