Revert "phy: qualcomm: usb28nm: Add MDM9607 init sequence"

  This reverts commit 557a28811c7e0286d3816842032db5eb7bb5f156.

This commit introduced an init sequence from downstream DT [1] in the
driver.  As mentioned by the comment above the HSPHY_INIT_CFG macro for
this sequence:

    /*
     * The macro is used to define an initialization sequence.  Each tuple
     * is meant to program 'value' into phy register at 'offset' with 'delay'
     * in us followed.
     */

Instead of corresponding to offsets into the phy register, the sequence
read by the downstream driver [2] is passed into ulpi_write [3] which
crafts the address-value pair into a new value and writes it into the
same register at USB_ULPI_VIEWPORT [4].  In other words, this init
sequence is programmed into the hardware in a totally different way than
downstream and is unlikely to achieve the desired result, if the hsphy
is working at all.

An alternative method needs to be found to write these init values at
the desired location.  Fortunately mdm9607 did not land upstream yet [5]
and should have its compatible revised to use the generic one, instead
of a compatible that writes wrong data to the wrong registers.

[1]: https://android.googlesource.com/kernel/msm/+/android-7.1.0_r0.2/arch/arm/boot/dts/qcom/mdm9607.dtsi#585
[2]: https://android.googlesource.com/kernel/msm/+/android-7.1.0_r0.2/drivers/usb/phy/phy-msm-usb.c#4183
[3]: https://android.googlesource.com/kernel/msm/+/android-7.1.0_r0.2/drivers/usb/phy/phy-msm-usb.c#468
[4]: https://android.googlesource.com/kernel/msm/+/android-7.1.0_r0.2/drivers/usb/phy/phy-msm-usb.c#418
[5]: https://lore.kernel.org/linux-arm-msm/20210805222812.40731-1-konrad.dybcio@somainline.org/

Reported-by: Michael Srba <Michael.Srba@seznam.cz>
Signed-off-by: Marijn Suijten <marijn.suijten@somainline.org>
---
 .../devicetree/bindings/phy/qcom,usb-hs-28nm.yaml   |  1 -
 drivers/phy/qualcomm/phy-qcom-usb-hs-28nm.c         | 13 -------------
 2 files changed, 14 deletions(-)
  

On 14/12/2022 22:37, Marijn Suijten wrote:
> This reverts commit 557a28811c7e0286d3816842032db5eb7bb5f156.
> 
> This commit introduced an init sequence from downstream DT [1] in the
> driver.  As mentioned by the comment above the HSPHY_INIT_CFG macro for
> this sequence:
> 
>      /*
>       * The macro is used to define an initialization sequence.  Each tuple
>       * is meant to program 'value' into phy register at 'offset' with 'delay'
>       * in us followed.
>       */
> 
> Instead of corresponding to offsets into the phy register, the sequence
> read by the downstream driver [2] is passed into ulpi_write [3] which
> crafts the address-value pair into a new value and writes it into the
> same register at USB_ULPI_VIEWPORT [4].  In other words, this init
> sequence is programmed into the hardware in a totally different way than
> downstream and is unlikely to achieve the desired result, if the hsphy
> is working at all.
> 
> An alternative method needs to be found to write these init values at
> the desired location.  Fortunately mdm9607 did not land upstream yet [5]
> and should have its compatible revised to use the generic one, instead
> of a compatible that writes wrong data to the wrong registers.
> 
> [1]: https://android.googlesource.com/kernel/msm/+/android-7.1.0_r0.2/arch/arm/boot/dts/qcom/mdm9607.dtsi#585
> [2]: https://android.googlesource.com/kernel/msm/+/android-7.1.0_r0.2/drivers/usb/phy/phy-msm-usb.c#4183
> [3]: https://android.googlesource.com/kernel/msm/+/android-7.1.0_r0.2/drivers/usb/phy/phy-msm-usb.c#468
> [4]: https://android.googlesource.com/kernel/msm/+/android-7.1.0_r0.2/drivers/usb/phy/phy-msm-usb.c#418
> [5]: https://lore.kernel.org/linux-arm-msm/20210805222812.40731-1-konrad.dybcio@somainline.org/
> 
> Reported-by: Michael Srba <Michael.Srba@seznam.cz>
> Signed-off-by: Marijn Suijten <marijn.suijten@somainline.org>

Hmm.

