Re: [PATCH v8 08/10] gpu: host1x: mipi: Keep MIPI clock enabled till calibration is done

[Dmitry Osipenko]

05.08.2020 20:04, Sowjanya Komatineni пишет:
> 
> On 8/5/20 9:57 AM, Dmitry Osipenko wrote:
>> 05.08.2020 19:50, Sowjanya Komatineni пишет:
>>> On 8/5/20 9:47 AM, Dmitry Osipenko wrote:
>>>> 05.08.2020 19:33, Sowjanya Komatineni пишет:
>>>>> On 8/5/20 7:19 AM, Dmitry Osipenko wrote:
>>>>>> 05.08.2020 17:05, Dmitry Osipenko пишет:
>>>>>>> 05.08.2020 16:46, Thierry Reding пишет:
>>>>>>>> On Mon, Aug 03, 2020 at 08:42:24AM -0700, Sowjanya Komatineni
>>>>>>>> wrote:
>>>>>>>>> With the split of MIPI calibration into tegra_mipi_calibrate() and
>>>>>>>>> tegra_mipi_wait(), MIPI clock is not kept enabled till the
>>>>>>>>> calibration
>>>>>>>>> is done.
>>>>>>>>>
>>>>>>>>> So, this patch skips disabling MIPI clock after triggering
>>>>>>>>> start of
>>>>>>>>> calibration and disables it only after waiting for done status
>>>>>>>>> from
>>>>>>>>> the calibration logic.
>>>>>>>>>
>>>>>>>>> This patch renames tegra_mipi_calibrate() as
>>>>>>>>> tegra_mipi_start_calibration()
>>>>>>>>> and tegra_mipi_wait() as tegra_mipi_finish_calibration() to be
>>>>>>>>> inline
>>>>>>>>> with their usage.
>>>>>>>>>
>>>>>>>>> As MIPI clock is left enabled and in case of any failures with CSI
>>>>>>>>> input
>>>>>>>>> streaming tegra_mipi_finish_calibration() will not get invoked.
>>>>>>>>> So added new API tegra_mipi_cancel_calibration() which disables
>>>>>>>>> MIPI clock
>>>>>>>>> and consumer drivers can call this in such cases.
>>>>>>>>>
>>>>>>>>> Reviewed-by: Dmitry Osipenko <digetx@xxxxxxxxx>
>>>>>>>>> Signed-off-by: Sowjanya Komatineni <skomatineni@xxxxxxxxxx>
>>>>>>>>> ---
>>>>>>>>>     drivers/gpu/drm/tegra/dsi.c |  4 ++--
>>>>>>>>>     drivers/gpu/host1x/mipi.c   | 19 ++++++++++---------
>>>>>>>>>     include/linux/host1x.h      |  5 +++--
>>>>>>>>>     3 files changed, 15 insertions(+), 13 deletions(-)
>>>>>>>>>
>>>>>>>>> diff --git a/drivers/gpu/drm/tegra/dsi.c
>>>>>>>>> b/drivers/gpu/drm/tegra/dsi.c
>>>>>>>>> index 3820e8d..a7864e9 100644
>>>>>>>>> --- a/drivers/gpu/drm/tegra/dsi.c
>>>>>>>>> +++ b/drivers/gpu/drm/tegra/dsi.c
>>>>>>>>> @@ -694,11 +694,11 @@ static int tegra_dsi_pad_calibrate(struct
>>>>>>>>> tegra_dsi *dsi)
>>>>>>>>>             DSI_PAD_PREEMP_PD(0x03) | DSI_PAD_PREEMP_PU(0x3);
>>>>>>>>>         tegra_dsi_writel(dsi, value, DSI_PAD_CONTROL_3);
>>>>>>>>>     -    err = tegra_mipi_calibrate(dsi->mipi);
>>>>>>>>> +    err = tegra_mipi_start_calibration(dsi->mipi);
>>>>>>>>>         if (err < 0)
>>>>>>>>>             return err;
>>>>>>>>>     -    return tegra_mipi_wait(dsi->mipi);
>>>>>>>>> +    return tegra_mipi_finish_calibration(dsi->mipi);
>>>>>>>>>     }
>>>>>>>>>       static void tegra_dsi_set_timeout(struct tegra_dsi *dsi,
>>>>>>>>> unsigned long bclk,
>>>>>>>>> diff --git a/drivers/gpu/host1x/mipi.c b/drivers/gpu/host1x/mipi.c
>>>>>>>>> index e606464..b15ab6e 100644
>>>>>>>>> --- a/drivers/gpu/host1x/mipi.c
>>>>>>>>> +++ b/drivers/gpu/host1x/mipi.c
>>>>>>>>> @@ -293,17 +293,19 @@ int tegra_mipi_disable(struct
>>>>>>>>> tegra_mipi_device *dev)
>>>>>>>>>     }
