Re: [PATCH tip/core/rcu 22/30] rcu: Don't flag non-starting GPs before GP kthread is running

From: Steven Rostedt
Date: Mon Feb 17 2020 - 17:21:43 EST


On Mon, 17 Feb 2020 14:03:56 -0800
"Paul E. McKenney" <paulmck@xxxxxxxxxx> wrote:

> And what is a day without micro-optimization of a slowpath? :-)

A day you have off, but still find yourself working ;-)

>
> OK, let's see...
>
> Grace-period kthread wakeups are normally mediated by rcu_start_this_gp(),
> which uses a funnel lock to consolidate concurrent requests to start
> a grace period. If a grace period is already in progress, it refrains
> from doing a wakeup because that means that the grace-period kthread
> will check for another grace period being needed at the end of the
> current grace period.
>
> Exceptions include:
>
> o The wakeup reporting the last quiescent state of the current
> grace period.
>
> o Emergency situations such as callback overloads and RCU CPU stalls.
>
> So on a busy system that is not overloaded, the common case is that
> rcu_gp_kthread_wake() is invoked only once per grace period because there
> is no emergency and there is a grace period in progress. If this system
> has short idle periods and a fair number of quiescent states, a reasonable
> amount of idle time, then the last quiescent state will not normally be
> detected by the grace-period kthread. But workloads can of course vary.
>
> The "!t" holds only during early boot. So we could put a likely() around
> the "t". But more to the point, at runtime, "!t" would always be false,
> so it really should be last in the list of "||" clauses. This isn't
> enough of a fastpath for a static branch to make sense.

Hey! Does that mean we can add a static branch for that check?


struct static_key rcu_booting = STATIC_KEY_INIT_TRUE;

[...]

if (READ_ONCE(rcu_state.gp_flags) ||
(current == t && !in_irq() && !in_serving_softirq())
return;

if (static_branch_unlikely(&rcu_booting) && !t)
return;

At end of boot:

static_key_disable(&rcu_booting);

That way we can really micro-optimize the slow path, and it basically
becomes a nop!

-- Steve

>
> The "!READ_ONCE(rcu_state.gp_flags)" will normally hold, though it is
> false often enough to pay for itself. Or has been in the past, anyway.
> I suspect that access to the global variable rcu_state.gp_flags is not
> always fast either.
>
> So I am having difficulty talking myself into modifying this one given
> the frequency of operations.
>
> Thanx, Paul