Re: [PATCH 01/51] CPU hotplug: Provide lockless versions of callback registration functions

[Toshi Kani <toshi.kani@hp.com>]

On Tue, 2014-02-11 at 09:27 +0000, Srivatsa S. Bhat wrote:
> On 02/11/2014 06:56 AM, Toshi Kani wrote:
> > On Thu, 2014-02-06 at 03:34 +0530, Srivatsa S. Bhat wrote:
> >  :
> [...]
> >>
> >> Also, since cpu_maps_update_begin/done() is like a super-set of
> >> get/put_online_cpus(), the former naturally protects the critical sections
> >> from concurrent hotplug operations.
> > 
> > get/put_online_cpus() is a reader-lock and concurrent executions are
> > allowed among the readers.  They won't be serialized until a cpu
> > online/offline operation begins.  By replacing this lock with
> > cpu_maps_update_begin/done(), we now serialize all readers.  Isn't that
> > too restrictive?
> 
> That's an excellent line of thought! It doesn't really hurt at the moment
> because the for_each_online_cpu() kind of loop that the initcalls of various
> subsystems run (before registering the notifier) are really tiny (typically
> the loop runs for just 1 cpu, the boot-cpu). In other words, this change
> represents a tiny increase in the critical section size; so its effect
> shouldn't be noticeable. (Note that in the old model, register_cpu_notifier()
> already takes the cpu_add_remove_lock, so they will be serialized at that
> point, and this is necessary).
> 
> However, going forward, when we start using more aggressive CPU onlining
> techniques during boot (such as parallel CPU hotplug), the issue you pointed
> out can become a real bottleneck, since for_each_online_cpu() can become
> quite a large loop, and hence explicit (and unnecessary) mutual exclusion
> will start hurting.
>  
> >  Can we fix the issue with CPU_POST_DEAD and continue
> > to use get_online_cpus()?
> > 
> 
> We don't want to get rid of CPU_POST_DEAD, so unfortunately we can't continue
> to use get_online_cpus(). However, I am thinking of introducing a Reader-Writer
> semaphore for this purpose, so that the registration routines can run in
> parallel most of the time. (Basically, the rw-semaphore is like
> get/put_online_cpus(), except that it protects the full hotplug critical section,
> including the CPU_POST_DEAD stage.)

I agree that introducing a reader-writer semaphore allows concurrent
executions.  Adding yet another hotplug lock is a bit unfortunate,
though.

This may be a dumb question, but can't we simply do this way?

        get_online_cpus();

        for_each_online_cpu(cpu)
                init_cpu(cpu);

        put_online_cpus();

        register_cpu_notifier(&foobar_cpu_notifier);

Thanks,
-Toshi