Delayed processing for ProcessManager.pidToProcessInfo

[christos68k]

Summary

• Renamed SymbolizationComplete to ProcessedUntil and moved to processinfo.go
• Dropped no longer needed "symbolize now" remnants
• Implemented delayed ProcessManager.pidToProcessInfo cleanup

Leverages #307 to ensure that process metadata is not discarded before all relevant trace events have been processed.

Fixes #278.

You may find reviewing commit-by-commit to be simpler.

[christos68k]

I kept this cleanup here as there's no immediate need to postpone cleaning up the eBPF map until traceCaptureKTime >= pidExitKtime (unlike pidToProcessInfo). This also speeds up execution of ProcessedUntil compared to having the map cleanup take place there.

[christos68k]

Moved this outside the lock for improved accuracy (there's a debug log in ProcessedUntil that prints exit latency).

[christos68k]

We don't want to attempt a repeat cleanup for the same PID, if we've previously performed it.

[christos68k]

Moved to processinfo.go for consistency (all pidToProcessInfo accessors in one place), renamed to ProcessedUntil and updated to also cleanup pidToProcessInfo.

[christos68k]

This is now taking place in ProcessedUntil, delayed until traceCaptureKTime >= exitKTime.

[christos68k]

Moved without changes to processinfo.go for consistency.

[christos68k]

Essentially same logic as before with these additions:

1. Don't add exitKTime to pm.exitEvents if it already exists.
2. Also add exitKTime to pm.exitEvents if pm.pidToProcessInfo[pid] exists, as we want to cleanup the latter in delayed fashion.

[christos68k]

Same logic as before with this single-line addition.

[florianl]

first look with some comments

[florianl]

To keep the global read & write lock as short as possible, the if !pidExists {..} part should be moved before if pidExists || (pm.interpreterTracerEnabled && len(pm.interpreters[pid]) > 0) {..}.

[rockdaboot]

That would prevent executing if _, pidExited = ... in case (pm.interpreterTracerEnabled && len(pm.interpreters[pid]) > 0 is true.

[christos68k]

As Tim wrote, this would alter the logic. I tried to keep as much of the original semantics the same to avoid introducing new races. Maybe here it's possible to safely say that if !pidExists then it's OK not to write exitKTime in pm.exitEvents but we'd need to carefully examine all subsystem interactions, check for race conditions etc.

[florianl]

follow up is done in #325

[florianl]

I think, pidExited should be renamed to pidExitProcessed so something similar, this would it make obvious, that we want to avoid duplicate work.

[christos68k]

Renamed

[florianl]

keep the lock holding as short as possible:

Suggested change

	pm.mu.Lock()
	defer pm.mu.Unlock()
	nowKTime := times.GetKTime()
	log.Debugf("ProcessedUntil captureKT: %v latency: %v ms",
	traceCaptureKTime, (nowKTime-traceCaptureKTime)/1e6)
	nowKTime := times.GetKTime()
	log.Debugf("ProcessedUntil captureKT: %v latency: %v ms",
	traceCaptureKTime, (nowKTime-traceCaptureKTime)/1e6)
	pm.mu.Lock()
	defer pm.mu.Unlock()

[christos68k]

This can affect the latency measurement, since we're timing before the lock.