Re: CPU Hotplug rework

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

On Thu, Mar 22, 2012 at 02:55:04PM +1030, Rusty Russell wrote:
> On Wed, 21 Mar 2012 10:01:59 +0100, Peter Zijlstra <a.p.zijlstra@chello.nl> wrote:
> > On Wed, 2012-03-21 at 09:30 +1030, Rusty Russell wrote:
> > > > > (2) Do something more efficient with userspace threads than migrating
> > > > >     them one at a time.
> > > > 
> > > > Sadly that can't really be done. We need to pick up every task
> > > > (userspace, but also running kernel threads) and update their state.
> > > 
> > > What if we had an "orphan" runqueue which everyone pulled from?  Then we
> > > could grab the lock, move them all to the fake rq, then let stuff happen
> > > normally.
> > 
> > Well, we could simply let them sit where they are and fudge load-balance
> > to consider it a source but not a destination until its empty, but it
> > might be somewhat tricky to make it fast enough to not introduce
> > noticable latencies. Also, you really don't want everyone to pull,
> > that's a serialization/scalability problem.
> > 
> > Also, since we really only move the currently runnable tasks it
> > shouldn't be too many in the first place. Is it really that expensive?
> 
> Good question, requires measurement to answer.
> 
> > > Maybe that's crap, but at least we could move the migration out of the
> > > hotplug callback somehow. 
> > 
> > Thing is, if its really too much for some people, they can orchestrate
> > it such that its not. Just move everybody in a cpuset, clear the to be
> > offlined cpu from the cpuset's mask -- this will migrate everybody away.
> > Then hotplug will find an empty runqueue and its fast, no?
> 
> I like this solution better.

As long as we have some way to handle kthreads that are algorithmically
tied to a given CPU.  There are coding conventions to handle this, for
example, do everything with preemption disabled and just after each
preempt_disable() verify that you are in fact running on the correct
CPU, but it is easy to imagine improvements.

							Thanx, Paul