Re: [linux-pm] [GIT PULL] One more power management fix for 2.6.37

[Dominik Brodowski <linux@dominikbrodowski.net>]

On Thu, Nov 04, 2010 at 06:04:05AM +0100, Rafael J. Wysocki wrote:
> On Wednesday, November 03, 2010, Dominik Brodowski wrote:
> > > There's apparently an ordering problem with dpm_list_mtx and
> > > socket->skt_mutex. Lockdep details appended.
> > > 
> > > Dominik, Rafael? What's the proper locking order here, and
> > > how do we fix this?
> > 
> > Thanks for noting this; let's see:
> > 
> > - We add a PCMCIA device holding skt_mutex, therefore we have the ordering
> >   (1) skt_mutex -> (2) dpm_list_mtx
> > 
> > - If we're suspending, dpm_list_mtx is held, but we need to acquire
> >   skt_mutex as we modify some data being protected by skt_mutex
> >   (1) dpm_list_mtx -> (2) skt_mutex
> > 
> > Rafael, any idea on how to solve this? How do other subsystems handle such
> > an issue? Do they call device_add() with no locks held at all?
> 
> They usually do from what I can tell.
> 
> Also only a few of them implement the ->suspend_noirq() callback, which is the
> one executed under dpm_list_mtx.
> 
> What exactly is protected by skt_mutex ?

e.g.
struct pcmcia_socket {
	...
        u_int                   suspended_state;
        int                     resume_status;
	...
}

Furthermore, one has to acquire skt_mutex first before obtaining ops_mutex,
which protects many more fields (and asserts exclusion for some code paths),
see Documentation/pcmcia/locking.txt for details.

Best,
	Dominik