Re: [PATCH/RFC 0/4] Add stop_machine_get/put_threads to stop_machine infrastructrue.

[Rusty Russell <rusty@rustcorp.com.au>]

On Wednesday 08 October 2008 21:14:50 Heiko Carstens wrote:
> On Wed, Oct 08, 2008 at 10:27:04AM +1000, Rusty Russell wrote:
> > OK, idea #2.  Let's just always have a kstopmachine thread running on
> > every online cpu.  Is there a sane way to reuse the workqueue threads for
> > this?
>
> That's a very good idea and what the patch below does. It even simplifies
> the stop_machine code and it does work on an otherwise idle system.
> The only thing that needs to be addressed is that workqueue threads aka
> stop_machine threads are no real time threads now.
> We would need something like create_workqueue_prio() or
> create_workqueue_rt(). Would that be acceptable?

Hmm, I was hoping to reuse the kevent threads rather than create YA set
of threads.  But hey, everyone else is doing it.

> +static struct workqueue_struct *stop_machine_wq;
> +static struct work_struct *stop_machine_work;
> +static struct stop_machine_data active, idle;
> +static cpumask_t active_cpus;

Hmm, please make active cpus a const cpumask_t pointer.  I'm trying to
get rid of these kind of decls in another patch series :)

>  /* This is the actual thread which stops the CPU.  It exits by itself rather
>   * than waiting for kthread_stop(), because it's easier for hotplug CPU. */

This comment is no longer valid...

> +static int __init stop_machine_init(void)
> +{
> +	stop_machine_wq = create_workqueue("kstop");
> +	stop_machine_work = kcalloc(NR_CPUS, sizeof(struct work_struct),
> +				    GFP_KERNEL);

Perhaps make stop_machine_work a per-cpu array of struct work_struct
instead of initializing it here.  Or at least make it a percpu pointer and
only alloc possible cpus.

Does it break cpu hotplug BTW?  That's usually the problem.

But it looks nice!
Rusty.