Re: [PATCH v2] ALSA: hda/tegra: enable clock during probe

[Thierry Reding <thierry.reding@gmail.com>]

[-- Attachment #1: Type: text/plain, Size: 4723 bytes --]

On Wed, Jan 30, 2019 at 05:40:42PM +0100, Takashi Iwai wrote:
> On Wed, 30 Jan 2019 13:45:49 +0100,
> Jon Hunter wrote:
> > 
> > 
> > On 25/01/2019 11:06, Sameer Pujar wrote:
> > > If CONFIG_PM is disabled or runtime PM calls are forbidden, the clocks
> > > will not be ON. This could cause issue during probe, where hda init
> > > setup is done. This patch enables clocks unconditionally during probe.
> > > 
> > > Along with above, follwoing changes are done.
> > >   * enable runtime PM before exiting from probe work. This helps to avoid
> > >     usage of pm_runtime_get_sync/pm_runtime_put() in probe work.
> > >   * hda_tegra_disable_clocks() is moved out of CONFIG_PM_SLEEP check.
> > >   * runtime PM callbacks moved out of CONFIG_PM check
> > > 
> > > Signed-off-by: Sameer Pujar <spujar@nvidia.com>
> > > Reviewed-by: Ravindra Lokhande <rlokhande@nvidia.com>
> > > Reviewed-by: Jon Hunter <jonathanh@nvidia.com>
> > > ---
> > >  sound/pci/hda/hda_tegra.c | 26 +++++++++++++++++---------
> > >  1 file changed, 17 insertions(+), 9 deletions(-)
> > > 
> > > diff --git a/sound/pci/hda/hda_tegra.c b/sound/pci/hda/hda_tegra.c
> > > index c8d18dc..ba6175f 100644
> > > --- a/sound/pci/hda/hda_tegra.c
> > > +++ b/sound/pci/hda/hda_tegra.c
> > > @@ -219,7 +219,6 @@ static int hda_tegra_enable_clocks(struct hda_tegra *data)
> > >  	return rc;
> > >  }
> > >  
> > > -#ifdef CONFIG_PM_SLEEP
> > >  static void hda_tegra_disable_clocks(struct hda_tegra *data)
> > >  {
> > >  	clk_disable_unprepare(data->hda2hdmi_clk);
> > > @@ -227,6 +226,7 @@ static void hda_tegra_disable_clocks(struct hda_tegra *data)
> > >  	clk_disable_unprepare(data->hda_clk);
> > >  }
> > >  
> > > +#ifdef CONFIG_PM_SLEEP
> > >  /*
> > >   * power management
> > >   */
> > > @@ -257,7 +257,6 @@ static int hda_tegra_resume(struct device *dev)
> > >  }
> > >  #endif /* CONFIG_PM_SLEEP */
> > >  
> > > -#ifdef CONFIG_PM
> > >  static int hda_tegra_runtime_suspend(struct device *dev)
> > >  {
> > >  	struct snd_card *card = dev_get_drvdata(dev);
> > > @@ -283,7 +282,7 @@ static int hda_tegra_runtime_resume(struct device *dev)
> > >  	int rc;
> > >  
> > >  	rc = hda_tegra_enable_clocks(hda);
> > > -	if (rc != 0)
> > > +	if (rc)
> > >  		return rc;
> > >  	if (chip && chip->running) {
> > >  		hda_tegra_init(hda);
> > > @@ -292,7 +291,6 @@ static int hda_tegra_runtime_resume(struct device *dev)
> > >  
> > >  	return 0;
> > >  }
> > > -#endif /* CONFIG_PM */
> > >  
> > >  static const struct dev_pm_ops hda_tegra_pm = {
> > >  	SET_SYSTEM_SLEEP_PM_OPS(hda_tegra_suspend, hda_tegra_resume)
> > > @@ -551,9 +549,9 @@ static int hda_tegra_probe(struct platform_device *pdev)
> > >  
> > >  	dev_set_drvdata(&pdev->dev, card);
> > >  
> > > -	pm_runtime_enable(hda->dev);
> > > -	if (!azx_has_pm_runtime(chip))
> > > -		pm_runtime_forbid(hda->dev);
> > > +	err = hda_tegra_enable_clocks(hda);
> > > +	if (err)
> > > +		goto out_free;
> > 
> > We also need to think about power-domains here. Enabling the clocks
> > might not be enough as the appropriate power-domain needs to be enabled.
> > For 64-bit Tegra runtime-pm will handle the power-domains (assuming they
> > are populated in device-tree). So I still think it is better we call
> > pm_runtime_get_sync() at some point rather than just replying on
> > enabling the clocks.
> 
> If I understand correctly the code, the pm domain is already activated
> at calling driver's probe callback.

As far as I can tell, the domain will also be powered off again after
probe finished, unless the device grabs a runtime PM reference. This is
what happens via the dev->pm_domain->sync() call after successful probe
of a driver.

It seems to me like it's not a very well defined case what to do when a
device needs to be powered up but runtime PM is not enabled.

Adding Rafael and linux-pm, maybe they can provide some guidance on what
to do in these situations.

To summarize, what we're debating here is how to handle powering up a
device if the pm_runtime infrastructure doesn't take care of it. Jon's
proposal here was, and we use this elsewhere, to do something like this:

	pm_runtime_enable(dev);
	if (!pm_runtime_enabled(dev)) {
		err = foo_runtime_resume(dev);
		if (err < 0)
			goto fail;
	}

So basically when runtime PM is not available, we explicitly "resume"
the device to power it up.

It seems to me like that's a fairly common problem, so I'm wondering if
there's something that the runtime PM core could do to help with this.
Or perhaps there's already a way to achieve this that we're all
overlooking?

Rafael, any suggestions?

Thanks,
Thierry

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 833 bytes --]