Re: [PATCH v2 6/8] ASoC: cs42l42: Add Soundwire support

[Richard Fitzgerald]

On 20/01/2023 19:55, Pierre-Louis Bossart wrote:
> On 1/19/23 09:35, Richard Fitzgerald wrote:
> > On 19/1/23 14:48, Pierre-Louis Bossart wrote:
> > > > > > +static int cs42l42_sdw_dai_startup(struct snd_pcm_substream
> > > > > > *substream,
> > > > > > +                   struct snd_soc_dai *dai)
> > > > > > +{
> > > > > > +    struct cs42l42_private *cs42l42 =
> > > > > > snd_soc_component_get_drvdata(dai->component);
> > > > > > +
> > > > > > +    if (!cs42l42->init_done)
> > > > > > +        return -ENODEV;
> > > > >
> > > > > Can this happen? IIRC the ASoC framework would use
> > > > > pm_runtime_resume_and_get() before .startup, which would guarantee that
> > > > > the device is initialized, no?
> > > >
> > > > Yes, this can happen. Because of the way that the SoundWire enumeration
> > > > was implemented in the core code, it isn't a probe event so we cannot
> > > > call snd_soc_register_component() on enumeration because -EPROBE_DEFER
> > > > wouldn't be handled. So the snd_soc_register_component() must be called
> > > > from probe(). This leaves a limbo situation where we've registered the
> > > > driver but in fact don't yet have any hardware. ALSA/ASoC doesn't know
> > > > that we've registered before we are functional so they are happy to
> > > > go ahead and try to use the soundcard. If for some reason the hardware
> > > > failed to enumerate we can get here without having enumerated.
> > >
> > > Humm, yes, but you've also made the regmap cache-only, so is there
> > > really a problem?
> >
> > It's true that normally we go past these stages in cache-only, but that
> > is because normally (non-Soundwire) we already initialized the hardware
> > to good state during probe().
> > If we just carry on when it hasn't enumerated and we haven't initialized
> > it yet, who knows what will happen if it enumerates some time later.
> >
> > We could just ignore it and see if anyone has a problem but for the sake
> > of a couple of lines of code I feel like I'd rather check for it.
> >
> > > FWIW I don't see a startup callback in any other codec driver. It may be
> > > wrong but it's also a sign that this isn't a problem we've seen so far
> > > on existing Intel-based platforms.
> >
> > It's nicer to do the check in startup() because then the application
> > open() will fail cleanly. We could delay until prepare - which is the
> > point we really need the hardware to be accessible - and hope the
> > hardware enumerated and initialized by that time. But that's not so
> > nice from the app point of view.
>
> Another way to avoid problems is to rely on the codec component .probe
> to check if the SoundWire device is initialized before registering a card.
>
> I just tried with a system where the ACPI info exposes a codec which is
> not connected, it fails nicely. That avoids the pitfalls of creating a
> card which isn't functional since all dependencies are not met.
>
> [   64.616530] snd_soc_sof_sdw:mc_probe: sof_sdw sof_sdw: Entry
> [   64.616549] snd_soc_sof_sdw:log_quirks: sof_sdw sof_sdw: quirk
> SOF_SDW_PCH_DMIC enabled
> [   64.616559] snd_soc_sof_sdw:sof_card_dai_links_create: sof_sdw
> sof_sdw: sdw 2, ssp 0, dmic 2, hdmi 0
> [   64.616587] snd_soc_sof_sdw:init_dai_link: sof_sdw sof_sdw: create
> dai link SDW0-Playback, id 0
> [   64.616600] snd_soc_sof_sdw:init_dai_link: sof_sdw sof_sdw: create
> dai link SDW0-Capture, id 1
> [   64.616607] snd_soc_sof_sdw:init_dai_link: sof_sdw sof_sdw: create
> dai link dmic01, id 2
> [   64.616614] snd_soc_sof_sdw:init_dai_link: sof_sdw sof_sdw: create
> dai link dmic16k, id 3
> [   69.757115] rt5682 sdw:0:025d:5682:00: Initialization not complete,
> timed out
> [   69.757128] rt5682 sdw:0:025d:5682:00: ASoC: error at
> snd_soc_component_probe on sdw:0:025d:5682:00: -110
> [   69.757224] sof_sdw sof_sdw: ASoC: failed to instantiate card -110
> [   69.757734] sof_sdw sof_sdw: snd_soc_register_card failed -110
>
> see
> https://elixir.bootlin.com/linux/latest/source/sound/soc/codecs/rt5682.c#L2927
>
> I think this is compatible with the device model and bind/unbind, but it
> could be improved with the removal of the wait if we had a way to return
> -EPROBEDEFER, and have a mechanism to force the deferred probe work to
> be triggered when a device actually shows up. It's a generic problem
> that the probe cannot always be a synchronous function but may complete
> 'later'.

I see what you've done in your patch, but I had already experimented
with this idea and found that the wait_for_completion() can deadlock the
Soundwire core.