Handle ioctl EINVAL error for set_power_state when FW is too old

[kikkiakki]

When using a BH FW version that didn't support new power management features, I ran into this error when attempting to use set_power_state:

(.venv) palexson@yyz-syseng-06:~/tt-burnin$ tt-burnin
 Detected Chips: 1
 Detecting ARC: |
 Detecting DRAM: |
 [] [14/14] ETH: |
Traceback (most recent call last):
  File "/home/palexson/tt-burnin/.venv/bin/tt-burnin", line 8, in <module>
    sys.exit(main())
  File "/home/palexson/tt-burnin/.venv/lib/python3.10/site-packages/tt_burnin/main.py", line 272, in main
    devs, devices = detect_and_group_devices()
  File "/home/palexson/tt-burnin/.venv/lib/python3.10/site-packages/tt_burnin/main.py", line 260, in detect_and_group_devices
    device.set_power_state("high")
Exception: Ioctl failed with EINVAL: Invalid argument

Joel suggests: "Ideally Luwen would handle old KMD/old FW without an error. It can just swallow the error from KMD."

[kaplanz]

My follow up questions are:

• Do we always want to unconditionally swallow EINVAL, or only for old KMD?
• Are there other reasons we'd get EINVAL other than old KMD/old FW?
  • If so, how do we differentiate this specific user error?
  • If not, are we sure going forward that'll remain the case?

Also worth clarifying is what does new KMD return to us with old FW?

[joelsmithTT]

1. This is up to you, but my personal view is detecting KMD version to do this vs that is a bit ugly. In some places probably unavoidable, and if the number of unavoidable places gets too large it suggests a need for some kind of capability negotiation API. If you do decide to detect KMD version you could avoid using the API entirely. O_APPEND is harmless/ignored on old KMD.
2. You'll get EINVAL for bad input but also failure to send the message to FW and receive a good response (for whatever reason - could be that the card is hung). The KMD could probably do better on this latter part, because with old FW you'll get EINVAL (because the message will fail)... but you'll also get EINVAL if the KMD is old and doesn't understand the ioctl. I could change the former if it helps, but ultimately if the power ioctl doesn't work ignoring the failure is probably fine: either it's unsupported and wouldn't have done anything (old KMD/old FW) or the card is in a bad enough state that you'll have other problems anyway.