Re: [PATCH v5 tip/core/rcu 15/16] rcu: Make RCU-tasks wait for idle tasks

[Peter Zijlstra <peterz@infradead.org>]

[-- Attachment #1: Type: text/plain, Size: 3875 bytes --]

On Wed, Aug 13, 2014 at 11:20:30AM -0700, Paul E. McKenney wrote:
> cpuidle_idle_call() does a fastpath irq-enable/exit if need-resched,
> then does stop_critical_timings() and rcu_idle_enter().  Then we
> have the buried complexity with cpuidle_select(), but a negative
> return says to check need-resched and enable interrupts or to
> invoke arch_cpu_idle(), which executes various sleep instructions
> on various architectures.  Some notable variants:


> And various other architectures seem to work similarly, but lots of
> hair here.  So Steven, you OK with the underlying arch_cpu_idle()
> functions being off-limits to tracing?

I didn't find anything particularly hairy in the arch_cpu_idle()
implementations, lots of simple 'go sleep' or 'spin' like things.

> Now, if cpuidle_select() returns non-negative, we are dealing with
> the CPU-idle governor, which is invoked at the later cpuidle_enter().
> 
> Hmmm...  On the CPU-idle drivers...
> 
> o	apm_idle_driver puts the idle loop into the ->enter() function,
> 	apm_cpu_idle().

Yes, this one is creative. The best I came up with is
adding CPUIDLE_FLAG_RCU_IDLE which indicates that the driver will do the
rcu_idle calls and place them in apm_do_idle() around the
apm_bios_call_simple() thing.

Now, that apm_bios_call_simple() thing uses on_cpu0(), which schedules
work on cpu0, which to me seems to guarantee this won't be used on any
SMP system, because that simply _cannot_ work for idle.

And on UP its a few more function calls, we could sprinkle some
__always_inline()s around if we really care I suppose.

> o	ACPI puts the idle loop in acpi_idle_do_entry(), and does call
> 	stop_critical_timings(), but not rcu_idle_enter().
> 	So presumably stop_critical_timings() can nest?  Not clear
> 	from the code.

Yeah, so I'm not sure I see that they nest properly..

Still ACPI does a lot of weird crap in the busmaster idle function,
again I'd suggest that CPUIDLE_FLAG_RCU_IDLE which would let the driver
do rcu_idle itself, and place it in appropriate sites.

Not too hard I think in this case.

> o	The CPS driver is even stranger...  Is cps_gen_entry_code()
> 	really depositing assembly instructions into a buffer that is
> 	passed back as a function?

I had not yet looked at this one; its got that cpu_pm_{enter,exit}()
thing going.. we could do the same and place the manual RCU_IDLE around
cps_pm_enter_state()

> o	The intel_idle driver is the one with mwait_idle_with_hints(),
> 	so you covered it below.

Yeah, fairly straight fwd driver that, _lots_ saner than the ACPI one.

> Your patch covers the cpuidle_enter() transition, which means
> that functions like cpuidle_enter(), acpi_idle_enter_c1(), and
> acpi_idle_do_entry() would be off-limits to trampolining.  In the case
> of CPS, quite a bit of code.

So I think we can do this; sure lots of code, but typically 'simpler'
than RCU stuff.

> > We should push the rcu_idle_{enter,exit}() down to around
> > mwait_idle_with_hints(), so we don't call half the word with RCU
> > disabled.
> 
> That would be for the intel_idle.c CPU-idle driver.  The other drivers
> also need rcu_idle_{enter,exit}().

Right, so simple drivers can use the generic rcu_idle bits from
kernel/sched/idle.c and difficult drivers can use CPUIDLE_FLAG_RCU_IDLE
and do some manual cleverness.

> > > I have already said that I will be happy to rip out the wakeup code
> > > when it is no longer needed, and I agree that it would be way better if
> > > not needed.
> > 
> > I'd prefer to dtrt now and not needing to fix it later.
> 
> Once it works, I might consider it "right" and adjust accordingly.
> At the moment, speculation.

I think its simpler than doing RCU, maybe a little more work, but hey,
I'm the idiot that does full arch/ sweeps on a semi regular basis.


[-- Attachment #2: Type: application/pgp-signature, Size: 836 bytes --]