Potential issue after threadDeath

[leelamv]

Timers can send signals even after they are deleted. It isn't clear how this happen (Not a kernel expert). However, when a signal is handled and try to read JNIEnv, this can be dead thread already from the jvm perspective. When we call JVM function here at , we will essentially asking jvm to get stack trace for a dead thread.

How to prove my theory ?
a) Maintain a flag which states thread death status and update that flag when OnThreadEnd is called.
b) When signal is handled here , build an assert check to confirm if it isnt a dead thread.

To increase the possibility of this issue happen:

• Create a JNI Thread with lot of pthread_key and also have destructors with short sleeps.
• Attach the jni thread to jvm.

Some suggestions for fix

• Reset the JNIEnv in OnThreadEnd callback.
• You may also use flag and turn the invariant mentioned above into a condition .

[aalexand]

@leelamv Is there a reproducer for this issue? Note that for the per-thread timers to be armed, -cprof_cpu_use_per_thread_timers needs to be set. #34 does not mention that flag.

[aalexand]

Fixed in #37.