Re: [PATCH v4 3/7] soundwire: Add support to lock across bus instances

[Shreyas NC <shreyas.nc@intel.com>]

On Tue, Jun 26, 2018 at 10:34:17AM +0200, Takashi Iwai wrote:
> On Tue, 26 Jun 2018 10:22:01 +0200,
> Shreyas NC wrote:
> > 
> > > > +/**
> > > > + * sdw_acquire_bus_lock: Acquire bus lock for all Master runtime(s)
> > > > + *
> > > > + * @stream: SoundWire stream
> > > > + *
> > > > + * Acquire bus_lock for each of the master runtime(m_rt) part of this
> > > > + * stream to reconfigure the bus.
> > > > + */
> > > > +static void sdw_acquire_bus_lock(struct sdw_stream_runtime *stream)
> > > > +{
> > > > +	struct sdw_master_runtime *m_rt = NULL;
> > > > +	struct sdw_bus *bus = NULL;
> > > > +
> > > > +	/* Iterate for all Master(s) in Master list */
> > > > +	list_for_each_entry(m_rt, &stream->master_list, stream_node) {
> > > > +		bus = m_rt->bus;
> > > > +
> > > > +		mutex_lock(&bus->bus_lock);
> > > > +	}
> > > > +}
> > > 
> > > So it's nested locks?  Then you'd need some more trick to deal with
> > > the lockdep.  I guess you'll get the false-positive deadlock detection
> > > by this code when the mutex lock debug is enabled.
> > > 
> > > Also, is the linked order assured not to lead to a real deadlock?
> > >
> > 
> > Hi Takashi,
> > 
> > Thanks for the review :)
> > 
> > A multi link SoundWire stream consists of a list of Master runtimes and
> > more importantly only one master runtime per SoundWire bus instance.
> > 
> > So, these mutexes are actually different mutex locks(one per bus instance)
> > and are not nested.
> 
> You take a mutex lock inside a mutex lock, so they are nested.
> If they take the very same lock, it's called a "deadlock" instead.
> 

Ok, myy bad, I misunderstood the comment :(

I forgot to add that I did check with mutex debug enabled and lockdep did
not complain though :)

> > In SDW we have a bus instance per Master (link). In multi-link case, a
> > stream may have multiple Masters, thus we need to lock all bus instances
> > before we operate on them.
> > 
> > Now since these are invoked from a stream (pcm ops) they will be always
> > serialized and DPCM ensures we are never racing.
> > 
> > We did add this note here and in Documentation to make it explicit.
> 
> Well, my question is whether the order to take the multiple locks is
> always assured.  You're calling like:
> 
> 	list_for_each_entry(m_rt, &stream->master_list, stream_node)
> 		mutex_lock();
> 
> And it's a linked-list.  If a stream has a link of masters like
> M1->M2->M3 while another stream has a link like M2->M1->M3, it'll lead
> to a deadlock with the concurrent calls above.
> 

These are called from PCM stream ops context and the DPCM holds
lock(fe->card->mutex) which serializes these operations.
So, in the scenario you have mentioned, we would not have
concurrent calls to this function.

> > > > +/**
> > > > + * sdw_release_bus_lock: Release bus lock for all Master runtime(s)
> > > > + *
> > > > + * @stream: SoundWire stream
> > > > + *
> > > > + * Release the previously held bus_lock after reconfiguring the bus.
> > > > + */
> > > > +static void sdw_release_bus_lock(struct sdw_stream_runtime *stream)
> > > > +{
> > > > +	struct sdw_master_runtime *m_rt = NULL;
> > > > +	struct sdw_bus *bus = NULL;
> > > > +
> > > > +	/* Iterate for all Master(s) in Master list */
> > > > +	list_for_each_entry(m_rt, &stream->master_list, stream_node) {
> > > > +		bus = m_rt->bus;
> > > > +		mutex_unlock(&bus->bus_lock);
> > > > +	}
> > > 
> > > ... and this looks bad.  The loop for unlocking should be traversed
> > > reversely.
> > > 
> > 
> > Yes in principle I agree locking should be in reverse, but as explained
> > above in this case, it does not matter much :)
> 
> It does matter when you dealing with the multiple nested mutexes...
> 

Ok

--Shreyas

--