Re: a case for a common efuse API?

[Alexandre Belloni <alexandre.belloni@free-electrons.com>]

Hi,

On 10/07/2014 at 16:26:16 +0200, Maxime Ripard wrote :
> On Wed, Jul 09, 2014 at 04:32:03PM -0700, Stephen Boyd wrote:
> > On 07/09/14 01:35, Maxime Ripard wrote:
> > > Hi Stephen,
> > >
> > > On Tue, Jul 08, 2014 at 01:00:23PM -0700, Stephen Boyd wrote:
> > >> Hi,
> > >>
> > >> On MSM chips we have some efuses (called qfprom) where we store things
> > >> like calibration data, speed bins, etc. We need to read out data from
> > >> the efuses in various drivers like the cpufreq, thermal, etc. This
> > >> essentially boils down to a bunch of readls on the efuse from a handful
> > >> of different drivers. In devicetree this looks a little odd because
> > >> these drivers end up having an extra reg property (or two) that points
> > >> to a register in the efuse and some length, i.e you see this:
> > >>
> > >> 	thermal-sensor@34000 {
> > >> 		compatible = "sensor";
> > >> 		reg = <0x34000 0x1000>, <0x10018 0xc>;
> > >> 		reg-names = "sensor", "efuse_calib";
> > >> 	}
> > >>
> > >>
> > >> I imagine in DT we want something more like this:
> > >>
> > >> 	efuse: efuse@10000 {
> > >> 		compatible = "efuse";
> > >> 		reg = <0x10000 0x1000>;
> > >> 	}
> > >>
> > >> 	thermal-sensor@34000 {
> > >> 		compatible = "sensor";
> > >> 		reg = <0x34000 0x1000>;
> > >> 		efuse = <&efuse 0x18>;
> > >> 	}
> > > We have pretty much the same things in the Allwinner SoCs. We have an
> > > efuse directly mapped into memory, with a few informations like a MAC
> > > address, the SoC ID, the serial number, some RSA keys for the device,
> > > etc.
> > >
> > > The thing is, some boards expose these informations in an external
> > > EEPROM as well.
> > >
> > > I started working and went quite far to create an "eeprom" framework
> > > to handle these cases, with a dt representation similar to what you
> > > were exposing.
> > >
> > > https://github.com/mripard/linux/tree/eeprom-framework-at24
> > >
> > > It was working quite well, I was about to send it, but was told that I
> > > should all be moved to MTD, and given up on it.
> > 
> > Did anything ever get merged? Or the whole thing was dropped?
> 
> Nope, I just never posted it. I could send it as an RFC though, and
> see what are the feedbacks.
> 
> > That branch looks like what I want, assuming we could get an agreement
> > on the binding. It looks like pretty much every SoC has this, and there
> > isn't any API or binding for it that I've seen. The only thing I see is
> > Documentation/devicetree/bindings/eeprom.txt and that doesn't cover the
> > client aspect at all.
> > 
> > Taking a quick peek at the code, it might be better to change the read
> > API to take a buffer and length, so that the caller doesn't need to free
> > the data allocated by the eeprom layer. It also makes it symmetrical
> > with the write API. We'd probably also need to make it work really early
> > for SoC's like Tegra where we want to read the SoC revision early. So
> > probably split off the device registration part to a later time to allow
> > register() to be called early.
> 
> I guess that the kind of things we could discuss after posting these
> patches, but yep, it looks reasonnable.
> 
> I'll try to get things a bit cleaner, and post them in the next days.
> 

Be aware that some SoCs are storing their OPPs there so this would be
useful if that framework is available early enough to register those to
cpufreq.


-- 
Alexandre Belloni, Free Electrons
Embedded Linux, Kernel and Android engineering
http://free-electrons.com