Right now, the kvm module is kept alive by VFS via fops_get/fops_put, but there
may be cases in which a kvm_get_kvm's matching kvm_put_kvm happens after
the file descriptor is closed.  One case that will be introduced soon is
when work is delegated to the system work queue; the worker might be
a bit late and the file descriptor can be closed in the meantime.  Ensure
that the module has not gone away by tying a module reference explicitly
to the lifetime of the struct kvm.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
---
 virt/kvm/kvm_main.c | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
index 64eb99444688..e3f37fc2ebf1 100644
--- a/virt/kvm/kvm_main.c
+++ b/virt/kvm/kvm_main.c
@@ -1131,6 +1131,9 @@ static struct kvm *kvm_create_vm(unsigned long type)
 	preempt_notifier_inc();
 	kvm_init_pm_notifier(kvm);
 
+	/* This is safe, since we have a reference from open(). */
+	__module_get(THIS_MODULE);
+
 	return kvm;
 
 out_err:
@@ -1220,6 +1223,7 @@ static void kvm_destroy_vm(struct kvm *kvm)
 	preempt_notifier_dec();
 	hardware_disable_all();
 	mmdrop(mm);
+	module_put(THIS_MODULE);
 }
 
 void kvm_get_kvm(struct kvm *kvm)
-- 
2.31.1

On Wed, Mar 02, 2022, Paolo Bonzini wrote:
> Right now, the kvm module is kept alive by VFS via fops_get/fops_put, but there
> may be cases in which a kvm_get_kvm's matching kvm_put_kvm happens after
> the file descriptor is closed.  One case that will be introduced soon is
> when work is delegated to the system work queue; the worker might be
> a bit late and the file descriptor can be closed in the meantime.  Ensure
> that the module has not gone away by tying a module reference explicitly
> to the lifetime of the struct kvm.
> 
> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
> ---
>  virt/kvm/kvm_main.c | 4 ++++
>  1 file changed, 4 insertions(+)
> 
> diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
> index 64eb99444688..e3f37fc2ebf1 100644
> --- a/virt/kvm/kvm_main.c
> +++ b/virt/kvm/kvm_main.c
> @@ -1131,6 +1131,9 @@ static struct kvm *kvm_create_vm(unsigned long type)
>  	preempt_notifier_inc();
>  	kvm_init_pm_notifier(kvm);
>  
> +	/* This is safe, since we have a reference from open(). */
> +	__module_get(THIS_MODULE);

This isn't sufficient.  For x86, it only grabs a reference to kvm.ko, not the
vendor module.  Instead, we can do:

	if (!try_module_get(kvm_chardev_ops.owner))
		return ERR_PTR(-EINVAL);

And then on top, revert commit revert ("KVM: set owner of cpu and vm file operations").
vCPUs file descriptors hold reference to the VM, which means they indirectly hold a
reference to the module. So once the "real" bug of struct kvm not holding a reference
to the module is fixed, grabbing a reference when a VM/vCPU inode is opened becomes
unnecessary.

On 3/2/22 18:47, Sean Christopherson wrote:
>> +	/* This is safe, since we have a reference from open(). */
>> +	__module_get(THIS_MODULE);
> This isn't sufficient.  For x86, it only grabs a reference to kvm.ko, not the
> vendor module.  Instead, we can do:
> 
> 	if (!try_module_get(kvm_chardev_ops.owner))
> 		return ERR_PTR(-EINVAL);
> 
> And then on top, revert commit revert ("KVM: set owner of cpu and vm file operations").
> vCPUs file descriptors hold reference to the VM, which means they indirectly hold a
> reference to the module. So once the "real" bug of struct kvm not holding a reference
> to the module is fixed, grabbing a reference when a VM/vCPU inode is opened becomes
> unnecessary.
> 

Got it, I'll wait for David to post the right thing. :)

Paolo