> All cases could be handled by first making a local copy of the time
   > value to be modified, update that and store it back into the global
   > variable. To do this, we need two operations like: ATOMIC_READTIMEVAL
   > and ATOMIC_WRITETIMEVAL.
   
   That will only work if there is only one writer (or any updating
   is always done with some lock held).

there is only one writer - see kern_clock.c:hardclock():

#if defined(MULTIPROCESSOR)
        /*
	 * If we are not the primary CPU, we're not allowed to do
	 * any more work.
	 */
	if (CPU_IS_PRIMARY(ci) == 0)
		return;
#endif  
   
   Statistics counters can (probably) be allowed to miscount.

one way to avoid having counters lose is to count them on a per-cpu
basis and have programs tally them when required, or perhaps have
something like hardclock() tally into the normal global every tick..


.mrg.