media: imx: mipi csi-2: Don't fail if initial state times-out

From: Ezequiel Garcia <[email protected]>

Not all sensors will be able to guarantee a proper initial state. This may be either because the driver is not properly written, or (probably unlikely) because the hardware won't support it.

While the right solution in the former case is to fix the sensor driver, the real world not always allows right solutions, due to lack of available documentation and support on these sensors.

Let's relax this requirement, and allow the driver to support stream start, even if the sensor initial sequence wasn't the expected.

Also improve the warning message to better explain the problem and provide a hint that the sensor driver needs to be fixed.

Signed-off-by Ezequiel Garcia <[email protected]>

Signed-off-by Fabio Estevam <[email protected]>

---
Changes since v2:
- Reduce the warning message (Steve)
Changes since v1:
- Changed the warning message to better explain the problem and provide a hint that the sensor driver needs to be fixed. (Phillip)
- Keep printing the phy_state information (Phillip)
- Do not change csi2_dphy_wait_clock_lane() (Phillip/Steve)

  drivers/staging/media/imx/imx6-mipi-csi2.c | 12 ++++--------
  1 file changed, 4 insertions(+), 8 deletions(-)

drivers/staging/media/imx/imx6-mipi-csi2.c +4 -8

index f29e28df36ed..bfa4b254c4e4 100644
@@ -243,7 +243,7 @@ static int __maybe_unused csi2_dphy_wait_ulp(struct csi2_dev *csi2)
  }
 
  /* Waits for low-power LP-11 state on data and clock lanes. */
-static int csi2_dphy_wait_stopstate(struct csi2_dev *csi2)
+static void csi2_dphy_wait_stopstate(struct csi2_dev *csi2)
  {
  	u32 mask, reg;
  	int ret;

@@ -254,11 +254,9 @@ static int csi2_dphy_wait_stopstate(struct csi2_dev *csi2)
  	ret = readl_poll_timeout(csi2->base + CSI2_PHY_STATE, reg,
  				 (reg & mask) == mask, 0, 500000);
  	if (ret) {
-		v4l2_err(&csi2->sd, "LP-11 timeout, phy_state = 0x%08x\n", reg);
-		return ret;
+		v4l2_warn(&csi2->sd, "LP-11 wait timeout, likely a sensor driver bug, expect capture failures.\n");
+		v4l2_warn(&csi2->sd, "phy_state = 0x%08x\n", reg);
  	}
-
-	return 0;
  }
 
  /* Wait for active clock on the clock lane. */

@@ -316,9 +314,7 @@ static int csi2_start(struct csi2_dev *csi2)
  	csi2_enable(csi2, true);
 
  	/* Step 5 */
-	ret = csi2_dphy_wait_stopstate(csi2);
-	if (ret)
-		goto err_assert_reset;
+	csi2_dphy_wait_stopstate(csi2);
 
  	/* Step 6 */
  	ret = v4l2_subdev_call(csi2->src_sd, video, s_stream, 1);

From: Ezequiel Garcia <[email protected]>

Not all sensors will be able to guarantee a proper initial state. This may be either because the driver is not properly written, or (probably unlikely) because the hardware won't support it.

While the right solution in the former case is to fix the sensor driver, the real world not always allows right solutions, due to lack of available documentation and support on these sensors.

Let's relax this requirement, and allow the driver to support stream start, even if the sensor initial sequence wasn't the expected.

Also improve the warning message to better explain the problem and provide a hint that the sensor driver needs to be fixed.

Signed-off-by Ezequiel Garcia <[email protected]>

Signed-off-by Fabio Estevam <[email protected]>

Quoted 13 lines · Fabio Estevam

From: Ezequiel Garcia <[email protected]>

Not all sensors will be able to guarantee a proper initial state. This may be either because the driver is not properly written, or (probably unlikely) because the hardware won't support it.

While the right solution in the former case is to fix the sensor driver, the real world not always allows right solutions, due to lack of available documentation and support on these sensors.

Let's relax this requirement, and allow the driver to support stream start, even if the sensor initial sequence wasn't the expected.

Quoted 5 lines

Also improve the warning message to better explain the problem and provide a hint that the sensor driver needs to be fixed.

Signed-off-by Ezequiel Garcia <[email protected]>

Signed-off-by Fabio Estevam <[email protected]>

Reviewed-by Philipp Zabel <[email protected]>

Hey Hans,

Quoted 17 lines · Philipp Zabel

Quoted 16 lines · Fabio Estevam

From: Ezequiel Garcia <[email protected]>

Not all sensors will be able to guarantee a proper initial state. This may be either because the driver is not properly written, or (probably unlikely) because the hardware won't support it.

While the right solution in the former case is to fix the sensor driver, the real world not always allows right solutions, due to lack of available documentation and support on these sensors.

Let's relax this requirement, and allow the driver to support stream start, even if the sensor initial sequence wasn't the expected.

