The current system of measuring CPU used by a process/the system is 
crude (at best).  Basically, if you get a timer tick when your 
application is running, then you get a whole timer tick's work of CPU 
added to your process's usage.  That's actually just fine for most 
applications, but some applications need tight bounds and more accurate 
accounting.  This patch addresses this concern.

This patch adds a configuration option that causes the kernel to measure 
the CPU used by a process at microsecond resolution.  This affects 
general CPU usage measurement, rusage, and virtual itimers.  The patch 
is against 2.4.17.  It touches some pretty fundamental data structures 
and ties in to all the exception (including interrupt and system call) 
handlers, so it's pretty invasive.  It shouldn't slow things down when 
the patch is turned off, but it will slow down all exceptions when it is 
turned on.  The patch currently works on PowerPC and x86 (although it 
won't work on earlier x86 processors due to lack of a high-resolution 
clock source).  It has seen extensive testing (around 9 months of heavy 
usage) in the lab on an older kernel.

I'm curious if this patch is interesting enough to work to include it in 
the mainstream kernel release.  If so, I'll gladly work to get it in to 
shape to go into 2.5 or whatever.  (I'll also gladly take comments on 
how to improve it :-).

-Corey