I got a bit concerned qcs405 was broken too, which would be worrying 
since I remember testing this code, though not specifically turning off 
the PHY init, then again, there's a near zero chance the USB PHY would 
work after reset without the specified init seq.

The original upstreamed code for qcs404/405

https://android.googlesource.com/kernel/msm/+/refs/heads/android-msm-coral-4.14-android10/arch/arm64/boot/dts/qcom/qcs405-usb.dtsi

https://android.googlesource.com/kernel/msm/+/refs/heads/android-msm-coral-4.14-android10/drivers/usb/phy/phy-qcom-snps-28nm-hs.c

Does a relative write to an offset of the PHY CSR

CSR base is 0x7a000
https://android.googlesource.com/kernel/msm/+/refs/heads/android-msm-coral-4.14-android10/arch/arm64/boot/dts/qcom/qcs405-usb.dtsi#74

https://android.googlesource.com/kernel/msm/+/refs/heads/android-msm-coral-4.14-android10/drivers/usb/phy/phy-qcom-snps-28nm-hs.c#285

which will result in 0x7a000 + 0xc0 = 0x01, etc

which for us upstream then is

https://github.com/torvalds/linux/blob/master/arch/arm64/boot/dts/qcom/qcs404.dtsi#L336

and

https://github.com/torvalds/linux/blob/master/drivers/phy/qualcomm/phy-qcom-usb-hs-28nm.c#L388

Writing 0x7a000 + 0xc0 = 0x01, etc - a writel() in this case

so that bit is fine.

mdm9607

https://android.googlesource.com/kernel/msm/+/android-7.1.0_r0.2/arch/arm/boot/dts/qcom/mdm9607.dtsi#569 
- compat string is

CSR is @ 0x6c000

https://android.googlesource.com/kernel/msm/+/android-7.1.0_r0.2/drivers/usb/phy/phy-msm-usb.c#664

downstream PHY init seq

https://android.googlesource.com/kernel/msm/+/android-7.1.0_r0.2/arch/arm/boot/dts/qcom/mdm9607.dtsi#585

and downstream driver PHY init write

https://android.googlesource.com/kernel/msm/+/android-7.1.0_r0.2/drivers/usb/phy/phy-msm-usb.c#664

Which does

writel_relaxed(ULPI_RUN | ULPI_WRITE |
	       ULPI_ADDR(reg) | ULPI_DATA(val),
	       USB_ULPI_VIEWPORT);

Yep, you're right, we can't do a simple writeb(csr + reg, val); for 
mdm9607 but, also the qcs404 => compat = "qcom,usb-hs-28nm-femtophy" is 
doing the right thing.

Reviewed-by: Bryan O'Donoghue <bryan.odonoghue@linaro.org>
  

On Wed, Dec 14, 2022 at 11:37:32PM +0100, Marijn Suijten wrote:
> This reverts commit 557a28811c7e0286d3816842032db5eb7bb5f156.
> 
> This commit introduced an init sequence from downstream DT [1] in the
> driver.  As mentioned by the comment above the HSPHY_INIT_CFG macro for
> this sequence:
> 
>     /*
>      * The macro is used to define an initialization sequence.  Each tuple
>      * is meant to program 'value' into phy register at 'offset' with 'delay'
>      * in us followed.
>      */
> 
> Instead of corresponding to offsets into the phy register, the sequence
> read by the downstream driver [2] is passed into ulpi_write [3] which
> crafts the address-value pair into a new value and writes it into the
> same register at USB_ULPI_VIEWPORT [4].  In other words, this init
> sequence is programmed into the hardware in a totally different way than
> downstream and is unlikely to achieve the desired result, if the hsphy
> is working at all.
> 

Thanks for sending this, I've also been meaning to do this for quite
some time :)

Reviewed-by: Stephan Gerhold <stephan@gerhold.net>

> An alternative method needs to be found to write these init values at
> the desired location.  Fortunately mdm9607 did not land upstream yet [5]
> and should have its compatible revised to use the generic one, instead
> of a compatible that writes wrong data to the wrong registers.
> 

There is a simple solution to write the init values correctly: You can
just use the ULPI PHY driver for MSM8916 (phy-qcom-usb-hs.c), which
writes those init values correctly.

