Re: [PATCH -mm resend] cpuhotplug: introduce try_get_online_cpus() take 3

[Andrew Morton <akpm@linux-foundation.org>]

On Wed, 10 Jun 2009 09:13:58 +0800 Lai Jiangshan <laijs@cn.fujitsu.com> wrote:

> It's for -mm tree.
> 
> It also works for mainline if you apply this at first:
> http://lkml.org/lkml/2009/2/17/58
> 
> Subject: [PATCH -mm] cpuhotplug: introduce try_get_online_cpus() take 3
> 
> get_online_cpus() is a typically coarsely granular lock.
> It's a source of ABBA or ABBCCA... deadlock.
> 
> Thanks to the CPU notifiers, Some subsystem's global lock will
> be required after cpu_hotplug.lock. Subsystem's global lock
> is coarsely granular lock too, thus a lot's of lock in kernel
> should be required after cpu_hotplug.lock(if we need
> cpu_hotplug.lock held too)
> 
> Otherwise it may come to a ABBA deadlock like this:
> 
> thread 1                                      |        thread 2
> _cpu_down()                                   |  Lock a-kernel-lock.
>   cpu_hotplug_begin()                         |
>     mutex_lock(&cpu_hotplug.lock)             |
>   __raw_notifier_call_chain(CPU_DOWN_PREPARE) |  get_online_cpus()
> ------------------------------------------------------------------------
>     Lock a-kernel-lock.(wait thread2)         |    mutex_lock(&cpu_hotplug.lock)
>                                                    (wait thread 1)

uh, OK.

> But CPU online/offline are happened very rarely, get_online_cpus()
> returns success quickly in all probability.
> So it's an asinine behavior that get_online_cpus() is not allowed
> to be required after we had held "a-kernel-lock".
> 
> To dispel the ABBA deadlock, this patch introduces
> try_get_online_cpus(). It returns fail very rarely. It gives the
> caller a chance to select an alternative way to finish works,
> instead of sleeping or deadlock.

I still think we should really avoid having to do this.  trylocks are
nasty things.

Looking at the above, one would think that a correct fix would be to fix
the bug in "thread 2": take the locks in the correct order?  As
try_get_online_cpus() doesn't actually have any callers, it's hard to
take that thought any further.