pci-exynos.c phy_init() usage

[Bjorn Helgaas]

On Tue, Jun 28, 2022 at 10:27:31AM +0200, Krzysztof Kozlowski wrote:
> On 28/06/2022 10:13, Marek Szyprowski wrote:
> > On 27.06.2022 12:47, Krzysztof Kozlowski wrote:
> >> On 27/06/2022 12:30, Marek Szyprowski wrote:
> >>> On 24.06.2022 20:07, Krzysztof Kozlowski wrote:
> >>>> On 24/06/2022 19:35, Bjorn Helgaas wrote:
> >>>>> In exynos_pcie_host_init() [1], we call:
> >>>>>
> >>>>>     phy_reset(ep->phy);
> >>>>>     phy_power_on(ep->phy);
> >>>>>     phy_init(ep->phy);
> >>>>>
> >>>>> The phy_init() function comment [2] says it must be called before
> >>>>> phy_power_on().  Is exynos doing this backwards?
> >>>> Looks like. I don't have Exynos hardware with a PCI, so cannot
> >>>> test/fix/verify.
> >>>>
> >>>> Luckily for Exynos ;-) it's not alone in this pattern:
> >>>> drivers/net/ethernet/marvell/sky2.c
> >>>> drivers/usb/dwc2/platform.c
> >>> I've checked that on the real hardware. Swapping the order of
> >>> phy_power_on and phy_init breaks driver operation.
> >>>
> >>> However pci-exynos is the only driver that uses the phy-exynos-pcie, so
> >>> we can simply swap the content of the init and power_on in the phy
> >>> driver to adjust the code to the right order. power_on/init and
> >>> exit/power_off are also called one after the other in pci-exynos,
> >>> without any activity between them, so we can also simply move all
> >>> operation to one pair of the callback, like power_on/off.
> >>>
> >>> Krzysztof, which solution would you prefer?
> >> I think the real problem is that the Exynos PCIe phy init
> >> (exynos5433_pcie_phy_init) performs parts of power on procedure, so the
> >> code is mixed. Probably also the phy init could not happen earlier due
> >> to gated clocks (ungated in exynos5433_pcie_phy_power_on).
> >>
> >> I would prefer to clean it up while ordering init+power_on, so figure
> >> out more or less correct procedure.
> >>
> >> You can also look at Artpec-8 PHY - it seems using correct order
> >> (init+reset):
> >> https://lore.kernel.org/all/20220614011616epcms2p7dcaa67c53b7df5802dd7a697e2d472d7@epcms2p7/
> > 
> > I've played a bit with those register writes in exynos_pcie_phy and 
> > frankly speaking the currenly used (power_on + init) is the only 
> > sequence that works properly. I'm leaning to move everything to 
> > phy_init/exit. I really don't see how to split it into init + power_on 
> > callbacks.
> 
> I was afraid it will be like this. I imagine that certain (not
> explicitly documented) init operations cannot even happen before power
> on, so this would be a lot of tries.
> 
> I am fine with it. Thanks for doing it.

If nothing can be improved, a comment to this effect might make it
look less like a mistake.

Bjorn