Re: rcu-bh design

From: Paul E. McKenney
Date: Fri May 04 2018 - 13:30:54 EST


On Fri, May 04, 2018 at 12:30:50PM -0400, Steven Rostedt wrote:
> On Fri, 04 May 2018 16:20:11 +0000
> Joel Fernandes <joelaf@xxxxxxxxxx> wrote:
>
> > Hi Paul, everyone,
> >
> > I had some question(s) about rcu-bh design.
> > I am trying to understand the reasoning or need of it. I see that rcu-bh
> > will disable softirqs across read-side sections. But I am wondering why
> > this is needed. __do_softirq already disables softirq when a softirq
> > handler is running. The only reason I can see is, rcu-bh helps in
> > situations where - a softirq interrupts a preemptible RCU read-section and
> > prevents that read section from completing. But this problem would happen
> > if anyone where to use rcu-preempt - then does rcu-preempt even make sense
> > to use and shouldn't everyone be using rcu-bh?
>
> I thought rcu-bh uses softirqs as a quiescent state. Thus, blocking
> softirqs from happening makes sense. I don't think an
> rcu_read_lock_bh() makes sense in a softirq.

Agreed, any place in the code where bottom halves are enabled is an RCU-bh
quiescent state.

> > The other usecase for rcu-bh seems to be if context-switch is used as a
> > quiescent state, then softirq flood can prevent that from happening and
> > cause rcu grace periods from completing.
>
> > But preemptible RCU *does not* use context-switch as a quiescent state.
>
> It doesn't?

It does, but only sort of.

A context switch really is always an RCU-preempt quiescent state from
the perspective of the CPU. However, from the perspective of the task,
context switch is a quiescent state only if the task is not in an
RCU-preempt read-side critical section at the time.

> > So in that case rcu-bh would make
> > sense only in a configuration where we're not using preemptible-rcu at all
> > and are getting flooded by softirqs. Is that the reason rcu-bh needs to
> > exist?
>
> Maybe I'm confused by what you are asking.

The existence and use of RCU-bh in no way degrades the preemptibility of
RCU-preempt read-side critical sections. The reason is that RCU-bh and
RCU-preempt are completely separate, represented by different rcu_state
structure instances.

Thanx, Paul