diff mbox series

[11/12] media: intel-cio2-bridge: Add a runtime-pm device-link between VCM and sensor

Message ID 20230627175643.114778-12-hdegoede@redhat.com (mailing list archive)
State New, archived
Headers show
Series media: intel-cio2-bridge: Add shared intel-cio2-bridge code, rework VCM instantiation | expand

Commit Message

Hans de Goede June 27, 2023, 5:56 p.m. UTC
In most cases when a VCM is used there is a single integrated module
with the sensor + VCM + lens. This means that the sensor and VCM often
share regulators and possibly also something like a powerdown pin.

In the ACPI tables this is modelled as a single ACPI device with
multiple I2cSerialBus resources.

On atomisp devices the regulators and clks are modelled as ACPI
power-resources, which are controlled by the (ACPI) power state
of the sensor. So the sensor must be in D0 power state for the VCM
to work.

To make this work add a device-link with DL_FLAG_PM_RUNTIME flag
so that the sensor will automatically be runtime-resumed whenever
the VCM is runtime-resumed.

This requires the probing of the VCM and thus the creation of the VCM
I2C-client to be delayed till after the sensor driver has bound.

Move the instantiation of the VCM I2C-client to the v4l2_async_notifier
bound op, so that it is done after the sensor driver has bound; and
add code to add the device-link.

This fixes the problem with the shared ACPI power-resources on atomisp2
and this avoids the need for VCM related workarounds on IPU3.

E.g. until now the dw9719 driver needed to get and control a Vsio
(V sensor IO) regulator since that needs to be enabled to enable I2C
pass-through on the PMIC on the sensor module. So the driver was
controlling this regulator even though the actual dw9719 chip has no
Vsio pin / power-plane.

This also removes the need for intel_cio2_bridge_init() to return
-EPROBE_DEFER since the VCM is now instantiated later.

Signed-off-by: Hans de Goede <hdegoede@redhat.com>
---
 drivers/media/common/intel-cio2-bridge.c      | 160 +++++++++++-------
 drivers/media/pci/intel/ipu3/ipu3-cio2-main.c |   5 +
 drivers/media/pci/intel/ipu3/ipu3-cio2.h      |   3 +
 include/media/intel-cio2-bridge.h             |   4 +-
 4 files changed, 111 insertions(+), 61 deletions(-)

Comments

Andy Shevchenko June 27, 2023, 9:03 p.m. UTC | #1
On Tue, Jun 27, 2023 at 07:56:41PM +0200, Hans de Goede wrote:
> In most cases when a VCM is used there is a single integrated module
> with the sensor + VCM + lens. This means that the sensor and VCM often
> share regulators and possibly also something like a powerdown pin.
> 
> In the ACPI tables this is modelled as a single ACPI device with
> multiple I2cSerialBus resources.
> 
> On atomisp devices the regulators and clks are modelled as ACPI
> power-resources, which are controlled by the (ACPI) power state
> of the sensor. So the sensor must be in D0 power state for the VCM
> to work.
> 
> To make this work add a device-link with DL_FLAG_PM_RUNTIME flag
> so that the sensor will automatically be runtime-resumed whenever
> the VCM is runtime-resumed.
> 
> This requires the probing of the VCM and thus the creation of the VCM
> I2C-client to be delayed till after the sensor driver has bound.
> 
> Move the instantiation of the VCM I2C-client to the v4l2_async_notifier
> bound op, so that it is done after the sensor driver has bound; and
> add code to add the device-link.
> 
> This fixes the problem with the shared ACPI power-resources on atomisp2
> and this avoids the need for VCM related workarounds on IPU3.
> 
> E.g. until now the dw9719 driver needed to get and control a Vsio
> (V sensor IO) regulator since that needs to be enabled to enable I2C
> pass-through on the PMIC on the sensor module. So the driver was
> controlling this regulator even though the actual dw9719 chip has no
> Vsio pin / power-plane.
> 
> This also removes the need for intel_cio2_bridge_init() to return
> -EPROBE_DEFER since the VCM is now instantiated later.

...

> +struct intel_cio2_bridge_instantiate_vcm_work_data {

Hmm... A bit long name :-)

> +	struct work_struct work;
> +	struct device *sensor;
>  	char name[16];
> +	struct i2c_board_info board_info;
> +};

...

> +	if (IS_ERR(vcm_client)) {
> +		dev_err(work->sensor, "Error instantiating VCM client: %ld\n",

or even %pe

> +			PTR_ERR(vcm_client));

...

> +out:

out_put_pm: ?

> +	pm_runtime_put(work->sensor);
> +	put_device(work->sensor);
> +	if (put_fwnode)
> +		fwnode_handle_put(work->board_info.fwnode);
> +	kfree(work);

...

> +int intel_cio2_bridge_instantiate_vcm(struct device *sensor)
> +{
> +	struct intel_cio2_bridge_instantiate_vcm_work_data *work;
> +	struct acpi_device *adev = ACPI_COMPANION(sensor);
> +	struct fwnode_handle *vcm_fwnode;
> +	struct i2c_client *vcm_client;
> +	char *sep;


I would rather split assignment, so below won't look a bit puzzling.

	struct acpi_device *adev;
	...
	adev = ACPI_COMPANION(sensor);

> +	if (!adev)
> +		return 0;
> +
> +	vcm_fwnode = fwnode_find_reference(dev_fwnode(sensor), "lens-focus", 0);
> +	if (IS_ERR(vcm_fwnode))
> +		return 0;
> +
> +	/* When reloading modules the client will already exist */
> +	vcm_client = i2c_find_device_by_fwnode(vcm_fwnode);
> +	if (vcm_client) {
> +		fwnode_handle_put(vcm_fwnode);
> +		put_device(&vcm_client->dev);
> +		return 0;
> +	}
> +
> +	work = kzalloc(sizeof(*work), GFP_KERNEL);
> +	if (!work) {
> +		fwnode_handle_put(vcm_fwnode);
> +		return -ENOMEM;
> +	}
> +
> +	INIT_WORK(&work->work, intel_cio2_bridge_instantiate_vcm_work);
> +	work->sensor = get_device(sensor);
> +	snprintf(work->name, sizeof(work->name), "%s-VCM",
> +		 acpi_dev_name(adev));
> +	work->board_info.dev_name = work->name;
> +	work->board_info.fwnode = vcm_fwnode;
> +	strscpy(work->board_info.type, fwnode_get_name(vcm_fwnode),
> +		I2C_NAME_SIZE);
> +	/* Strip "-<link>" postfix */
> +	sep = strchrnul(work->board_info.type, '-');
> +	*sep = 0;
> +
> +	queue_work(system_long_wq, &work->work);
> +
> +	return 0;
> +}
Bingbu Cao Oct. 30, 2023, 8:30 a.m. UTC | #2
Hans,

