On Wed, Mar 13, 2019 at 08:19:41AM -0700, Guenter Roeck wrote:
> On Wed, Mar 13, 2019 at 12:31:53PM +0000, Grönke, Christian wrote:
> > Hello all,
> >
> > I am currently working on a hwmon/pmbus driver for a PMBus capable 3Y
> > Power/FSP power supply (YH5301-1EAR).
> >
> > The datasheet has limited information. But with the help of some online
> > sources, other datasheets from the vendor and the pmbus sub-system I could
> > figure out most of it.
> >
> > However, I did run in some troubles with the VOUT values. To my
> > understanding (and from what I could validate) the device does encode _all_
> > values as "LINEAR11". Meaning all values have a mantissa of 11 bit and the
> > exponent in the other 5 bits of the register word. This causes some
> > troubles with the read out of the VOUT registers (e.g. READ_VOUT). The
> > pmbus subsystem expects these values to be in "LINEAR16" encoding. Hence
> > the full word register is the mantissa and the exponent is supposed to be
> > in VOUT_MODE.
> > Sadly, the VOUT_MODE register isn't supported. It reads 0xFF.
> >
> ... violating its own specification...
>
> > In my first attempt to work around this, I provided a custom read_word_data
> > function that would fixup the value. However, that did lead to problems
> > with negative values and also had a precision loss (12,09V -> 12V). I tried
> > to compensate by faking the values as 'direct' and adjusting the m/b/R
> > values to match. This is also not perfect, as it is messy and also seems to
> > report the wrong values in some cases.
>
> Messy is relative. Polluting generic code is just as messy.
> What are those cases where a wrong value is reported ?
>
> >
> > I think the best solution would be to prevent pmbus (more specifically
> > 'pmbus_reg2data_linear') to treat the obtained value as LINEAR16. A quick
> > hack shows me that this would work with all the reported values of the
> > device. However, this means proving da vendor/device specific workaround by
> > changing a generic component.
> >
> > I am not sure how you feel about this. Ultimately, I'd like to upstream the
> > driver (and potential fix), so I would like to find a solution that is fine
> > with you all.
> >
> > My current proposal would be to introduce a flag in
> > pmbus_platform_data.flags that allows to disable the LINEAR16 switch in
> > case the sensor has the class PSC_VOLTAGE_OUT. At least this seems the
> > change with the smallest impact. I am not sure how to name the flag but to
> > propose something I'd say 'PMBUS_VOUT_IS_LINEAR11'
> >
> > What do you think?
>
> Not sure if that is less messy.
>
> Have you tried to fake the VOUT_MODE command and to provide an exponent
> that makes sense ? While LINEAR16 does not support negative values,
> I don't immediately see that as a practical problem, unless the power
> supply indeed reports negative output voltages.
>
As a follow-up: If the above is not feasible (eg the exponent changes
all the time, or the reported voltages can indeed be negative), an
alternative would be to implement support for IEEE half precision
floating point format, as specified in PMBus 1.3.1.
Downside is that a device using IEEE half precision floating point
values must not use LINEAR{11,16} at all. But that would still be
better than hacking the PMBus core code.
Guenter
> Thanks,
> Guenter
>
> P.s.: I asked for more information from the vendor, but I don't really expect
> to
> get anything. Worth trying, though.
