BMR491 reports NaN for power

[mkeeter]

While investigating https://github.com/oxidecomputer/hardware-gimlet/issues/1988 , we noticed that the BMR491 power sensor (112 0x70 power 4 F - 0x67 bmr491 V12_SYS_A2) is reading NaN. That seems weird!

[cbiffle]

I assume this is not obviously reproduceable. :\

Poking at the pmbus code now just to see if I find anything obvious.

[cbiffle]

If you recall, @mkeeter, which actual quantity was reading as NaN? The BMR491 has various voltage and current outputs, but does not have a power output per se (as in something measured in watts), so I want to make sure I'm looking at the right thing.

I realize it's been a minute.

[cbiffle]

Alright, caught up with Matt on this in chat.

I think this is a bug but not a Gimlet bug. Here's what it looks like is happening.

1. app/gimlet/base.toml defines one power sensor channel on the BMR491.
2. The BMR491 RON in pmbus defines zero power sensor channels.
3. Accordingly, the sensors task allocates space for a power sensor whose readings are never delivered.
4. The sensors task uses f32::NAN as an initialization value for the data_value array, and leaks that at the API boundary.
5. For reasons I'm still investigating, humility sensors apparently doesn't work on a released image, so the engineers were using humility readvar.
6. This exposed the sentinel value.

Fortunately this means it's a lower severity issue. I'd argue it's still a bug, or possibly three smaller bugs in a trenchcoat. The things this makes me want to go investigate are:

• Why does humility sensors not work on at least some released Gimlet images? (guess: it only tries udprpc)
• Can we use a different sentinel value internally? (I have strong opinions about the correct use of NaN, having been burned by it before -- I think using it as a sentinel value is bad because it sends people like me hunting for bad floating point arithmetic. The bmr491 has 22,821 non-blank non-comment lines of code to check. Thank goodness for rust-analyzer. Anyway -- I think a different sentinel would be good.)

[cbiffle]

humility sensors, for the record, uses the hiffy generic Idol call interface. So it should work on release image with a dongle attached, and is not expected to work on a release image over IP. To get sensor data over the management network, we'd either need to use a control plane oriented service or add to the gimlet-inspector for debugging.