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

["Paul E. McKenney" <paulmck@kernel.org>]

On Tue, Nov 21, 2023 at 05:52:09PM -0500, Steven Rostedt wrote:
> On Tue, 21 Nov 2023 14:26:33 -0800
> "Paul E. McKenney" <paulmck@kernel.org> wrote:
> 
> > On Tue, Nov 21, 2023 at 04:38:34PM -0500, Steven Rostedt wrote:
> > > On Tue, 21 Nov 2023 13:14:16 -0800
> > > "Paul E. McKenney" <paulmck@kernel.org> wrote:
> > >   
> > > > On Tue, Nov 21, 2023 at 09:30:49PM +0100, Peter Zijlstra wrote:  
> > > > > On Tue, Nov 21, 2023 at 11:25:18AM -0800, Paul E. McKenney wrote:    
> > > > > > #define preempt_enable() \
> > > > > > do { \
> > > > > > 	barrier(); \
> > > > > > 	if (!IS_ENABLED(CONFIG_PREEMPT_RCU) && raw_cpu_read(rcu_data.rcu_urgent_qs) && \
> > > > > > 	    (preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK | HARDIRQ_MASK | NMI_MASK) == PREEMPT_OFFSET) &&
> > > > > > 	    !irqs_disabled()) \  
> > > 
> > > Could we make the above an else case of the below if ?  
> > 
> > Wouldn't that cause the above preempt_count() test to always fail?
> 
> preempt_count_dec_and_test() returns true if preempt_count() is zero, which
> happens only if NEED_RESCHED is set, and the rest of preempt_count() is not
> set. (NEED_RESCHED bit in preempt_count() is really the inverse of
> NEED_RESCHED). Do we need to call rcu_all_qs() when we call the scheduler?
> Isn't scheduling a quiescent state for most RCU flavors?
> 
> I thought this was to help move along the quiescent states without added
> cond_resched() around, which has:
> 
> int __sched __cond_resched(void)
> {
> 	if (should_resched(0)) {
> 		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
> 	return 0;
> }
> 
> Where if we schedule, we don't call rcu_all_qs().

True enough, but in this __cond_resched() case we know that we are in
an RCU quiescent state regardless of what should_resched() says.

In contrast, with preempt_enable(), we are only in a quiescent state
if __preempt_count_dec_and_test() returns true, and even then only if
interrupts are enabled.

> I stand by that being in the else statement. It looks like that would keep
> the previous work flow.

Ah, because PREEMPT_NEED_RESCHED is zero when we need to reschedule,
so that when __preempt_count_dec_and_test() returns false, we might
still be in an RCU quiescent state in the case where there was no need
to reschedule.  Good point!

In which case...

#define preempt_enable() \
do { \
	barrier(); \
	if (unlikely(preempt_count_dec_and_test())) \
		__preempt_schedule(); \
	else if (!sched_feat(FORCE_PREEMPT) && \
		 (preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK | HARDIRQ_MASK | NMI_MASK) == PREEMPT_OFFSET) && \
		 !irqs_disabled()) \
) \
			rcu_all_qs(); \
} while (0)

Keeping rcu_all_qs() pretty much as is.  Or some or all of the "else if"
condition could be pushed down into rcu_all_qs(), depending on whether
Peter's objection was call-site object code size, execution path length,
or both.  ;-)

If the objection is both call-site object code size and execution path
length, then maybe all but the preempt_count() check should be pushed
into rcu_all_qs().

Was that what you had in mind, or am I missing your point?

							Thanx, Paul