Re: [PATCH v3 13/19] KVM: arm64: Add support KVM_SYSTEM_EVENT_SUSPEND to PSCI SYSTEM_SUSPEND

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

 



On Thu, 24 Feb 2022 20:05:59 +0000,
Oliver Upton <oupton@xxxxxxxxxx> wrote:
> 
> On Thu, Feb 24, 2022 at 03:40:15PM +0000, Marc Zyngier wrote:
> > > diff --git a/arch/arm64/kvm/psci.c b/arch/arm64/kvm/psci.c
> > > index 2bb8d047cde4..a7de84cec2e4 100644
> > > --- a/arch/arm64/kvm/psci.c
> > > +++ b/arch/arm64/kvm/psci.c
> > > @@ -245,6 +245,11 @@ static int kvm_psci_system_suspend(struct kvm_vcpu *vcpu)
> > >  		return 1;
> > >  	}
> > >  
> > > +	if (kvm->arch.system_suspend_exits) {
> > > +		kvm_vcpu_set_system_event_exit(vcpu, KVM_SYSTEM_EVENT_SUSPEND);
> > > +		return 0;
> > > +	}
> > > +
> > 
> > So there really is a difference in behaviour here. Userspace sees the
> > WFI behaviour before reset (it implements it), while when not using
> > the SUSPEND event, reset occurs before anything else.
> > 
> > They really should behave in a similar way (WFI first, reset next).
> 
> I mentioned this on the other patch, but I think the conversation should
> continue here as UAPI context is in this one.
> 
> If SUSPEND exits are disabled and SYSTEM_SUSPEND is implemented in the
> kernel, userspace cannot observe any intermediate state. I think it is
> necessary for migration, otherwise if userspace were to save the vCPU
> post-WFI, pre-reset the pending reset would get lost along the way.
> 
> As far as userspace is concerned, I think the WFI+reset operation is
> atomic. SUSPEND exits just allow userspace to intervene before said
> atomic operation.
>
> Perhaps I'm missing something: assuming SUSPEND exits are disabled, what
> value is provided to userspace if it can see WFI behavior before the
> reset?

Signals get in the way, and break the notion of atomicity. Userspace
*will* observe this.

I agree that save/restore is an important point, and that snapshoting
the guest at this stage should capture the reset value. But it is the
asymmetry of the behaviours that I find jarring:

- if you ask for userspace exit, no reset value is applied and you
  need to implement the reset in userspace

- if you *don't* ask for a userspace exit, the reset values are
  applied, and a signal while in WFI will result in this reset being
  observed

Why can't the userspace exit path also apply the reset values *before*
exiting? After all, you can model this exit to userspace as
reset+WFI+'spurious exit from WFI'. This would at least unify the two
behaviours.

I still dislike the reset state being applied early, but consistency
(and save/restore) trumps taste here. I know I'm being pedantic here,
but we've been burned with loosely defined semantics in the past, and
I want to get this right. Or less wrong.

Thanks,

	M.

-- 
Without deviation from the norm, progress is not possible.



[Index of Archives]     [KVM ARM]     [KVM ia64]     [KVM ppc]     [Virtualization Tools]     [Spice Development]     [Libvirt]     [Libvirt Users]     [Linux USB Devel]     [Linux Audio Users]     [Yosemite Questions]     [Linux Kernel]     [Linux SCSI]     [XFree86]

  Powered by Linux