[PATCH] device: separate all subsys mutexes (was: Re: [BUG] potential deadlock led by cpu_hotplug lock (memcg involved))

[Michal Hocko <mhocko@suse.cz>]

[CCing Greg and Kay]
On Tue 12-03-13 12:07:50, Michal Hocko wrote:
> [Let's CC Ingo and Peter]
> 
> On Tue 12-03-13 11:15:55, Michal Hocko wrote:
> > On Tue 12-03-13 14:36:46, Li Zefan wrote:
> > > Seems a new bug in 3.9 kernel?
> > > 
> > > 
> > > [  207.271924] ======================================================
> > > [  207.271932] [ INFO: possible circular locking dependency detected ]
> > > [  207.271942] 3.9.0-rc1-0.7-default+ #34 Not tainted
> > > [  207.271948] -------------------------------------------------------
> > 
> > 1) load_module -> subsys_interface_register -> mc_deveice_add (*) -> subsys->p->mutex -> link_path_walk -> lookup_slow -> i_mutex
> > 2) sys_write -> _cpu_down -> cpu_hotplug_begin -> cpu_hotplug.lock -> mce_cpu_callback -> mce_device_remove(**) -> device_unregister -> bus_remove_device -> subsys mutex
> > 3) vfs_readdir -> i_mutex -> filldir64 -> might_fault -> might_lock_read(mmap_sem) -> page_fault -> mmap_sem -> drain_all_stock -> cpu_hotplug.lock
> > 
> > 1) takes cpu_subsys subsys (*) but 2) takes mce_device subsys (**) so
> > the deadlock is not possible AFAICS.

Thanks to Jiri Kosina, who pointed out that the root cause is
bus_register which uses a static key and both mce and cpu subsys are
registered by subsys_system_register so they use the same key.  Maybe
something like the following (compile tested with both LOCKDEP on/off)
should work:
---