Re: 3.0.23-rt39 BUG: scheduling while atomic, _cpu_down()

On 03/09/12 18:03, Steven Rostedt wrote:

On Fri, 2012-03-09 at 17:51 -0800, Frank Rowand wrote:

quoted

Hi Thomas,
(and Steve, since you've been mucking about in the hotplug code,)

Yet another bit of fallout related to hotplug.

This appears to affect 3.0-rt1, 3.0.23-rt29, and 3.2.9-rt17.
(Thus probably all 3.x versions, but I haven't been masochistic
enough to check.)

If not defined CONFIG_PREEMPT_RT_FULL then

So you're saying that there's a bug with PREEMPT_RT off?

Yes, that is the problem.  But it is due to the RT patches,
not present in mainline.

quoted

  - hotplug_lock() is a mutex_lock()
  - cpu_hotplug_begin() calls hotplug_lock()
  - _cpu_down() calls cpu_hotplug_begin() with preemption
    disabled (disabled by the migrate_disable())

The call to cpu_hotplug_begin() is not in a preempt disabled
region until hotplug-use-migrate-disable.patch.  This patch
moves the preempt_enable() (aka migrate_enable()) from just
before calling cpu_hotplug_begin() to after "out_cancel:".

The BUG can be triggered by:

   echo 0 > /sys/devices/system/cpu/cpu1/online 

Note, there's a lot more wrong with hotplug in rt than this. The hotplug
code really needs a total rewrite for RT to work with it. It probably
should have a total rewrite in mainline anyway.

Yep, I've been watching all the hotplug related conversations with interest...

-Frank