If you look closely at downstream you can see that all targets with the
"SNPS Femto PHY" (at least MSM8909, MDM9607, MSM8976) actually use a
mixture of the code currently implemented in the mainline ULPI PHY
driver (phy-qcom-usb-hs.c) and the actual Femto PHY driver
(phy-qcom-usb-hs-28nm.c):

 - The device trees contains "qcom,dp-manual-pullup", which enables the
   ULPI_MISC_A_VBUSVLDEXT magic that is currently implemented partly in
   phy-qcom-usb-hs.c and partly in ci_hdrc_msm.c: 
     - Downstream: https://git.codelinaro.org/clo/la/kernel/msm-4.9/-/blob/29f81c613f4c853f4125ef189ede1f6d610653c0/drivers/usb/gadget/ci13xxx_msm.c#L113
     - Mainline: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/phy/qualcomm/phy-qcom-usb-hs.c?h=v6.1#n92
       https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/usb/chipidea/ci_hdrc_msm.c?h=v6.1#n117

  - The ULPI init sequence is also implemented by phy-qcom-usb-hs.c in
    mainline: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/phy/qualcomm/phy-qcom-usb-hs.c?h=v6.1#n144

  - The PHY CSR registers implemented by phy-qcom-usb-hs-28nm.c are only
    used to enter retention mode downstream (I guess some low-power mode
    where an USB device/host could still wake up the other host/device).

We don't have support for proper support for retention on mainline
(phy-qcom-usb-hs-28nm.c never enables the retention mode without
 fully powering off the PHY immediately after using the regulators).

So I suggest that we simply use the ULPI PHY driver for MSM8916 at least
for MSM8909 and MDM9607 (haven't checked MSM8976 in detail). I have
tried it and it works without any problems on both MSM8909 and MDM9607.
No driver changes needed!

Thanks,
Stephan
  

On 14-12-22, 23:37, Marijn Suijten wrote:
> This reverts commit 557a28811c7e0286d3816842032db5eb7bb5f156.
> 
> This commit introduced an init sequence from downstream DT [1] in the
> driver.  As mentioned by the comment above the HSPHY_INIT_CFG macro for
> this sequence:
> 
>     /*
>      * The macro is used to define an initialization sequence.  Each tuple
>      * is meant to program 'value' into phy register at 'offset' with 'delay'
>      * in us followed.
>      */
> 
> Instead of corresponding to offsets into the phy register, the sequence
> read by the downstream driver [2] is passed into ulpi_write [3] which
> crafts the address-value pair into a new value and writes it into the
> same register at USB_ULPI_VIEWPORT [4].  In other words, this init
> sequence is programmed into the hardware in a totally different way than
> downstream and is unlikely to achieve the desired result, if the hsphy
> is working at all.
> 
> An alternative method needs to be found to write these init values at
> the desired location.  Fortunately mdm9607 did not land upstream yet [5]
> and should have its compatible revised to use the generic one, instead
> of a compatible that writes wrong data to the wrong registers.

