Re: [PATCH v2 0/4] s32g: Use a syscon for GPR

[Dan Carpenter <dan.carpenter@linaro.org>]

On Mon, Dec 15, 2025 at 04:07:48PM -0500, Frank Li wrote:
> On Mon, Dec 15, 2025 at 11:11:03PM +0300, Dan Carpenter wrote:
> > On Mon, Dec 15, 2025 at 02:28:43PM -0500, Frank Li wrote:
> > > On Mon, Dec 15, 2025 at 09:33:54PM +0300, Dan Carpenter wrote:
> > > > On Mon, Dec 15, 2025 at 10:56:49AM -0500, Frank Li wrote:
> > > > > On Mon, Dec 15, 2025 at 05:41:43PM +0300, Dan Carpenter wrote:
> > > > > > The s32g devices have a GPR register region which holds a number of
> > > > > > miscellaneous registers.  Currently only the stmmac/dwmac-s32.c uses
> > > > > > anything from there and we just add a line to the device tree to
> > > > > > access that GMAC_0_CTRL_STS register:
> > > > > >
> > > > > >                         reg = <0x4033c000 0x2000>, /* gmac IP */
> > > > > >                               <0x4007c004 0x4>;    /* GMAC_0_CTRL_STS */
> > > > > >
> > > > > > We still have to maintain backwards compatibility to this format,
> > > > > > of course, but it would be better to access these through a syscon.
> > > > > > First of all, putting all the registers together is more organized
> > > > > > and shows how the hardware actually is implemented.  Secondly, in
> > > > > > some versions of this chipset those registers can only be accessed
> > > > > > via SCMI, if the registers aren't grouped together each driver will
> > > > > > have to create a whole lot of if then statements to access it via
> > > > > > IOMEM or via SCMI,
> > > > >
> > > > > Does SCMI work as regmap? syscon look likes simple, but missed abstract
> > > > > in overall.
> > > > >
> > > >
> > > > The SCMI part of this is pretty complicated and needs discussion.  It
> > > > might be that it requires a vendor extension.  Right now, the out of
> > > > tree code uses a nvmem vendor extension but that probably won't get
> > > > merged upstream.
> > > >
> > > > But in theory, it's fairly simple, you can write a regmap driver and
> > > > register it as a syscon and everything that was accessing nxp,phy-sel
> > > > accesses the same register but over SCMI.
> > >
> > > nxp,phy-sel is not standard API. Driver access raw register value. such
> > > as write 1 to offset 0x100.
> > >
> > > After change to SCMI, which may mapped to difference command. Even change
> > > to other SOC, value and offset also need be changed. It is not standilzed
> > > as what you expected.
> >
> > We're writing to an offset in a syscon.  Right now the device tree
> > says that the syscon is an MMIO syscon.  But for SCMI devices we
> > would point the phandle to a custom syscon.  The phandle and the offset
> > would stay the same, but how the syscon is implemented would change.
> 
> Your SCMI syscon driver will convert some private hard code to some
> function, such previous example's '1' as SEL_RGMII. It is hard maintained
> in long term.
> 

No, there isn't any conversion needed.  It's exactly the same as writing
to the register except it goes through SCMI.

> >
> > >
> > > >
> > > > > You still use regmap by use MMIO. /* GMAC_0_CTRL_STS */
> > > > >
> > > > > regmap = devm_regmap_init_mmio(dev, sts_offset, &regmap_config);
> > > > >
> > > >
> > > > You can use have an MMIO syscon, or you can create a custom driver
> > > > and register it as a syscon using of_syscon_register_regmap().
> > >
> > > My means is that it is not necessary to create nxp,phy-sel, especially
> > > there already have <0x4007c004 0x4>;    /* GMAC_0_CTRL_STS */
> > >
> >
> > Right now the out of tree dwmac-s32cc.c driver does something like
> > this:
> >
> >     89          if (gmac->use_nvmem) {
> >     90                  ret = write_nvmem_cell(gmac->dev, "gmac_phy_intf_sel", intf_sel);
> >     91                  if (ret)
> >     92                          return ret;
> >     93          } else {
> >     94                  writel(intf_sel, gmac->ctrl_sts);
> >     95          }
> >
> > Which is quite complicated, but with a syscon, then it's just:
> >
> > 	regmap_write(gmac->sts_regmap, gmac->sts_offset, S32_PHY_INTF_SEL_RGMII);
> >
> > Even without SCMI, the hardware has all these registers grouped together
> > it just feels cleaner to group them together in the device tree as well.
> 
> Why not implement standard phy interface,
> phy_set_mode_ext(PHY_MODE_ETHERNET, RGMII);
> 
> For example:  drivers/pci/controller/dwc/pci-imx6.c
> 
> In legency platform, it use syscon to set some registers. It becomes mess
> when more platform added.  And it becomes hard to convert because avoid
> break compatibltiy now.
> 
> It doesn't become worse since new platforms switched to use standard
> inteface, (phy, reset ...).
> 

This happens below that layer, this is just saying where the registers
are found.  The GMAC_0_CTRL_STS is just one register in the GPR region,
most of the others are unrelated to PHY.

regards,
dan carpenter