Re: [RFC] workqueue: avoiding unbounded wq on isolated CPUs by default

[Mike Galbraith <umgwanakikbuti@gmail.com>]

On Sat, 2015-07-18 at 15:36 +0200, Frederic Weisbecker wrote:
> On Fri, Jul 17, 2015 at 07:15:48PM +0200, Mike Galbraith wrote:
> > On Fri, 2015-07-17 at 11:27 -0400, Tejun Heo wrote:
> > 
> > > I'm just curious whether there was any specific reason we didn't do
> > > this before (ISTR people discussing it back then too).
> > 
> > I'm dead set against all this auto-presume nonsense fwtw  Allocating a
> > pool of no_hz_full _capable_ CPUs should not entice the kernel to make
> > any rash assumptions.  Let users do the button poking, they know what
> > they want, and when they want it.
> 
> We need to make a choice then. Either we do all the affinity tuning from
> userspace with a common tool, which is what I had wished before everybody
> asked for pre-settings.

Giving userspace what they need to do what they want seems right to me.

> Or we do it in the kernel, now we should define some kind of CONFIG_ISOLATION
> to make that proper and rule the various kinds of isolation people are
> interested in.
> 
> But we can't leave it half-way like it is currently with everything preset on
> top of nohz: rcu nocb mask, watchdog mask, cpu_isolation_map and exclude workqueue.

Yeah.  Hell, maybe I'm wrong.  Maybe people really want this rigidity
and hand-holding by the kernel, but it just seems dainbramaged to me.
ATM, you pay a high price (the overhead) for the capability, but until
that auto-assume isolcpus landed, those CPUs weren't forever more
specialists, they were CPUs with an extra (costly) capability, could be
disconnected/reconnected to load balancing on the fly, and used however
the user saw fit.

I can imagine an auto-everything kernel having a bit of trouble with an
SGI beast from hell.  Too bad I don't have access to one, I'd try to
boot a tune for maximum hand holding kernel.

	-Mike