Let's start with the reproducer (HelloSpin.java attached):

java -XX:NativeMemoryTracking=detail HelloSpin
<wait 10-15 seconds>
jcmd VM.native_memory baseline
jcmd HelloSpin GC.heap_dump heap.hprof
jcmd VM.native_memory detail.diff

The detail.diff will among other things output an entry for this:

[0x00007f8f6dfc29dd] HeapDumper::dump(char const*)+0x3d
                            (malloc=8192KB +8192KB, #1 +1)

The entry reflects the buffer allocated in heapDumper.cpp:DumpWriter::DumpWriter(const char* path):

_buffer = (char*)os::malloc(_size, mtInternal);

where size is io_buffer_size = 8MB.


However, the _buffer is *not* leaked - it is free:d up immediately when the heap dumper has done its job, so the information displayed is incorrect.

If I trigger a GC using:

jcmd HelloSpin GC.run

The entry goes away, which I assume means that the information is just not updated in a timely fashion.

This is on Linux/x64. I can reproduce the same problem on both 7u40 and 8.

ILW=MMM=P3