In q6v5_regulator_enable(), when any operation fails for regulator at
index 'i', the error cleanup path unconditionally calls
regulator_disable() starting from index 'i'. However, regulator 'i'
was never successfully enabled at this point, resulting in an
unbalanced disable.

There are three distinct failure points:
- regulator_set_voltage() failure: voltage was never set, load was
never set, regulator was never enabled.
- regulator_set_load() failure: voltage was set, but regulator was
never enabled.
- regulator_enable() failure: voltage and load were set, but
regulator was never enabled.

Fix this by introducing three separate error labels to handle each
failure point correctly. For the failing regulator at index 'i',
only reset the resources that were actually configured, without
calling regulator_disable(). Then roll back all previously enabled
regulators using 'i--' in the for loop initializer to skip the
never-enabled regulator.

Fixes: 19f902b53b47 ("remoteproc: qcom: Initialize and enable proxy and active regulators.")
Cc: stable@vger.kernel.org
Signed-off-by: Sailesh Nandanavanam <saileshnandanavanam@gmail.com>
---
 drivers/remoteproc/qcom_q6v5_mss.c | 16 +++++++++++-----
 1 file changed, 11 insertions(+), 5 deletions(-)

diff --git a/drivers/remoteproc/qcom_q6v5_mss.c b/drivers/remoteproc/qcom_q6v5_mss.c
index ae78f5c7c1b6..9a17aa065f50 100644
--- a/drivers/remoteproc/qcom_q6v5_mss.c
+++ b/drivers/remoteproc/qcom_q6v5_mss.c
@@ -311,7 +311,7 @@ static int q6v5_regulator_enable(struct q6v5 *qproc,
 				dev_err(qproc->dev,
 					"Failed to request voltage for %d.\n",
 						i);
-				goto err;
+				goto err_set_voltage;
 			}
 		}
 
@@ -321,20 +321,26 @@ static int q6v5_regulator_enable(struct q6v5 *qproc,
 			if (ret < 0) {
 				dev_err(qproc->dev,
 					"Failed to set regulator mode\n");
-				goto err;
+				goto err_set_load;
 			}
 		}
 
 		ret = regulator_enable(regs[i].reg);
 		if (ret) {
 			dev_err(qproc->dev, "Regulator enable failed\n");
-			goto err;
+			goto err_enable;
 		}
 	}
 
 	return 0;
-err:
-	for (; i >= 0; i--) {
+err_enable:
+	if (regs[i].uA > 0)
+		regulator_set_load(regs[i].reg, 0);
+err_set_load:
+	if (regs[i].uV > 0)
+		regulator_set_voltage(regs[i].reg, 0, INT_MAX);
+err_set_voltage:
+	for (i--; i >= 0; i--) {
 		if (regs[i].uV > 0)
 			regulator_set_voltage(regs[i].reg, 0, INT_MAX);
 
-- 
2.34.1

On 7/10/26 9:46 PM, Sailesh Nandanavanam wrote:
> In q6v5_regulator_enable(), when any operation fails for regulator at
> index 'i', the error cleanup path unconditionally calls
> regulator_disable() starting from index 'i'. However, regulator 'i'
> was never successfully enabled at this point, resulting in an
> unbalanced disable.
> 
> There are three distinct failure points:
> - regulator_set_voltage() failure: voltage was never set, load was
> never set, regulator was never enabled.
> - regulator_set_load() failure: voltage was set, but regulator was
> never enabled.
> - regulator_enable() failure: voltage and load were set, but
> regulator was never enabled.
> 
> Fix this by introducing three separate error labels to handle each
> failure point correctly. For the failing regulator at index 'i',
> only reset the resources that were actually configured, without
> calling regulator_disable(). Then roll back all previously enabled
> regulators using 'i--' in the for loop initializer to skip the
> never-enabled regulator.
> 
> Fixes: 19f902b53b47 ("remoteproc: qcom: Initialize and enable proxy and active regulators.")
> Cc: stable@vger.kernel.org
> Signed-off-by: Sailesh Nandanavanam <saileshnandanavanam@gmail.com>
> ---

[...]

> -err:
> -	for (; i >= 0; i--) {
> +err_enable:
> +	if (regs[i].uA > 0)
> +		regulator_set_load(regs[i].reg, 0);
> +err_set_load:
> +	if (regs[i].uV > 0)
> +		regulator_set_voltage(regs[i].reg, 0, INT_MAX);

The first two labels only unwind a single regulator

Konrad

On 7/17/26 3:01 PM, Konrad Dybcio wrote:
> The first two labels only unwind a single regulator

Thanks for taking a look. Could you clarify whether this is a
correctness concern (e.g. the fallthrough from err_enable/err_set_load
into err_set_voltage not doing what you'd expect), or more a
structural/style preference (e.g. avoiding three chained labels in
favor of a different approach)? Happy to send a v2 once I understand
what you'd like changed.

Thanks,
Sailesh


On Fri, Jul 17, 2026 at 3:01 PM Konrad Dybcio
<konrad.dybcio@oss.qualcomm.com> wrote:
>
> On 7/10/26 9:46 PM, Sailesh Nandanavanam wrote:
> > In q6v5_regulator_enable(), when any operation fails for regulator at
> > index 'i', the error cleanup path unconditionally calls
> > regulator_disable() starting from index 'i'. However, regulator 'i'
> > was never successfully enabled at this point, resulting in an
> > unbalanced disable.
> >
> > There are three distinct failure points:
> > - regulator_set_voltage() failure: voltage was never set, load was
> > never set, regulator was never enabled.
> > - regulator_set_load() failure: voltage was set, but regulator was
> > never enabled.
> > - regulator_enable() failure: voltage and load were set, but
> > regulator was never enabled.
> >
> > Fix this by introducing three separate error labels to handle each
> > failure point correctly. For the failing regulator at index 'i',
> > only reset the resources that were actually configured, without
> > calling regulator_disable(). Then roll back all previously enabled
> > regulators using 'i--' in the for loop initializer to skip the
> > never-enabled regulator.
> >
> > Fixes: 19f902b53b47 ("remoteproc: qcom: Initialize and enable proxy and active regulators.")
> > Cc: stable@vger.kernel.org
> > Signed-off-by: Sailesh Nandanavanam <saileshnandanavanam@gmail.com>
> > ---
>
> [...]
>
> > -err:
> > -     for (; i >= 0; i--) {
> > +err_enable:
> > +     if (regs[i].uA > 0)
> > +             regulator_set_load(regs[i].reg, 0);
> > +err_set_load:
> > +     if (regs[i].uV > 0)
> > +             regulator_set_voltage(regs[i].reg, 0, INT_MAX);
>
> The first two labels only unwind a single regulator
>
> Konrad