Also improve the warning message to better explain the problem and provide a hint that the sensor driver needs to be fixed.

Quoted 5 lines

Quoted 2 lines

Signed-off-by Ezequiel Garcia <[email protected]>

Signed-off-by Fabio Estevam <[email protected]>

Reviewed-by Philipp Zabel <[email protected]>

This seems ready to pick and it has Philipp's and Steve's RB.

Quoted 20 lines · Ezequiel Garcia

Hey Hans,

Quoted 17 lines · Philipp Zabel

Quoted 16 lines · Fabio Estevam

From: Ezequiel Garcia <[email protected]>

Not all sensors will be able to guarantee a proper initial state. This may be either because the driver is not properly written, or (probably unlikely) because the hardware won't support it.

While the right solution in the former case is to fix the sensor driver, the real world not always allows right solutions, due to lack of available documentation and support on these sensors.

Let's relax this requirement, and allow the driver to support stream start, even if the sensor initial sequence wasn't the expected.

Also improve the warning message to better explain the problem and provide a hint that the sensor driver needs to be fixed.

Quoted 7 lines

Quoted 5 lines

Quoted 2 lines

Signed-off-by Ezequiel Garcia <[email protected]>

Signed-off-by Fabio Estevam <[email protected]>

Reviewed-by Philipp Zabel <[email protected]>

This seems ready to pick and it has Philipp's and Steve's RB.

Hi Ezequiel,

In general the LP-11 condition should be detected by hardware (or firmware) in such a way that it's detected even if a transmitter that holds the state just a short period of time. In other words, software is not supposed to be even testing for it.

Have you checked how it works if you simply leave out this test?

Hi Sakari,

Thanks for reviewing the patch.

Quoted 36 lines · Sakari Ailus

Quoted 20 lines · Ezequiel Garcia

Hey Hans,

Quoted 17 lines · Philipp Zabel

Quoted 16 lines · Fabio Estevam

From: Ezequiel Garcia <[email protected]>

Not all sensors will be able to guarantee a proper initial state. This may be either because the driver is not properly written, or (probably unlikely) because the hardware won't support it.

While the right solution in the former case is to fix the sensor driver, the real world not always allows right solutions, due to lack of available documentation and support on these sensors.

Let's relax this requirement, and allow the driver to support stream start, even if the sensor initial sequence wasn't the expected.

Also improve the warning message to better explain the problem and provide a hint that the sensor driver needs to be fixed.

Quoted 7 lines

Quoted 5 lines

Quoted 2 lines

Signed-off-by Ezequiel Garcia <[email protected]>

Signed-off-by Fabio Estevam <[email protected]>

Reviewed-by Philipp Zabel <[email protected]>

This seems ready to pick and it has Philipp's and Steve's RB.

Hi Ezequiel,

In general the LP-11 condition should be detected by hardware (or firmware) in such a way that it's detected even if a transmitter that holds the state just a short period of time. In other words, software is not supposed to be even testing for it.

Quoted 2 lines

Have you checked how it works if you simply leave out this test?

The current change relaxes a condition, which we observed was too strict. Some drivers might be unable to enter LP-11 state, but I don't think that's a reason to fail capture.

We had to fix at least OV5645 and OV5640 recently because of this, and I can imagine more drivers will have the same issue.

Quoted 1 line

Quoted 1 line

Have you checked how it works if you simply leave out this test?

Some devices can be commanded to enter LP-11 state but some will just briefly visit that state. The LP-11 state is mandatory but software should not be involved in detecting it if at all possible.

So if the hardware does not require further initialisation to be done in LP-11 state, you should remove the check.

Quoted 2 lines

We had to fix at least OV5645 and OV5640 recently because of this, and I can imagine more drivers will have the same issue.

This is actually an issue in the IMX driver (or hardware), not in the sensor driver. It may be that sometimes it's easier to work around it in the sensor driver.

So, I'd like to know whether the check itself is a driver bug, or something that the hardware requires. The fact that you're sending this patch suggests the former.

Hi Sakari,

On Thu, 2019-08-08 at 11:26 +0300, Sakari Ailus wrote:
[...]

Quoted 1 line

Quoted 1 line

Quoted 1 line

Have you checked how it works if you simply leave out this test?

Whether this works or not depends on the sensor used, and for some sensor/drivers may depend on timing (or random factors influencing it). See below.

[...]

Quoted 3 lines

Some devices can be commanded to enter LP-11 state but some will just briefly visit that state. The LP-11 state is mandatory but software should not be involved in detecting it if at all possible.

This is a good point. Devices that can be set to LP-11 state immediately, but that don't stay there long enough (either because they wait for less than the required by spec 100µs, or because system load causes this check to be executed too late) may end up working reliably even though the warning fires.

Quoted 6 lines

So if the hardware does not require further initialisation to be done in LP-11 state, you should remove the check.

