Re: [PATCH V5 2/2] tracing, perf : add cpu hotplug trace events

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Thu, Feb 24, 2011 at 09:58:32PM +0100, Peter Zijlstra wrote:
> On Thu, 2011-02-24 at 21:47 +0100, Thomas Gleixner wrote:
> > On Thu, 24 Feb 2011, Peter Zijlstra wrote:
> > 
> > > On Thu, 2011-02-24 at 15:24 -0500, Nicolas Pitre wrote:
> > > 
> > > > Most SMP ARM processors are going to use it soon.  Powering down idle 
> > > > cores provides substantial power saving.
> > > 
> > > And why can't regular idle paths be used? CPU hotplug is a massively
> > > expensive operation.
> > 
> > To achieve the same result from idle, you need to exclude the core
> > from any unwanted wakeup. At the moment cpu unplug is the only way to
> > achieve that.
> 
> Right, everything is a nail because all we have is a hammer like.
> 
> > If you want to do the same from idle, then we need the isolation
> > features Frederic is working on for RT/HPC.
> > 
> > They allow us to isolate cores completely for totaly different
> > reasons, but it could be resused to provide full isolation of a core
> > in a very deep power state.
> 
> Exactly.
> 
> > That would solve the problem w/o going through kstompmachine
> 
> Right, kstopmachine is a large part of the problem, but cpu hotplug
> really does an insane amount of work if all you want is to idle the
> core.

You do indeed need to know that the CPU will be powered off for quite
some time to be worth the extra work.  And a number of embedded devices
do have this level of foreknowledge -- for example, they might use the
second CPU only when the user is doing some computationally intensive
task, so that the device would normally be using only a single CPU.

							Thanx, Paul