Re: [PATCH] KVM: s390: reset crypto attributes for all vcpus

[Tony Krowiak]

On 04/20/2018 08:26 AM, David Hildenbrand wrote:
> On 19.04.2018 23:13, Tony Krowiak wrote:
> > Introduces a new function to reset the crypto attributes for all
> > vcpus whether they are running or not. Each vcpu in KVM will
> > be removed from SIE prior to resetting the crypto attributes in its
> > SIE state description. After all vcpus have had their crypto attributes
> > reset the vcpus will be restored to SIE.
> >
> > This function is incorporated into the kvm_s390_vm_set_crypto(kvm)
> > function to fix a reported issue whereby the crypto key wrapping
> > attributes could potentially get out of synch for running vcpus.
> >
> > Reported-by: Halil Pasic <pasic@xxxxxxxxxxxxxxxxxx>
> A reported-by for a code refactoring is strange.

I was asked to include this.

> > Signed-off-by: Tony Krowiak <akrowiak@xxxxxxxxxxxxxxxxxx>
> > ---
> >   arch/s390/kvm/kvm-s390.c |   18 ++++++++++++++----
> >   arch/s390/kvm/kvm-s390.h |   13 +++++++++++++
> >   2 files changed, 27 insertions(+), 4 deletions(-)
> >
> > diff --git a/arch/s390/kvm/kvm-s390.c b/arch/s390/kvm/kvm-s390.c
> > index fa355a6..4fa3037 100644
> > --- a/arch/s390/kvm/kvm-s390.c
> > +++ b/arch/s390/kvm/kvm-s390.c
> > @@ -789,6 +789,19 @@ static int kvm_s390_set_mem_control(struct kvm *kvm, struct kvm_device_attr *att
> >   	return ret;
> >   }
> >   
> > +void kvm_s390_vcpu_crypto_reset_all(struct kvm *kvm)
> > + {
> > +	int i;
> > +	struct kvm_vcpu *vcpu;
> > +
> > +	kvm_s390_vcpu_block_all(kvm);
> > +
> > +	kvm_for_each_vcpu(i, vcpu, kvm)
> > +	        kvm_s390_vcpu_crypto_setup(vcpu);
> > +
> > +	kvm_s390_vcpu_unblock_all(kvm);
> This code has to be protected by kvm->lock. Can that be guaranteed by
> the caller?

I can hold the kvm->lock in this function, but if you look at the 
function from which it
is called, kvm_s390_vm_set_crypto(vcpu) below, the kvm->lock is already 
held by that
function to do other work. I suppose the kvm_s390_vm_set_crypto(vcpu) 
instruction could
have released the lock prior to calling 
kvm_s390_vcpu_crypto_reset_all(kvm), but since
this function is called within a block of crypto work, it made sense to 
me to place
the responsibility for locking in the caller and include a comment in 
the comments for
the function definition:

Note: The kvm->lock must be held while calling this function



> > +}
> > +
> >   static void kvm_s390_vcpu_crypto_setup(struct kvm_vcpu *vcpu);
> >   
> >   static int kvm_s390_vm_set_crypto(struct kvm *kvm, struct kvm_device_attr *attr)
> > @@ -832,10 +845,7 @@ static int kvm_s390_vm_set_crypto(struct kvm *kvm, struct kvm_device_attr *attr)
> >   		return -ENXIO;
> >   	}
> >   
> > -	kvm_for_each_vcpu(i, vcpu, kvm) {
> > -		kvm_s390_vcpu_crypto_setup(vcpu);
> > -		exit_sie(vcpu);
> > -	}
> > +	kvm_s390_vcpu_crypto_reset_all(kvm);
> >   	mutex_unlock(&kvm->lock);
> >   	return 0;
> >   }
> > diff --git a/arch/s390/kvm/kvm-s390.h b/arch/s390/kvm/kvm-s390.h
> > index 1b5621f..981e3ba 100644
> > --- a/arch/s390/kvm/kvm-s390.h
> > +++ b/arch/s390/kvm/kvm-s390.h
> > @@ -410,4 +410,17 @@ static inline int kvm_s390_use_sca_entries(void)
> >   }
> >   void kvm_s390_reinject_machine_check(struct kvm_vcpu *vcpu,
> >   				     struct mcck_volatile_info *mcck_info);
> > +
> > +/**
> > + * kvm_s390_vcpu_crypto_reset_all
> > + *
> > + * Reset the crypto attributes for each vcpu. This can be done while the vcpus
> > + * are running as each vcpu will be removed from SIE before resetting the crypt
> > + * attributes and restored to SIE afterward.
> > + *
> > + * Note: The kvm->lock must be held while calling this function
> > + *
> > + * @kvm: the KVM guest
> > + */
> > +void kvm_s390_vcpu_crypto_reset_all(struct kvm *kvm);
> >   #endif