Skip to contentSkip to conversation

ALSA: HDA: Early Forbid of runtime PM

PATCH In review Harsha Priya 3 participants 5 messages Aug 27, 2020 – Sep 18, 2020 1 message not in the archive Sound

lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 ALSA: HDA: Early Forbid of runtime PM

For certain codecs (like Realtek), pm_runtime_forbid() is invoked in the probe function after build_controls(). In a stress test, its observed occasionally that runtime PM calls are invoked before controls are built. This causes the codec to be runtime suspended before probe completes. Because of this, not all controls are enumerated correctly, and audio does not work until system is rebooted.

This issue being common across all codecs, pm_runtime_forbid() is called when the codec object is created to fix this issue. A codec enables or disables runtime pm in its own probe function.

Multiple stress tests of 2000+ cycles has been done to test the fix.

---
 sound/pci/hda/hda_codec.c | 3 +++
 1 file changed, 3 insertions(+)
sound/pci/hda/hda_codec.c +3 -0
index e96a87f1b611..a356c21edb90 100644
@@ -1000,6 +1000,9 @@ int snd_hda_codec_device_new(struct hda_bus *bus, struct snd_card *card,
 	if (err < 0)
 		goto error;
 
+	/* PM runtime needs to be enabled later after binding codec */
+	pm_runtime_forbid(&codec->core.dev);
+
 	return 0;
 
  error:
Signature · 1 line
2.17.1
#1 lore raw
Quoted 13 lines · Harsha Priya

For certain codecs (like Realtek), pm_runtime_forbid() is invoked in the probe function after build_controls(). In a stress test, its observed occasionally that runtime PM calls are invoked before controls are built. This causes the codec to be runtime suspended before probe completes. Because of this, not all controls are enumerated correctly, and audio does not work until system is rebooted.

This issue being common across all codecs, pm_runtime_forbid() is called when the codec object is created to fix this issue. A codec enables or disables runtime pm in its own probe function.

Quoted 5 lines

Multiple stress tests of 2000+ cycles has been done to test the fix.

Thanks. The only concern is about the influence on the relevant ASoC code, especially hdac_hda.c.

Kai, could you check whether this still works?

Takashi

Quoted 21 lines
---
 sound/pci/hda/hda_codec.c | 3 +++
 1 file changed, 3 insertions(+)
sound/pci/hda/hda_codec.c +3 -0
index e96a87f1b611..a356c21edb90 100644
@@ -1000,6 +1000,9 @@ int snd_hda_codec_device_new(struct hda_bus *bus, struct snd_card *card,
 	if (err < 0)
 		goto error;
 
+	/* PM runtime needs to be enabled later after binding codec */
+	pm_runtime_forbid(&codec->core.dev);
+
 	return 0;
 
  error:
Signature · 1 line
2.17.1
#2 lore raw

Hey,

Quoted 9 lines · Takashi Iwai
Quoted 3 lines · Harsha Priya

This issue being common across all codecs, pm_runtime_forbid() is called when the codec object is created to fix this issue. A codec enables or disables runtime pm in its own probe function.

Thanks. The only concern is about the influence on the relevant ASoC code, especially hdac_hda.c.

Kai, could you check whether this still works?

I believe Harsha is testing mostly with ASoC, so hdac_hda.c should be covered.

I did queue a SOF CI job for this v2 patch and I'm seeing some failures in module load/unload test that might be related and need checking before we merge:

https://sof-ci.01.org/linuxpr/PR2403/build4400/devicetest/

The actual runtime-PM tests in our test set pass, but module load/unload has failures on some platforms. I'll follow-up when I have a better understanding what goes wrong.

Br, Kai

A message that is not in the archive (sent off-list or not yet received): BN6PR1101MB2275262388388F71D62390299C520@BN6PR1101MB2275.namprd11.prod.outlook.com

#4 lore raw
Quoted 25 lines · Emmanuel
Quoted 25 lines

Subject: Re: [PATCH v2] ALSA: HDA: Early Forbid of runtime PM

Hey,

Quoted 9 lines · Takashi Iwai
Quoted 3 lines · Harsha Priya

This issue being common across all codecs, pm_runtime_forbid() is called when the codec object is created to fix this issue. A codec enables or disables runtime pm in its own probe function.

Thanks. The only concern is about the influence on the relevant ASoC code, especially hdac_hda.c.

Kai, could you check whether this still works?

I believe Harsha is testing mostly with ASoC, so hdac_hda.c should be covered.

I did queue a SOF CI job for this v2 patch and I'm seeing some failures in module load/unload test that might be related and need checking before we merge:

https://sof-ci.01.org/linuxpr/PR2403/build4400/devicetest/

Quoted 8 lines
Quoted 4 lines

The actual runtime-PM tests in our test set pass, but module load/unload has failures >on some platforms. I'll follow-up when I have a better understanding what goes wrong.

Br, Kai

I am just trying to think ahead to get something that may work. We can add an if check around the pm_runtime_forbid like below If (!codec->auto_runtime_pm)

	Pm_runtime_forbid(&codec->core.dev)

Do you think that will work??

snd_hda_codec_device_new() is called before setting the flag, so it's no use for adding the conditional.

thanks,

Takashi

#5 lore raw
In reply to Kai Vehmanen

Hi,

Quoted 11 lines · Kai Vehmanen
Quoted 4 lines · Takashi Iwai

Thanks. The only concern is about the influence on the relevant ASoC code, especially hdac_hda.c.

Kai, could you check whether this still works?

I did queue a SOF CI job for this v2 patch and I'm seeing some failures in module load/unload test that might be related and need checking before we merge:

https://sof-ci.01.org/linuxpr/PR2403/build4400/devicetest/

ok this took a bit longer than expected, but we've been debugging this with Emmanuel and found the rootcause.

So if we take this patch in and forbid runtime PM in device_new, it then becomes mandatory for all controllers to call set_default_power_save(). In SOF, this is only added recently to some machine drivers (notably the widely used generic HDA machine driver), which explains why tests pass on many targets. However, on platforms that use other machine drivers, set_default_power_save()/snd_hda_set_power_save() is not called, and device will never hit suspend (and this fails the tests).

To further complicate the matter, recently merged "ASoC: Intel:
skl_hda_dsp_generic: Fix NULLptr dereference in autosuspend delay" will cause tests to fail on more platforms as HDMI/DP codec's powersave settings are not set.

The original patch is ok, but to merge it without a lot of regressions, we need to rework how powersave initial settings are done in SOF and other ASoC controllers using HDAC. I do think this is the right solution.

Br, Kai