I'm afraid that there is some inaccuracy in the way cpu.c reports performance 
statistics, at least on Linux 2.6 systems.
Basically, the reported CPU usage (user, including nice, plus system) is 
overestimated since the elapsed time is underestimated, mainly due to the I/O 
wait time not being considered at all (this can be noticed when creating a 
big tar file).

Please find attached a cpu.c that I believe more accurate on Linux 2.6 systems;
for the code change, check the compilation key RS(=true).