Tidy up GO_GC handling?

[dtrudg]

Type of issue

technical debt

Description of issue

In #8 we disabled go garbage collection for the starter. This was done to avoid the issue in #7 - where if the timing of a GC cycle was unfortunate, we could end up closing internal Go runtime netpoll fds in StartProcess, causing a crash.

When runc addressed a security issue by closing Fds, they faced the same need to avoid closing internal Go netpoll fds. Rather than disable GC so the fds do not appear, they used an internal Go function to identify the fds and avoid closing them:

opencontainers/runc@a9833ff#diff-6dc5f3f1e98fc4e379c98c5c301256dc1950dcb04fbce8d280bf12c41fadc1aaR73

We could consider using this approach.

[dtrudg]

@tri-adam - curious about your take on this? Better to disable GC... or better to link against and use an internal Go function?

Both are pretty hacky...

[tri-adam]

The approach taken in the runc repo looks like a good one, other than the obvious downside of relying on poll.IsPollDescriptor. Assuming that function doesn't go anywhere, it's probably the better of the two evils? One thing that makes me nervous about it is IsPollDescriptor is only used in tests as far as I can see. On the other hand, the test code hasn't been modified in four years.

With the assumption that a change or removal of poll.IsPollDescriptor in some future version of Go would break our code base in an obvious way, I'd be in favour of this change.