Performance monitoring API

[rmn30]

We need some performance monitoring APIs so that we can understand the performance of CHERIoT applications. This issue is to track ideas and development.

Here are some ideas for requirements / features:

• need to be able to enable / disable performance monitoring APIs at compile time via a flag / cpp definition so that monitoring will have no impact (performance or security) on release builds. Maybe per component?
• performance and space impact should be very low when enabled
• need to be able to retrieve statistics (count, cycles, instructions?) about significant events, probably per thread and / or per compartment. Min / max of these statistics?
• Events we may want to monitor:
  • switcher: compartment calls, faults, scheduler time?
  • scheduler: interrupts (per interrupt type?), interrupt latency?, thread transitions? accuracy of timer interrupt?
  • allocator: allocations (bin by size?), revocation sweeps, time spent waiting for revocation
  • networking
  • sonata cache: instruction / data accesses, hits, misses, miss cycles

It would be nice to have an extensible API that could accommodate all of these but they may be different enough that that won't work.

Can anyone point to existing performance monitoring APIs we could clone or use as reference?

Security consideration: do we anticipate including any of this in release builds? If so what what can we do about the side-channels?

[davidchisnall]

@hlef