Re: [PATCH 2/3] powerpc: Instrument Hypervisor Calls: add wrappers

[Mike Kravetz <kravetz@us.ibm.com>]

On Thu, Jun 15, 2006 at 08:33:55AM +1000, Paul Mackerras wrote:
> Nathan Lynch writes:
> 
> > Without disabling preemption around the mfspr ... update_stats section
> > in these hcall wrappers, you risk updating the stats on the wrong cpu.
> 
> I think we are only looking for total counts and times anyway, so it
> doesn't really matter which cpu updates the stats, as long as the time
> gets accounted on some cpu.  The use of per-cpu counters is just for
> better cache behaviour.

Thought about this a little more.

We do an mfspr to get a 'timestamp' before the actual hcall.  Then, make
the hcall and do another mfspr after.  Isn't it possible to be preempted
and perform the mfspr's on separate CPUs?  Worse yet, wouldn't this possibly
add 'time preempted' to the hcall time?.  Do we have the same (skewed time)
issue with interrupts?

Thinking that enable/disable around the call might be worth the overhead.
But, disabling interrupts would not be worth it.

-- 
Mike