>>>>>>>>>     EXPORT_SYMBOL(tegra_mipi_disable);
>>>>>>>>>     -int tegra_mipi_wait(struct tegra_mipi_device *device)
>>>>>>>>> +void tegra_mipi_cancel_calibration(struct tegra_mipi_device
>>>>>>>>> *device)
>>>>>>>>> +{
>>>>>>>>> +    clk_disable(device->mipi->clk);
>>>>>>>> Do we need to do anything with the MIPI_CAL_CTRL and
>>>>>>>> MIPI_CAL_STATUS
>>>>>>>> registers here? We don't clear the START bit in the former when the
>>>>>>>> calibration has successfully finished, but I suspect that's because
>>>>>>>> the bit is self-clearing. But I wonder if we still need to clear it
>>>>>>>> upon cancellation to make sure the calibration does indeed stop.
>>>>>>> Apparently there is no way to explicitly stop calibration other
>>>>>>> than to
>>>>>>> reset MIPI calibration block, but Sowjanya says this is unnecessary.
>>>>>>>
>>>>>>> Perhaps having a fixed delay before disabling clock could be
>>>>>>> enough to
>>>>>>> ensure that calibration is stopped before the clock is disabled?
>>>>>>>
>>>>>> Actually, there is a MIPI_CAL_ACTIVE bit in the status register.
>>>>>> Maybe
>>>>>> it needs to be polled until it's unset?
>>>>> Confirmed with HW design team during this patch update.
>>>>>
>>>>> SW does not need to clear START bit and only write 1 takes effect to
>>>>> that bit.
>>>>>
>>>>> Also, no need to have delay or do any other register settings
>>>>> unclear as
>>>>> its FSM and there's nothing to get stuck.
>>>>>
>>>>> Also it goes thru small finite set of codes and by the time sensor
>>>>> programming happens for enabling streaming FSM will finish its
>>>>> calibration sequence much early and it will only wait for pads LP-11.
>>>>>
>>>>> So, during cancel we only need disable MIPI clock.
>>>>>
>>>> But there is no guarantee that cancel_calibration() couldn't be invoked
>>>> in the middle of the calibration process, hence FSM could freeze in an
>>>> intermediate state if it's running on the disabled MIPI clock, this
>>>> doesn't sound good.
>>> Actual calibration logic uses UART_FST_CAL clock which is always enabled
>> What enables the UART_FST_CAL clock? I don't see this clock used
>> anywhere.
> 
> UART_FST_MIPI_CAL is shared with uart and is always enabled all the time.
> 
> I don't see mipi driver handling this and I think that's because this
> clock is enabled all the time as its used for UART as well. Probably
> thierry can comment on this clock.

It's not only the MIPI driver, the clock isn't defined at all neither in
the clk driver, nor in clk DT bindings.

It could be fragile to assume that it's always enabled.

> Also regarding cancel calibration, as FSM goes thru only finite sequence
> codes by the time csi stream and sensor stream happens which is where we
> check for calibration to complete for sure calibration will be finished
> and calibration logic will only wait for pads to be in LP-11 to apply
> results. So nothing special needed during cancel except to turn clock
> off to balance its usage count.

I guess it should be okay for the case of the VI driver, but
in general please don't assume that code can't change in the future. The
common API always should be made reliable for all possible situations.