Hi,
	Thanks to Andrew Morton who tried some things that I probably
should have but didn't, here's the second version of a patch to give
microstate accounting capability to Linux 2.5.73

I've also included a sample program to use the new system call, and to
print out the times spent on active and expired queues, time spent
sleeping, and time spent interrupted, etc.

To save you having to dig up my previous email, here's what it does
again:
	-- The results of getrusage() are unreliable in two ways:
	   1. Interrupt handlers are charged to the task that was
	      interrupted, giving too high a system time.
	   2. User and system times are accrued at scheduler_tick
	      time.  If a task is I/O bound, then it tends to do a
	      little bit of work (less than a tick) then sleep,
	      so it's almost never running when the tick happens, so
	      the user time is too low.

	-- To diagnose some of the interactivity problems we've been
	seeing in 2.5, you really want to know if the problem is that
	a task is spending too much time on the EXPIRED queue, is being
	starved because of interrupt load, or whatever. The attached
	patch allows one to measure these directly.

The Microstate accounting patch (MSA) measures with the precision of
the in-chip clock (TSC on IA32, ITC on IA64) the time spent in the
various states, and exports the timers via a new system call (msa())
and through a new file /proc/pid/msa

It gives misleading results on my laptop, where the power management
seems to change the TSC frequency rather often, but on the UP P4
desktop machine, and the single and dual Itanium-2 boxes I tried, the
results correlate extremely well with other timing information that I
have (e.g., perfmon).  I tried using the monotonic clock but that had
the same problem (i.e., when the machine is less than 100% utilised,
the clock speed drops and the INTERRUPTIBLE SLEEP timer reads low)

The measured overhead of having MSA in place is available on
http://www.gelato.unsw.edu.au/IA64wiki/MicroStateAccounting 
To summarise: the cost of a context switch goes up by under 5%; but
kernbench shows no significant time differences between with and without the
patch in and enabled.