Quoted 2 lines

We had to fix at least OV5645 and OV5640 recently because of this, and I can imagine more drivers will have the same issue.

Quoted 7 lines

This is actually an issue in the IMX driver (or hardware), not in the sensor driver. It may be that sometimes it's easier to work around it in the sensor driver.

So, I'd like to know whether the check itself is a driver bug, or something that the hardware requires. The fact that you're sending this patch suggests the former.

This is something that the hardware requires, according to the reference manual. See the comment in drivers/staging/media/imx/imx6-mipi-csi2.c, especially step 5.:

/*
 * The required sequence of MIPI CSI-2 startup as specified in the i.MX6
 * reference manual is as follows:
 *
 * 1. Deassert presetn signal (global reset).
 *        It's not clear what this "global reset" signal is (maybe APB *        global reset), but in any case this step would be probably *        be carried out during driver load in csi2_probe().
 *
 * 2. Configure MIPI Camera Sensor to put all Tx lanes in LP-11 state. *        This must be carried out by the MIPI sensor's s_power(ON) subdev *        op.
 *
 * 3. D-PHY initialization.
 * 4. CSI2 Controller programming (Set N_LANES, deassert PHY_SHUTDOWNZ, *    deassert PHY_RSTZ, deassert CSI2_RESETN).
 * 5. Read the PHY status register (PHY_STATE) to confirm that all data and *    clock lanes of the D-PHY are in LP-11 state.
 * 6. Configure the MIPI Camera Sensor to start transmitting a clock on the *    D-PHY clock lane.
 * 7. CSI2 Controller programming - Read the PHY status register (PHY_STATE) *    to confirm that the D-PHY is receiving a clock on the D-PHY clock lane. */

I read this as the hardware needing to see the LP-11 -> HS transition after the DPHY reset has been released, and before the CSI2 RX controller is programmed.

If not all lanes are in LP-11 state at step 5., we can't guarantee that the DPHY has already seen this transition nor that it will see it in time before we start enabling the CSI2 RX.
Since this can lead to anything from sporadic to 100% startup failure, depending on timing or whether the sensor is configured correctly to produce this transition at all, I'd like this warning to stay. It has been helpful before when developing sensor drivers.

Hi Sakari,

On Thu, 2019-08-08 at 11:26 +0300, Sakari Ailus wrote:
[...]

Quoted 1 line

Quoted 1 line

Quoted 1 line

Have you checked how it works if you simply leave out this test?

Whether this works or not depends on the sensor used, and for some sensor/drivers may depend on timing (or random factors influencing it). See below.

[...]

Quoted 3 lines

Some devices can be commanded to enter LP-11 state but some will just briefly visit that state. The LP-11 state is mandatory but software should not be involved in detecting it if at all possible.

This is a good point. Devices that can be set to LP-11 state immediately, but that don't stay there long enough (either because they wait for less than the required by spec 100µs, or because system load causes this check to be executed too late) may end up working reliably even though the warning fires.

Quoted 5 lines

So if the hardware does not require further initialisation to be done in LP-11 state, you should remove the check.

Quoted 2 lines

We had to fix at least OV5645 and OV5640 recently because of this, and I can imagine more drivers will have the same issue.

Quoted 7 lines

This is actually an issue in the IMX driver (or hardware), not in the sensor driver. It may be that sometimes it's easier to work around it in the sensor driver.

So, I'd like to know whether the check itself is a driver bug, or something that the hardware requires. The fact that you're sending this patch suggests the former.

This is something that the hardware requires, according to the reference manual. See the comment in drivers/staging/media/imx/imx6-mipi-csi2.c, especially step 5.:

/*
 * The required sequence of MIPI CSI-2 startup as specified in the i.MX6
 * reference manual is as follows:
 *
 * 1. Deassert presetn signal (global reset).
 *        It's not clear what this "global reset" signal is (maybe APB *        global reset), but in any case this step would be probably *        be carried out during driver load in csi2_probe().
 *
 * 2. Configure MIPI Camera Sensor to put all Tx lanes in LP-11 state. *        This must be carried out by the MIPI sensor's s_power(ON) subdev *        op.
 *
 * 3. D-PHY initialization.
 * 4. CSI2 Controller programming (Set N_LANES, deassert PHY_SHUTDOWNZ, *    deassert PHY_RSTZ, deassert CSI2_RESETN).
 * 5. Read the PHY status register (PHY_STATE) to confirm that all data and *    clock lanes of the D-PHY are in LP-11 state.
 * 6. Configure the MIPI Camera Sensor to start transmitting a clock on the *    D-PHY clock lane.
 * 7. CSI2 Controller programming - Read the PHY status register (PHY_STATE) *    to confirm that the D-PHY is receiving a clock on the D-PHY clock lane. */

I read this as the hardware needing to see the LP-11 -> HS transition after the DPHY reset has been released, and before the CSI2 RX controller is programmed.

