Re: [PATCH] lglock: make lg_lock_global() actually lock globally

[Nick Piggin <npiggin@kernel.dk>]

On Thu, Aug 26, 2010 at 10:55:21AM +0200, Tejun Heo wrote:
> Hello,
> 
> On 08/25/2010 10:00 PM, Linus Torvalds wrote:
> >> lg_lock_global() currently only acquires spinlocks for online CPUs, but
> >> it's meant to lock all possible CPUs.  At Nick's suggestion, change
> >> for_each_online_cpu() to for_each_possible_cpu() to get the expected
> >> behavior.
> > 
> > Can you say what this actually matters for? Don't we do stop-machine
> > for CPU hotplug anyway? And if we don't, shouldn't we? Exactly because
> > otherwise "for_each_online_cpu()" is always racy (and that has nothing
> > to do with the lglock).
> 
> We only do stop-machine for cpu downs not ups, so code running w/
> preemption disabled is guaranteed that no cpu goes down while it's
> running but not the other way around.  There are two ways to achieve
> synchronization against cpu up/down operations.  One is explicitly
> using get/put_online_cpus() and the other is via cpu notifiers with
> proper synchronization.

Oh, I thought we quiesce / preempt all online cpus before adding
another one. That sucks if we don't because then you need a big
heavy get_online_cpus when a simple preempt_disable would have
worked.

Why is that? Don't tell me realtime people want some latency "guarantee"
while onlining CPUs? :)

> 
> So, yeah, given that there's no cpu notifier implemented, the use of
> for_each_online_cpu for brlock seems fishy to me.  It probably should
> use for_each_possible_cpu().

It should if that's the case, yes.

Thanks,
Nick