On 6/28/23 1:56 AM, Hans de Goede wrote:
> In most cases when a VCM is used there is a single integrated module
> with the sensor + VCM + lens. This means that the sensor and VCM often
> share regulators and possibly also something like a powerdown pin.
> 
> In the ACPI tables this is modelled as a single ACPI device with
> multiple I2cSerialBus resources.
> 
> On atomisp devices the regulators and clks are modelled as ACPI
> power-resources, which are controlled by the (ACPI) power state
> of the sensor. So the sensor must be in D0 power state for the VCM
> to work.
> 
> To make this work add a device-link with DL_FLAG_PM_RUNTIME flag
> so that the sensor will automatically be runtime-resumed whenever
> the VCM is runtime-resumed.
> 
> This requires the probing of the VCM and thus the creation of the VCM
> I2C-client to be delayed till after the sensor driver has bound.
> 
> Move the instantiation of the VCM I2C-client to the v4l2_async_notifier
> bound op, so that it is done after the sensor driver has bound; and
> add code to add the device-link.
> 
> This fixes the problem with the shared ACPI power-resources on atomisp2
> and this avoids the need for VCM related workarounds on IPU3.
> 
> E.g. until now the dw9719 driver needed to get and control a Vsio
> (V sensor IO) regulator since that needs to be enabled to enable I2C
> pass-through on the PMIC on the sensor module. So the driver was
> controlling this regulator even though the actual dw9719 chip has no
> Vsio pin / power-plane.
> 
> This also removes the need for intel_cio2_bridge_init() to return
> -EPROBE_DEFER since the VCM is now instantiated later.
> 
> Signed-off-by: Hans de Goede <hdegoede@redhat.com>
> ---
>  drivers/media/common/intel-cio2-bridge.c      | 160 +++++++++++-------
>  drivers/media/pci/intel/ipu3/ipu3-cio2-main.c |   5 +
>  drivers/media/pci/intel/ipu3/ipu3-cio2.h      |   3 +
>  include/media/intel-cio2-bridge.h             |   4 +-
>  4 files changed, 111 insertions(+), 61 deletions(-)
> 
> diff --git a/drivers/media/common/intel-cio2-bridge.c b/drivers/media/common/intel-cio2-bridge.c
> index 2f4256f9152c..997581bcc3b9 100644
> --- a/drivers/media/common/intel-cio2-bridge.c
> +++ b/drivers/media/common/intel-cio2-bridge.c
> @@ -5,7 +5,9 @@
>  #include <linux/acpi.h>
>  #include <linux/device.h>
>  #include <linux/i2c.h>
> +#include <linux/pm_runtime.h>
>  #include <linux/property.h>
> +#include <linux/workqueue.h>
>  
>  #include <media/intel-cio2-bridge.h>
>  #include <media/v4l2-fwnode.h>
> @@ -189,30 +191,111 @@ static void cio2_bridge_create_connection_swnodes(
>  	cio2_bridge_init_swnode_group(sensor);
>  }
>  
> -static void cio2_bridge_instantiate_vcm_i2c_client(
> -	struct intel_cio2_sensor *sensor)
> -{
> -	struct i2c_board_info board_info = { };
> +/*
> + * The actual instantiation must be done from a workqueue to avoid
> + * a deadlock on taking list_lock from v4l2-async twice.
> + */
> +struct intel_cio2_bridge_instantiate_vcm_work_data {
> +	struct work_struct work;
> +	struct device *sensor;
>  	char name[16];
> +	struct i2c_board_info board_info;
> +};
>  
> -	if (!sensor->vcm_type)
> -		return;
> +static void intel_cio2_bridge_instantiate_vcm_work(struct work_struct *_work)
> +{
> +	struct intel_cio2_bridge_instantiate_vcm_work_data *work =
> +		container_of(_work,
> +			     struct intel_cio2_bridge_instantiate_vcm_work_data,
> +			     work);
> +	struct acpi_device *adev = ACPI_COMPANION(work->sensor);
> +	struct i2c_client *vcm_client;
> +	bool put_fwnode = true;
> +	int ret;
>  
> -	snprintf(name, sizeof(name), "%s-VCM", acpi_dev_name(sensor->adev));
> -	board_info.dev_name = name;
> -	strscpy(board_info.type, sensor->vcm_type, ARRAY_SIZE(board_info.type));
> -	board_info.swnode = &sensor->swnodes[SWNODE_VCM];
> -
> -	sensor->vcm_i2c_client =
> -		i2c_acpi_new_device_by_fwnode(acpi_fwnode_handle(sensor->adev),
> -					      1, &board_info);
> -	if (IS_ERR(sensor->vcm_i2c_client)) {
> -		dev_warn(&sensor->adev->dev, "Error instantiation VCM i2c-client: %ld\n",
> -			 PTR_ERR(sensor->vcm_i2c_client));
> -		sensor->vcm_i2c_client = NULL;
> +	/*
> +	 * The client may get probed before the device_link gets added below
> +	 * make sure the sensor is powered-up during probe.
> +	 */
> +	ret = pm_runtime_get_sync(work->sensor);
> +	if (ret < 0) {
> +		dev_err(work->sensor, "Error %d runtime-resuming sensor, cannot instantiate VCM\n",
> +			ret);
> +		goto out;
>  	}

One question here: how do we make sure that the runtime PM of the sensor
is enabled during the .bound callback? Or is it a mandatory requirement
of driver of such camera sensors?

