Re: [PATCH v5 06/11] drm/display: add CEC helpers code

[Dmitry Baryshkov <dmitry.baryshkov@oss.qualcomm.com>]

On 14/04/2025 17:58, Maxime Ripard wrote:
> On Mon, Apr 07, 2025 at 06:11:03PM +0300, Dmitry Baryshkov wrote:
>> +static void drm_connector_hdmi_cec_adapter_unregister(struct drm_connector *connector)
>> +{
>> +	struct drm_connector_hdmi_cec_data *data = connector->cec.data;
>> +
>> +	cec_delete_adapter(data->adapter);
>> +
>> +	if (data->funcs->uninit)
>> +		data->funcs->uninit(connector);
>> +
>> +	kfree(data);
>> +	connector->cec.data = NULL;
>> +}
>>
>> [...]
>>
>> +int drm_connector_hdmi_cec_register(struct drm_connector *connector,
>> +				    const struct drm_connector_hdmi_cec_funcs *funcs,
>> +				    const char *name,
>> +				    u8 available_las,
>> +				    struct device *dev)
>> +{
>> +	struct drm_connector_hdmi_cec_data *data;
>> +	struct cec_connector_info conn_info;
>> +	struct cec_adapter *cec_adap;
>> +	int ret;
>> +
>> +	if (!funcs->init || !funcs->enable || !funcs->log_addr || !funcs->transmit)
>> +		return -EINVAL;
>> +
>> +	data = kzalloc(sizeof(*data), GFP_KERNEL);
>> +	if (!data)
>> +		return -ENOMEM;
>> +
>> +	data->funcs = funcs;
>> +
>> +	cec_adap = cec_allocate_adapter(&drm_connector_hdmi_cec_adap_ops, connector, name,
>> +					CEC_CAP_DEFAULTS | CEC_CAP_CONNECTOR_INFO,
>> +					available_las ? : CEC_MAX_LOG_ADDRS);
>> +	ret = PTR_ERR_OR_ZERO(cec_adap);
>> +	if (ret < 0)
>> +		goto err_free;
>> +
>> +	cec_fill_conn_info_from_drm(&conn_info, connector);
>> +	cec_s_conn_info(cec_adap, &conn_info);
>> +
>> +	data->adapter = cec_adap;
>> +
>> +	mutex_lock(&connector->cec.mutex);
>> +
>> +	connector->cec.data = data;
>> +	connector->cec.funcs = &drm_connector_hdmi_cec_adapter_funcs;
>> +
>> +	ret = funcs->init(connector);
>> +	if (ret < 0)
>> +		goto err_delete_adapter;
>> +
>> +	ret = cec_register_adapter(cec_adap, dev);
>> +	if (ret < 0)
>> +		goto err_delete_adapter;
> 
> I'm a bit concerned about the respective lifetimes of CEC adapters and
> DRM connectors.
> 
> When you register the CEC adapter, its associated structure is
> kzalloc'd, and freed when the DRM connector is freed (so when nobody has
> any reference to it anymore: either when the device is torn down, or a
> DP-MST hotplug scenario).
> 
> The CEC adapter however will only be freed when its own users will close
> their file descriptor. So we can have a scenario when the CEC adapter is
> still live but the DRM connector has been unregistered. Thus, the CEC
> adapter data will have been kfree'd.

If I understand correctly, CEC core will handle this thanks to the 
cec_is_registered() calls in the important places. Nevertheless it's 
worth adding a comment and a set of drm_dev_enter() / _exit() calls.

> You might consider safe because $REASONS, but those need to be properly
> detailed and explained.
> 
> That's another reason why I think that just putting the connector
> pointer as data is better: connectors are refcounted, so we know those
> aren't an issue.

Not quite. CEC adapter itself doesn't have a refcount on the connector. 
And if add one, we'd create a loop, preventing connector from being 
unregistered.

-- 
With best wishes
Dmitry