Applied after adding missing subsystem tag, thanks
  

On 2023-01-13 23:16:15, Vinod Koul wrote:
> On 14-12-22, 23:37, Marijn Suijten wrote:
> > This reverts commit 557a28811c7e0286d3816842032db5eb7bb5f156.
> > 
> > This commit introduced an init sequence from downstream DT [1] in the
> > driver.  As mentioned by the comment above the HSPHY_INIT_CFG macro for
> > this sequence:
> > 
> >     /*
> >      * The macro is used to define an initialization sequence.  Each tuple
> >      * is meant to program 'value' into phy register at 'offset' with 'delay'
> >      * in us followed.
> >      */
> > 
> > Instead of corresponding to offsets into the phy register, the sequence
> > read by the downstream driver [2] is passed into ulpi_write [3] which
> > crafts the address-value pair into a new value and writes it into the
> > same register at USB_ULPI_VIEWPORT [4].  In other words, this init
> > sequence is programmed into the hardware in a totally different way than
> > downstream and is unlikely to achieve the desired result, if the hsphy
> > is working at all.
> > 
> > An alternative method needs to be found to write these init values at
> > the desired location.  Fortunately mdm9607 did not land upstream yet [5]
> > and should have its compatible revised to use the generic one, instead
> > of a compatible that writes wrong data to the wrong registers.
> 
> Applied after adding missing subsystem tag, thanks

Thanks, it wasn't clear to me whether to suffix the title when already
included in the Revert: "phy: qualcomm: ..." title :)

- Marijn
  

On 16-01-23, 21:35, Marijn Suijten wrote:
> On 2023-01-13 23:16:15, Vinod Koul wrote:
> > On 14-12-22, 23:37, Marijn Suijten wrote:
> > > This reverts commit 557a28811c7e0286d3816842032db5eb7bb5f156.
> > > 
> > > This commit introduced an init sequence from downstream DT [1] in the
> > > driver.  As mentioned by the comment above the HSPHY_INIT_CFG macro for
> > > this sequence:
> > > 
> > >     /*
> > >      * The macro is used to define an initialization sequence.  Each tuple
> > >      * is meant to program 'value' into phy register at 'offset' with 'delay'
> > >      * in us followed.
> > >      */
> > > 
> > > Instead of corresponding to offsets into the phy register, the sequence
> > > read by the downstream driver [2] is passed into ulpi_write [3] which
> > > crafts the address-value pair into a new value and writes it into the
> > > same register at USB_ULPI_VIEWPORT [4].  In other words, this init
> > > sequence is programmed into the hardware in a totally different way than
> > > downstream and is unlikely to achieve the desired result, if the hsphy
> > > is working at all.
> > > 
> > > An alternative method needs to be found to write these init values at
> > > the desired location.  Fortunately mdm9607 did not land upstream yet [5]
> > > and should have its compatible revised to use the generic one, instead
> > > of a compatible that writes wrong data to the wrong registers.
> > 
> > Applied after adding missing subsystem tag, thanks
> 
> Thanks, it wasn't clear to me whether to suffix the title when already
> included in the Revert: "phy: qualcomm: ..." title :)

A revert patch is a patch as well so the patch rules apply there as well,
so should say "subsystem tag: other tags: Revert foo..."

Thanks
  

On 2023-01-17 11:06:12, Vinod Koul wrote:
<snip>
> > Thanks, it wasn't clear to me whether to suffix the title when already
> > included in the Revert: "phy: qualcomm: ..." title :)
> 
> A revert patch is a patch as well so the patch rules apply there as well,
> so should say "subsystem tag: other tags: Revert foo..."

Ack, but then /keep/ "subsystem tag: other tags:" /within/ the Revert
string, so "phy: qualcomm: Revert "phy: qualcomm: ...""?

- Marijn
  

On 17-01-23, 10:19, Marijn Suijten wrote:
> On 2023-01-17 11:06:12, Vinod Koul wrote:
> <snip>
> > > Thanks, it wasn't clear to me whether to suffix the title when already
> > > included in the Revert: "phy: qualcomm: ..." title :)
> > 
> > A revert patch is a patch as well so the patch rules apply there as well,
> > so should say "subsystem tag: other tags: Revert foo..."
> 
> Ack, but then /keep/ "subsystem tag: other tags:" /within/ the Revert
> string, so "phy: qualcomm: Revert "phy: qualcomm: ...""?

Ideally yes!