[sheepdog] [PATCH v2 1/3] sheep: make {register, unregister}_event thread safe

[Hitoshi Mitake]

At Tue, 17 Dec 2013 15:24:01 +0800,
Liu Yuan wrote:
> 
> On Tue, Dec 17, 2013 at 03:11:30PM +0800, Liu Yuan wrote:
> > On Tue, Dec 17, 2013 at 03:58:07PM +0900, Hitoshi Mitake wrote:
> > > At Tue, 17 Dec 2013 14:50:19 +0800,
> > > Liu Yuan wrote:
> > > > 
> > > > On Tue, Dec 17, 2013 at 03:42:29PM +0900, Hitoshi Mitake wrote:
> > > > > At Tue, 17 Dec 2013 14:31:56 +0800,
> > > > > Liu Yuan wrote:
> > > > > > 
> > > > > > This allow us to call even handling functions in worker thread
> > > > > > 
> > > > > > Signed-off-by: Liu Yuan <namei.unix at gmail.com>
> > > > > > ---
> > > > > 
> > > > > I think this change is dangerous. This permits worker threads to
> > > > > unregister events even if these events are processed in the main
> > > > > thread. Making a new work queue and delegate it to register/unregister
> > > > > events would be safer.
> > > > 
> > > > This scheme is not pratical for async request.
> > > > 
> > > > This is just internal API, which are supported to be called by programmers
> > > > and check it is correct.
> > > 
> > > The checking will cost us lots of time. The problem is a race
> > > condition caused by multiple threads.
> > 
> > Why we would face this kind of problem? we can make sure at any time, there will
> > be a single thread manipulate it, be it worker or main thread.
> > 
> > I think we are talking about different issues, you are supposed that events
> > handling will be generically thread-safe for multiple threads. But this wouldn't
> > happen I think. Instead, assumption 'one event will be only be manipulated by
> > single entity (thus no multiple threads case)' will hold true in the long run.
> 
> I want to clarify that I'm making register/ungister thread safe, not 'event'
> itself thread safe, just allow us to call register/ungister in the worker thread
> besides main thread and we oursevles should make sure the calling of register
> and unregister are safe and correct.

I understand it. But I think "make sure the calling of register and unregister
are safe and correct" is really hard without automated methods. The new
register_event() and unregister_event() seem not to have the above limitation.

Thanks,
Hitoshi