Re: [RFC PATCH v3 1/9] CPU hotplug: Provide APIs to prevent CPU offline from atomic context

[Tejun Heo <tj@kernel.org>]

Hello,

On Tue, Dec 11, 2012 at 07:32:13PM +0530, Srivatsa S. Bhat wrote:
> On 12/11/2012 07:17 PM, Tejun Heo wrote:
> > Hello, Srivatsa.
> > 
> > On Tue, Dec 11, 2012 at 06:43:54PM +0530, Srivatsa S. Bhat wrote:
> >> This approach (of using synchronize_sched()) also looks good. It is simple,
> >> yet effective, but unfortunately inefficient at the writer side (because
> >> he'll have to wait for a full synchronize_sched()).
> > 
> > While synchornize_sched() is heavier on the writer side than the
> > originally posted version, it doesn't stall the whole machine and
> > wouldn't introduce latencies to others.  Shouldn't that be enough?
> > 
> 
> Short answer: Yes. But we can do better, with almost comparable code
> complexity. So I'm tempted to try that out.
> 
> Long answer:
> Even in the synchronize_sched() approach, we still have to identify the
> readers who need to be converted to use the new get/put_online_cpus_atomic()
> APIs and convert them. Then, if we can come up with a scheme such that
> the writer has to wait only for those readers to complete, then why not?
> 
> If such a scheme ends up becoming too complicated, then I agree, we
> can use synchronize_sched() itself. (That's what I meant by saying that
> we'll use this as a fallback).
> 
> But even in this scheme which uses synchronize_sched(), we are
> already half-way through (we already use 2 types of sync schemes -
> counters and rwlocks). Just a little more logic can get rid of the
> unnecessary full-wait too.. So why not give it a shot?

It's not really about the code complexity but making the reader side
as light as possible.  Please keep in mind that reader side is still
*way* more hotter than the writer side.  Before, the writer side was
heavy to the extent which causes noticeable disruptions on the whole
system and I think that's what we're trying to hunt down here.  If we
can shave of memory barriers from reader side by using
synchornized_sched() on writer side, that is the *better* result, not
worse.

Thanks.

-- 
tejun