A number of error reporting steps require memory. When we are out of memory, those steps may fail.

A prominent example is NMT: NMT detailed reports would be super helpful in analyzing native OOMs (see also: JDK-8227031). However, to create that report NMT needs memory from C-Heap. If C-Heap is exhausted, report fails or crashes. 

Other examples include stacktrace printing, which on some platforms invokes ElfDecoder to print symbol names, which needs C-Heap as well.

In a perfect world we would harden everything destined to be running inside error handling, e.g. to work with pre-allocated buffers, or to not allocate memory at all. But that is difficult (not impossible) and increases complexity, which is undesired.

A pragmatic solution would be:

- allocate a "ballast" buffer from C-Heap on VM startup
- on native OOM, "drop" that ballast by free()ing the buffer. This returns the memory back to the C-lib allocator, reduces the memory pressure and hopefully enables the few error reporting steps which need C-Heap to finish successfully.

Of course there is absolutely no guarantee that this works - code running concurrently may gobble the memory up the instant we release it, for instance.
 
However, this is pragmatic and dead simple to implement and in practice works surprisingly well. A hack such as this did often help us to get an NMT detail report in an OOM situation where otherwise we would have gotten nothing.