On Thu, Jul 29, 2021 at 12:01:37AM +0200, Frederic Weisbecker wrote:
> On Wed, Jul 28, 2021 at 08:34:14PM +0100, Valentin Schneider wrote:
> > On 28/07/21 01:08, Frederic Weisbecker wrote:
> > > On Wed, Jul 21, 2021 at 12:51:17PM +0100, Valentin Schneider wrote:
> > >> Signed-off-by: Valentin Schneider <valentin.schneider@xxxxxxx>
> > >> ---
> > >> kernel/rcu/tree_plugin.h | 3 +--
> > >> 1 file changed, 1 insertion(+), 2 deletions(-)
> > >>
> > >> diff --git a/kernel/rcu/tree_plugin.h b/kernel/rcu/tree_plugin.h
> > >> index ad0156b86937..6c3c4100da83 100644
> > >> --- a/kernel/rcu/tree_plugin.h
> > >> +++ b/kernel/rcu/tree_plugin.h
> > >> @@ -70,8 +70,7 @@ static bool rcu_rdp_is_offloaded(struct rcu_data *rdp)
> > >> !(lockdep_is_held(&rcu_state.barrier_mutex) ||
> > >> (IS_ENABLED(CONFIG_HOTPLUG_CPU) && lockdep_is_cpus_held()) ||
> > >> rcu_lockdep_is_held_nocb(rdp) ||
> > >> - (rdp == this_cpu_ptr(&rcu_data) &&
> > >> - !(IS_ENABLED(CONFIG_PREEMPT_COUNT) && preemptible())) ||
> > >> + (rdp == this_cpu_ptr(&rcu_data) && is_pcpu_safe()) ||
> > >
> > > I fear that won't work. We really need any caller of rcu_rdp_is_offloaded()
> > > on the local rdp to have preemption disabled and not just migration disabled,
> > > because we must protect against concurrent offloaded state changes.
> > >
> > > The offloaded state is changed by a workqueue that executes on the target rdp.
> > >
> > > Here is a practical example where it matters:
> > >
> > > CPU 0
> > > -----
> > > // =======> task rcuc running
> > > rcu_core {
> > > rcu_nocb_lock_irqsave(rdp, flags) {
> > > if (!rcu_segcblist_is_offloaded(rdp->cblist)) {
> > > // is not offloaded right now, so it's going
> > > // to just disable IRQs. Oh no wait:
> > > // preemption
> > > // ========> workqueue running
> > > rcu_nocb_rdp_offload();
> > > // ========> task rcuc resume
> > > local_irq_disable();
> > > }
> > > }
> > > ....
> > > rcu_nocb_unlock_irqrestore(rdp, flags) {
> > > if (rcu_segcblist_is_offloaded(rdp->cblist)) {
> > > // is offloaded right now so:
> > > raw_spin_unlock_irqrestore(rdp, flags);
> > >
> > > And that will explode because that's an impaired unlock on nocb_lock.
> >
> > Harumph, that doesn't look good, thanks for pointing this out.
> >
> > AFAICT PREEMPT_RT doesn't actually require to disable softirqs here (since
> > it forces RCU callbacks on the RCU kthreads), but disabled softirqs seem to
> > be a requirement for much of the underlying functions and even some of the
> > callbacks (delayed_put_task_struct() ~> vfree() pays close attention to
> > in_interrupt() for instance).
> >
> > Now, if the offloaded state was (properly) protected by a local_lock, do
> > you reckon we could then keep preemption enabled?
>
> I guess we could take such a local lock on the update side
> (rcu_nocb_rdp_offload) and then take it on rcuc kthread/softirqs
> and maybe other places.
>
> But we must make sure that rcu_core() is preempt-safe from a general perspective
> in the first place. From a quick glance I can't find obvious issues...yet.
>
> Paul maybe you can see something?
Let's see...
o Extra context switches in rcu_core() mean extra quiescent
states. It therefore might be necessary to wrap rcu_core()
in an rcu_read_lock() / rcu_read_unlock() pair, because
otherwise an RCU grace period won't wait for rcu_core().
Actually, better have local_bh_disable() imply
rcu_read_lock() and local_bh_enable() imply rcu_read_unlock().
But I would hope that this already happened.
o The rcu_preempt_deferred_qs() check should still be fine,
unless there is a raw_bh_disable() in -rt.
o The set_tsk_need_resched() and set_preempt_need_resched()
might preempt immediately. I cannot think of a problem
with that, but careful testing is clearly in order.
o The values checked by rcu_check_quiescent_state() could now
change while this function is running. I don't immediately
see a problematic sequence of events, but here be dragons.
I therefore suggest disabling preemption across this function.
Or if that is impossible, taking a very careful look at the
proposed expansion of the state space of this function.
o I don't see any new races in the grace-period/callback check.
New callbacks can appear in interrupt handlers, after all.
o The rcu_check_gp_start_stall() function looks similarly
unproblematic.
o Callback invocation can now be preempted, but then again it
recently started being concurrent, so this should be no
added risk over offloading/de-offloading.
o I don't see any problem with do_nocb_deferred_wakeup().
o The CONFIG_RCU_STRICT_GRACE_PERIOD check should not be
impacted.
So some adjustments might be needed, but I don't see a need for
major surgery.
This of course might be a failure of imagination on my part, so it
wouldn't hurt to double-check my observations.
> > From a naive outsider PoV, rdp->nocb_lock looks like a decent candidate,
> > but it's a *raw* spinlock (I can't tell right now whether changing this is
> > a horrible idea or not), and then there's
>
> Yeah that's not possible, nocb_lock is too low level and has to be called with
> IRQs disabled. So if we take that local_lock solution, we need a new lock.
No argument here!
Thanx, Paul
