@@ -1329,6 +1329,12 @@ static void kvm_destroy_devices(struct kvm *kvm)
* We do not need to take the kvm->lock here, because nobody else
* has a reference to the struct kvm at this point and therefore
* cannot access the devices list anyhow.
+ *
+ * The device list is generally managed as an rculist, but list_del()
+ * is used intentionally here. If a bug in KVM introduced a reader that
+ * was not backed by a reference on the kvm struct, the hope is that
+ * it'd consume the poisoned forward pointer instead of suffering a
+ * use-after-free, even though this cannot be guaranteed.
*/
list_for_each_entry_safe(dev, tmp, &kvm->devices, vm_node) {
list_del(&dev->vm_node);
@@ -4725,7 +4731,8 @@ static int kvm_device_release(struct inode *inode, struct file *filp)
if (dev->ops->release) {
mutex_lock(&kvm->lock);
- list_del(&dev->vm_node);
+ list_del_rcu(&dev->vm_node);
+ synchronize_rcu();
dev->ops->release(dev);
mutex_unlock(&kvm->lock);
}
@@ -4808,7 +4815,7 @@ static int kvm_ioctl_create_device(struct kvm *kvm,
kfree(dev);
return ret;
}
- list_add(&dev->vm_node, &kvm->devices);
+ list_add_rcu(&dev->vm_node, &kvm->devices);
mutex_unlock(&kvm->lock);
if (ops->init)
@@ -4819,7 +4826,8 @@ static int kvm_ioctl_create_device(struct kvm *kvm,
if (ret < 0) {
kvm_put_kvm_no_destroy(kvm);
mutex_lock(&kvm->lock);
- list_del(&dev->vm_node);
+ list_del_rcu(&dev->vm_node);
+ synchronize_rcu();
if (ops->release)
ops->release(dev);
mutex_unlock(&kvm->lock);
@@ -366,6 +366,8 @@ static int kvm_vfio_create(struct kvm_device *dev, u32 type)
struct kvm_device *tmp;
struct kvm_vfio *kv;
+ lockdep_assert_held(&dev->kvm->lock);
+
/* Only one VFIO "device" per VM */
list_for_each_entry(tmp, &dev->kvm->devices, vm_node)
if (tmp->ops == &kvm_vfio_ops)
A subsequent change to KVM/arm64 will necessitate walking the device list outside of the kvm->lock. Prepare by converting to an rculist. This has zero effect on the VM destruction path, as it is expected every reader is backed by a reference on the kvm struct. On the other hand, ensure a given device is completely destroyed before dropping the kvm->lock in the release() path, as certain devices expect to be a singleton (e.g. the vfio-kvm device). Cc: Paolo Bonzini <pbonzini@redhat.com> Cc: Sean Christopherson <seanjc@google.com> Signed-off-by: Oliver Upton <oliver.upton@linux.dev> --- virt/kvm/kvm_main.c | 14 +++++++++++--- virt/kvm/vfio.c | 2 ++ 2 files changed, 13 insertions(+), 3 deletions(-)