Re: [PATCH 06/13] phy: qcom-qmp-pcie: drop bogus register update

[Dmitry Baryshkov <dmitry.baryshkov@linaro.org>]

On 29/09/2022 10:12, Johan Hovold wrote:
> On Wed, Sep 28, 2022 at 10:48:40PM +0300, Dmitry Baryshkov wrote:
>> On 28/09/2022 22:10, Dmitry Baryshkov wrote:
>>> On 28/09/2022 18:28, Johan Hovold wrote:
>>>> Since commit 0d58280cf1e6 ("phy: Update PHY power control sequence") the
>>>> PHY is powered on before configuring the registers and only the MSM8996
>>>> PCIe PHY, which includes the POWER_DOWN_CONTROL register in its PCS
>>>> initialisation table, may possibly require a second update afterwards.
>>>>
>>>> To make things worse, the POWER_DOWN_CONTROL register lies at a
>>>> different offset on more recent SoCs so that the second update, which
>>>> still used a hard-coded offset, would write to an unrelated register
>>>> (e.g. a revision-id register on SC8280XP).
>>>>
>>>> As the MSM8996 PCIe PHY is now handled by a separate driver, simply drop
>>>> the bogus register update.
>>>>
>>>> Fixes: e4d8b05ad5f9 ("phy: qcom-qmp: Use proper PWRDOWN offset for
>>>> sm8150 USB") added support
>>>
>>> I'm not sure about the particular fixes tag. Backporting from the split
>>> driver into old qmp driver would be a complete pain.
>>>
>>>> Signed-off-by: Johan Hovold <johan+linaro@kernel.org>
>>>
>>> Reviewed-by: Dmitry Baryshkov <dmitry.baryshkov@linaro.org>
>>
>> After digging some more, I stumbled upon the commit 0d58280cf1e6 ("phy:
>> Update PHY power control sequence"), which puts explicit register write
>> here, telling that 'PCIe PHYs need an extra power control before
>> deasserts reset state'.
> 
> That's the commit I'm referring to above.
> 
>> I can confirm this with the register tables from downstream dtsi. E.g.
>> consider sdm845-pcie.dts, pcie0 table. The PCS_POWER_DOWN_CONTROL is the
>> register 0x804.
>>
>> The programmings starts with <0x804 0x1 0x0>, writing 1 to
>> PCS_POWER_DOWN_CONTROL (which if I'm not mistaken we do not do at this
>> moment). Then after writing all the serdes/tx/rx/pcs/pcs_misc tables
>> comes the write <0x804 0x3 0x0> (which you are trying to remove here).
> 
> The PHY would already have been powered on with the mainline driver, that
> write has already happened.
> 
> Whether or not PCIe support for SDM845 has been broken since it was
> first mainlined almost three years ago is a separate issue. I assume
> Bjorn tested it before sending it upstream.
> 
> 	421c9a0e9731 ("phy: qcom: qmp: Add SDM845 PCIe QMP PHY support")

On SDM845 PCIe0 is used for the WiFi, e.g. on the RB3 (dragonboard845) 
device. The PHY definitely works with the upstream kernels.

> 
>> Same sequence applies to the PCIe PHY on msm8998.
>>
>> Most newer PHYs have the expected sequence (of writing 0x3 to
>> PCS_POWER_DOWN_CONTROL) before writing all registers.
>>
>> As a short summary: unless we get any additional information that 8998
>> and sdm845 tables are incorrect, I'd suggest adding a conditional here
>> (ugh) and using it here and in qmp_pcie_init() call.
> 
> I see little point in doing that unless you dig out an SDM845, confirm
> that it has never worked with upstream, and update the init sequence
> first.

Digging out an sdm845 is not a problem per se. It works, but it also has 
an additional regwrite that this patch tries to remove. I will try 
checking whether it works with this register write being removed or not.

> 
>> Vinod, Bjorn, do you have any additional info?
> 
> An explanation of the split POWER_DOWN_CONTROL updates on MSM8996 would
> be good to have either way

-- 
With best wishes
Dmitry