On 02/02/2023 16:12:33+0100, Johan Hovold wrote:
> On Fri, Jan 27, 2023 at 05:05:03PM +0100, Alexandre Belloni wrote:
> > On 27/01/2023 16:51:27+0100, Johan Hovold wrote:
>
> > > > > @@ -380,9 +478,23 @@ static int pm8xxx_rtc_probe(struct platform_device *pdev)
> > > > > rtc_dd->allow_set_time = of_property_read_bool(pdev->dev.of_node,
> > > > > "allow-set-time");
> > > > >
> > > > > + rtc_dd->nvmem_cell = devm_nvmem_cell_get(&pdev->dev, "offset");
> > > >
> > > > Maybe we should get something more specific than just "offset" so this
> > > > could be parsed in the RTC core at some point (this is the second RTC to
> > > > behave like this)
> > >
> > > Yes, that thought crossed my mind, but it's an nvmem cell name (label)
> > > and not a generic devicetree property. If you look at the binding
> > > document I think the name makes sense given the current description, and
> > > I'm not sure changing to something like 'base' would be much of an
> > > improvement.
> > >
> > > I also don't expect there to be more broken RTCs out there like these
> > > ones. Hopefully Qualcomm will even get this fixed at some point
> > > themselves.
> > >
> > > And I assume you were think of the old Atmel driver which uses a timer
> > > counter and a scratch register as a base? That one is also a bit
> > > different in that the timer can be reset, just not set.
> >
> > Nope, I'm thinking about the gamecube one and probably the nintendo
> > switch one which seems to behave similarly (no driver in the kernel
> > though).
>
> Found the gamecube one now (misread you comment above to imply that it
> was also out of tree).
>
> That one is also different in that the timer in that RTC can also be
> set (e.g. like the atmel one), but for consistency with some firmware an
> offset also needs to be read from SRAM (not NVRAM) and applied. That
> offset is also never updated by Linux.

Yeah, I deally, the gamecube counter shouldn't be updated, the switch
one doesn't seem to be updatable. I guess the idea being that games need
to be able to know if you are messing with the RTC to get an advantage.

--
Alexandre Belloni, co-owner and COO, Bootlin
Embedded Linux and Kernel engineering
https://bootlin.com

2023-02-09 16:59:31

by Ard Biesheuvel

[permalink] [raw]

Subject: Re: [RFC 18/24] dt-bindings: rtc: qcom-pm8xxx: add uefi-variable offset

On Wed, 1 Feb 2023 at 17:09, Johan Hovold <[email protected]> wrote:
>
> [ +CC: Ard ]
>
> On Mon, Jan 30, 2023 at 12:49:44PM -0600, Rob Herring wrote:
> > On Thu, Jan 26, 2023 at 03:20:51PM +0100, Johan Hovold wrote:
> > > On many Qualcomm platforms the PMIC RTC control and time registers are
> > > read-only so that the RTC time can not be updated. Instead an offset
> > > needs be stored in some machine-specific non-volatile memory, which a
> > > driver can take into account.
> > >
> > > Add a 'qcom,uefi-rtc-info' boolean flag which indicates that the RTC
> > > offset is stored in a Qualcomm specific UEFI variable so that the RTC
> > > time can be updated on such platforms.
> > >
> > > The UEFI variable is
> > >
> > > 882f8c2b-9646-435f-8de5-f208ff80c1bd-RTCInfo
> > >
> > > and holds a 12-byte structure where the first four bytes is a GPS time
> > > offset in little-endian byte order.
> >
> > Can't you just try to read the UEFI variable and use it if that
> > succeeds?
>
> Generally, yes. The problem here is that this UEFI variable is not used
> on all devices using these PMICs and I need a way to determine whether
> to wait for the UEFI variables to become available or not (e.g. when
> efivars support is built as module, yes, that's a thing now...).
>

Could we read this variable at boot and pass it to the kernel in a
different way? That way, we only have to defer the ability to set the
RTC, right?

> > I don't like this in DT because what if lots of devices start storing
> > lots of things in vendor specific UEFI variables. It doesn't scale.
>
> I hope we won't see that even if we already have some devices for x86
> platforms storing MAC addresses and such in UEFI variables. They
> currently access the UEFI firmware directly (i.e. not using the efivars
> abstraction) and simply assume UEFI is always there.
>
> With the Google SMI efivars implementation or the new Qualcomm SMC-based
> one, we need a way to determine whether to wait for efivars to become
> registered. For drivers where efivars is always needed we can just probe
> defer, but in this case we should not wait unless the DT indicates that
> the RTC offset is stored in UEFI on this particular machine.
>
> Just as the nvmem-cell property indicates that the offset is stored in
> some abstract nvmem, it seems reasonable to describe the offset being
> stored in UEFI when that is the case (even if it is indeed generally
> possible to probe for the latter).
>
> An alternative might be to describe the efivars fw dependency in DT too
> (e.g. for device links), but I believe you have already expressed some
> concerns over that:
>
> https://lore.kernel.org/lkml/[email protected]/
>
> Johan