Re: [Qestion] Is preempt_disable/enable needed in non-preemption code path

From: Paul E. McKenney
Date: Fri Apr 16 2021 - 13:26:32 EST


On Fri, Apr 16, 2021 at 06:51:10PM +0800, Xu, Yanfei wrote:
>
>
> On 4/16/21 1:07 AM, Paul E. McKenney wrote:
> > [Please note: This e-mail is from an EXTERNAL e-mail address]
> >
> > On Fri, Apr 16, 2021 at 12:18:42AM +0800, Xu, Yanfei wrote:
> > >
> > >
> > > On 4/15/21 11:43 PM, Paul E. McKenney wrote:
> > > > [Please note: This e-mail is from an EXTERNAL e-mail address]
> > > >
> > > > On Thu, Apr 15, 2021 at 11:04:05PM +0800, Xu, Yanfei wrote:
> > > > > Hi experts,
> > > > >
> > > > > I am learning rcu mechanism and its codes. When looking at the
> > > > > rcu_blocking_is_gp(), I found there is a pair preemption disable/enable
> > > > > operation in non-preemption code path. And it has been a long time. I can't
> > > > > understand why we need it? Is there some thing I missed? If not, can we
> > > > > remove the unnecessary operation like blow?
> > > >
> > > > Good point, you are right that preemption is disabled anyway in that block
> > > > of code. However, preempt_disable() and preempt_enable() also prevent the
> > > > compiler from moving that READ_ONCE() around. So my question to you is
> > > > whether it is safe to remove those statements entirely or whether they
> > > > should instead be replaced by barrier() or similar.
> > >
> > > Thanks for your reply! :)
> > >
> > > Yes, preempt_disable() and preempt_enable() defined in !preemption are
> > > barrier(). barrier can prevent from reordering that READ_ONCE(), but base on
> > > my current understanding, volatile in READ_ONCE can also tell the compiler
> > > not to reorder it. So, I think it's safe?
> >
> > Maybe.
> >
> > Please keep in mind that although the compiler is prohibited from
> > reordering volatile accesses with each other, there is nothing stopping
> > it from reordering volatile accesses with non-volatile accesses.
>
> Thanks for your patient explanation!
>
> I am trying to absorb what you said. Blow are my understanding:
> 1. "the compiler is prohibited from reordering volatile accesses with each
> other" means these situations:
> int a;
> foo()
> {
> for(;;)
> READ_ONCE(a);
> }
>
> or
>
> int a,b;
> foo()
> {
> int c,d;
> c = READ_ONCE(a);
> d = READ_ONCE(b);
> }

Yes, in both cases the load instructions emitted for the READ_ONCE()
macros must be emitted in order. The underlying hardware is free
to reorder.

> 2. "volatile accesses with non-volatile accesses" means d=b may happen
> before c=READ_ONCE(a) :
> int a;
> foo()
> {
> int b = 2
> int c,d;
> c = READ_ONCE(a);
> d = b;
> }
> if we want to keep the ordering of volatile access "c=READ_ONCE(a)" and
> non-volatile access "d=b", we should use stronger barrier like barrier().

Or an additional READ_ONCE() for b or a WRITE_ONCE() for d. But again,
this would constrain only the compiler, not the hardware.

But this wouldn't matter in most cases, because both b and d are local
variables whose addresses were never taken. So someone would need to
be using something crazy to poke into others' stacks for this to matter.

> Hope I didn't misunderstand.

It looks like you have most of it.

> Back to rcu_blocking_is_gp(), I find this link today
> https://www.spinics.net/lists/rcu/msg03985.html
> With the content in this link, I still haven't got the meaning of these two
> barrier(). I think I should learn knowledge about cpu-hotplug and things
> which talked in the link first to make sure if I am missing something, and
> then consult you. :)

That sounds like a very good approach!

Keep in mind that I am worried not just about the current state of
the code and compilers, but also their possible future states.

Thanx, Paul

> Best regards,
> Yanfei
>
> >
> > Thanx, Paul
> >
> > > Best regards,
> > > Yanfei
> > >
> > > >
> > > > Thanx, Paul
> > > >
> > > > > diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
> > > > > index da6f5213fb74..c6d95a00715e 100644
> > > > > --- a/kernel/rcu/tree.c
> > > > > +++ b/kernel/rcu/tree.c
> > > > > @@ -3703,7 +3703,6 @@ static int rcu_blocking_is_gp(void)
> > > > > if (IS_ENABLED(CONFIG_PREEMPTION))
> > > > > return rcu_scheduler_active == RCU_SCHEDULER_INACTIVE;
> > > > > might_sleep(); /* Check for RCU read-side critical section. */
> > > > > - preempt_disable();
> > > > > /*
> > > > > * If the rcu_state.n_online_cpus counter is equal to one,
> > > > > * there is only one CPU, and that CPU sees all prior accesses
> > > > > @@ -3718,7 +3717,6 @@ static int rcu_blocking_is_gp(void)
> > > > > * Those memory barriers are provided by CPU-hotplug code.
> > > > > */
> > > > > ret = READ_ONCE(rcu_state.n_online_cpus) <= 1;
> > > > > - preempt_enable();
> > > > > return ret;
> > > > > }
> > > > >
> > > > >
> > > > >
> > > > > Best regards,
> > > > > Yanfei