ASoC: SOF: Fix suspend while paused corner case #5054

ujfalusi · 2024-06-11T07:53:55Z

Hi,

Summary: this PR fixes the suspend while paused corner case for both IPC3 and IPC4 (was locking up Intel platforms)

The suspend while we have paused stream has been broken likely for several years (?). I cannot pin-point the exact commit, but there are several underlying issues contributing to the broken corner case:

wrong order of stopping the link DMA on the paused stream (set_hw_params_upon_resume did that)
IPC3 will not stop the host DMA on suspend if there were a paused stream.
IPC4 pipelines were not reset because the pipelines were in paused state, so unbind/free fails and the whole stack got confused.

This PR replaces #5040 with a more complicated and in SOF fix:

replace the set_hw_params_upon_resume() with suspend_early() and call it before we tear down the remaining pipelines:
the stream is paused, so we need to follow the same sequence as we have when we suspend while we have active audio
reset the started/paused count when we suspend the paused pipelines
stop the host DMA with IPC3 if we suspend while we have paused stream.

With this PR the system is no longer locks up and the paused aplay/arecord can be PAUSE_RELEASEd after the system resume.
Since we don't support the RESUME trigger, the stream is going to be re-started (like with active audio), but things work now correctly.

Fixes: #5035

…in dai_suspend The TRIGGER_SUSPEND sequence is: pre_trigger() trigger() post_trigger() hda_link_dma_cleanup() In hda_dai_suspend() we only call the post_trigger() and the link cleanup, do it in the same order as it would have been during a trigger. Signed-off-by: Peter Ujfalusi <[email protected]>

…/SUSPEND The same rule applies to the paused_count as for the started_count: on STOP/SUSPEND both needs to be reset to 0. Signed-off-by: Peter Ujfalusi <[email protected]>

The new suspend_early() can be used by platform code to prepare for the system suspend and execute tasks needed to be ready to handle the suspend call. Signed-off-by: Peter Ujfalusi <[email protected]>

…set_hw_params_upon_resume The function 'hda_dsp_set_hw_params_upon_resume' has lost it's purpose as we are not remotely doing anything resembling what the function implies. The only thing we do is to stop any paused streams (paused streams will not receive suspend trigger and would block the suspend in hardware). The call sequence in pm.c is also wrong for the set_hw_params_upon_resume() since we would need to make sure that the link DMA is stopped _before_ we tear down the pipelines belonging to the paused stream. Signed-off-by: Peter Ujfalusi <[email protected]>

The set_hw_params_upon_resume() was used by Intel platforms. Over time it's purpose got changed and finally the Intel code is converted to use the suspend_early() callback instead. Signed-off-by: Peter Ujfalusi <[email protected]>

…spend Introduce a new flag to mark a stream during system suspend and if it is set, send the STOP trigger for the platform driver to stop the DMA. This is needed in case the platform_stop_during_hw_free is not true (IPC3) and the suspend happens when a stream is in paused state. Signed-off-by: Peter Ujfalusi <[email protected]>

…on system suspend The suspending flag will force the generic code to stop the platform DMA of the paused stream. Signed-off-by: Peter Ujfalusi <[email protected]>

…ams correctly During system suspend while we have paused streams we need to make sure that the pipelines are properly reset. Since the stream is in paused state, the pipelines are also in paused state, which is correct for the RESET transition. Reset the started/paused counters to indicate that the pipeline will need to be initialized after resume and PAUSE_RELEASE. Signed-off-by: Peter Ujfalusi <[email protected]>

plbossart · 2024-06-11T08:38:58Z

sound/soc/sof/intel/hda-dai-ops.c

 		 */
 		swidget->spipe->started_count = 0;
+		swidget->spipe->paused_count = 0;


what would be the behavior on resume then? The stream would restart? How can we keep the stream paused on resume?

we cannot keep the streams paused, we don't support RESUME trigger, so all streams would be restarted anyways, paused or not paused.
In practice: on resume the audio will remain paused (will not run), but when you PAUSE_RELEASE it, then we will restart it. We need to re-initialize everything to get working audio.

how will the audio remain paused if the pause_count remains zero? I don't understand how this counter is used.

