Re: [PATCH v4 2/5] arm64: KVM: use mutex_trylock_nest_lock when locking all vCPUs

[Peter Zijlstra]

On Thu, May 01, 2025 at 01:44:28PM +0100, Marc Zyngier wrote:
> On Thu, 01 May 2025 12:15:52 +0100,
> Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> > 
> > > > + */
> > > > +int kvm_trylock_all_vcpus(struct kvm *kvm)
> > > > +{
> > > > +	struct kvm_vcpu *vcpu;
> > > > +	unsigned long i, j;
> > > > +
> > > > +	kvm_for_each_vcpu(i, vcpu, kvm)
> > > > +		if (!mutex_trylock_nest_lock(&vcpu->mutex, &kvm->lock))
> > 
> > This one includes an assertion that kvm->lock is actually held.
> 
> Ah, cunning. Thanks.
> 
> > That said, I'm not at all sure what the purpose of all this trylock
> > stuff is here.
> > 
> > Can someone explain? Last time I asked someone said something about
> > multiple VMs, but I don't know enough about kvm to know what that means.
> 
> Multiple VMs? That'd be real fun. Not.
> 
> > Are those vcpu->mutex another class for other VMs? Or what gives?
> 
> Nah. This is firmly single VM.
> 
> The purpose of this contraption is that there are some rare cases
> where we need to make sure that if we update some global state, all
> the vcpus of a VM need to see, or none of them.
> 
> For these cases, the guarantee comes from luserspace, and it gives the
> pinky promise that none of the vcpus are running at that point. But
> being of a suspicious nature, we assert that this is true by trying to
> take all the vcpu mutexes in one go. This will fail if a vcpu is
> running, as KVM itself takes the vcpu mutex before doing anything.
> 
> Similar requirement exists if we need to synthesise some state for
> userspace from all the individual vcpu states.

Ah, okay. Because x86 is simply doing mutex_lock() instead of
mutex_trylock() -- which would end up waiting for this activity to
subside I suppose.

Hence the use of the killable variant I suppose, for when they get tired
of waiting.

If all the architectures are basically doing the same thing, it might
make sense to unify this particular behaviour. But what do I know.