> +
> +	/*
> +	 * Note the client is created only once and then kept around
> +	 * even after a rmmod, just like the software-nodes.
> +	 */
> +	vcm_client = i2c_acpi_new_device_by_fwnode(acpi_fwnode_handle(adev),
> +						   1, &work->board_info);
> +	if (IS_ERR(vcm_client)) {
> +		dev_err(work->sensor, "Error instantiating VCM client: %ld\n",
> +			PTR_ERR(vcm_client));
> +		goto out;
> +	}
> +
> +	device_link_add(&vcm_client->dev, work->sensor, DL_FLAG_PM_RUNTIME);
> +
> +	dev_info(work->sensor, "Instantiated %s VCM\n", work->board_info.type);
> +	put_fwnode = false; /* Ownership has passed to the i2c-client */
> +
> +out:
> +	pm_runtime_put(work->sensor);
> +	put_device(work->sensor);
> +	if (put_fwnode)
> +		fwnode_handle_put(work->board_info.fwnode);
> +	kfree(work);
>  }
>  
> +int intel_cio2_bridge_instantiate_vcm(struct device *sensor)
> +{
> +	struct intel_cio2_bridge_instantiate_vcm_work_data *work;
> +	struct acpi_device *adev = ACPI_COMPANION(sensor);
> +	struct fwnode_handle *vcm_fwnode;
> +	struct i2c_client *vcm_client;
> +	char *sep;
> +
> +	if (!adev)
> +		return 0;
> +
> +	vcm_fwnode = fwnode_find_reference(dev_fwnode(sensor), "lens-focus", 0);
> +	if (IS_ERR(vcm_fwnode))
> +		return 0;
> +
> +	/* When reloading modules the client will already exist */
> +	vcm_client = i2c_find_device_by_fwnode(vcm_fwnode);
> +	if (vcm_client) {
> +		fwnode_handle_put(vcm_fwnode);
> +		put_device(&vcm_client->dev);
> +		return 0;
> +	}
> +
> +	work = kzalloc(sizeof(*work), GFP_KERNEL);
> +	if (!work) {
> +		fwnode_handle_put(vcm_fwnode);
> +		return -ENOMEM;
> +	}
> +
> +	INIT_WORK(&work->work, intel_cio2_bridge_instantiate_vcm_work);
> +	work->sensor = get_device(sensor);
> +	snprintf(work->name, sizeof(work->name), "%s-VCM",
> +		 acpi_dev_name(adev));
> +	work->board_info.dev_name = work->name;
> +	work->board_info.fwnode = vcm_fwnode;
> +	strscpy(work->board_info.type, fwnode_get_name(vcm_fwnode),
> +		I2C_NAME_SIZE);
> +	/* Strip "-<link>" postfix */
> +	sep = strchrnul(work->board_info.type, '-');
> +	*sep = 0;
> +
> +	queue_work(system_long_wq, &work->work);
> +
> +	return 0;
> +}
> +EXPORT_SYMBOL(intel_cio2_bridge_instantiate_vcm);
> +
>  static void cio2_bridge_unregister_sensors(struct intel_cio2_bridge *bridge)
>  {
>  	struct intel_cio2_sensor *sensor;
> @@ -222,7 +305,6 @@ static void cio2_bridge_unregister_sensors(struct intel_cio2_bridge *bridge)
>  		sensor = &bridge->sensors[i];
>  		software_node_unregister_node_group(sensor->group);
>  		acpi_dev_put(sensor->adev);
> -		i2c_unregister_device(sensor->vcm_i2c_client);
>  	}
>  }
>  
> @@ -273,8 +355,6 @@ static int cio2_bridge_connect_sensor(
>  		primary = acpi_fwnode_handle(adev);
>  		primary->secondary = fwnode;
>  
> -		cio2_bridge_instantiate_vcm_i2c_client(sensor);
> -
>  		dev_info(bridge->dev, "Found supported sensor %s\n",
>  			 acpi_dev_name(adev));
>  
> @@ -311,41 +391,6 @@ static int cio2_bridge_connect_sensors(struct intel_cio2_bridge *bridge)
>  	return ret;
>  }
>  
> -/*
> - * The VCM cannot be probed until the PMIC is completely setup. We cannot rely
> - * on -EPROBE_DEFER for this, since the consumer<->supplier relations between
> - * the VCM and regulators/clks are not described in ACPI, instead they are
> - * passed as board-data to the PMIC drivers. Since -PROBE_DEFER does not work
> - * for the clks/regulators the VCM i2c-clients must not be instantiated until
> - * the PMIC is fully setup.
> - *
> - * The sensor/VCM ACPI device has an ACPI _DEP on the PMIC, check this using the
> - * acpi_dev_ready_for_enumeration() helper, like the i2c-core-acpi code does
> - * for the sensors.
> - */
> -static int cio2_bridge_sensors_are_ready(
> -	const struct intel_cio2_sensor_config *supported_sensors)
> -{
> -	struct acpi_device *adev;
> -	bool ready = true;
> -	unsigned int i;
> -
> -	for (i = 0; supported_sensors[i].hid; i++) {
> -		const struct intel_cio2_sensor_config *cfg =
> -			&supported_sensors[i];
> -
> -		for_each_acpi_dev_match(adev, cfg->hid, NULL, -1) {
> -			if (!adev->status.enabled)
> -				continue;
> -
> -			if (!acpi_dev_ready_for_enumeration(adev))
> -				ready = false;
> -		}
> -	}
> -
> -	return ready;
> -}
> -
>  int intel_cio2_bridge_init(
>  	struct device *dev,
>  	int (*parse_sensor_fwnode)(struct acpi_device *adev,
> @@ -358,9 +403,6 @@ int intel_cio2_bridge_init(
>  	unsigned int i;
>  	int ret;
>  
> -	if (!cio2_bridge_sensors_are_ready(supported_sensors))
> -		return -EPROBE_DEFER;
> -
>  	bridge = kzalloc(sizeof(*bridge), GFP_KERNEL);
>  	if (!bridge)
>  		return -ENOMEM;
> diff --git a/drivers/media/pci/intel/ipu3/ipu3-cio2-main.c b/drivers/media/pci/intel/ipu3/ipu3-cio2-main.c
> index afc3255b5eb8..a22d213d66e9 100644
> --- a/drivers/media/pci/intel/ipu3/ipu3-cio2-main.c
> +++ b/drivers/media/pci/intel/ipu3/ipu3-cio2-main.c
> @@ -1386,10 +1386,15 @@ static int cio2_notifier_bound(struct v4l2_async_notifier *notifier,
>  	struct cio2_device *cio2 = to_cio2_device(notifier);
>  	struct sensor_async_subdev *s_asd = to_sensor_asd(asd);
>  	struct cio2_queue *q;
> +	int ret;
>  
>  	if (cio2->queue[s_asd->csi2.port].sensor)
>  		return -EBUSY;
>  
> +	ret = intel_cio2_bridge_instantiate_vcm(sd->dev);
> +	if (ret)
> +		return ret;
> +
>  	q = &cio2->queue[s_asd->csi2.port];
>  
>  	q->csi2 = s_asd->csi2;
> diff --git a/drivers/media/pci/intel/ipu3/ipu3-cio2.h b/drivers/media/pci/intel/ipu3/ipu3-cio2.h
> index 7ff7915f9823..208d60364f44 100644
> --- a/drivers/media/pci/intel/ipu3/ipu3-cio2.h
> +++ b/drivers/media/pci/intel/ipu3/ipu3-cio2.h
> @@ -12,6 +12,7 @@
>  
>  #include <asm/page.h>
>  
> +#include <media/intel-cio2-bridge.h>
>  #include <media/media-device.h>
>  #include <media/media-entity.h>
>  #include <media/v4l2-async.h>
> @@ -463,6 +464,8 @@ static inline struct cio2_queue *vb2q_to_cio2_queue(struct vb2_queue *vq)
>  int cio2_bridge_init(struct device *dev);
>  #else
>  static inline int cio2_bridge_init(struct device *dev) { return 0; }
> +static inline int
> +intel_cio2_bridge_instantiate_vcm(struct device *dev) { return 0; }
>  #endif
>  
>  #endif
> diff --git a/include/media/intel-cio2-bridge.h b/include/media/intel-cio2-bridge.h
> index 90fde180faef..faa60a5b1cd8 100644
> --- a/include/media/intel-cio2-bridge.h
> +++ b/include/media/intel-cio2-bridge.h
> @@ -10,7 +10,6 @@
>  #include <media/v4l2-fwnode.h>
>  
>  struct acpi_device;
> -struct i2c_client;
>  
>  #define INTEL_CIO2_HID					"INT343E"
>  #define INTEL_CIO2_MAX_LANES				4
> @@ -56,7 +55,6 @@ struct intel_cio2_sensor {
>  	/* append link(u8) in "-%u" format as suffix of HID */
>  	char name[ACPI_ID_LEN + 4];
>  	struct acpi_device *adev;
> -	struct i2c_client *vcm_i2c_client;
>  
>  	/* SWNODE_COUNT + 1 for terminating NULL */
>  	const struct software_node *group[SWNODE_COUNT + 1];
> @@ -99,6 +97,8 @@ int intel_cio2_bridge_init(
>  				   const struct intel_cio2_sensor_config *cfg),
>  	const struct intel_cio2_sensor_config *supported_sensors);
>  
> +int intel_cio2_bridge_instantiate_vcm(struct device *sensor);
> +
>  enum v4l2_fwnode_orientation
>  intel_cio2_bridge_get_orientation(struct acpi_device *adev);
>  
>
Sakari Ailus Oct. 30, 2023, 8:55 a.m. UTC | #3
Hi Bingbu,

