Re: [PATCHv2 1/1] locking/qspinlock/x86: Avoid test-and-set when PV_DEDICATED is set

[Eduardo Valentin <eduval@amazon.com>]

On Fri, Nov 03, 2017 at 11:09:53AM +0100, Paolo Bonzini wrote:
> On 02/11/2017 19:43, Eduardo Valentin wrote:
> > On Thu, Nov 02, 2017 at 07:24:16PM +0100, Paolo Bonzini wrote:
> >> On 02/11/2017 19:08, Eduardo Valentin wrote:
> >>> On Thu, Nov 02, 2017 at 06:56:46PM +0100, Paolo Bonzini wrote:
> >>>> On 02/11/2017 18:45, Eduardo Valentin wrote:
> >>>>> Currently, the existing qspinlock implementation will fallback to
> >>>>> test-and-set if the hypervisor has not set the PV_UNHALT flag.
> >>>>>
> >>>>> This patch gives the opportunity to guest kernels to select
> >>>>> between test-and-set and the regular queueu fair lock implementation
> >>>>> based on the PV_DEDICATED KVM feature flag. When the PV_DEDICATED
> >>>>> flag is not set, the code will still fall back to test-and-set,
> >>>>> but when the PV_DEDICATED flag is set, the code will use
> >>>>> the regular queue spinlock implementation.
> >>>>
> >>>> Have you seen Waiman's series that lets you specify this on the guest
> >>>> command line instead?  Would this be acceptable for your use case?
> >>>
> >>> No, can you please share a link to it? is it already merged to tip/master?
> >>
> >> [PATCH-tip v2 0/2] x86/paravirt: Enable users to choose PV lock type
> >> https://lkml.org/lkml/2017/11/1/655
> >>
> >>>> (In other words, is there a difference for you between making the host
> >>>> vs. guest administrator toggle the feature?  "@amazon.com" means you are
> >>>> the host admin, how would you use it?)
> >>>
> >>> The way I think of this is this is a flag set by host side so the
> >>> guest adapts accordingly.
> >>>
> >>> If the admin in guest side wants to ignore what the host is
> >>> flagging, that is a different story.
> >>
> >> Okay, this makes sense.  But perhaps it should be a separate CPUID leaf,
> >> such as "configuration hints", rather than properly a feature.
> > 
> > Oh OK, you don't think this starts to deviate from the feature concept.
> > But would the PV_UNHALT also go to "configuration hints" bucket?
> 
> PV_UNHALT says whether the pvqspinlock API is available, PV_DEDICATED
> says whether it should be used.
> 
> > Another way to see this is we have three locking feature options to select from,
> > so we need at least two bits here.
> 
> PV_DEDICATED = 1, PV_UNHALT = anything: default is qspinlock
> PV_DEDICATED = 0, PV_UNHALT = 1: default is pvqspinlock
> PV_DEDICATED = 0, PV_UNHALT = 0: default is tas
> 
> What do you think?

Sounds reasonable, and it is almost what the patch does. But to achieve the above
table, we need in include the following chunk:
diff --git a/arch/x86/kernel/kvm.c b/arch/x86/kernel/kvm.c
index 8bb9594..dacd7cf 100644
--- a/arch/x86/kernel/kvm.c
+++ b/arch/x86/kernel/kvm.c
@@ -642,6 +642,8 @@ void __init kvm_spinlock_init(void)
 {
        if (!kvm_para_available())
                return;
+       if (kvm_para_has_feature(KVM_FEATURE_PV_DEDICATED))
+               return;
        /* Does host kernel support KVM_FEATURE_PV_UNHALT? */
        if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
                return;

Now we get:
PV_DEDICATED	PV_UNHALT	IMPLEMENTATION
	1	X		qspinlock
	0	1		pvspinlock
	0	0		tas

Do you still think PV_DEDICATED goes as a "configuration hint", given that it would take precedence on a feature (PV_UNHALT)?
Or do we keep everything as features (two bits to represent selection of three features)?

BR,


> 
> Paolo

-- 
All the best,
Eduardo Valentin