[PATCH] soundwire: qcom: wait for the interrupt thread before gating the clock

[PATCH] soundwire: qcom: wait for the interrupt thread before gating the clock

[Oleg Keri]

A broadcast write returns as soon as the interrupt thread signals
SPECIAL_CMD_ID_FINISHED, while the thread still has to clear and re-read
the interrupt status. swrm_runtime_suspend() gates hclk right after the
ClockStopNow broadcast, so those accesses race against the clock being
removed; on x1e80100 and glymur hclk is the SoundWire clock itself. The
next clock-stop exit then fails at random: bus clash, parity errors
reported by the peripherals or AUTO_ENUM_FAILED, each leaving one side of
the speakers detached until reboot.

Wait for the interrupt thread to finish before gating the clock.

Fixes: 74e79da9fd46 ("soundwire: qcom: add runtime pm support")
Signed-off-by: Oleg Keri <okerixx@gmail.com>
---
 drivers/soundwire/qcom.c | 1 +
 1 file changed, 1 insertion(+)

diff --git a/drivers/soundwire/qcom.c b/drivers/soundwire/qcom.c
index 6e0c5fb10efa..76f940f4eb31 100644
--- a/drivers/soundwire/qcom.c
+++ b/drivers/soundwire/qcom.c
@@ -1887,6 +1887,7 @@ static int __maybe_unused swrm_runtime_suspend(struct device *dev)
 		}
 	}
 
+	synchronize_irq(ctrl->irq);
 	clk_disable_unprepare(ctrl->hclk);
 
 	usleep_range(300, 305);
-- 
2.55.0

[PATCH v2] soundwire: qcom: disable the interrupt before gating the clock

[Oleg Keri]

swrm_runtime_suspend() gates hclk straight after the ClockStopNow
broadcast, while the interrupt thread that completed it is still
clearing the status and CLK_STOP_FINISHED is about to fire. Both touch
registers with the clock going away; on x1e80100 and glymur hclk is the
SoundWire clock itself. The next clock-stop exit then fails at random
(bus clash, parity errors, AUTO_ENUM_FAILED) and one side of the
speakers stays detached until reboot.

Disable the interrupt before gating the clock and enable it again once
the clock is back.

Fixes: 74e79da9fd46 ("soundwire: qcom: add runtime pm support")
Signed-off-by: Oleg Keri <okerixx@gmail.com>
---
v2: disable_irq() instead of synchronize_irq(), which left a window for
CLK_STOP_FINISHED; about one boot in four still hit it.

 drivers/soundwire/qcom.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/drivers/soundwire/qcom.c b/drivers/soundwire/qcom.c
index 6e0c5fb10efa..147e55b7a600 100644
--- a/drivers/soundwire/qcom.c
+++ b/drivers/soundwire/qcom.c
@@ -1801,6 +1801,7 @@ static int __maybe_unused swrm_runtime_resume(struct device *dev)
 	}
 
 	clk_prepare_enable(ctrl->hclk);
+	enable_irq(ctrl->irq);
 
 	if (ctrl->clock_stop_not_supported) {
 		reinit_completion(&ctrl->enumeration);
@@ -1887,6 +1888,7 @@ static int __maybe_unused swrm_runtime_suspend(struct device *dev)
 		}
 	}
 
+	disable_irq(ctrl->irq);
 	clk_disable_unprepare(ctrl->hclk);
 
 	usleep_range(300, 305);
-- 
2.55.0

Re: [PATCH v2] soundwire: qcom: disable the interrupt before gating the clock

[Dmitry Baryshkov]

On Fri, Oct 09, 2026 at 10:26:39AM +0200, Oleg Keri wrote:
> swrm_runtime_suspend() gates hclk straight after the ClockStopNow
> broadcast, while the interrupt thread that completed it is still
> clearing the status and CLK_STOP_FINISHED is about to fire. Both touch
> registers with the clock going away; on x1e80100 and glymur hclk is the
> SoundWire clock itself. The next clock-stop exit then fails at random
> (bus clash, parity errors, AUTO_ENUM_FAILED) and one side of the
> speakers stays detached until reboot.
> 
> Disable the interrupt before gating the clock and enable it again once
> the clock is back.
> 
> Fixes: 74e79da9fd46 ("soundwire: qcom: add runtime pm support")
> Signed-off-by: Oleg Keri <okerixx@gmail.com>
> ---
> v2: disable_irq() instead of synchronize_irq(), which left a window for
> CLK_STOP_FINISHED; about one boot in four still hit it.
> 
>  drivers/soundwire/qcom.c | 2 ++
>  1 file changed, 2 insertions(+)
> 

Reviewed-by: Dmitry Baryshkov <dmitry.baryshkov@oss.qualcomm.com>


-- 
With best wishes
Dmitry