On Wed, Sep 04, 2019 at 01:28:19PM +0200, Peter Zijlstra wrote:
> @@ -196,6 +198,17 @@ static int membarrier_register_global_expedited(void)
> */
> smp_mb();
> } else {
> + struct task_struct *g, *t;
> +
> + read_lock(&tasklist_lock);
> + do_each_thread(g, t) {
> + if (t->mm == mm) {
> + atomic_or(MEMBARRIER_STATE_GLOBAL_EXPEDITED,
> + &t->membarrier_state);
> + }
> + } while_each_thread(g, t);
> + read_unlock(&tasklist_lock);
> +
> /*
> * For multi-mm user threads, we need to ensure all
> * future scheduler executions will observe the new
Arguably, because this is exposed to unpriv users and a potential
preemption latency issue, we could do it in 3 passes:
- RCU, mark all found lacking, count
- RCU, mark all found lacking, count
- if count of last pass, tasklist_lock
That way, it becomes much harder to trigger the bad case.
Do we worry about that?
