Prevent endless loop in lock cycle detection. The following scenario would previously fail:

[copybara-service]

Prevent endless loop in lock cycle detection. The following scenario would previously fail:

• T1: Enter lockOrDetectPotentialLocksCycle:
  • Lock CAF.class, add itself to lockThreadIsWaitingOn, unlock CAF.class
  • Lock the cycle detecting lock (CDL)
  • Lock CAF.class, mark itself as lockOwnerThread, remove itself from lockThreadIsWaitingOn
  • Exit lockOrDetectPotentialLocksCycle
• T1: Re-enter lockOrDetectPotentialLocksCycle:
  • Lock CAF.class, add itself to lockThreadIsWaitingOn, unlock CAF.class
    T2: Enter lockOrDetectPotentialLocksCycle
  • Lock CAF.class, invoke detectPotentialLocksCycle.

At this point, detectPotentialLocksCycle will now loop forever, because the lockOwnerThread is also in lockThreadIsWaitingOn. During the course of looping forever, it will OOM, because it's building up an in-memory structure of what it thinks are cycles.

The solution is to avoid the re-entrant T1 from adding itself to lockThreadIsWaitingOn if it's already the lockOwnerThread. It's guaranteed that it won't relinquish the lock concurrently, because it's the exact same thread that owns it.

Fixes #1635 and Fixes #1510