On Wed, Oct 28, 2020 at 09:02:43PM +0100, Peter Zijlstra wrote:
> On Wed, Oct 28, 2020 at 03:54:28PM +0100, Peter Zijlstra wrote:
> > On Wed, Oct 28, 2020 at 12:07:13PM +0100, Peter Zijlstra wrote:
> > > AFAICT we only need/use irq_work_queue_on() on remote CPUs, since we
> > > can directly access local state. So avoid the IRQ_WORK dependency and
> > > use the unconditionally available irq_work_queue_remote().
> > >
> > > This survives a number of TREE01 runs.
> >
> > OK, Paul mentioned on IRC that while it is extremely unlikely, this code
> > does not indeed guarantee it will not try to IPI self.
> >
> > I'll try again.
>
> This is the best I could come up with.. :/
>
> ---
> Subject: rcu/tree: Use irq_work_queue_remote()
> From: Peter Zijlstra <pet...@infradead.org>
> Date: Wed Oct 28 11:53:40 CET 2020
>
> All sites that consume rcu_iw_gp_seq seem to have rcu_node lock held,
> so setting it probably should too. Also the effect of self-IPI here
> would be setting rcu_iw_gp_seq to the value we just set it to
> (pointless) and clearing rcu_iw_pending, which we just set, so don't
> set it.
>
> Passes TREE01.
>
> Signed-off-by: Peter Zijlstra (Intel) <pet...@infradead.org>
> ---
> kernel/rcu/tree.c | 10 ++++++----
> 1 file changed, 6 insertions(+), 4 deletions(-)
>
> --- a/kernel/rcu/tree.c
> +++ b/kernel/rcu/tree.c
> @@ -1308,14 +1308,16 @@ static int rcu_implicit_dynticks_qs(stru
> resched_cpu(rdp->cpu);
> WRITE_ONCE(rdp->last_fqs_resched, jiffies);
> }
> -#ifdef CONFIG_IRQ_WORK
> + raw_spin_lock_rcu_node(rnp);
The caller of rcu_implicit_dynticks_qs() already holds this lock.
Please see the force_qs_rnp() function and its second call site,
to which rcu_implicit_dynticks_qs() is passed as an argument.
But other than that, this does look plausible. And getting rid of
that #ifdef is worth something. ;-)
Thanx, Paul
> if (!rdp->rcu_iw_pending && rdp->rcu_iw_gp_seq != rnp->gp_seq &&
> (rnp->ffmask & rdp->grpmask)) {
> - rdp->rcu_iw_pending = true;
> rdp->rcu_iw_gp_seq = rnp->gp_seq;
> - irq_work_queue_on(&rdp->rcu_iw, rdp->cpu);
> + if (likely(rdp->cpu != smp_processor_id())) {
> + rdp->rcu_iw_pending = true;
> + irq_work_queue_remote(rdp->cpu, &rdp->rcu_iw);
> + }
> }
> -#endif
> + raw_spin_unlock_rcu_node(rnp);
> }
>
> return 0;
