lkml.org 
[lkml]   [2011]   [Aug]   [5]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: 3.0-git15 Atomic scheduling in pidmap_init
On Fri, Aug 05, 2011 at 10:22:45AM -0400, Josh Boyer wrote:
> On Thu, Aug 04, 2011 at 11:56:46PM -0700, Paul E. McKenney wrote:
> > > > I could be missing something obvious, but I don't see a way to avoid
> > > > using GFP_KERNEL without a lot of rip-up in the rest of the init path.
> > >
> > > As an aside, I bisected this back to:
> > >
> > > e8f7c70f44f sched: Make sleeping inside spinlock detection working in
> > > !CONFIG_PREEMPT
> >
> > OK, added Frederic on CC.
> >
> > > However, that doesn't seem all that helpful. The
> > > CONFIG_DEBUG_SPINLOCK_SLEEP option later got renamed to
> > > DEBUG_ATOMIC_SLEEP, and all it's doing is selecting PREEMPT_COUNT. At
> > > first glance, it seems this commit just allowed an issue that's been
> > > around for a while (benign or otherwise) to finally show up.
> > >
> > > (The Fedora kernel configs have CONFIG_PREEMPT_VOLUNTARY set, but not
> > > CONFIG_PREEMPT so PREEMPT_COUNT wasn't getting selected until this
> > > option did so.)
> >
> > Understood. So my question is "what is the real way to fix this?"
> > Within RCU, I would probably wrapper the calls to set_need_resched()
> > so that it checks for the scheduler being fully alive. Except for the
> > call from rcu_enter_nohz(), of course -- if that one is called before
> > the scheduler is ready, then that is a bug that needs to be fixed.
>
> By scheduler being fully alive, do you mean when rcu_scheduler_starting
> is called? Or do you mean the actual scheduler, because sched_init is
> called well before any of this happens.
>
> > Nevertheless, I am wondering if all of this isn't really papering over
> > some real problem somewhere. The way we get to this place is from people
> > registering RCU callbacks during early boot, which is OK in and of itself,
> > at least in moderation. But if someone is expecting those callbacks to be
> > invoked before the scheduler is fully set up and running multiple tasks,
> > they are going to be disappointed.
>
> Is there a way to dump what callbacks have been registered? As far as I
> can see, we call rcu_check_callbacks unconditionally when a timer
> interrupt is taken and that calls rcu_pending unconditionally as well.
> Before that, rcu_init is called which eventually sets up the per-cpu
> data via rcu_init_percpu_data and that sets rdp->qs_pending = 1.
> Until a quiescent state is reached __rcu_pending is going to try and
> force it, which is where the set_need_resched is called.
>
> Basically, I took what you said about wrapping set_need_resched and came
> up with the patch below. It gets rid of the oops from pidmap_init, but
> I need to test it a bit more. Would be happy to have feedback.
>
> josh
>
> ---
>
> diff --git a/kernel/rcutree.c b/kernel/rcutree.c
> index ba06207..8c6cb6e 100644
> --- a/kernel/rcutree.c
> +++ b/kernel/rcutree.c
> @@ -1681,8 +1681,14 @@ static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
> rdp->n_rp_qs_pending++;
> if (!rdp->preemptible &&
> ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs) - 1,
> - jiffies))
> - set_need_resched();
> + jiffies)) {
> + /* Make sure we're ready to mark the task as needing
> + * rescheduling otherwise we can trigger oopes early
> + * in the init path
> + */
> + if (rcu_scheduler_active)
> + set_need_resched();

What about we avoid setting rdp->qs_pending = 1 for the CPU
that handles the boot?

> + }
> } else if (rdp->qs_pending && rdp->passed_quiesc) {
> rdp->n_rp_report_qs++;
> return 1;


\
 
 \ /
  Last update: 2011-08-05 19:11    [W:0.189 / U:0.020 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site