Re: [RFC PATCH 48/86] rcu: handle quiescent states for PREEMPT_RCU=n

[Paul E. McKenney]

On Tue, Nov 21, 2023 at 02:13:53PM +0800, Z qiang wrote:
> >
> > On Mon, Nov 20, 2023 at 09:17:57PM -0800, Paul E. McKenney wrote:
> > > On Mon, Nov 20, 2023 at 07:26:05PM -0800, Ankur Arora wrote:
> > > >
> > > > Paul E. McKenney <paulmck@xxxxxxxxxx> writes:
> > > > > On Tue, Nov 07, 2023 at 01:57:34PM -0800, Ankur Arora wrote:
> > > > >> cond_resched() is used to provide urgent quiescent states for
> > > > >> read-side critical sections on PREEMPT_RCU=n configurations.
> > > > >> This was necessary because lacking preempt_count, there was no
> > > > >> way for the tick handler to know if we were executing in RCU
> > > > >> read-side critical section or not.
> > > > >>
> > > > >> An always-on CONFIG_PREEMPT_COUNT, however, allows the tick to
> > > > >> reliably report quiescent states.
> > > > >>
> > > > >> Accordingly, evaluate preempt_count() based quiescence in
> > > > >> rcu_flavor_sched_clock_irq().
> > > > >>
> > > > >> Suggested-by: Paul E. McKenney <paulmck@xxxxxxxxxx>
> > > > >> Signed-off-by: Ankur Arora <ankur.a.arora@xxxxxxxxxx>
> > > > >> ---
> > > > >>  kernel/rcu/tree_plugin.h |  3 ++-
> > > > >>  kernel/sched/core.c      | 15 +--------------
> > > > >>  2 files changed, 3 insertions(+), 15 deletions(-)
> > > > >>
> > > > >> diff --git a/kernel/rcu/tree_plugin.h b/kernel/rcu/tree_plugin.h
> > > > >> index f87191e008ff..618f055f8028 100644
> > > > >> --- a/kernel/rcu/tree_plugin.h
> > > > >> +++ b/kernel/rcu/tree_plugin.h
> > > > >> @@ -963,7 +963,8 @@ static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
> > > > >>   */
> > > > >>  static void rcu_flavor_sched_clock_irq(int user)
> > > > >>  {
> > > > >> -        if (user || rcu_is_cpu_rrupt_from_idle()) {
> > > > >> +        if (user || rcu_is_cpu_rrupt_from_idle() ||
> > > > >> +            !(preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK))) {
> > > > >
> > > > > This looks good.
> > > > >
> > > > >>                  /*
> > > > >>                   * Get here if this CPU took its interrupt from user
> > > > >> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> > > > >> index bf5df2b866df..15db5fb7acc7 100644
> > > > >> --- a/kernel/sched/core.c
> > > > >> +++ b/kernel/sched/core.c
> > > > >> @@ -8588,20 +8588,7 @@ int __sched _cond_resched(void)
> > > > >>                  preempt_schedule_common();
> > > > >>                  return 1;
> > > > >>          }
> > > > >> -        /*
> > > > >> -         * In preemptible kernels, ->rcu_read_lock_nesting tells the tick
> > > > >> -         * whether the current CPU is in an RCU read-side critical section,
> > > > >> -         * so the tick can report quiescent states even for CPUs looping
> > > > >> -         * in kernel context.  In contrast, in non-preemptible kernels,
> > > > >> -         * RCU readers leave no in-memory hints, which means that CPU-bound
> > > > >> -         * processes executing in kernel context might never report an
> > > > >> -         * RCU quiescent state.  Therefore, the following code causes
> > > > >> -         * cond_resched() to report a quiescent state, but only when RCU
> > > > >> -         * is in urgent need of one.
> > > > >> -         *      /
> > > > >> -#ifndef CONFIG_PREEMPT_RCU
> > > > >> -        rcu_all_qs();
> > > > >> -#endif
> > > > >
> > > > > But...
> > > > >
> > > > > Suppose we have a long-running loop in the kernel that regularly
> > > > > enables preemption, but only momentarily.  Then the added
> > > > > rcu_flavor_sched_clock_irq() check would almost always fail, making
> > > > > for extremely long grace periods.
> > > >
> > > > So, my thinking was that if RCU wants to end a grace period, it would
> > > > force a context switch by setting TIF_NEED_RESCHED (and as patch 38 mentions
> > > > RCU always uses the the eager version) causing __schedule() to call
> > > > rcu_note_context_switch().
> > > > That's similar to the preempt_schedule_common() case in the
> > > > _cond_resched() above.
> > >
> > > But that requires IPIing that CPU, correct?
> > >
> > > > But if I see your point, RCU might just want to register a quiescent
> > > > state and for this long-running loop rcu_flavor_sched_clock_irq() does
> > > > seem to fall down.
> > > >
> > > > > Or did I miss a change that causes preempt_enable() to help RCU out?
> > > >
> > > > Something like this?
> > > >
> > > > diff --git a/include/linux/preempt.h b/include/linux/preempt.h
> > > > index dc5125b9c36b..e50f358f1548 100644
> > > > --- a/include/linux/preempt.h
> > > > +++ b/include/linux/preempt.h
> > > > @@ -222,6 +222,8 @@ do { \
> > > >         barrier(); \
> > > >         if (unlikely(preempt_count_dec_and_test())) \
> > > >                 __preempt_schedule(); \
> > > > +       if (!(preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK))) \
> > > > +               rcu_all_qs(); \
> > > >  } while (0)
> > >
> > > Or maybe something like this to lighten the load a bit:
> > >
> > > #define preempt_enable() \
> > > do { \
> > >       barrier(); \
> > >       if (unlikely(preempt_count_dec_and_test())) { \
> > >               __preempt_schedule(); \
> > >               if (raw_cpu_read(rcu_data.rcu_urgent_qs) && \
> > >                   !(preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK))) \
> > >                       rcu_all_qs(); \
> > >       } \
> > > } while (0)
> > >
> > > And at that point, we should be able to drop the PREEMPT_MASK, not
> > > that it makes any difference that I am aware of:
> > >
> > > #define preempt_enable() \
> > > do { \
> > >       barrier(); \
> > >       if (unlikely(preempt_count_dec_and_test())) { \
> > >               __preempt_schedule(); \
> > >               if (raw_cpu_read(rcu_data.rcu_urgent_qs) && \
> > >                   !(preempt_count() & SOFTIRQ_MASK)) \
> > >                       rcu_all_qs(); \
> > >       } \
> > > } while (0)
> > >
> > > Except that we can migrate as soon as that preempt_count_dec_and_test()
> > > returns.  And that rcu_all_qs() disables and re-enables preemption,
> > > which will result in undesired recursion.  Sigh.
> > >
> > > So maybe something like this:
> > >
> > > #define preempt_enable() \
> > > do { \
> > >       if (raw_cpu_read(rcu_data.rcu_urgent_qs) && \
> > >           !(preempt_count() & SOFTIRQ_MASK)) \
> >
> > Sigh.  This needs to include (PREEMPT_MASK | SOFTIRQ_MASK),
> > but check for equality to something like (1UL << PREEMPT_SHIFT).
> >
> 
> For PREEMPT_RCU=n and CONFIG_PREEMPT_COUNT=y kernels
> for report QS in preempt_enable(),  we can refer to this:
> 
> void rcu_read_unlock_strict(void)
> {
>         struct rcu_data *rdp;
> 
>         if (irqs_disabled() || preempt_count() || !rcu_state.gp_kthread)
>                 return;
>         rdp = this_cpu_ptr(&rcu_data);
>         rdp->cpu_no_qs.b.norm = false;
>         rcu_report_qs_rdp(rdp);
>         udelay(rcu_unlock_delay);
> }
> 
> The rcu critical section may be in the NMI handler  needs to be considered.

