Re: [PATCH] cpuidle-haltpoll: allow force loading on hosts without the REALTIME hint

From: Marcelo Tosatti <mtosatti@redhat.com>
To: "Maciej S. Szmigiero" <mail@maciej.szmigiero.name>,
	Paolo Bonzini <pbonzini@redhat.com>
Cc: "Rafael J. Wysocki" <rjw@rjwysocki.net>,
	Daniel Lezcano <daniel.lezcano@linaro.org>,
	Joao Martins <joao.m.martins@oracle.com>,
	Boris Ostrovsky <boris.ostrovsky@oracle.com>,
	linux-pm@vger.kernel.org, linux-kernel@vger.kernel.org,
	kvm@vger.kernel.org
Subject: Re: [PATCH] cpuidle-haltpoll: allow force loading on hosts without the REALTIME hint
Date: Fri, 6 Mar 2020 06:56:17 -0300	[thread overview]
Message-ID: <20200306095617.GD32190@fuller.cnet> (raw)
In-Reply-To: <114f7b8d-6f88-222a-d1fa-abcfc0e6a1f2@maciej.szmigiero.name>

On Mon, Feb 24, 2020 at 09:10:18PM +0100, Maciej S. Szmigiero wrote:
> (CC'ing also Marcelo as the cpuidle-haltpoll driver author and the KVM ML).

Hi Maciej,

> On 21.02.2020 18:43, Maciej S. Szmigiero wrote:
> > From: "Maciej S. Szmigiero" <maciej.szmigiero@oracle.com>
> > 
> > Before commit 1328edca4a14 ("cpuidle-haltpoll: Enable kvm guest polling
> > when dedicated physical CPUs are available") the cpuidle-haltpoll driver
> > could also be used in scenarios when the host does not advertise the
> > KVM_HINTS_REALTIME hint.
> > 
> > While the behavior introduced by the aforementioned commit makes sense as
> > the default 

It makes sense for the pCPU overcommitted case only.

> > there are cases where the old behavior is desired, for example,
> > when other kernel changes triggered by presence by this hint are unwanted,
> > for some workloads where the latency benefit from polling overweights the
> > loss from idle CPU capacity that otherwise would be available, or just when
> > running under older Qemu versions that lack this hint.
> > 
> > Let's provide a typical "force" module parameter that allows restoring the
> > old behavior.
> > 
> > Signed-off-by: Maciej S. Szmigiero <maciej.szmigiero@oracle.com>

I would rather: 

1) Switch back to the default non-overcommitted case.

or even better (but requires more investment)

2) Make on the flight dynamic configuration (after all pCPU
overcommitment=true/false is a property that changes during
the day, depending on system load).

But its up to Paolo to decide, really.

> > ---
> >  drivers/cpuidle/cpuidle-haltpoll.c | 12 +++++++++++-
> >  1 file changed, 11 insertions(+), 1 deletion(-)
> > 
> > diff --git a/drivers/cpuidle/cpuidle-haltpoll.c b/drivers/cpuidle/cpuidle-haltpoll.c
> > index b0ce9bc78113..07e5b36076bb 100644
> > --- a/drivers/cpuidle/cpuidle-haltpoll.c
> > +++ b/drivers/cpuidle/cpuidle-haltpoll.c
> > @@ -18,6 +18,11 @@
> >  #include <linux/kvm_para.h>
> >  #include <linux/cpuidle_haltpoll.h>
> >  
> > +static bool force __read_mostly;
> > +module_param(force, bool, 0444);
> > +MODULE_PARM_DESC(force,
> > +		 "Load even if the host does not provide the REALTIME hint");
> > +
> >  static struct cpuidle_device __percpu *haltpoll_cpuidle_devices;
> >  static enum cpuhp_state haltpoll_hp_state;
> >  
> > @@ -90,6 +95,11 @@ static void haltpoll_uninit(void)
> >  	haltpoll_cpuidle_devices = NULL;
> >  }
> >  
> > +static bool haltpool_want(void)
> > +{
> > +	return kvm_para_has_hint(KVM_HINTS_REALTIME) || force;
> > +}
> > +
> >  static int __init haltpoll_init(void)
> >  {
> >  	int ret;
> > @@ -102,7 +112,7 @@ static int __init haltpoll_init(void)
> >  	cpuidle_poll_state_init(drv);
> >  
> >  	if (!kvm_para_available() ||
> > -		!kvm_para_has_hint(KVM_HINTS_REALTIME))
> > +	    !haltpool_want())
> >  		return -ENODEV;
> >  
> >  	ret = cpuidle_register_driver(drv);
> > 