>> Basically we need to protect two things:
>> - The BE links
>> - The concurrent accesses to BEs
>> The former belongs to each FE that holds the links, and the FE stream
>> lock would cover.  The latter is rather a per-BE business.
>>
>> An oft-seen risk of multiple locks is deadlocking, but in this case,
>> as long as we keep the lock order FE->BE, nothing wrong can happen.
> 
> famous last words  "nothing wrong can happen." :-)
> 
> I already added a helper to do this FE lock, I can easily replace the
> implementation to remove the spin_lock and use the FE PCM lock.
> we might even add the lock in the definition of for_each_dpcm_be() to
> avoid misses.
> 
> Let me try this out today, thanks for the suggestions.

well, it's not successful at all...

When I replace the existing dpcm_lock with the FE PCM lock as you
suggested, without any additional changes, speaker-test produces valid
audio on the endpoints, but if I try a Ctrl-C or limit the number of
loops with the '-l' option, I hear an endless loop on the same buffer
and I have to power cycle my test device to stop the sound.

See 2 patches attached, the first patch with the introduction of the
helper works fine, the second with the use of the FE PCM lock doesn't.
In hindsight I am glad I worked on minimal patches, one after the other,
to identify problems.

And when I add the BE lock, then nothing happens. Device stuck and no
audio...

There must be something we're missing related to the locking...

My work version is at
https://github.com/plbossart/sound/tree/fix/dpcm-lock5 if anyone wants
to take a look.