On Mon, Oct 30, 2023 at 04:30:39PM +0800, Bingbu Cao wrote:
> > +static void intel_cio2_bridge_instantiate_vcm_work(struct work_struct *_work)
> > +{
> > +	struct intel_cio2_bridge_instantiate_vcm_work_data *work =
> > +		container_of(_work,
> > +			     struct intel_cio2_bridge_instantiate_vcm_work_data,
> > +			     work);
> > +	struct acpi_device *adev = ACPI_COMPANION(work->sensor);
> > +	struct i2c_client *vcm_client;
> > +	bool put_fwnode = true;
> > +	int ret;
> >  
> > -	snprintf(name, sizeof(name), "%s-VCM", acpi_dev_name(sensor->adev));
> > -	board_info.dev_name = name;
> > -	strscpy(board_info.type, sensor->vcm_type, ARRAY_SIZE(board_info.type));
> > -	board_info.swnode = &sensor->swnodes[SWNODE_VCM];
> > -
> > -	sensor->vcm_i2c_client =
> > -		i2c_acpi_new_device_by_fwnode(acpi_fwnode_handle(sensor->adev),
> > -					      1, &board_info);
> > -	if (IS_ERR(sensor->vcm_i2c_client)) {
> > -		dev_warn(&sensor->adev->dev, "Error instantiation VCM i2c-client: %ld\n",
> > -			 PTR_ERR(sensor->vcm_i2c_client));
> > -		sensor->vcm_i2c_client = NULL;
> > +	/*
> > +	 * The client may get probed before the device_link gets added below
> > +	 * make sure the sensor is powered-up during probe.
> > +	 */
> > +	ret = pm_runtime_get_sync(work->sensor);
> > +	if (ret < 0) {
> > +		dev_err(work->sensor, "Error %d runtime-resuming sensor, cannot instantiate VCM\n",
> > +			ret);
> > +		goto out;
> >  	}
> 
> One question here: how do we make sure that the runtime PM of the sensor
> is enabled during the .bound callback? Or is it a mandatory requirement
> of driver of such camera sensors?

The sensor driver needs to enable runtime PM in probe, otherwise this won't
work. But I don't see why a driver wouldn't? Of course otherwise it
wouldn't be a hard requirement.

This function gets called after the upstream notifier (CIO2 in this case)
driver's bound is called, the sensor's probe function has been run by this
time so this part should be fine.
Hans de Goede Oct. 30, 2023, 8:58 a.m. UTC | #4
Hi,

On 10/30/23 09:55, Sakari Ailus wrote:
> Hi Bingbu,
> 
> On Mon, Oct 30, 2023 at 04:30:39PM +0800, Bingbu Cao wrote:
>>> +static void intel_cio2_bridge_instantiate_vcm_work(struct work_struct *_work)
>>> +{
>>> +	struct intel_cio2_bridge_instantiate_vcm_work_data *work =
>>> +		container_of(_work,
>>> +			     struct intel_cio2_bridge_instantiate_vcm_work_data,
>>> +			     work);
>>> +	struct acpi_device *adev = ACPI_COMPANION(work->sensor);
>>> +	struct i2c_client *vcm_client;
>>> +	bool put_fwnode = true;
>>> +	int ret;
>>>  
>>> -	snprintf(name, sizeof(name), "%s-VCM", acpi_dev_name(sensor->adev));
>>> -	board_info.dev_name = name;
>>> -	strscpy(board_info.type, sensor->vcm_type, ARRAY_SIZE(board_info.type));
>>> -	board_info.swnode = &sensor->swnodes[SWNODE_VCM];
>>> -
>>> -	sensor->vcm_i2c_client =
>>> -		i2c_acpi_new_device_by_fwnode(acpi_fwnode_handle(sensor->adev),
>>> -					      1, &board_info);
>>> -	if (IS_ERR(sensor->vcm_i2c_client)) {
>>> -		dev_warn(&sensor->adev->dev, "Error instantiation VCM i2c-client: %ld\n",
>>> -			 PTR_ERR(sensor->vcm_i2c_client));
>>> -		sensor->vcm_i2c_client = NULL;
>>> +	/*
>>> +	 * The client may get probed before the device_link gets added below
>>> +	 * make sure the sensor is powered-up during probe.
>>> +	 */
>>> +	ret = pm_runtime_get_sync(work->sensor);
>>> +	if (ret < 0) {
>>> +		dev_err(work->sensor, "Error %d runtime-resuming sensor, cannot instantiate VCM\n",
>>> +			ret);
>>> +		goto out;
>>>  	}
>>
>> One question here: how do we make sure that the runtime PM of the sensor
>> is enabled during the .bound callback? Or is it a mandatory requirement
>> of driver of such camera sensors?
> 
> The sensor driver needs to enable runtime PM in probe, otherwise this won't
> work. But I don't see why a driver wouldn't? Of course otherwise it
> wouldn't be a hard requirement.

Right. I believe that having runtime pm support is more or less a hard
requirement for sensor drivers used on x86. This is also necessary in
cases where the power on/off sequence is (partially) implemented in
ACPI AML code in the _PS0 and _PS3 methods of the sensor's ACPI device.

This happens on [atom]ISP2 based devices and AFAIK also on ChromeOS
devices.

So I think it is safe to assume that sensor drivers implement runtime
pm support and in cases where the driver does not we need to fix the driver.

Regards,

Hans
Sakari Ailus Oct. 30, 2023, 9:23 a.m. UTC | #5
Hi Hans,

On Mon, Oct 30, 2023 at 09:58:09AM +0100, Hans de Goede wrote:
> Hi,
> 
> On 10/30/23 09:55, Sakari Ailus wrote:
> > Hi Bingbu,
> > 
> > On Mon, Oct 30, 2023 at 04:30:39PM +0800, Bingbu Cao wrote:
> >>> +static void intel_cio2_bridge_instantiate_vcm_work(struct work_struct *_work)
> >>> +{
> >>> +	struct intel_cio2_bridge_instantiate_vcm_work_data *work =
> >>> +		container_of(_work,
> >>> +			     struct intel_cio2_bridge_instantiate_vcm_work_data,
> >>> +			     work);
> >>> +	struct acpi_device *adev = ACPI_COMPANION(work->sensor);
> >>> +	struct i2c_client *vcm_client;
> >>> +	bool put_fwnode = true;
> >>> +	int ret;
> >>>  
> >>> -	snprintf(name, sizeof(name), "%s-VCM", acpi_dev_name(sensor->adev));
> >>> -	board_info.dev_name = name;
> >>> -	strscpy(board_info.type, sensor->vcm_type, ARRAY_SIZE(board_info.type));
> >>> -	board_info.swnode = &sensor->swnodes[SWNODE_VCM];
> >>> -
> >>> -	sensor->vcm_i2c_client =
> >>> -		i2c_acpi_new_device_by_fwnode(acpi_fwnode_handle(sensor->adev),
> >>> -					      1, &board_info);
> >>> -	if (IS_ERR(sensor->vcm_i2c_client)) {
> >>> -		dev_warn(&sensor->adev->dev, "Error instantiation VCM i2c-client: %ld\n",
> >>> -			 PTR_ERR(sensor->vcm_i2c_client));
> >>> -		sensor->vcm_i2c_client = NULL;
> >>> +	/*
> >>> +	 * The client may get probed before the device_link gets added below
> >>> +	 * make sure the sensor is powered-up during probe.
> >>> +	 */
> >>> +	ret = pm_runtime_get_sync(work->sensor);
> >>> +	if (ret < 0) {
> >>> +		dev_err(work->sensor, "Error %d runtime-resuming sensor, cannot instantiate VCM\n",
> >>> +			ret);
> >>> +		goto out;
> >>>  	}
> >>
> >> One question here: how do we make sure that the runtime PM of the sensor
> >> is enabled during the .bound callback? Or is it a mandatory requirement
> >> of driver of such camera sensors?
> > 
> > The sensor driver needs to enable runtime PM in probe, otherwise this won't
> > work. But I don't see why a driver wouldn't? Of course otherwise it
> > wouldn't be a hard requirement.

I meant to write "wasn't earlier a hard requirement". This could also be
documented, I can write a patch for this.