You are quite right, though one advantage of leveraging preempt_enable()
is that it cannot really enable preemption in an NMI handler.
But yes, that might need to be accounted for in the comparison with
preempt_count().

The actual condition needs to also allow for the possibility that
this preempt_enable() happened in a kernel built with preemptible RCU.
And probably a few other things that I have not yet thought of.

For one thing, rcu_implicit_dynticks_qs() might need adjustment.
Though I am currently hoping that it will still be able to enlist the
help of other things, for example, preempt_enable() and local_bh_enable().

Yes, it is the easiest thing in the world to just whip out the
resched_cpu() hammer earlier in the grace period, and maybe that is the
eventual solution.  But I would like to try avoiding the extra IPIs if
that can be done reasonably.  ;-)

							Thanx, Paul

> Thanks
> Zqiang
> 
> 
> 
> >
> > Clearly time to sleep.  :-/
> >
> >                                                         Thanx, Paul
> >
> > >               rcu_all_qs(); \
> > >       barrier(); \
> > >       if (unlikely(preempt_count_dec_and_test())) { \
> > >               __preempt_schedule(); \
> > >       } \
> > > } while (0)
> > >
> > > Then rcu_all_qs() becomes something like this:
> > >
> > > void rcu_all_qs(void)
> > > {
> > >       unsigned long flags;
> > >
> > >       /* Load rcu_urgent_qs before other flags. */
> > >       if (!smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs)))
> > >               return;
> > >       this_cpu_write(rcu_data.rcu_urgent_qs, false);
> > >       if (unlikely(raw_cpu_read(rcu_data.rcu_need_heavy_qs))) {
> > >               local_irq_save(flags);
> > >               rcu_momentary_dyntick_idle();
> > >               local_irq_restore(flags);
> > >       }
> > >       rcu_qs();
> > > }
> > > EXPORT_SYMBOL_GPL(rcu_all_qs);
> > >
> > > > Though I do wonder about the likelihood of hitting the case you describe
> > > > and maybe instead of adding the check on every preempt_enable()
> > > > it might be better to instead force a context switch in the
> > > > rcu_flavor_sched_clock_irq() (as we do in the PREEMPT_RCU=y case.)
> > >
> > > Maybe.  But rcu_all_qs() is way lighter weight than a context switch.
> > >
> > >                                                       Thanx, Paul