[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [PATCH 00/36] cpuidle,rcu: Cleanup the mess



On Tue, Jun 14, 2022 at 12:19:29PM +0100, Mark Rutland wrote:
> On Wed, Jun 08, 2022 at 04:27:23PM +0200, Peter Zijlstra wrote:
> > Hi All! (omg so many)
> 
> Hi Peter,
> 
> Sorry for the delay; my plate has also been rather full recently. I'm 
> beginning
> to page this in now.

No worries; we all have too much to do ;-)

> > These here few patches mostly clear out the utter mess that is cpuidle vs 
> > rcuidle.
> > 
> > At the end of the ride there's only 2 real RCU_NONIDLE() users left
> > 
> >   arch/arm64/kernel/suspend.c:            RCU_NONIDLE(__cpu_suspend_exit());
> >   drivers/perf/arm_pmu.c:                 RCU_NONIDLE(armpmu_start(event, 
> > PERF_EF_RELOAD));
> 
> The latter of these is necessary because apparently PM notifiers are called
> with RCU not watching. Is that still the case today (or at the end of this
> series)? If so, that feels like fertile land for more issues (yaey...). If 
> not,
> we should be able to drop this.

That should be fixed; fingers crossed :-)

> >   kernel/cfi.c:   RCU_NONIDLE({
> > 
> > (the CFI one is likely dead in the kCFI rewrite) and there's only a hand 
> > full
> > of trace_.*_rcuidle() left:
> > 
> >   kernel/trace/trace_preemptirq.c:                        
> > trace_irq_enable_rcuidle(CALLER_ADDR0, CALLER_ADDR1);
> >   kernel/trace/trace_preemptirq.c:                        
> > trace_irq_disable_rcuidle(CALLER_ADDR0, CALLER_ADDR1);
> >   kernel/trace/trace_preemptirq.c:                        
> > trace_irq_enable_rcuidle(CALLER_ADDR0, caller_addr);
> >   kernel/trace/trace_preemptirq.c:                        
> > trace_irq_disable_rcuidle(CALLER_ADDR0, caller_addr);
> >   kernel/trace/trace_preemptirq.c:                
> > trace_preempt_enable_rcuidle(a0, a1);
> >   kernel/trace/trace_preemptirq.c:                
> > trace_preempt_disable_rcuidle(a0, a1);
> > 
> > All of them are in 'deprecated' code that is unused for GENERIC_ENTRY.
> 
> I think those are also unused on arm64 too?
> 
> If not, I can go attack that.

My grep spots:

arch/arm64/kernel/entry-common.c:               trace_hardirqs_on();
arch/arm64/include/asm/daifflags.h:     trace_hardirqs_off();
arch/arm64/include/asm/daifflags.h:             trace_hardirqs_off();

The _on thing should be replaced with something like:

        trace_hardirqs_on_prepare();
        lockdep_hardirqs_on_prepare();
        instrumentation_end();
        rcu_irq_exit();
        lockdep_hardirqs_on(CALLER_ADDR0);

(as I think you know, since you have some of that already). And
something similar for the _off thing, but with _off_finish().

> > I've touched a _lot_ of code that I can't test and likely broken some of it 
> > :/
> > In particular, the whole ARM cpuidle stuff was quite involved with OMAP 
> > being
> > the absolute 'winner'.
> > 
> > I'm hoping Mark can help me sort the remaining ARM64 bits as he moves that 
> > to
> > GENERIC_ENTRY.
> 
> Moving to GENERIC_ENTRY as a whole is going to take a tonne of work
> (refactoring both arm64 and the generic portion to be more amenable to each
> other), but we can certainly move closer to that for the bits that matter 
> here.

I know ... been there etc.. :-)

> Maybe we want a STRICT_ENTRY option to get rid of all the deprecated stuff 
> that
> we can select regardless of GENERIC_ENTRY to make that easier.

Possible yeah.

> > I've also got a note that says ARM64 can probably do a WFE based
> > idle state and employ TIF_POLLING_NRFLAG to avoid some IPIs.
> 
> Possibly; I'm not sure how much of a win that'll be given that by default 
> we'll
> have a ~10KHz WFE wakeup from the timer, but we could take a peek.

Ohh.. I didn't know it woke up *that* often. I just know Will made use
of it in things like smp_cond_load_relaxed() which would be somewhat
similar to a very shallow idle state that looks at the TIF word.



 


Rackspace

Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.