> 
> Right. I believe that having runtime pm support is more or less a hard
> requirement for sensor drivers used on x86. This is also necessary in
> cases where the power on/off sequence is (partially) implemented in
> ACPI AML code in the _PS0 and _PS3 methods of the sensor's ACPI device.
> 
> This happens on [atom]ISP2 based devices and AFAIK also on ChromeOS
> devices.
> 
> So I think it is safe to assume that sensor drivers implement runtime
> pm support and in cases where the driver does not we need to fix the driver.

I agree.
Bingbu Cao Nov. 1, 2023, 6:26 a.m. UTC | #6
Sakari and Hans,

------------------------------------------------------------------------
BRs,  
Bingbu Cao 

>-----Original Message-----
>From: Sakari Ailus <sakari.ailus@linux.intel.com>
>Sent: Monday, October 30, 2023 5:24 PM
>To: Hans de Goede <hdegoede@redhat.com>
>Cc: Bingbu Cao <bingbu.cao@linux.intel.com>; Laurent Pinchart
><laurent.pinchart@ideasonboard.com>; Daniel Scally
><dan.scally@ideasonboard.com>; Mauro Carvalho Chehab <mchehab@kernel.org>;
>Andy Shevchenko <andy@kernel.org>; Kate Hsuan <hpa@redhat.com>; Cao, Bingbu
><bingbu.cao@intel.com>; linux-media@vger.kernel.org
>Subject: Re: [PATCH 11/12] media: intel-cio2-bridge: Add a runtime-pm
>device-link between VCM and sensor
>
>Hi Hans,
>
>On Mon, Oct 30, 2023 at 09:58:09AM +0100, Hans de Goede wrote:
>> Hi,
>>
>> On 10/30/23 09:55, Sakari Ailus wrote:
>> > Hi Bingbu,
>> >
>> > On Mon, Oct 30, 2023 at 04:30:39PM +0800, Bingbu Cao wrote:
>> >>> +static void intel_cio2_bridge_instantiate_vcm_work(struct
>> >>> +work_struct *_work) {
>> >>> +	struct intel_cio2_bridge_instantiate_vcm_work_data *work =
>> >>> +		container_of(_work,
>> >>> +			     struct
>intel_cio2_bridge_instantiate_vcm_work_data,
>> >>> +			     work);
>> >>> +	struct acpi_device *adev = ACPI_COMPANION(work->sensor);
>> >>> +	struct i2c_client *vcm_client;
>> >>> +	bool put_fwnode = true;
>> >>> +	int ret;
>> >>>
>> >>> -	snprintf(name, sizeof(name), "%s-VCM", acpi_dev_name(sensor->adev));
>> >>> -	board_info.dev_name = name;
>> >>> -	strscpy(board_info.type, sensor->vcm_type,
>ARRAY_SIZE(board_info.type));
>> >>> -	board_info.swnode = &sensor->swnodes[SWNODE_VCM];
>> >>> -
>> >>> -	sensor->vcm_i2c_client =
>> >>> -		i2c_acpi_new_device_by_fwnode(acpi_fwnode_handle(sensor-
>>adev),
>> >>> -					      1, &board_info);
>> >>> -	if (IS_ERR(sensor->vcm_i2c_client)) {
>> >>> -		dev_warn(&sensor->adev->dev, "Error instantiation VCM i2c-
>client: %ld\n",
>> >>> -			 PTR_ERR(sensor->vcm_i2c_client));
>> >>> -		sensor->vcm_i2c_client = NULL;
>> >>> +	/*
>> >>> +	 * The client may get probed before the device_link gets added below
>> >>> +	 * make sure the sensor is powered-up during probe.
>> >>> +	 */
>> >>> +	ret = pm_runtime_get_sync(work->sensor);
>> >>> +	if (ret < 0) {
>> >>> +		dev_err(work->sensor, "Error %d runtime-resuming sensor,
>cannot instantiate VCM\n",
>> >>> +			ret);
>> >>> +		goto out;
>> >>>  	}
>> >>
>> >> One question here: how do we make sure that the runtime PM of the
>> >> sensor is enabled during the .bound callback? Or is it a mandatory
>> >> requirement of driver of such camera sensors?
>> >
>> > The sensor driver needs to enable runtime PM in probe, otherwise
>> > this won't work. But I don't see why a driver wouldn't? Of course
>> > otherwise it wouldn't be a hard requirement.
>
>I meant to write "wasn't earlier a hard requirement". This could also be
>documented, I can write a patch for this.

The problem here is that some driver doesn't enable the runtime PM before
v4l2_async_register_subdev_sensor(), that may cause race condition and
VCM instantiation failure as the runtime PM is not enabled yet. Then there
is no chance for VCM instantiation to succeed. 

Does this 'requirement' or limit really make sense?

>
>>
>> Right. I believe that having runtime pm support is more or less a hard
>> requirement for sensor drivers used on x86. This is also necessary in
>> cases where the power on/off sequence is (partially) implemented in
>> ACPI AML code in the _PS0 and _PS3 methods of the sensor's ACPI device.
>>
>> This happens on [atom]ISP2 based devices and AFAIK also on ChromeOS
>> devices.
>>
>> So I think it is safe to assume that sensor drivers implement runtime
>> pm support and in cases where the driver does not we need to fix the
>driver.
>
>I agree.
>
>--
>Regards,
>
>Sakari Ailus
Sakari Ailus Nov. 1, 2023, 7:12 a.m. UTC | #7
Hi Bingbu,

