I think the only solution is to follow the example of the PCM case, where the type of lock depends on the FE types, with the assumption that there are no mixed atomic/non-atomic FE configurations.
Yes, and I guess we can simply replace those all card->dpcm_lock with FE's stream lock. It'll solve the atomicity problem, too, and the FE stream lock can be applied outside the loop of dpcm_be_disconnect() gracefully.
And, this should solve the race with dpcm_be_dai_trigger() as well, because it's called from dpcm_fe_dai_trigger() that is called already inside the FE's stream lock held by PCM core. A PoC is something like below. (I replaced the superfluous *_irqsave with *_irq there)
No I don't think so. The code starts from an FE and loops for all the BEs connected to that FE, but we want to serialize at the BE level! we really need a dpcm lock at the card level, not the FE/stream level.
I am testing a quick set of changes at https://github.com/thesofproject/linux/pull/3201, where I also replaced irqsave by irq btw.