If not all lanes are in LP-11 state at step 5., we can't guarantee that the DPHY has already seen this transition nor that it will see it in time before we start enabling the CSI2 RX.
Since this can lead to anything from sporadic to 100% startup failure, depending on timing or whether the sensor is configured correctly to produce this transition at all, I'd like this warning to stay. It has been helpful before when developing sensor drivers.

Quoted 59 lines · Philipp Zabel

Hi Sakari,

On Thu, 2019-08-08 at 11:26 +0300, Sakari Ailus wrote:
[...]

Quoted 1 line

Quoted 1 line

Quoted 1 line

Have you checked how it works if you simply leave out this test?

Whether this works or not depends on the sensor used, and for some sensor/drivers may depend on timing (or random factors influencing it). See below.

[...]

Quoted 3 lines

Some devices can be commanded to enter LP-11 state but some will just briefly visit that state. The LP-11 state is mandatory but software should not be involved in detecting it if at all possible.

This is a good point. Devices that can be set to LP-11 state immediately, but that don't stay there long enough (either because they wait for less than the required by spec 100µs, or because system load causes this check to be executed too late) may end up working reliably even though the warning fires.

Quoted 5 lines

So if the hardware does not require further initialisation to be done in LP-11 state, you should remove the check.

Quoted 2 lines

We had to fix at least OV5645 and OV5640 recently because of this, and I can imagine more drivers will have the same issue.

Quoted 7 lines

This is actually an issue in the IMX driver (or hardware), not in the sensor driver. It may be that sometimes it's easier to work around it in the sensor driver.

So, I'd like to know whether the check itself is a driver bug, or something that the hardware requires. The fact that you're sending this patch suggests the former.

This is something that the hardware requires, according to the reference manual. See the comment in drivers/staging/media/imx/imx6-mipi-csi2.c, especially step 5.:

/*
  * The required sequence of MIPI CSI-2 startup as specified in the i.MX6
  * reference manual is as follows:
  *
  * 1. Deassert presetn signal (global reset).
  *        It's not clear what this "global reset" signal is (maybe APB *        global reset), but in any case this step would be probably *        be carried out during driver load in csi2_probe().
  *
  * 2. Configure MIPI Camera Sensor to put all Tx lanes in LP-11 state. *        This must be carried out by the MIPI sensor's s_power(ON) subdev *        op.
  *
  * 3. D-PHY initialization.
  * 4. CSI2 Controller programming (Set N_LANES, deassert PHY_SHUTDOWNZ, *    deassert PHY_RSTZ, deassert CSI2_RESETN).
  * 5. Read the PHY status register (PHY_STATE) to confirm that all data and *    clock lanes of the D-PHY are in LP-11 state.
  * 6. Configure the MIPI Camera Sensor to start transmitting a clock on the *    D-PHY clock lane.
  * 7. CSI2 Controller programming - Read the PHY status register (PHY_STATE) *    to confirm that the D-PHY is receiving a clock on the D-PHY clock lane. */

Quoted 3 lines

I read this as the hardware needing to see the LP-11 -> HS transition after the DPHY reset has been released, and before the CSI2 RX controller is programmed.

I think that's a fair assumption, and there's another paragraph at the end of that comment above that adds more evidence to that:

...
* All steps 3 through 7 are carried out by csi2_s_stream(ON) here. Step
* 6 is accomplished by calling the source subdev's s_stream(ON) between
* steps 5 and 7.
*/

So the driver is expecting that the LP-11 state persists until step 6, at which the LP-11 -> HS transition occurs when streaming is started at the transmitter.

But if the transmitter only stays in LP-11 state for the minimum 100 usec after it is powered on, and then _automatically_ transitions to HS, it's quite possible the LP-11 -> HS transition will happen long before the DPHY is taken out of reset. That's because the transmitter's s_power(ON) is called when the capture device is opened (via v4l2_pipeline_pm_use()), but the steps above are carried out when streaming starts, so userland would have to issue VIDIOC_STREAMON well within the 100 usec expires after open()'ing the device, for the receiver hardware to see the transition.

Perhaps that would be an argument for moving steps 1 - 5 into the driver's s_power(ON) call, which would first call s_power(ON) to the transmitter and then immediately go through steps 1 - 5. Steps 6,7 would then remain in s_stream(ON).

Quoted 59 lines · Philipp Zabel

Hi Sakari,

On Thu, 2019-08-08 at 11:26 +0300, Sakari Ailus wrote:
[...]

Quoted 1 line

Quoted 1 line

Quoted 1 line

Have you checked how it works if you simply leave out this test?

Whether this works or not depends on the sensor used, and for some sensor/drivers may depend on timing (or random factors influencing it). See below.

[...]

Quoted 3 lines

Some devices can be commanded to enter LP-11 state but some will just briefly visit that state. The LP-11 state is mandatory but software should not be involved in detecting it if at all possible.

