Re: [PATCH] rcu: Fix dynticks_nmi_nesting underflow check in rcu_is_cpu_rrupt_from_idle

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

On Tue, Jan 05, 2021 at 02:42:32PM +0100, Peter Zijlstra wrote:
> On Wed, Dec 23, 2020 at 07:12:31AM -0800, Paul E. McKenney wrote:
> > On Wed, Dec 23, 2020 at 02:09:37PM +0530, Neeraj Upadhyay wrote:
> > > For the smp_call_function() optimization, where callbacks can run from
> > > idle context, in commit 806f04e9fd2c ("rcu: Allow for smp_call_function()
> > > running callbacks from idle"), an additional check is added in
> > > rcu_is_cpu_rrupt_from_idle(), for dynticks_nmi_nesting value being 0,
> > > for these smp_call_function() callbacks running from idle loop.
> > > However, this commit missed updating a preexisting underflow check
> > > of dynticks_nmi_nesting, which checks for a non zero positive value.
> > > Fix this warning and while at it, read the counter only once.
> > > 
> > > Signed-off-by: Neeraj Upadhyay <neeraju@codeaurora.org>
> > > ---
> > > 
> > > Hi,
> > > 
> > > I was not able to get this warning, with scftorture.
> > > 
> > >   RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nmi_nesting) <= 0,
> > >     "RCU dynticks_nmi_nesting counter underflow/zero!");
> > > 
> > > Not sure if idle loop smp_call_function() optimization is already present
> > > in mainline?
> > 
> > Now that you mention it, I don't see it.
> 
> kernel/sched/idle.c:do_idle() calls flush_smp_call_function_from_idle().
> 
> (nothing x86 specific about it)

Got it, thank you!

The reason Neeraj was unable to trigger the problematic warning from
scftorture is that its smp_call_function() handlers do not invoke
rcu_is_cpu_rrupt_from_idle().  Without adding this to those handlers
(which would be a good change to make), the only way to trigger this is
for an expedited RCU grace period to IPI a CPU that goes idle while the
IPI is in flight, which is not the easiest thing to make happen.

> > > Another thing, which I am not sure of is, maybe lockdep gets disabled
> > > in the idle loop contexts, where rcu_is_cpu_rrupt_from_idle() is called?
> > > Was this the original intention, to keep the lockdep based
> > > RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nmi_nesting) <= 0
> > > check separate from idle task context nesting value
> > > WARN_ON_ONCE(!nesting && !is_idle_task(current)) check?
> > 
> > An easy way to test lockdep is to create a pair of locks, acquire them
> > in one order then release them both, and finally acquire them in the
> > other order and then release them both.  If lockdep is configured and
> > enabled, it will complain.
> 
> IIRC (and this is after not staring at the computer for 2 weeks) lockdep
> should work just fine in idle, except of course that RCU will be stopped
> so actually taking locks will scream bloody murder due to tracing etc..

Fair enough...

And I hope that the time off went well for you and yours!

> > The only reason I used RCU_LOCKDEP_WARN() was that people were complaining
> > to me about idle-entry overhead back at that time.  So without lockdep,
> > there is zero overhead.  Maybe people have become more tolerant of idle
> > delays, or perhaps they are not so worried about an extra check of a
> > cache-hot quantity.
> 
> Not having checks also saves on $I and branches, in general I think
> having checks depend on DEBUG features, esp. those we don't really
> expect to trigger is still sane.

OK, so should we convert the WARN_ON_ONCE() to RCU_LOCKDEP_WARN() while
we are in the area?

> > I am tempted to pull this in as is, given the current logical
> > inconsistency in the checks.  Thoughts?
> 
> Patch looks ok, although I've seen compilers do CSE on
> __this_cpu_read() (on x86).

True, but the compilers might might have a harder time of this on other
architectures.

							Thanx, Paul