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

From: Paul E. McKenney
Date: Thu Apr 15 2021 - 13:15:14 EST


On Fri, Apr 16, 2021 at 01:01:17AM +0800, Xu, Yanfei wrote:
>
>
> On 4/16/21 12:18 AM, 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?
> >
> > Best regards,
> > Yanfei
>
> Hi Paul,
> I objdump the function rcu_blocking_is_gp():
>
> after dropping the barrier():
> ffffffff81107c50 <rcu_blocking_is_gp>:
> ffffffff81107c50: e8 7b 2a f5 ff callq ffffffff8105a6d0
> <__fentry__>
> ffffffff81107c55: 8b 05 41 fe 7c 01 mov 0x17cfe41(%rip),%eax
> # ffffffff828d7a9c <rcu_state+0x221c>
> ffffffff81107c5b: 55 push %rbp
> ffffffff81107c5c: 48 89 e5 mov %rsp,%rbp
> ffffffff81107c5f: 5d pop %rbp
> ffffffff81107c60: 83 f8 01 cmp $0x1,%eax
> ffffffff81107c63: 0f 9e c0 setle %al
> ffffffff81107c66: 0f b6 c0 movzbl %al,%eax
> ffffffff81107c69: c3 retq
> ffffffff81107c6a: 66 0f 1f 44 00 00 nopw 0x0(%rax,%rax,1)
>
> the original codes:
> ffffffff81107ba0 <rcu_blocking_is_gp>:
> ffffffff81107ba0: e8 2b 2b f5 ff callq ffffffff8105a6d0
> <__fentry__>
> ffffffff81107ba5: 55 push %rbp
> ffffffff81107ba6: 48 89 e5 mov %rsp,%rbp
> ffffffff81107ba9: 8b 05 ed fe 7c 01 mov 0x17cfeed(%rip),%eax
> # ffffffff828d7a9c <rcu_state+0x221c>
> ffffffff81107baf: 83 f8 01 cmp $0x1,%eax
> ffffffff81107bb2: 5d pop %rbp
> ffffffff81107bb3: 0f 9e c0 setle %al
> ffffffff81107bb6: 0f b6 c0 movzbl %al,%eax
> ffffffff81107bb9: c3 retq
> ffffffff81107bba: 66 0f 1f 44 00 00 nopw 0x0(%rax,%rax,1)
>
> umm... It did been reordered by compiler after dropping the barrier(),
> however, I think the result will not be effected. Right?

Indeed, the compiler is free to reorder volatile accesses with
non-volatile accesses.

The result might not be affected by your compiler using your particular
settings for optimizations, but is that true for all optimization settings
for all compilers, both now and in the future?

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