[FEATURE] Instantiate and run MPS in pKVM

[podhrmic]

Summary

Run the MPS in pKVM

• Guest VM
  • has a baseline.img (R/W) and MPS.img (R)
  • automatically launches the MPS software
• Host VM
  • has a baseline.img (R/W) and a config.img (R) that contains the JSON/XML config file, and the respective binaries
  • automatically launches the guest
• Base
  • runs the test harness
  • if the test harness is not attached, the MPS runs in an interactive mode, and can be commanded through UART

Do

• Run the original MPS (unmodified HARDENS) code in the guest VM, including the test harness (follow mps-test and mps-build CI jobs)
• Pipe UART from the Guest VM to the Base platform, run MPS in the Guest VM and the test harness in Base platform
• Modify the test harness and the MPS code to change the actuator values via GPIO, on top of printing it to UART

[spernsteiner]

I've started on this on the 35-mps-in-pkvm branch. So far I've written a runner that will start on boot, mount an application partition, and start up processes or VMs as indicated in the application's config file. I've run a simple "hello world" test, and next I'm planning to test "hello world" in nested VMs, mimicking the setup we plan to use for MPS.

[spernsteiner]

I successfully connected MPS in the guest via UARTs to a Unix socket in the base system, and I was able to run one of the MPS test cases against it in that configuration. Some more work is needed to be able to reset the MPS between test cases without shutting down and restarting the entire VM.

[spernsteiner]

Some more work is needed to be able to reset the MPS between test cases

I've implemented a new MPS command R that causes the MPS to re-exec itself, which resets it to the initial state for the next test.

I've gotten many of the tests to pass with MPS running inside the VM. I had some issues with the test suite at first, but eventually discovered that the tests require a specific build configuration: make -C src SENSORS=NotSimulated SELF_TEST=Disabled

I'm currently investigating some issues with the QEMU UART that cause large delays (> 10 seconds) between sending the data and the MPS receiving it. These delays increase over time as the test suite runs until eventually the tests start timing out.

[podhrmic]

Regarding the test build - here is how MPS is built and tested in the CI

[spernsteiner]

With #53, the test suite passes with the tests running on the base system and the MPS running inside the guest VM, with the two communicating via an emulated UART. My guess is that the communication delays I was seeing previously were due to contention on the display lock causing the refresh display (D) command to take a long time to run.

I'm now working on cleaning up the setup code for this so I can open a PR.

[spernsteiner]

I tried running the MPS test suite without QUICK=1 as a final test of the MPS-on-pKVM setup, and it ran into more issues with UART delays. It's failing in scenarios/normal_4, though not in the same place as #56 (it fails on case 8/9 instead of 272/273).

However, the test suite does work when both the tests and MPS are run inside the guest VM. All tests succeed up to the errors mentioned in #56. I will add some automation for this alternative setup and open a PR for that tomorrow.

[spernsteiner]

I got the test suite working with MPS-in-pKVM (modulo #56) by reducing the frequency of update-display commands to avoid overwhelming the MPS process: 7b3777b