This is a good point. Devices that can be set to LP-11 state immediately, but that don't stay there long enough (either because they wait for less than the required by spec 100µs, or because system load causes this check to be executed too late) may end up working reliably even though the warning fires.

Quoted 5 lines

So if the hardware does not require further initialisation to be done in LP-11 state, you should remove the check.

Quoted 2 lines

We had to fix at least OV5645 and OV5640 recently because of this, and I can imagine more drivers will have the same issue.

Quoted 7 lines

This is actually an issue in the IMX driver (or hardware), not in the sensor driver. It may be that sometimes it's easier to work around it in the sensor driver.

So, I'd like to know whether the check itself is a driver bug, or something that the hardware requires. The fact that you're sending this patch suggests the former.

This is something that the hardware requires, according to the reference manual. See the comment in drivers/staging/media/imx/imx6-mipi-csi2.c, especially step 5.:

/*
 * The required sequence of MIPI CSI-2 startup as specified in the i.MX6
 * reference manual is as follows:
 *
 * 1. Deassert presetn signal (global reset).
 *        It's not clear what this "global reset" signal is (maybe APB *        global reset), but in any case this step would be probably *        be carried out during driver load in csi2_probe().
 *
 * 2. Configure MIPI Camera Sensor to put all Tx lanes in LP-11 state. *        This must be carried out by the MIPI sensor's s_power(ON) subdev *        op.
 *
 * 3. D-PHY initialization.
 * 4. CSI2 Controller programming (Set N_LANES, deassert PHY_SHUTDOWNZ, *    deassert PHY_RSTZ, deassert CSI2_RESETN).
 * 5. Read the PHY status register (PHY_STATE) to confirm that all data and *    clock lanes of the D-PHY are in LP-11 state.
 * 6. Configure the MIPI Camera Sensor to start transmitting a clock on the *    D-PHY clock lane.
 * 7. CSI2 Controller programming - Read the PHY status register (PHY_STATE) *    to confirm that the D-PHY is receiving a clock on the D-PHY clock lane. */

Quoted 3 lines

I read this as the hardware needing to see the LP-11 -> HS transition after the DPHY reset has been released, and before the CSI2 RX controller is programmed.

I think that's a fair assumption, and there's another paragraph at the end of that comment above that adds more evidence to that:

...
* All steps 3 through 7 are carried out by csi2_s_stream(ON) here. Step
* 6 is accomplished by calling the source subdev's s_stream(ON) between
* steps 5 and 7.
*/

So the driver is expecting that the LP-11 state persists until step 6, at which the LP-11 -> HS transition occurs when streaming is started at the transmitter.

But if the transmitter only stays in LP-11 state for the minimum 100 usec after it is powered on, and then _automatically_ transitions to HS, it's quite possible the LP-11 -> HS transition will happen long before the DPHY is taken out of reset. That's because the transmitter's s_power(ON) is called when the capture device is opened (via v4l2_pipeline_pm_use()), but the steps above are carried out when streaming starts, so userland would have to issue VIDIOC_STREAMON well within the 100 usec expires after open()'ing the device, for the receiver hardware to see the transition.

Perhaps that would be an argument for moving steps 1 - 5 into the driver's s_power(ON) call, which would first call s_power(ON) to the transmitter and then immediately go through steps 1 - 5. Steps 6,7 would then remain in s_stream(ON).

Hi Steve,

Quoted 77 lines · Steve Longerbeam

Quoted 58 lines · Philipp Zabel

Hi Sakari,

On Thu, 2019-08-08 at 11:26 +0300, Sakari Ailus wrote:
[...]

Quoted 1 line

Quoted 1 line

Quoted 1 line

Have you checked how it works if you simply leave out this test?

Whether this works or not depends on the sensor used, and for some sensor/drivers may depend on timing (or random factors influencing it). See below.

[...]

Quoted 3 lines

Some devices can be commanded to enter LP-11 state but some will just briefly visit that state. The LP-11 state is mandatory but software should not be involved in detecting it if at all possible.

This is a good point. Devices that can be set to LP-11 state immediately, but that don't stay there long enough (either because they wait for less than the required by spec 100µs, or because system load causes this check to be executed too late) may end up working reliably even though the warning fires.

Quoted 5 lines

So if the hardware does not require further initialisation to be done in LP-11 state, you should remove the check.

Quoted 2 lines

We had to fix at least OV5645 and OV5640 recently because of this, and I can imagine more drivers will have the same issue.

Quoted 7 lines

This is actually an issue in the IMX driver (or hardware), not in the sensor driver. It may be that sometimes it's easier to work around it in the sensor driver.

So, I'd like to know whether the check itself is a driver bug, or something that the hardware requires. The fact that you're sending this patch suggests the former.