This counter is used internally to ipc4 (to track the individual pipeline states) , ALSA keeps track of the PCM state. On suspend we need to stop everything, so we can put the DSP to off.
So, when we suspend all pipelines will be reset.

@ujfalusi this is too confusing. We increment the paused_count when we pause the stream. So lets assume this sequence:

Start the stream : started_count = 1

Pause the stream : paused_count = 1

Suspend the system: The stream doesnt get the suspend trigger, so the started/paused_counts remain at 1

hda_dai_suspend() gets invoked so both counts get reset to 0.

Resume the system

Pause_release the stream: this will decrement the paused_count to -1?
Am I understanding this correctly?

plbossart · 2024-06-11T08:40:41Z

sound/soc/sof/pm.c

@@ -210,6 +210,16 @@ static int sof_suspend(struct device *dev, bool runtime_suspend)
 	if (runtime_suspend && !sof_ops(sdev)->runtime_suspend)
 		return 0;

+	/* Prepare the DSP for system suspend */


What would be the difference with a prepare callback, as we do for SoundWire to go top down and make sure all parent/children are properly pm_runtime resumed before the system suspend.

the prepare callback happens before ALSA sends the suspend trigger (it is also sent in prepare callback), we use pm callbacks, not ASoC component callbacks..

i was talking about the pm .prepare callback, which is called before suspend. I was not referring to the prepare as a followup to hw_params - naming is confusing.

Yes, I have tried that exactly as first thing, it does not work, it comes before the trigger suspend and breaks suspend with active audio.

fair enough. There should still be a more detailed description of what this callback should implement vs, what needs to remain in the suspend callback proper.
Last time we introduced the probe_early, it was rather straightforward as it included all the features that could not be done in a workqueue. A suspend_early isn't very clear.

plbossart · 2024-06-11T09:20:44Z

sound/soc/sof/ipc4-pcm.c

+					 pipeline->state, spipe->started_count, spipe->paused_count);
+
+			spipe->started_count = 0;
+			spipe->paused_count = 0;


I am still not clear on when this paused_count would be restored when resuming back to a PAUSED state.

Right, so we have moved the pipelines to PAUSED state, it's started_count and paused_count is not 0, then we suspend the system.
We need to move the pipelines to RESET state to be able to UNBIND and then DELETE them. This is not optional, this is a must.
When the DSP is powered down we will not going to be able to PAUSE_RELEASE (not that we can do that correctly by PAUSE PUSH/RELEASE alone), so we need to reset these counters to 0, nothing will be running or being paused when the system resumes in firmware.

As for the PAUSE_RELEASE, ALSA will notice that we don't support RESUME, so it will re-start the stream.

I think this has been semi broken always. We had errors on suspend while paused, then we had errors on pause release and more errors when user space gave up, then we rpm suspend and reset the counters and on next try things would work.

My question was "when is paused_count increased again on resume" ?

If we receive PAUSE_PUSH. After system resume ALSA will not touch a paused PCM, from ALSA pow PAUSED == SUSPENDED and a PAUSED stream can only be changed via PAUSE_RELEASE (even a stop needs intermediate release).

Since we don't support RESUME (from suspend), when user sends the trigger PAUSE_RELEASE then ALSA will return with error and the stream is restarted.

The point is: when we resume from suspend we don't have any active pipelines in DSP, so no pipeline can have a started/paused counter other then 0.

so the pipeline is paused but not paused, depending on which state machine you're looking at, and you're relying on userspace to deal with errors?
Humm...this seems a bit weird if I am honest, we end-up with a non-symmetrical state after resume.

The PCM stream (as in ALSA terms) is paused but the state of the pipelines (as in IPC4 terms) are stopped (stopped and paused count == 0).
If the PCM is paused and the system suspends we will not get any triggers to react, so we end up with stale counters. We don't have PAUSED->SUSPENDED state transition handler (nether ALSA), but:
RUNNING->SUSPENDED is started_count-- (paused_count is 0) and if both 0 then we send the IPCs and stuff
PAUSED->RUNNING is paused_count-- (started_count is > 0)
so
PAUSE->SUSPENDED is rightfully started_count--, paused_count--, unless they were already 0, hrm, what if we have multiple paused streams with common pipelines? The first one will clear the shared portion, the second will clear it's own.

