Messages in this thread | | | Date | Mon, 3 Jan 2011 16:49:18 +0100 | From | Oleg Nesterov <> | Subject | Re: [RFC][PATCH 08/17] sched: Drop the rq argument to sched_class::select_task_rq() |
| |
On 01/03, Peter Zijlstra wrote: > > On Mon, 2011-01-03 at 15:59 +0100, Oleg Nesterov wrote: > > On 01/03, Peter Zijlstra wrote: > > > > > > On Wed, 2010-12-29 at 22:31 +0800, Yong Zhang wrote: > > > > > - /* > > > > > - * select_task_rq() can race against ->cpus_allowed > > > > > - */ > > > > > - if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed) && > > > > > - likely(cpu_active(dest_cpu)) && migrate_task(p, rq)) { > > > > > + if (likely(cpu_active(dest_cpu)) && need_migrate_task(p)) { > > > > > > > > If we drop rq_lock, need_migrate_task() maybe return true but > > > > p is already running on other cpu. Thus we do a wrong migration > > > > call. > > > > > > Yeah, too bad.. ;-) exec load balancing is more an optimistic thing > > > anyway, if it got rebalanced under out feet we don't care. > > > > I don't understand this need_migrate_task() at all (with or without > > the patch). This task is current/running, it should always return T. > > This is true for the sched_exec() case, yes. > > > I guess, migrate_task() was needed before to initialize migration_req. > > But afaict you can call set_cpus_allowed_ptr() on !self.
Ah, sorry for the confusion, I only meant sched_exec() case. set_cpus_allowed_ptr() does need need_migrate_task(), of course.
As for set_cpus_allowed_ptr()->need_migrate_task() path, I have another question,
static bool need_migrate_task(struct task_struct *p) { /* * If the task is not on a runqueue (and not running), then * the next wake-up will properly place the task. */ smp_rmb(); /* finish_lock_switch() */ return p->on_rq || p->on_cpu; }
I don't understand this smp_rmb(). Yes, finish_lock_switch() does wmb() before it clears ->on_cpu, but how these 2 barriers can pair?
In fact, I am completely confused. I do not understand why do we check task_running() at all. If we see on_rq == 0 && on_cpu == 1, then this task is going to clear its on_cpu soon, once it finishes context_switch().
Probably, this check was needed before, try_to_wake_up() could activate the task_running() task without migrating. But, at first glance, this is no longer possible after this series?
Oleg.
| |