This is something that the hardware requires, according to the reference manual. See the comment in drivers/staging/media/imx/imx6-mipi-csi2.c, especially step 5.:

/*
 * The required sequence of MIPI CSI-2 startup as specified in the i.MX6
 * reference manual is as follows:
 *
 * 1. Deassert presetn signal (global reset).
 *        It's not clear what this "global reset" signal is (maybe APB *        global reset), but in any case this step would be probably *        be carried out during driver load in csi2_probe().
 *
 * 2. Configure MIPI Camera Sensor to put all Tx lanes in LP-11 state. *        This must be carried out by the MIPI sensor's s_power(ON) subdev *        op.
 *
 * 3. D-PHY initialization.
 * 4. CSI2 Controller programming (Set N_LANES, deassert PHY_SHUTDOWNZ, *    deassert PHY_RSTZ, deassert CSI2_RESETN).
 * 5. Read the PHY status register (PHY_STATE) to confirm that all data and *    clock lanes of the D-PHY are in LP-11 state.
 * 6. Configure the MIPI Camera Sensor to start transmitting a clock on the *    D-PHY clock lane.
 * 7. CSI2 Controller programming - Read the PHY status register (PHY_STATE) *    to confirm that the D-PHY is receiving a clock on the D-PHY clock lane. */

Quoted 3 lines

I read this as the hardware needing to see the LP-11 -> HS transition after the DPHY reset has been released, and before the CSI2 RX controller is programmed.

I think that's a fair assumption, and there's another paragraph at the end of that comment above that adds more evidence to that:

...
* All steps 3 through 7 are carried out by csi2_s_stream(ON) here. Step
* 6 is accomplished by calling the source subdev's s_stream(ON) between
* steps 5 and 7.
*/

So the driver is expecting that the LP-11 state persists until step 6, at which the LP-11 -> HS transition occurs when streaming is started at the transmitter.

Quoted 8 lines

But if the transmitter only stays in LP-11 state for the minimum 100 usec after it is powered on, and then _automatically_ transitions to HS, it's quite possible the LP-11 -> HS transition will happen long before the DPHY is taken out of reset. That's because the transmitter's s_power(ON) is called when the capture device is opened (via v4l2_pipeline_pm_use()), but the steps above are carried out when streaming starts, so userland would have to issue VIDIOC_STREAMON well within the 100 usec expires after open()'ing the device, for the receiver hardware to see the transition.

After powering on the sensor it sensor should stay in (software) standby mode until streaming is started. It should still have its lanes in LP-11 mode if the sensor supports it (as documented) until the sensor switches to the streaming mode.

Quoted 5 lines

Perhaps that would be an argument for moving steps 1 - 5 into the driver's s_power(ON) call, which would first call s_power(ON) to the transmitter and then immediately go through steps 1 - 5. Steps 6,7 would then remain in s_stream(ON).

Not all sensor drivers power on the sensor before starting streaming.

Perhaps we could add a prepare_streaming() callback (in absence of a better proposal) to address that? We'd also need a corresponding unprepare_streaming() callback as well --- to power off the sensor. I think this only should be done if the sensor can be switched to LP-11 explicitly; many simply don't support that.

Hi Sakari,

Quoted 88 lines · Sakari Ailus

Hi Steve,

Quoted 77 lines · Steve Longerbeam

Quoted 58 lines · Philipp Zabel

Hi Sakari,

On Thu, 2019-08-08 at 11:26 +0300, Sakari Ailus wrote:
[...]

Quoted 1 line

Quoted 1 line

Quoted 1 line

Have you checked how it works if you simply leave out this test?

Whether this works or not depends on the sensor used, and for some sensor/drivers may depend on timing (or random factors influencing it). See below.

[...]

Quoted 3 lines

Some devices can be commanded to enter LP-11 state but some will just briefly visit that state. The LP-11 state is mandatory but software should not be involved in detecting it if at all possible.

This is a good point. Devices that can be set to LP-11 state immediately, but that don't stay there long enough (either because they wait for less than the required by spec 100µs, or because system load causes this check to be executed too late) may end up working reliably even though the warning fires.

Quoted 5 lines

So if the hardware does not require further initialisation to be done in LP-11 state, you should remove the check.

Quoted 2 lines

We had to fix at least OV5645 and OV5640 recently because of this, and I can imagine more drivers will have the same issue.

Quoted 7 lines

This is actually an issue in the IMX driver (or hardware), not in the sensor driver. It may be that sometimes it's easier to work around it in the sensor driver.

So, I'd like to know whether the check itself is a driver bug, or something that the hardware requires. The fact that you're sending this patch suggests the former.