Right, multiple paused stream with common pipelines indeed is broken. (hw:0,0 and hw:0,31 on HDA machine)

But to be fair without this PR the whole system just locks up, so a bit better than what we have atm.

So, back again to the same page: #5040 is fixing all of these issues without any side effect.
single paused stream
two paused streams with shared pipeline section

plbossart · 2024-06-11T09:21:45Z

sound/soc/sof/intel/hda-dai.c

+						   hext_stream,
+						   cpu_dai);
+			if (ret < 0)
+				return ret;


I can't recall if this was an editing mistake or intentional. @ranj063 would you happen to remember this?

@plbossart I dont think the order really matters in this case. All we do in the post trigger op for suspend is clear the count for pipeline.

sound/soc/sof/intel/hda-common-ops.c

ujfalusi · 2024-06-12T11:51:03Z

Closing, replaced by: #5058

ujfalusi added 8 commits June 11, 2024 10:29

ASoC: SOF: Intel: hda-dai-ops: ipc4: Reset the paused counter on STOP…

9a7c7ab

…/SUSPEND The same rule applies to the paused_count as for the started_count: on STOP/SUSPEND both needs to be reset to 0. Signed-off-by: Peter Ujfalusi <[email protected]>

ASoC: SOF: Add new optional suspend_early() dsp_ops callback function

365cd78

The new suspend_early() can be used by platform code to prepare for the system suspend and execute tasks needed to be ready to handle the suspend call. Signed-off-by: Peter Ujfalusi <[email protected]>

ASoC: SOF: ipc3-topology: Set the suspending flag for paused streams …

c3aa4bd

…on system suspend The suspending flag will force the generic code to stop the platform DMA of the paused stream. Signed-off-by: Peter Ujfalusi <[email protected]>

ujfalusi requested a review from ssavati June 11, 2024 07:53

ujfalusi requested review from bardliao, ranj063, lgirdwood, plbossart, kv2019i and dbaluta as code owners June 11, 2024 07:53

plbossart reviewed Jun 11, 2024

View reviewed changes

ujfalusi mentioned this pull request Jun 12, 2024

ASoC: SOF: sof-pcm/pm: Stop paused streams before the system suspend #5058

Open

ujfalusi closed this Jun 12, 2024

kv2019i mentioned this pull request Jun 26, 2024

ASoC: SOF: Intel: hda: fix null deref on system suspend entry #5085

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

ASoC: SOF: Fix suspend while paused corner case #5054

ASoC: SOF: Fix suspend while paused corner case #5054

ujfalusi commented Jun 11, 2024 •

edited

Loading

plbossart Jun 11, 2024

ujfalusi Jun 11, 2024

plbossart Jun 11, 2024

ujfalusi Jun 11, 2024

ranj063 Jun 11, 2024

plbossart Jun 11, 2024

ujfalusi Jun 11, 2024

plbossart Jun 11, 2024

ujfalusi Jun 11, 2024

plbossart Jun 11, 2024

plbossart Jun 11, 2024

ujfalusi Jun 11, 2024

ujfalusi Jun 11, 2024

plbossart Jun 11, 2024

ujfalusi Jun 11, 2024

plbossart Jun 11, 2024

ujfalusi Jun 11, 2024

ujfalusi Jun 11, 2024

ujfalusi Jun 11, 2024

ujfalusi Jun 11, 2024 •

edited

Loading

plbossart Jun 11, 2024

ranj063 Jun 11, 2024

ujfalusi commented Jun 12, 2024

ASoC: SOF: Fix suspend while paused corner case #5054

ASoC: SOF: Fix suspend while paused corner case #5054

Conversation

ujfalusi commented Jun 11, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

ujfalusi Jun 11, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

ujfalusi commented Jun 12, 2024

ujfalusi commented Jun 11, 2024 •

edited

Loading

ujfalusi Jun 11, 2024 •

edited

Loading