On Wed, Nov 01, 2023 at 06:26:39AM +0000, Cao, Bingbu wrote:
> Sakari and Hans,
> 
> ------------------------------------------------------------------------
> BRs,  
> Bingbu Cao 
> 
> >-----Original Message-----
> >From: Sakari Ailus <sakari.ailus@linux.intel.com>
> >Sent: Monday, October 30, 2023 5:24 PM
> >To: Hans de Goede <hdegoede@redhat.com>
> >Cc: Bingbu Cao <bingbu.cao@linux.intel.com>; Laurent Pinchart
> ><laurent.pinchart@ideasonboard.com>; Daniel Scally
> ><dan.scally@ideasonboard.com>; Mauro Carvalho Chehab <mchehab@kernel.org>;
> >Andy Shevchenko <andy@kernel.org>; Kate Hsuan <hpa@redhat.com>; Cao, Bingbu
> ><bingbu.cao@intel.com>; linux-media@vger.kernel.org
> >Subject: Re: [PATCH 11/12] media: intel-cio2-bridge: Add a runtime-pm
> >device-link between VCM and sensor
> >
> >Hi Hans,
> >
> >On Mon, Oct 30, 2023 at 09:58:09AM +0100, Hans de Goede wrote:
> >> Hi,
> >>
> >> On 10/30/23 09:55, Sakari Ailus wrote:
> >> > Hi Bingbu,
> >> >
> >> > On Mon, Oct 30, 2023 at 04:30:39PM +0800, Bingbu Cao wrote:
> >> >>> +static void intel_cio2_bridge_instantiate_vcm_work(struct
> >> >>> +work_struct *_work) {
> >> >>> +	struct intel_cio2_bridge_instantiate_vcm_work_data *work =
> >> >>> +		container_of(_work,
> >> >>> +			     struct
> >intel_cio2_bridge_instantiate_vcm_work_data,
> >> >>> +			     work);
> >> >>> +	struct acpi_device *adev = ACPI_COMPANION(work->sensor);
> >> >>> +	struct i2c_client *vcm_client;
> >> >>> +	bool put_fwnode = true;
> >> >>> +	int ret;
> >> >>>
> >> >>> -	snprintf(name, sizeof(name), "%s-VCM", acpi_dev_name(sensor->adev));
> >> >>> -	board_info.dev_name = name;
> >> >>> -	strscpy(board_info.type, sensor->vcm_type,
> >ARRAY_SIZE(board_info.type));
> >> >>> -	board_info.swnode = &sensor->swnodes[SWNODE_VCM];
> >> >>> -
> >> >>> -	sensor->vcm_i2c_client =
> >> >>> -		i2c_acpi_new_device_by_fwnode(acpi_fwnode_handle(sensor-
> >>adev),
> >> >>> -					      1, &board_info);
> >> >>> -	if (IS_ERR(sensor->vcm_i2c_client)) {
> >> >>> -		dev_warn(&sensor->adev->dev, "Error instantiation VCM i2c-
> >client: %ld\n",
> >> >>> -			 PTR_ERR(sensor->vcm_i2c_client));
> >> >>> -		sensor->vcm_i2c_client = NULL;
> >> >>> +	/*
> >> >>> +	 * The client may get probed before the device_link gets added below
> >> >>> +	 * make sure the sensor is powered-up during probe.
> >> >>> +	 */
> >> >>> +	ret = pm_runtime_get_sync(work->sensor);
> >> >>> +	if (ret < 0) {
> >> >>> +		dev_err(work->sensor, "Error %d runtime-resuming sensor,
> >cannot instantiate VCM\n",
> >> >>> +			ret);
> >> >>> +		goto out;
> >> >>>  	}
> >> >>
> >> >> One question here: how do we make sure that the runtime PM of the
> >> >> sensor is enabled during the .bound callback? Or is it a mandatory
> >> >> requirement of driver of such camera sensors?
> >> >
> >> > The sensor driver needs to enable runtime PM in probe, otherwise
> >> > this won't work. But I don't see why a driver wouldn't? Of course
> >> > otherwise it wouldn't be a hard requirement.
> >
> >I meant to write "wasn't earlier a hard requirement". This could also be
> >documented, I can write a patch for this.
> 
> The problem here is that some driver doesn't enable the runtime PM before
> v4l2_async_register_subdev_sensor(), that may cause race condition and
> VCM instantiation failure as the runtime PM is not enabled yet. Then there
> is no chance for VCM instantiation to succeed. 

These drivers need to be fixed.

All initialisation needs to complete by the time the async sub-device is
registered as this is when the sensor may be bound to a notifier and also
accessible via the UAPI.

In most of the cases the rest of the probe completes even if runtime PM is
enabled after registering the sub-device (without IPU bridge) but it's
still not correct.

> 
> Does this 'requirement' or limit really make sense?
Bingbu Cao Nov. 1, 2023, 7:38 a.m. UTC | #8
Sakari,

------------------------------------------------------------------------
BRs,  
Bingbu Cao 

>-----Original Message-----
>From: Sakari Ailus <sakari.ailus@linux.intel.com>
>Sent: Wednesday, November 1, 2023 3:12 PM
>To: Cao, Bingbu <bingbu.cao@intel.com>
>Cc: Hans de Goede <hdegoede@redhat.com>; Bingbu Cao
><bingbu.cao@linux.intel.com>; Laurent Pinchart
><laurent.pinchart@ideasonboard.com>; Daniel Scally
><dan.scally@ideasonboard.com>; Mauro Carvalho Chehab <mchehab@kernel.org>;
>Andy Shevchenko <andy@kernel.org>; Kate Hsuan <hpa@redhat.com>; linux-
>media@vger.kernel.org
>Subject: Re: [PATCH 11/12] media: intel-cio2-bridge: Add a runtime-pm
>device-link between VCM and sensor
>
>Hi Bingbu,
>
>On Wed, Nov 01, 2023 at 06:26:39AM +0000, Cao, Bingbu wrote:
>> Sakari and Hans,
>>
>> ----------------------------------------------------------------------
>> --
>> BRs,
>> Bingbu Cao
>>
>> >-----Original Message-----
>> >From: Sakari Ailus <sakari.ailus@linux.intel.com>
>> >Sent: Monday, October 30, 2023 5:24 PM
>> >To: Hans de Goede <hdegoede@redhat.com>
>> >Cc: Bingbu Cao <bingbu.cao@linux.intel.com>; Laurent Pinchart
>> ><laurent.pinchart@ideasonboard.com>; Daniel Scally
>> ><dan.scally@ideasonboard.com>; Mauro Carvalho Chehab
>> ><mchehab@kernel.org>; Andy Shevchenko <andy@kernel.org>; Kate Hsuan
>> ><hpa@redhat.com>; Cao, Bingbu <bingbu.cao@intel.com>;
>> >linux-media@vger.kernel.org
>> >Subject: Re: [PATCH 11/12] media: intel-cio2-bridge: Add a runtime-pm
>> >device-link between VCM and sensor
>> >
>> >Hi Hans,
>> >
>> >On Mon, Oct 30, 2023 at 09:58:09AM +0100, Hans de Goede wrote:
>> >> Hi,
>> >>
>> >> On 10/30/23 09:55, Sakari Ailus wrote:
>> >> > Hi Bingbu,
>> >> >
>> >> > On Mon, Oct 30, 2023 at 04:30:39PM +0800, Bingbu Cao wrote:
>> >> >>> +static void intel_cio2_bridge_instantiate_vcm_work(struct
>> >> >>> +work_struct *_work) {
>> >> >>> +	struct intel_cio2_bridge_instantiate_vcm_work_data *work =
>> >> >>> +		container_of(_work,
>> >> >>> +			     struct
>> >intel_cio2_bridge_instantiate_vcm_work_data,
>> >> >>> +			     work);
>> >> >>> +	struct acpi_device *adev = ACPI_COMPANION(work->sensor);
>> >> >>> +	struct i2c_client *vcm_client;
>> >> >>> +	bool put_fwnode = true;
>> >> >>> +	int ret;
>> >> >>>
>> >> >>> -	snprintf(name, sizeof(name), "%s-VCM", acpi_dev_name(sensor-
>>adev));
>> >> >>> -	board_info.dev_name = name;
>> >> >>> -	strscpy(board_info.type, sensor->vcm_type,
>> >ARRAY_SIZE(board_info.type));
>> >> >>> -	board_info.swnode = &sensor->swnodes[SWNODE_VCM];
>> >> >>> -
>> >> >>> -	sensor->vcm_i2c_client =
>> >> >>> -
>	i2c_acpi_new_device_by_fwnode(acpi_fwnode_handle(sensor-
>> >>adev),
>> >> >>> -					      1, &board_info);
>> >> >>> -	if (IS_ERR(sensor->vcm_i2c_client)) {
>> >> >>> -		dev_warn(&sensor->adev->dev, "Error instantiation VCM
>i2c-
>> >client: %ld\n",
>> >> >>> -			 PTR_ERR(sensor->vcm_i2c_client));
>> >> >>> -		sensor->vcm_i2c_client = NULL;
>> >> >>> +	/*
>> >> >>> +	 * The client may get probed before the device_link gets
>added below
>> >> >>> +	 * make sure the sensor is powered-up during probe.
>> >> >>> +	 */
>> >> >>> +	ret = pm_runtime_get_sync(work->sensor);
>> >> >>> +	if (ret < 0) {
>> >> >>> +		dev_err(work->sensor, "Error %d runtime-resuming
>sensor,
>> >cannot instantiate VCM\n",
>> >> >>> +			ret);
>> >> >>> +		goto out;
>> >> >>>  	}
>> >> >>
>> >> >> One question here: how do we make sure that the runtime PM of
>> >> >> the sensor is enabled during the .bound callback? Or is it a
>> >> >> mandatory requirement of driver of such camera sensors?
>> >> >
>> >> > The sensor driver needs to enable runtime PM in probe, otherwise
>> >> > this won't work. But I don't see why a driver wouldn't? Of course
>> >> > otherwise it wouldn't be a hard requirement.
>> >
>> >I meant to write "wasn't earlier a hard requirement". This could also
>> >be documented, I can write a patch for this.
>>
>> The problem here is that some driver doesn't enable the runtime PM
>> before v4l2_async_register_subdev_sensor(), that may cause race
>> condition and VCM instantiation failure as the runtime PM is not
>> enabled yet. Then there is no chance for VCM instantiation to succeed.
>
>These drivers need to be fixed.
>
>All initialisation needs to complete by the time the async sub-device is
>registered as this is when the sensor may be bound to a notifier and also
>accessible via the UAPI.