This is something that the hardware requires, according to the reference manual. See the comment in drivers/staging/media/imx/imx6-mipi-csi2.c, especially step 5.:

/*
 * The required sequence of MIPI CSI-2 startup as specified in the i.MX6
 * reference manual is as follows:
 *
 * 1. Deassert presetn signal (global reset).
 *        It's not clear what this "global reset" signal is (maybe APB *        global reset), but in any case this step would be probably *        be carried out during driver load in csi2_probe().
 *
 * 2. Configure MIPI Camera Sensor to put all Tx lanes in LP-11 state. *        This must be carried out by the MIPI sensor's s_power(ON) subdev *        op.
 *
 * 3. D-PHY initialization.
 * 4. CSI2 Controller programming (Set N_LANES, deassert PHY_SHUTDOWNZ, *    deassert PHY_RSTZ, deassert CSI2_RESETN).
 * 5. Read the PHY status register (PHY_STATE) to confirm that all data and *    clock lanes of the D-PHY are in LP-11 state.
 * 6. Configure the MIPI Camera Sensor to start transmitting a clock on the *    D-PHY clock lane.
 * 7. CSI2 Controller programming - Read the PHY status register (PHY_STATE) *    to confirm that the D-PHY is receiving a clock on the D-PHY clock lane. */

Quoted 3 lines

I read this as the hardware needing to see the LP-11 -> HS transition after the DPHY reset has been released, and before the CSI2 RX controller is programmed.

I think that's a fair assumption, and there's another paragraph at the end of that comment above that adds more evidence to that:

...
* All steps 3 through 7 are carried out by csi2_s_stream(ON) here. Step
* 6 is accomplished by calling the source subdev's s_stream(ON) between
* steps 5 and 7.
*/

So the driver is expecting that the LP-11 state persists until step 6, at which the LP-11 -> HS transition occurs when streaming is started at the transmitter.

Quoted 8 lines

But if the transmitter only stays in LP-11 state for the minimum 100 usec after it is powered on, and then _automatically_ transitions to HS, it's quite possible the LP-11 -> HS transition will happen long before the DPHY is taken out of reset. That's because the transmitter's s_power(ON) is called when the capture device is opened (via v4l2_pipeline_pm_use()), but the steps above are carried out when streaming starts, so userland would have to issue VIDIOC_STREAMON well within the 100 usec expires after open()'ing the device, for the receiver hardware to see the transition.

Quoted 4 lines

After powering on the sensor it sensor should stay in (software) standby mode until streaming is started. It should still have its lanes in LP-11 mode if the sensor supports it (as documented) until the sensor switches to the streaming mode.

Yes that's my understanding too.

Quoted 5 lines

Quoted 5 lines

Perhaps that would be an argument for moving steps 1 - 5 into the driver's s_power(ON) call, which would first call s_power(ON) to the transmitter and then immediately go through steps 1 - 5. Steps 6,7 would then remain in s_stream(ON).

Quoted 7 lines

Not all sensor drivers power on the sensor before starting streaming.

Perhaps we could add a prepare_streaming() callback (in absence of a better proposal) to address that? We'd also need a corresponding unprepare_streaming() callback as well --- to power off the sensor. I think this only should be done if the sensor can be switched to LP-11 explicitly; many simply don't support that.

So you are saying a sensor can power on and place its lanes in LP-11 in this new prepare_streaming() callback, the reason being, this can be called closer in time, and just before, the receiver DPHY is brought out of reset, in order for the receiver DPHY to be able to detect a possible early transition to HS mode?

While I don't have any objection to adding a prepare_streaming() callback (it could be useful for other reasons), I don't see why the sensor could not simply make use of existing s_power to power on and place the lanes in LP-11. That is what it's intended for anyway.

There's really no drawback to the patch I posted that moves receiver DPHY init and reset-release to s_power time in the receiver driver. If given a choice, an isolated fix to the imx mipi csi-2 receiver driver should be preferable to implementing a new callback.

But perhaps one advantage to the prepare_streaming() callback idea, is that there is more of a guarantee of minimum time between the sensor-power-on-and-enter-LP-11 event and receiver-DPHY-reset-release event, since s_power is normally called from v4l2_pipeline_pm_use() which could add extra overhead between the time s_power is called at the transmitter and at the receiver (but not much - the transmitter is connected directly to receiver in the graph).

Hi Steve,

Quoted 113 lines · Steve Longerbeam

Hi Sakari,

Quoted 87 lines · Sakari Ailus

Hi Steve,

Quoted 76 lines · Steve Longerbeam

Quoted 56 lines · Philipp Zabel

Hi Sakari,

On Thu, 2019-08-08 at 11:26 +0300, Sakari Ailus wrote:
[...]

Quoted 1 line

Quoted 1 line

Quoted 1 line

Have you checked how it works if you simply leave out this test?

Whether this works or not depends on the sensor used, and for some sensor/drivers may depend on timing (or random factors influencing it). See below.

[...]

Quoted 3 lines

Some devices can be commanded to enter LP-11 state but some will just briefly visit that state. The LP-11 state is mandatory but software should not be involved in detecting it if at all possible.

This is a good point. Devices that can be set to LP-11 state immediately, but that don't stay there long enough (either because they wait for less than the required by spec 100µs, or because system load causes this check to be executed too late) may end up working reliably even though the warning fires.

Quoted 5 lines

So if the hardware does not require further initialisation to be done in LP-11 state, you should remove the check.

