I would prefer we implement Units in dbus-sensors.  Reasons below.

On Wed, Sep 16, 2020 at 09:24:19AM -0700, Ed Tanous wrote:
> On Wed, Sep 16, 2020 at 8:23 AM James Feist <james.feist@linux.intel.com> wrote:
> >
> > On 9/16/2020 6:28 AM, Andrei Kartashev wrote:
> > > Hi,
> > >
> > > We noticed that dbus-sensors doesn't fully implement
> > > xyz.openbmc_project.Sensor.Value interface: there is no Unit property
> > > for all the sensors, defined by dbus-sensors.
> > > I believe it was intentionally, but I still wondering what was the
> > > reason?
> >
> > It was originally as the information seemed redundant. If the
> > information is needed I'm fine with someone adding it, it just hasn't
> > seemed to be a high priority.
> 
> Considering we've gone this long with no impact (considering the path
> can be used to lookup the unit) I wonder if we should consider
> removing unit from the sensor Value API?  It doesn't seem used.

The redundancy comes about from this list: [1].  I don't really recall
why we went with that approach where sensors were required to be in a
particular place in the object hierarchy.

I think we only "got away" with "no impact" from the lack of
implementation because the particular implementations that work well
with dbus-sensors also didn't implement it because dbus-sensors didn't
provide it.  That's kind of circular logic as a reason to eliminate it.

> 
> >
> > > I noticed that in intel-ipmi-oem units are determined based on object
> > > paths, but that looks ugly since there is well-defined natural
> > > interface for units in dbus.
> > > Lack of the "Unit" property in the interface breaks some existing
> > > logic.
> > >
> 
> Technically the way the interfaces define it, both are valid to use to
> determine the Units, and both would need to be lookup tables.  Is
> using the path any more ugly than using the property?

There isn't really a programatic way to define and enforce object paths
presently.  There is a programatic way to define and enforce
enum values.  The enums don't actually need a lookup table, if you're
using the sdbusplus generated headers and bindings.

If I were to guess which of the requirements would be more likely to be
changed it would be the object path.  There are very few other places
where we have such strict requirements on object paths (though we do
have places where the object path has meaning).  The current definition
is a bit ambiguous by what is meant by "the correct hierarchy within the
sensors namespace[2]", but the current implementations seem to take this
to mean `/xyz/openbmc_project/sensors/...`[3].  I don't know that this is
particular convenient for a multi-host system or any case where a BMC
is aggregating sensors from other BMCs.

Unless we get some stronger definition on what is meant by [2], the only
completely correct code is to look for the type at all the various
segments of the object path (ex. "fan_tach" could be found at
/xyz/openbmc_project/sensors/fan_tach/node2/fan0 or
/node2/xyz/openbmc_project/sensors/fan_tach/fan0).  Using the enum value
is far easier in this case.

1. https://github.com/openbmc/phosphor-dbus-interfaces/blob/master/xyz/openbmc_project/Sensor/Value.interface.yaml#L37
2. https://github.com/openbmc/phosphor-dbus-interfaces/blob/master/xyz/openbmc_project/Sensor/Value.interface.yaml#L3
3. https://github.com/openbmc/bmcweb/blob/cb13a39253848ece442971301ade9c09d98bf08e/redfish-core/lib/sensors.hpp#L56

-- 
Patrick Williams