Ack. Thanks!

>
>In most of the cases the rest of the probe completes even if runtime PM is
>enabled after registering the sub-device (without IPU bridge) but it's
>still not correct.
>
>>
>> Does this 'requirement' or limit really make sense?
>
>--
>Regards,
>
>Sakari Ailus
Andy Shevchenko Nov. 2, 2023, 1:35 p.m. UTC | #9
On Wed, Nov 01, 2023 at 07:38:24AM +0000, Cao, Bingbu wrote:

Side note: Can you please Cc me in your v2, I believe I have some comments,
but let's see a new version for them?
Hans de Goede Nov. 2, 2023, 1:54 p.m. UTC | #10
Hi Andy,

On 11/2/23 14:35, Andy Shevchenko wrote:
> On Wed, Nov 01, 2023 at 07:38:24AM +0000, Cao, Bingbu wrote:
> 
> Side note: Can you please Cc me in your v2, I believe I have some comments,
> but let's see a new version for them?

I'm a bit confused, this is a series written by me not Bingbu
and this series has been merged for a while now...

Regards,

Hans
Bingbu Cao Nov. 3, 2023, 2:01 a.m. UTC | #11
Hans,

On 11/2/23 9:54 PM, Hans de Goede wrote:
> Hi Andy,
> 
> On 11/2/23 14:35, Andy Shevchenko wrote:
>> On Wed, Nov 01, 2023 at 07:38:24AM +0000, Cao, Bingbu wrote:
>>
>> Side note: Can you please Cc me in your v2, I believe I have some comments,
>> but let's see a new version for them?
> 
> I'm a bit confused, this is a series written by me not Bingbu
> and this series has been merged for a while now...

Yes and sorry. :)
I just borrow this thread to discuss more following your patch.

> 
> Regards,
> 
> Hans
> 
> 
> 
>
diff mbox series

Patch

diff --git a/drivers/media/common/intel-cio2-bridge.c b/drivers/media/common/intel-cio2-bridge.c
index 2f4256f9152c..997581bcc3b9 100644
--- a/drivers/media/common/intel-cio2-bridge.c
+++ b/drivers/media/common/intel-cio2-bridge.c
@@ -5,7 +5,9 @@ 
 #include <linux/acpi.h>
 #include <linux/device.h>
 #include <linux/i2c.h>
+#include <linux/pm_runtime.h>
 #include <linux/property.h>
+#include <linux/workqueue.h>
 
 #include <media/intel-cio2-bridge.h>
 #include <media/v4l2-fwnode.h>
@@ -189,30 +191,111 @@  static void cio2_bridge_create_connection_swnodes(
 	cio2_bridge_init_swnode_group(sensor);
 }
 
-static void cio2_bridge_instantiate_vcm_i2c_client(
-	struct intel_cio2_sensor *sensor)
-{
-	struct i2c_board_info board_info = { };
+/*
+ * The actual instantiation must be done from a workqueue to avoid
+ * a deadlock on taking list_lock from v4l2-async twice.
+ */
+struct intel_cio2_bridge_instantiate_vcm_work_data {
+	struct work_struct work;
+	struct device *sensor;
 	char name[16];
+	struct i2c_board_info board_info;
+};
 
-	if (!sensor->vcm_type)
-		return;
+static void intel_cio2_bridge_instantiate_vcm_work(struct work_struct *_work)
+{
+	struct intel_cio2_bridge_instantiate_vcm_work_data *work =
+		container_of(_work,
+			     struct intel_cio2_bridge_instantiate_vcm_work_data,
+			     work);
+	struct acpi_device *adev = ACPI_COMPANION(work->sensor);
+	struct i2c_client *vcm_client;
+	bool put_fwnode = true;
+	int ret;
 
-	snprintf(name, sizeof(name), "%s-VCM", acpi_dev_name(sensor->adev));
-	board_info.dev_name = name;
-	strscpy(board_info.type, sensor->vcm_type, ARRAY_SIZE(board_info.type));
-	board_info.swnode = &sensor->swnodes[SWNODE_VCM];
-
-	sensor->vcm_i2c_client =
-		i2c_acpi_new_device_by_fwnode(acpi_fwnode_handle(sensor->adev),
-					      1, &board_info);
-	if (IS_ERR(sensor->vcm_i2c_client)) {
-		dev_warn(&sensor->adev->dev, "Error instantiation VCM i2c-client: %ld\n",
-			 PTR_ERR(sensor->vcm_i2c_client));
-		sensor->vcm_i2c_client = NULL;
+	/*
+	 * The client may get probed before the device_link gets added below
+	 * make sure the sensor is powered-up during probe.
+	 */
+	ret = pm_runtime_get_sync(work->sensor);
+	if (ret < 0) {
+		dev_err(work->sensor, "Error %d runtime-resuming sensor, cannot instantiate VCM\n",
+			ret);
+		goto out;
 	}
+
+	/*
+	 * Note the client is created only once and then kept around
+	 * even after a rmmod, just like the software-nodes.
+	 */
+	vcm_client = i2c_acpi_new_device_by_fwnode(acpi_fwnode_handle(adev),
+						   1, &work->board_info);
+	if (IS_ERR(vcm_client)) {
+		dev_err(work->sensor, "Error instantiating VCM client: %ld\n",
+			PTR_ERR(vcm_client));
+		goto out;
+	}
+
+	device_link_add(&vcm_client->dev, work->sensor, DL_FLAG_PM_RUNTIME);
+
+	dev_info(work->sensor, "Instantiated %s VCM\n", work->board_info.type);
+	put_fwnode = false; /* Ownership has passed to the i2c-client */
+
+out:
+	pm_runtime_put(work->sensor);
+	put_device(work->sensor);
+	if (put_fwnode)
+		fwnode_handle_put(work->board_info.fwnode);
+	kfree(work);
 }
 