Quoted 2 lines

We had to fix at least OV5645 and OV5640 recently because of this, and I can imagine more drivers will have the same issue.

Quoted 7 lines

This is actually an issue in the IMX driver (or hardware), not in the sensor driver. It may be that sometimes it's easier to work around it in the sensor driver.

So, I'd like to know whether the check itself is a driver bug, or something that the hardware requires. The fact that you're sending this patch suggests the former.

This is something that the hardware requires, according to the reference manual. See the comment in drivers/staging/media/imx/imx6-mipi-csi2.c, especially step 5.:

/*
 * The required sequence of MIPI CSI-2 startup as specified in the i.MX6
 * reference manual is as follows:
 *
 * 1. Deassert presetn signal (global reset).
 *        It's not clear what this "global reset" signal is (maybe APB *        global reset), but in any case this step would be probably *        be carried out during driver load in csi2_probe().
 *
 * 2. Configure MIPI Camera Sensor to put all Tx lanes in LP-11 state. *        This must be carried out by the MIPI sensor's s_power(ON) subdev *        op.
 *
 * 3. D-PHY initialization.
 * 4. CSI2 Controller programming (Set N_LANES, deassert PHY_SHUTDOWNZ, *    deassert PHY_RSTZ, deassert CSI2_RESETN).
 * 5. Read the PHY status register (PHY_STATE) to confirm that all data and *    clock lanes of the D-PHY are in LP-11 state.
 * 6. Configure the MIPI Camera Sensor to start transmitting a clock on the *    D-PHY clock lane.
 * 7. CSI2 Controller programming - Read the PHY status register (PHY_STATE) *    to confirm that the D-PHY is receiving a clock on the D-PHY clock lane.

Quoted 5 lines

    */

I read this as the hardware needing to see the LP-11 -> HS transition after the DPHY reset has been released, and before the CSI2 RX controller is programmed.

I think that's a fair assumption, and there's another paragraph at the end of that comment above that adds more evidence to that:

...
* All steps 3 through 7 are carried out by csi2_s_stream(ON) here. Step
* 6 is accomplished by calling the source subdev's s_stream(ON) between
* steps 5 and 7.
*/

So the driver is expecting that the LP-11 state persists until step 6, at which the LP-11 -> HS transition occurs when streaming is started at the transmitter.

Quoted 8 lines

But if the transmitter only stays in LP-11 state for the minimum 100 usec after it is powered on, and then _automatically_ transitions to HS, it's quite possible the LP-11 -> HS transition will happen long before the DPHY is taken out of reset. That's because the transmitter's s_power(ON) is called when the capture device is opened (via v4l2_pipeline_pm_use()), but the steps above are carried out when streaming starts, so userland would have to issue VIDIOC_STREAMON well within the 100 usec expires after open()'ing the device, for the receiver hardware to see the transition.

Quoted 4 lines

After powering on the sensor it sensor should stay in (software) standby mode until streaming is started. It should still have its lanes in LP-11 mode if the sensor supports it (as documented) until the sensor switches to the streaming mode.

Yes that's my understanding too.

Quoted 11 lines

Quoted 4 lines

Perhaps that would be an argument for moving steps 1 - 5 into the driver's s_power(ON) call, which would first call s_power(ON) to the transmitter and then immediately go through steps 1 - 5. Steps 6,7 would then remain in s_stream(ON).

Not all sensor drivers power on the sensor before starting streaming.

Perhaps we could add a prepare_streaming() callback (in absence of a better proposal) to address that? We'd also need a corresponding unprepare_streaming() callback as well --- to power off the sensor. I think this only should be done if the sensor can be switched to LP-11 explicitly; many simply don't support that.

So you are saying a sensor can power on and place its lanes in LP-11 in this new prepare_streaming() callback, the reason being, this can be called closer in time, and just before, the receiver DPHY is brought out of reset, in order for the receiver DPHY to be able to detect a possible early transition to HS mode?

Quoted 4 lines

While I don't have any objection to adding a prepare_streaming() callback (it could be useful for other reasons), I don't see why the sensor could not simply make use of existing s_power to power on and place the lanes in LP-11. That is what it's intended for anyway.

The vast majority of new drivers no longer have s_power callbacks --- they use runtime PM instead.

Quoted 12 lines

There's really no drawback to the patch I posted that moves receiver DPHY init and reset-release to s_power time in the receiver driver. If given a choice, an isolated fix to the imx mipi csi-2 receiver driver should be preferable to implementing a new callback.

But perhaps one advantage to the prepare_streaming() callback idea, is that there is more of a guarantee of minimum time between the sensor-power-on-and-enter-LP-11 event and receiver-DPHY-reset-release event, since s_power is normally called from v4l2_pipeline_pm_use() which could add extra overhead between the time s_power is called at the transmitter and at the receiver (but not much - the transmitter is connected directly to receiver in the graph).