If you're really worried that you'll leak an entire region worth of data just allocate the region with malloc instead of mmap and then use Valgrind to tell you what regions you aren't destroying.
Yes, if your program just needs to allocate lots of memory, do some computations, then exit, this approach works. But the programs where security is most critical do not usually follow that pattern.
Look, I'm totally willing to accept that there might be flaws with this approach, but with the exception of adrusi, no one's objections have been all that reasonable. If there are legitimate objections (not just handwavy, oh, you'll probably implement the allocator wrong) I'd love to hear them. I use this pattern in my own code and I'll stop if there are legitimate flaws.
Also, Akamai released a patch for the OpenSSL allocator bug mentioned earlier. Guess what the patch used: mmaped regions.[2]
[1]: http://www.gnu.org/software/libc/manual/html_node/Obstacks.h...
[2]: http://thread.gmane.org/gmane.comp.encryption.openssl.user/5...
It is entirely possible to have no bugs in the <100 lines needed for this code, so I don't think the bug argument is valid. adrusi did raise a legitimate concern however.
The semantics are simple as well:
Edit: also, you're not reinventing malloc for no good reason. A region based allocator is much faster, possibly at the expense of using more space than the system allocator (because you might reserve more space than you actually need).