Re: [PATCH 2/2] stop_machine: dequeue work before signal completion

[Tejun Heo <tj@kernel.org>]

Hello, again.

On Fri, Feb 08, 2013 at 11:42:43AM +0800, Hillf Danton wrote:
> As checked with BUG_ON in the case of CPU_UP_PREPARE, we have to dequeue
> work first for further actions, then stopper reaches sane and clear state.

When a CPU is finally put down in either CPU_UP_CANCELLED or
CPU_POST_DEAD, cpu_stop_cpu_callback() signals immediate completion on
all cpu_stop_works still queued on the dead CPU; unfortunately, this
code is buggy in that it doesn't remove the canceled work items off
the stopper->works leaving it corrupted, which will trigger BUG_ON()
during CPU_UP_PREPARE if the CPU is brought back online.

This bug isn't easily triggered because CPU_DOWN has to race against
cpu_stop calls and most, if not all, cpu stop users pin target CPUs.

Fix it by popping each work item off stopper->works.

> Signed-off-by: Hillf Danton <dhillf@gmail.com>

Maybe

Cc: stable@vger.kernel.org

> --- a/kernel/stop_machine.c	Fri Feb  8 11:22:44 2013
> +++ b/kernel/stop_machine.c	Fri Feb  8 11:29:40 2013
> @@ -342,8 +342,12 @@ static int __cpuinit cpu_stop_cpu_callba
>  		kthread_stop(stopper->thread);
>  		/* drain remaining works */
>  		spin_lock_irq(&stopper->lock);
> -		list_for_each_entry(work, &stopper->works, list)
> +		while (!list_empty(&stopper->works)) {
> +			work = list_first_entry(&stopper->works,
> +					struct cpu_stop_work, list);
> +			list_del_init(&work->list);
>  			cpu_stop_signal_done(work->done, false);
> +		}

I think your previous version was better with @work declaration moved
inside the while() loop.

Thanks.

-- 
tejun