On Tue, Jun 27, 2017 at 03:08:32PM +0200, Paolo Bonzini wrote:
> On 27/06/2017 13:57, Claudio Imbrenda wrote:
> > If I'm not missing anything, in case kvm_create_vm_debugfs fails, we
> > will have a memory leak due to not freeing the kvm object.
> >
> > A call to kvm_put_kvm was accidentally removed from an error handling in
> > commit 506cfba9e726 ("KVM: don't use anon_inode_getfd() before possible failures")
> >
> > This patch simply restores the call to kvm_put_kvm, so that the kvm
> > object is destroyed before returning an error.
> >
> > Signed-off-by: Claudio Imbrenda <imbrenda@xxxxxxxxxxxxxxxxxx>
> > Fixes: 506cfba9e726 ("KVM: don't use anon_inode_getfd() before possible failures")
> > ---
> > virt/kvm/kvm_main.c | 1 +
> > 1 file changed, 1 insertion(+)
> >
> > diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
> > index f0fe9d0..257d2a8 100644
> > --- a/virt/kvm/kvm_main.c
> > +++ b/virt/kvm/kvm_main.c
> > @@ -3194,6 +3194,7 @@ static int kvm_dev_ioctl_create_vm(unsigned long type)
> > if (kvm_create_vm_debugfs(kvm, r) < 0) {
> > put_unused_fd(r);
> > fput(file);
> > + kvm_put_kvm(kvm);
> > return -ENOMEM;
> > }
> >
> >
>
> Queued, thanks.
It's broken. Look: once we are past the anon_inode_getfile(), the
reference we held on kvm is transferred into new struct file. After
that point we don't drop kvm - we drop file. And as long as that
file is held, it will keep holding what used to be our reference to
kvm. Once all references to file are gone, its ->release() will be
called and that's where kvm reference in it will be dropped.
IOW, this patch introduces a double-put.
