[PATCH 2/2] accel/ivpu: Fix PM related deadlocks in MS IOCTLs

[Lizhi Hou]

On 3/25/25 04:43, Maciej Falkowski wrote:
> From: Jacek Lawrynowicz <jacek.lawrynowicz at linux.intel.com>
>
> Prevent runtime resume/suspend while MS IOCTLs are in progress.
> Failed suspend will call ivpu_ms_cleanup() that would try to acquire
> file_priv->ms_lock, which is already held by the IOCTLs.
>
> Fixes: cdfad4db7756 ("accel/ivpu: Add NPU profiling support")
> Cc: <stable at vger.kernel.org> # v6.11+
> Signed-off-by: Jacek Lawrynowicz <jacek.lawrynowicz at linux.intel.com>
> Signed-off-by: Maciej Falkowski <maciej.falkowski at linux.intel.com>
> ---
>   drivers/accel/ivpu/ivpu_debugfs.c |  4 ++--
>   drivers/accel/ivpu/ivpu_ms.c      | 18 ++++++++++++++++++
>   2 files changed, 20 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/accel/ivpu/ivpu_debugfs.c b/drivers/accel/ivpu/ivpu_debugfs.c
> index 0825851656a2..f0dad0c9ce33 100644
> --- a/drivers/accel/ivpu/ivpu_debugfs.c
> +++ b/drivers/accel/ivpu/ivpu_debugfs.c
> @@ -332,7 +332,7 @@ ivpu_force_recovery_fn(struct file *file, const char __user *user_buf, size_t si
>   		return -EINVAL;
>   
>   	ret = ivpu_rpm_get(vdev);
> -	if (ret)
> +	if (ret < 0)
>   		return ret;
>   
>   	ivpu_pm_trigger_recovery(vdev, "debugfs");
> @@ -383,7 +383,7 @@ static int dct_active_set(void *data, u64 active_percent)
>   		return -EINVAL;
>   
>   	ret = ivpu_rpm_get(vdev);
> -	if (ret)
> +	if (ret < 0)
>   		return ret;
>   
>   	if (active_percent)
> diff --git a/drivers/accel/ivpu/ivpu_ms.c b/drivers/accel/ivpu/ivpu_ms.c
> index eb485cf15ad6..2a043baf10ca 100644
> --- a/drivers/accel/ivpu/ivpu_ms.c
> +++ b/drivers/accel/ivpu/ivpu_ms.c
> @@ -45,6 +45,10 @@ int ivpu_ms_start_ioctl(struct drm_device *dev, void *data, struct drm_file *fil
>   	    args->sampling_period_ns < MS_MIN_SAMPLE_PERIOD_NS)
>   		return -EINVAL;
>   
> +	ret = ivpu_rpm_get(vdev);
> +	if (ret < 0)
> +		return ret;
> +
>   	mutex_lock(&file_priv->ms_lock);
>   
>   	if (get_instance_by_mask(file_priv, args->metric_group_mask)) {
> @@ -97,6 +101,8 @@ int ivpu_ms_start_ioctl(struct drm_device *dev, void *data, struct drm_file *fil
>   	kfree(ms);
>   unlock:
>   	mutex_unlock(&file_priv->ms_lock);
> +
> +	ivpu_rpm_put(vdev);
>   	return ret;
>   }
>   
> @@ -161,6 +167,10 @@ int ivpu_ms_get_data_ioctl(struct drm_device *dev, void *data, struct drm_file *
>   	if (!args->metric_group_mask)
>   		return -EINVAL;
>   
> +	ret = ivpu_rpm_get(vdev);
> +	if (ret < 0)
> +		return ret;
> +
>   	mutex_lock(&file_priv->ms_lock);
>   
>   	ms = get_instance_by_mask(file_priv, args->metric_group_mask);
> @@ -188,6 +198,7 @@ int ivpu_ms_get_data_ioctl(struct drm_device *dev, void *data, struct drm_file *
>   unlock:
>   	mutex_unlock(&file_priv->ms_lock);
>   
> +	ivpu_rpm_put(vdev);
>   	return ret;
>   }
>   
> @@ -205,11 +216,17 @@ int ivpu_ms_stop_ioctl(struct drm_device *dev, void *data, struct drm_file *file
>   {
>   	struct ivpu_file_priv *file_priv = file->driver_priv;
>   	struct drm_ivpu_metric_streamer_stop *args = data;
> +	struct ivpu_device *vdev = file_priv->vdev;
>   	struct ivpu_ms_instance *ms;
> +	int ret;
>   
>   	if (!args->metric_group_mask)
>   		return -EINVAL;
>   
> +	ret = ivpu_rpm_get(vdev);
> +	if (ret < 0)
> +		return ret;
> +
>   	mutex_lock(&file_priv->ms_lock);
>   
>   	ms = get_instance_by_mask(file_priv, args->metric_group_mask);
> @@ -218,6 +235,7 @@ int ivpu_ms_stop_ioctl(struct drm_device *dev, void *data, struct drm_file *file
>   
>   	mutex_unlock(&file_priv->ms_lock);
>   
> +	ivpu_rpm_put(vdev);
Reviewed-by: Lizhi Hou <lizhi.hou at amd.com>
>   	return ms ? 0 : -EINVAL;
>   }
>