Re: [RFC][PATCH 0/4] Redesign cpu_hotplug locking.

[Andrew Morton <akpm@osdl.org>]

On Sun, 27 Aug 2006 23:19:46 +0530
Dipankar Sarma <dipankar@in.ibm.com> wrote:

> I don't see why this
> is needed - 
> 
> + break;
> +
> + case CPU_DOWN_PREPARE:
> + 	mutex_lock(&workqueue_mutex);
> + 	break;
> +
> + case CPU_DOWN_FAILED:
> + 	mutex_unlock(&workqueue_mutex);
> 	break;
> 
> This seems like some implicit code locking to me. Why is it not
> sufficient to hold the lock in the CPU_DEAD code while walking
> the workqueues ?

?

We need to hold workqueue_mutex to protect the per-cpu workqueue resources
while cpu_online_map is changing and while per-cpu memory is being
allocated or freed.

Look at cpu_down() and mentally replace the
blocking_notifier_call_chain(CPU_DOWN_PREPARE) with
mutex_lock(workqueue_mutex), etc.  The __stop_machine_run() in there
modifies the (ie: potentially frees) the workqueue code's per-cpu memory. 
So we take that resource's lock while doing so.