+int intel_cio2_bridge_instantiate_vcm(struct device *sensor)
+{
+	struct intel_cio2_bridge_instantiate_vcm_work_data *work;
+	struct acpi_device *adev = ACPI_COMPANION(sensor);
+	struct fwnode_handle *vcm_fwnode;
+	struct i2c_client *vcm_client;
+	char *sep;
+
+	if (!adev)
+		return 0;
+
+	vcm_fwnode = fwnode_find_reference(dev_fwnode(sensor), "lens-focus", 0);
+	if (IS_ERR(vcm_fwnode))
+		return 0;
+
+	/* When reloading modules the client will already exist */
+	vcm_client = i2c_find_device_by_fwnode(vcm_fwnode);
+	if (vcm_client) {
+		fwnode_handle_put(vcm_fwnode);
+		put_device(&vcm_client->dev);
+		return 0;
+	}
+
+	work = kzalloc(sizeof(*work), GFP_KERNEL);
+	if (!work) {
+		fwnode_handle_put(vcm_fwnode);
+		return -ENOMEM;
+	}
+
+	INIT_WORK(&work->work, intel_cio2_bridge_instantiate_vcm_work);
+	work->sensor = get_device(sensor);
+	snprintf(work->name, sizeof(work->name), "%s-VCM",
+		 acpi_dev_name(adev));
+	work->board_info.dev_name = work->name;
+	work->board_info.fwnode = vcm_fwnode;
+	strscpy(work->board_info.type, fwnode_get_name(vcm_fwnode),
+		I2C_NAME_SIZE);
+	/* Strip "-<link>" postfix */
+	sep = strchrnul(work->board_info.type, '-');
+	*sep = 0;
+
+	queue_work(system_long_wq, &work->work);
+
+	return 0;
+}
+EXPORT_SYMBOL(intel_cio2_bridge_instantiate_vcm);
+
 static void cio2_bridge_unregister_sensors(struct intel_cio2_bridge *bridge)
 {
 	struct intel_cio2_sensor *sensor;
@@ -222,7 +305,6 @@  static void cio2_bridge_unregister_sensors(struct intel_cio2_bridge *bridge)
 		sensor = &bridge->sensors[i];
 		software_node_unregister_node_group(sensor->group);
 		acpi_dev_put(sensor->adev);
-		i2c_unregister_device(sensor->vcm_i2c_client);
 	}
 }
 
@@ -273,8 +355,6 @@  static int cio2_bridge_connect_sensor(
 		primary = acpi_fwnode_handle(adev);
 		primary->secondary = fwnode;
 
-		cio2_bridge_instantiate_vcm_i2c_client(sensor);
-
 		dev_info(bridge->dev, "Found supported sensor %s\n",
 			 acpi_dev_name(adev));
 
@@ -311,41 +391,6 @@  static int cio2_bridge_connect_sensors(struct intel_cio2_bridge *bridge)
 	return ret;
 }
 
-/*
- * The VCM cannot be probed until the PMIC is completely setup. We cannot rely
- * on -EPROBE_DEFER for this, since the consumer<->supplier relations between
- * the VCM and regulators/clks are not described in ACPI, instead they are
- * passed as board-data to the PMIC drivers. Since -PROBE_DEFER does not work
- * for the clks/regulators the VCM i2c-clients must not be instantiated until
- * the PMIC is fully setup.
- *
- * The sensor/VCM ACPI device has an ACPI _DEP on the PMIC, check this using the
- * acpi_dev_ready_for_enumeration() helper, like the i2c-core-acpi code does
- * for the sensors.
- */
-static int cio2_bridge_sensors_are_ready(
-	const struct intel_cio2_sensor_config *supported_sensors)
-{
-	struct acpi_device *adev;
-	bool ready = true;
-	unsigned int i;
-
-	for (i = 0; supported_sensors[i].hid; i++) {
-		const struct intel_cio2_sensor_config *cfg =
-			&supported_sensors[i];
-
-		for_each_acpi_dev_match(adev, cfg->hid, NULL, -1) {
-			if (!adev->status.enabled)
-				continue;
-
-			if (!acpi_dev_ready_for_enumeration(adev))
-				ready = false;
-		}
-	}
-
-	return ready;
-}
-
 int intel_cio2_bridge_init(
 	struct device *dev,
 	int (*parse_sensor_fwnode)(struct acpi_device *adev,
@@ -358,9 +403,6 @@  int intel_cio2_bridge_init(
 	unsigned int i;
 	int ret;
 
-	if (!cio2_bridge_sensors_are_ready(supported_sensors))
-		return -EPROBE_DEFER;
-
 	bridge = kzalloc(sizeof(*bridge), GFP_KERNEL);
 	if (!bridge)
 		return -ENOMEM;
diff --git a/drivers/media/pci/intel/ipu3/ipu3-cio2-main.c b/drivers/media/pci/intel/ipu3/ipu3-cio2-main.c
index afc3255b5eb8..a22d213d66e9 100644
--- a/drivers/media/pci/intel/ipu3/ipu3-cio2-main.c
+++ b/drivers/media/pci/intel/ipu3/ipu3-cio2-main.c
@@ -1386,10 +1386,15 @@  static int cio2_notifier_bound(struct v4l2_async_notifier *notifier,
 	struct cio2_device *cio2 = to_cio2_device(notifier);
 	struct sensor_async_subdev *s_asd = to_sensor_asd(asd);
 	struct cio2_queue *q;
+	int ret;
 
 	if (cio2->queue[s_asd->csi2.port].sensor)
 		return -EBUSY;
 
+	ret = intel_cio2_bridge_instantiate_vcm(sd->dev);
+	if (ret)
+		return ret;
+
 	q = &cio2->queue[s_asd->csi2.port];
 
 	q->csi2 = s_asd->csi2;
diff --git a/drivers/media/pci/intel/ipu3/ipu3-cio2.h b/drivers/media/pci/intel/ipu3/ipu3-cio2.h
index 7ff7915f9823..208d60364f44 100644
--- a/drivers/media/pci/intel/ipu3/ipu3-cio2.h
+++ b/drivers/media/pci/intel/ipu3/ipu3-cio2.h
@@ -12,6 +12,7 @@ 
 
 #include <asm/page.h>
 
+#include <media/intel-cio2-bridge.h>
 #include <media/media-device.h>
 #include <media/media-entity.h>
 #include <media/v4l2-async.h>
@@ -463,6 +464,8 @@  static inline struct cio2_queue *vb2q_to_cio2_queue(struct vb2_queue *vq)
 int cio2_bridge_init(struct device *dev);
 #else
 static inline int cio2_bridge_init(struct device *dev) { return 0; }
+static inline int
+intel_cio2_bridge_instantiate_vcm(struct device *dev) { return 0; }
 #endif
 
 #endif
diff --git a/include/media/intel-cio2-bridge.h b/include/media/intel-cio2-bridge.h
index 90fde180faef..faa60a5b1cd8 100644
--- a/include/media/intel-cio2-bridge.h
+++ b/include/media/intel-cio2-bridge.h
@@ -10,7 +10,6 @@ 
 #include <media/v4l2-fwnode.h>
 
 struct acpi_device;
-struct i2c_client;
 
 #define INTEL_CIO2_HID					"INT343E"
 #define INTEL_CIO2_MAX_LANES				4
@@ -56,7 +55,6 @@  struct intel_cio2_sensor {
 	/* append link(u8) in "-%u" format as suffix of HID */
 	char name[ACPI_ID_LEN + 4];
 	struct acpi_device *adev;
-	struct i2c_client *vcm_i2c_client;
 
 	/* SWNODE_COUNT + 1 for terminating NULL */
 	const struct software_node *group[SWNODE_COUNT + 1];
@@ -99,6 +97,8 @@  int intel_cio2_bridge_init(
 				   const struct intel_cio2_sensor_config *cfg),
 	const struct intel_cio2_sensor_config *supported_sensors);
 
+int intel_cio2_bridge_instantiate_vcm(struct device *sensor);
+
 enum v4l2_fwnode_orientation
 intel_cio2_bridge_get_orientation(struct acpi_device *adev);