Re: task isolation discussion at Linux Plumbers

[Andy Lutomirski]

On Wed, Nov 9, 2016 at 8:52 PM, Paul E. McKenney
<paulmck@xxxxxxxxxxxxxxxxxx> wrote:
> On Wed, Nov 09, 2016 at 05:44:02PM -0800, Andy Lutomirski wrote:
>> On Wed, Nov 9, 2016 at 9:38 AM, Paul E. McKenney
>> <paulmck@xxxxxxxxxxxxxxxxxx> wrote:
>>
>> Are you planning on changing rcu_nmi_enter()?  It would make it easier
>> to figure out how they interact if I could see the code.
>
> It already calls rcu_dynticks_eqs_exit(), courtesy of the earlier
> consolidation patches.
>
>> > diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
>> > index dbf20b058f48..342c8ee402d6 100644
>> > --- a/kernel/rcu/tree.c
>> > +++ b/kernel/rcu/tree.c
>>
>>
>> >  /*
>> > @@ -305,17 +318,22 @@ static void rcu_dynticks_eqs_enter(void)
>> >  static void rcu_dynticks_eqs_exit(void)
>> >  {
>> >         struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
>> > +       int seq;
>> >
>> >         /*
>> > -        * CPUs seeing atomic_inc() must see prior idle sojourns,
>> > +        * CPUs seeing atomic_inc_return() must see prior idle sojourns,
>> >          * and we also must force ordering with the next RCU read-side
>> >          * critical section.
>> >          */
>> > -       smp_mb__before_atomic(); /* See above. */
>> > -       atomic_inc(&rdtp->dynticks);
>> > -       smp_mb__after_atomic(); /* See above. */
>> > +       seq = atomic_inc_return(&rdtp->dynticks);
>> >         WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
>> > -                    !(atomic_read(&rdtp->dynticks) & 0x1));
>> > +                    !(seq & RCU_DYNTICK_CTRL_CTR));
>>
>> I think there's still a race here.  Suppose we're running this code on
>> cpu n and...
>>
>> > +       if (seq & RCU_DYNTICK_CTRL_MASK) {
>> > +               rcu_eqs_special_exit();
>> > +               /* Prefer duplicate flushes to losing a flush. */
>> > +               smp_mb__before_atomic(); /* NMI safety. */
>>
>> ... another CPU changes the page tables and calls rcu_eqs_special_set(n) here.
>
> But then rcu_eqs_special_set() will return false because we already
> exited the extended quiescent state at the atomic_inc_return() above.
>
> That should tell the caller to send an IPI.
>
>> That CPU expects that we will flush prior to continuing, but we won't.
>> Admittedly it's highly unlikely that any stale TLB entries would be
>> created yet, but nothing rules it out.
>
> That said, 0day is having some heartburn from this, so I must have broken
> something somewhere.  My own tests of course complete just fine...
>
>> > +               atomic_and(~RCU_DYNTICK_CTRL_MASK, &rdtp->dynticks);
>> > +       }
>>
>> Maybe the way to handle it is something like:
>>
>> this_cpu_write(rcu_nmi_needs_eqs_special, 1);
>> barrier();
>>
>> /* NMI here will call rcu_eqs_special_exit() regardless of the value
>> in dynticks */
>>
>> atomic_and(...);
>> smp_mb__after_atomic();
>> rcu_eqs_special_exit();
>>
>> barrier();
>> this_cpu_write(rcu_nmi_needs_eqs_special, 0);
>>
>>
>> Then rcu_nmi_enter() would call rcu_eqs_special_exit() if the dynticks
>> bit is set *or* rcu_nmi_needs_eqs_special is set.
>>
>> Does that make sense?
>
> I believe that rcu_eqs_special_set() returning false covers this, but
> could easily be missing something.

I think you're right.  I'll stare at it some more when I do the actual
TLB flush patch.

--Andy