On Thu, Sep 22, 2022 at 03:23:38PM -0600, Alex Williamson wrote: > Well, we don't really need to have this behavior, we could choose to > implement the first two patches with the caller holding the > group_lock. Only one of the callers needs the duplicate test, no-iommu > creates its own iommu_group and therefore cannot have an existing > device. I think patches 1 & 2 would look like below*, with patch 3 > simply moving the change from vfio_group_get() to refcount_inc() into > the equivalent place in vfio_group_find_of_alloc(). Thanks, Yeah, this is nice I just rebased it onto Kevins series and the reason I did this patch has evaporated so let me check it again, maybe we don't need it. Jason