On Tue, 2019-11-26 at 09:14 -0800, Sean Christopherson wrote:
> On Tue, Nov 26, 2019 at 01:44:14PM -0300, Leonardo Bras wrote:
> > On Mon, 2019-10-21 at 15:58 -0700, Sean Christopherson wrote:
> 
> ...
> 
> > > diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
> > > index 67ef3f2e19e8..b8534c6b8cf6 100644
> > > --- a/virt/kvm/kvm_main.c
> > > +++ b/virt/kvm/kvm_main.c
> > > @@ -772,6 +772,18 @@ void kvm_put_kvm(struct kvm *kvm)
> > >  }
> > >  EXPORT_SYMBOL_GPL(kvm_put_kvm);
> > > 
> > > +/*
> > > + * Used to put a reference that was taken on behalf of an object
> > > associated
> > > + * with a user-visible file descriptor, e.g. a vcpu or device,
> > > if installation
> > > + * of the new file descriptor fails and the reference cannot be
> > > transferred to
> > > + * its final owner.  In such cases, the caller is still actively
> > > using @kvm and
> > > + * will fail miserably if the refcount unexpectedly hits zero.
> > > + */
> > > +void kvm_put_kvm_no_destroy(struct kvm *kvm)
> > > +{
> > > +	WARN_ON(refcount_dec_and_test(&kvm->users_count));
> > > +}
> > > +EXPORT_SYMBOL_GPL(kvm_put_kvm_no_destroy);
> > > 
> > >  static int kvm_vm_release(struct inode *inode, struct file
> > > *filp)
> > >  {
> > > @@ -2679,7 +2691,7 @@ static int kvm_vm_ioctl_create_vcpu(struct
> > > kvm
> > > *kvm, u32 id)
> > >  	kvm_get_kvm(kvm);
> > >  	r = create_vcpu_fd(vcpu);
> > >  	if (r < 0) {
> > > -		kvm_put_kvm(kvm);
> > > +		kvm_put_kvm_no_destroy(kvm);
> > >  		goto unlock_vcpu_destroy;
> > >  	}
> > > 
> > > @@ -3117,7 +3129,7 @@ static int kvm_ioctl_create_device(struct
> > > kvm
> > > *kvm,
> > >  	kvm_get_kvm(kvm);
> > >  	ret = anon_inode_getfd(ops->name, &kvm_device_fops, dev, O_RDWR
> > > > O_CLOEXEC);
> > >  	if (ret < 0) {
> > > -		kvm_put_kvm(kvm);
> > > +		kvm_put_kvm_no_destroy(kvm);
> > >  		mutex_lock(&kvm->lock);
> > >  		list_del(&dev->vm_node);
> > >  		mutex_unlock(&kvm->lock);
> > 
> > Hello,
> > 
> > I see what are you solving here, but would not this behavior cause
> > the
> > refcount to reach negative values?
> > 
> > If so, is not there a problem? I mean, in some archs (powerpc
> > included)
> > refcount_dec_and_test() will decrement and then test if the value
> > is
> > equal 0. If we ever reach a negative value, this will cause that
> > memory
> > to never be released. 
> > 
> > An example is that refcount_dec_and_test(), on other archs than
> > x86,
> > will call atomic_dec_and_test(), which on include/linux/atomic-
> > fallback.h will do:
> > 
> > return atomic_dec_return(v) == 0;
> > 
> > To change this behavior, it would mean change the whole
> > atomic_*_test
> > behavior, or do a copy function in order to change this '== 0' to 
> > '<= 0'. 
> > 
> > Does it make sense? Do you need any help on this?
> 
> I don't think so.  refcount_dec_and_test() will WARN on an underflow
> when
> the kernel is built with CONFIG_REFCOUNT_FULL=y.  I see no value in
> duplicating those sanity checks in KVM.
> 
> This new helper and WARN is to explicitly catch @users_count
> unexpectedly
> hitting zero, which is orthogonal to an underflow (although odds are
> good
> that a bug that triggers the WARN in kvm_put_kvm_no_destroy() will
> also
> lead to an underflow).  Leaking the memory is deliberate as the
> alternative
> is a guaranteed use-after-free, i.e. kvm_put_kvm_no_destroy() is
> intended
> to be used when users_count is guaranteed to be valid after it is
> decremented.


I agree an use-after-free more problem than a memory leak, but I think
that there is a way to solve this without leaking the memory also.

One option would be reordering the kvm_put_kvm(), like in this patch:
https://lkml.org/lkml/2019/11/26/517

And the other would be creating a new atomic operation that checks if
the counter is less than zero:

atomic_dec_and_test_negative(atomic_t *v)
{
	return atomic_dec_return(v) <= 0;
} 

And apply it to generic refcount.

Do you think that would work?

Best regards,

Leonardo Bras