On Wed, Aug 21, 2019 at 04:33:58PM -0700, Paul E. McKenney wrote:
> On Wed, Aug 21, 2019 at 06:19:04PM -0500, Scott Wood wrote:
> > A plain local_bh_disable() is documented as creating an RCU critical
> > section, and (at least) rcutorture expects this to be the case. However,
> > in_softirq() doesn't block a grace period on PREEMPT_RT, since RCU checks
> > preempt_count() directly. Even if RCU were changed to check
> > in_softirq(), that wouldn't allow blocked BH disablers to be boosted.
> >
> > Fix this by calling rcu_read_lock() from local_bh_disable(), and update
> > rcu_read_lock_bh_held() accordingly.
>
> Cool! Some questions and comments below.
>
> Thanx, Paul
>
> > Signed-off-by: Scott Wood <swood@xxxxxxxxxx>
> > ---
> > Another question is whether non-raw spinlocks are intended to create an
> > RCU read-side critical section due to implicit preempt disable.
>
> Hmmm... Did non-raw spinlocks act like rcu_read_lock_sched()
> and rcu_read_unlock_sched() pairs in -rt prior to the RCU flavor
> consolidation? If not, I don't see why they should do so after that
> consolidation in -rt.
May be I am missing something, but I didn't see the connection between
consolidation and this patch. AFAICS, this patch is so that
rcu_read_lock_bh_held() works at all on -rt. Did I badly miss something?
> > If they
> > are, then we'd need to add rcu_read_lock() there as well since RT doesn't
> > disable preemption (and rcutorture should explicitly test with a
> > spinlock). If not, the documentation should make that clear.
>
> True enough!
>
> > include/linux/rcupdate.h | 4 ++++
> > kernel/rcu/update.c | 4 ++++
> > kernel/softirq.c | 12 +++++++++---
> > 3 files changed, 17 insertions(+), 3 deletions(-)
> >
> > diff --git a/include/linux/rcupdate.h b/include/linux/rcupdate.h
> > index 388ace315f32..d6e357378732 100644
> > --- a/include/linux/rcupdate.h
> > +++ b/include/linux/rcupdate.h
> > @@ -615,10 +615,12 @@ static inline void rcu_read_unlock(void)
> > static inline void rcu_read_lock_bh(void)
> > {
> > local_bh_disable();
> > +#ifndef CONFIG_PREEMPT_RT_FULL
> > __acquire(RCU_BH);
> > rcu_lock_acquire(&rcu_bh_lock_map);
> > RCU_LOCKDEP_WARN(!rcu_is_watching(),
> > "rcu_read_lock_bh() used illegally while idle");
> > +#endif
>
> Any chance of this using "if (!IS_ENABLED(CONFIG_PREEMPT_RT_FULL))"?
> We should be OK providing a do-nothing __maybe_unused rcu_bh_lock_map
> for lockdep-enabled -rt kernels, right?
Since this function is small, I prefer if -rt defines their own
rcu_read_lock_bh() which just does the local_bh_disable(). That would be way
cleaner IMO. IIRC, -rt does similar things for spinlocks, but it has been
sometime since I look at the -rt patchset.
> > }
> >
> > /*
> > @@ -628,10 +630,12 @@ static inline void rcu_read_lock_bh(void)
> > */
> > static inline void rcu_read_unlock_bh(void)
> > {
> > +#ifndef CONFIG_PREEMPT_RT_FULL
> > RCU_LOCKDEP_WARN(!rcu_is_watching(),
> > "rcu_read_unlock_bh() used illegally while idle");
> > rcu_lock_release(&rcu_bh_lock_map);
> > __release(RCU_BH);
> > +#endif
>
> Ditto.
>
> > local_bh_enable();
> > }
> >
> > diff --git a/kernel/rcu/update.c b/kernel/rcu/update.c
> > index 016c66a98292..a9cdf3d562bc 100644
> > --- a/kernel/rcu/update.c
> > +++ b/kernel/rcu/update.c
> > @@ -296,7 +296,11 @@ int rcu_read_lock_bh_held(void)
> > return 0;
> > if (!rcu_lockdep_current_cpu_online())
> > return 0;
> > +#ifdef CONFIG_PREEMPT_RT_FULL
> > + return lock_is_held(&rcu_lock_map) || irqs_disabled();
> > +#else
> > return in_softirq() || irqs_disabled();
> > +#endif
>
> And globally.
And could be untangled a bit as well:
if (irqs_disabled())
return 1;
if (IS_ENABLED(CONFIG_PREEMPT_RT_FULL))
return lock_is_held(&rcu_lock_map);
return in_softirq();
> > }
> > EXPORT_SYMBOL_GPL(rcu_read_lock_bh_held);
> >
> > diff --git a/kernel/softirq.c b/kernel/softirq.c
> > index d16d080a74f7..6080c9328df1 100644
> > --- a/kernel/softirq.c
> > +++ b/kernel/softirq.c
> > @@ -115,8 +115,10 @@ void __local_bh_disable_ip(unsigned long ip, unsigned int cnt)
> > long soft_cnt;
> >
> > WARN_ON_ONCE(in_irq());
> > - if (!in_atomic())
> > + if (!in_atomic()) {
> > local_lock(bh_lock);
> > + rcu_read_lock();
> > + }
> > soft_cnt = this_cpu_inc_return(softirq_counter);
> > WARN_ON_ONCE(soft_cnt == 0);
> > current->softirq_count += SOFTIRQ_DISABLE_OFFSET;
> > @@ -151,8 +153,10 @@ void _local_bh_enable(void)
> > #endif
> >
> > current->softirq_count -= SOFTIRQ_DISABLE_OFFSET;
> > - if (!in_atomic())
> > + if (!in_atomic()) {
> > + rcu_read_unlock();
> > local_unlock(bh_lock);
> > + }
> > }
> >
> > void _local_bh_enable_rt(void)
> > @@ -185,8 +189,10 @@ void __local_bh_enable_ip(unsigned long ip, unsigned int cnt)
> > WARN_ON_ONCE(count < 0);
> > local_irq_enable();
> >
> > - if (!in_atomic())
> > + if (!in_atomic()) {
> > + rcu_read_unlock();
> > local_unlock(bh_lock);
> > + }
>
> The return from in_atomic() is guaranteed to be the same at
> local_bh_enable() time as was at the call to the corresponding
> local_bh_disable()?
>
> I could have sworn that I ran afoul of this last year. Might these
> added rcu_read_lock() and rcu_read_unlock() calls need to check for
> CONFIG_PREEMPT_RT_FULL?
Great point! I think they should be guarded but will let Scott answer that
one.
thanks,
- Joel
