On 07/24/2015 02:21 PM, Steven Rostedt wrote:

On Fri, 24 Jul 2015 21:02:26 +0200 (CEST)
Thomas Gleixner [off-list ref] wrote:

quoted

On Wed, 22 Jul 2015, Steven Rostedt wrote:

quoted

It's still there. Note, that when I pull -rt into stable, I do a
 git quiltimport, which renames the patch set. It's now called:

 0089-slub-delay-ctor-until-the-object-is-requested.patch

quoted

From the 4.0.8-rt6 announce:

- The delayed kmem_cache constructor caused problems. Steven Rostedt
  reported problems versus the signal handling code and Koehrer
  Mathias reported the same. The patch in question has been reverted
  and a patch currently sitting in -mm has been added which provides
  the same functionality (running the constructor with enabled
  interrupts). Patch provided by Thomas Gleixner. 

The stable trees which carry the delay-ctor patches should be updated
to this.

Reverting the patch at least prevents the BUG. Latencywise the slub patch

   mm-slub-move-slab-initialization-into-irq-enabled-re.patch

is the one which wants to be ported back.

OK, thanks.

Just to let people know. I'm going to be traveling next week and when I
get back in August, my top priority is to work on the stable patches.
For 3.18, I'll be looking at the patches that have been emailed to the
list, and I want to fix NO_HZ for 3.10-3.14.

In case it saves you some work, we've had some success getting NO_HZ to
work on 3.14 (x86_64) after back-porting the following commit:

3010279f0fc36f0388872203e63ca49912f648fd x86: Tell irq work about self
IPI support

-Gratian