- we should probably extend pthread_mutexattr_t with a field to create a
  single linked list of all instances.  This requires changing the
  pthread_mutexattr_* functions.


- a new attribute for mutexes: number of times we spin before calling
sys_futex



- test with threaded process terminating and semadj (?) being applied
  only after all threads are gone


- system() is cancelable.  What happens to the child process?  It probably
  should be killed.

  The signal handler setting is a process resource.  Setting and resetting
  it from multiple thread must be possible.


- the syscall wrapping code should really be in libc