Dnmalloc 1.0

Dnmalloc 1.0 has been released!

Thanks to Rainer Wichmann for bugfixing the code to make it release quality. He will also be shipping dnmalloc 1.0 with the next version of Samhain. He'll also be contributing possible future enhancements to dnmalloc and as such is now an official developer.

To facilitate this, we've set up an svn server: http://svn.fort-knox.org/code/dnmalloc/

You can check out the code from this server by doing:
svn co http://svn.fort-knox.org/code/dnmalloc/

Changes by Rainer since beta 5:
Compiler warnings fixed
Define REALLOC_ZERO_BYTES_FREES because it's what GNU libc does
(and what the standard says)
Removed unused code
Fix assert(aligned_OK(chunk(newp)));
-> assert(aligned_OK(chunk(oldp)));
Fix statistics in sYSMALLOc
Fix overwrite of av->top in sYSMALLOc
Provide own assert(), glibc assert() doesn't work (calls malloc)
Fix bug in mEMALIGn(), put remainder in hashtable before calling fREe
Remove cfree, independent_cmalloc, independent_comalloc (untested
public functions not covered by any standard)
Provide posix_memalign (that one is in the standard)
Move the malloc_state struct to mmapped memory protected by guard pages
Add arc4random function to initialize random canary on startup
Implement random canary at end of (re|m)alloced/memaligned buffer,
check at free/realloc
Remove code conditional on !HAVE_MMAP, since mmap is required anyway.
Use standard HAVE_foo macros (as generated by autoconf) instead of LACKS_foo

Profiling: Reorder branches in hashtable_add, next_chunkinfo,
prev_chunkinfo, hashtable_insert, mALLOc, fREe, request2size,
checked_request2size (gcc predicts if{} branch to be taken).
Use UNLIKELY macro (gcc __builtin_expect()) where branch
reordering would make the code awkward.

Portability: Hashtable always covers full 32bit address space to
avoid assumptions about memory layout.
Portability: Try hard to enforce mapping of mmapped memory into
32bit address space, even on 64bit systems.
Portability: Provide a dnmalloc_pthread_init() function, since
pthread locking on HP-UX only works if initialized
after the application has entered main().
Portability: On *BSD, pthread_mutex_lock is unusable since it
calls malloc, use spinlocks instead.
Portability: Dynamically detect whether the heap is within
32bit address range (e.g. on Linux x86_64, it isn't).
Don't use sbrk() if the heap is mapped to an address
outside the 32bit range, since this doesn't work with
the hashtable. New macro morecore32bit.

Success on: HP-UX 11.11/pthread, Linux/pthread (32/64 bit),
FreeBSD/pthread, and Solaris 10 i386/pthread.
Fail on: OpenBSD/pthread (in _thread_machdep_save_float_state),
might be related to OpenBSD pthread internals (??).
Non-treaded version (#undef USE_MALLOC_LOC)
works on OpenBSD.