Re: [PATCH] ASoC: codecs: lpass-macro: release macro and dcodec clocks in runtime suspend

[Val Packett <val@packett.cool>]

On 10/23/26 2:32 PM, Ramón Jesús Miranda wrote:
> The LPASS rx, tx, wsa, and va macro codec drivers prepare and enable their
> "macro" and "dcodec" clocks (which correspond to q6prm LPASS_HW_MACRO and
> LPASS_HW_DCODEC clock votes to the ADSP) in probe() and only release them
> in remove(). Their runtime PM callbacks only toggle mclk, npl, and fsgen.
>
> On Qualcomm X1E80100 systems where the ADSP is booted by Linux (such as
> running at EL2 on the Microsoft Surface Pro 11), keeping the LPASS_HW_MACRO
> and LPASS_HW_DCODEC votes active while audio is idle prevents the ADSP from
> entering low-power sleep during s2idle, which in turn blocks the SoC from
> reaching CX collapse (CXSD) and AOSS sleep (AOSD) (both counters remain at
> 0). Also, calling clk_disable_unprepare() unconditionally in remove() on
> clocks already disabled by runtime_suspend() causes clock enable count
> underflows if the driver is unloaded while runtime-suspended.
>
> Disable "dcodec" and "macro" in runtime_suspend(), re-enable them in
> runtime_resume(), and guard the clock teardown in remove() with
> !pm_runtime_status_suspended(). With this change, ADSP, CXSD, and AOSD
> reach >99.9% residency during s2idle (~0.20 W system sleep power on a
> Microsoft Surface Pro 11 OLED).
>
> Assisted-by: Claude:claude-opus-5-5
> Assisted-by: Antigravity:gemini-2.5-pro
Advertisements for particular slop-generators are discouraged since 
816d9992d9ed ("coding-assistants: simplify attribution") btw, and the 
tag is no longer required since 1e035297ffda ("docs: remove Assisted-by 
references").
> [..]
> --- a/sound/soc/codecs/lpass-tx-macro.c
> +++ b/sound/soc/codecs/lpass-tx-macro.c
> @@ -2390,11 +2390,13 @@ static void tx_macro_remove(struct platform_device *pdev)
>   {
>   	struct tx_macro *tx = dev_get_drvdata(&pdev->dev);
>   
> -	clk_disable_unprepare(tx->macro);
> -	clk_disable_unprepare(tx->dcodec);
> -	clk_disable_unprepare(tx->mclk);
> -	clk_disable_unprepare(tx->npl);
> -	clk_disable_unprepare(tx->fsgen);
> +	if (!pm_runtime_status_suspended(&pdev->dev)) {
> +		clk_disable_unprepare(tx->macro);
> +		clk_disable_unprepare(tx->dcodec);
> +		clk_disable_unprepare(tx->mclk);
> +		clk_disable_unprepare(tx->npl);
> +		clk_disable_unprepare(tx->fsgen);
> +	}
>   
>   	lpass_macro_pds_exit(tx->pds);
>   }
> [..]

You're working with a tree that's months out of date! This does not 
apply to linux-next at all.

Commit a5ee3180a2a7 ("ASoC: codecs: lpass-{tx,rx}-macro: switch to PM 
clock framework") back in July removed all of these calls, and the 
clocks are now automatically managed by the framework, which has solved 
the issue.

Several people in #aarch64-laptops have gotten CXSD state working 
recently, and no one touched sound/soc/codecs.

~val