On 8/14/19 12:16 AM, Sakari Ailus wrote:
Hi Steve, On Tue, Aug 13, 2019 at 04:27:06PM -0700, Steve Longerbeam wrote:Hi Sakari, On 8/13/19 1:28 AM, Sakari Ailus wrote:Hi Steve, On Thu, Aug 08, 2019 at 11:02:29AM -0700, Steve Longerbeam wrote:On 8/8/19 1:53 AM, Philipp Zabel wrote:Hi Sakari, On Thu, 2019-08-08 at 11:26 +0300, Sakari Ailus wrote: [...]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. [...]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.So if the hardware does not require further initialisation to be done in LP-11 state, you should remove the check.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.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.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.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.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? 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.
Ok I see that now. Most of them call pm_runtime_get at stream on, but some others in an internal_ops open() callback.
So never mind the patch I posted that moves receiver DPHY power on to .s_power(), that will actually make things worse for sensors that power on via runtime PM in their .s_stream op.
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).