In my previous patch that fixes a bunch of TLB issues,
a new bug was introduced. Since we now rely on mm->cpu_vm_mask
to indicate the true MMU owner, we need to update this flag
whenever there is ownership change.
It turns out that activate_mm() does the ownership as well.
Here is the patch that fixes this problem.
A big thank to Clausen for spotting this and tracing it to
a great depth.
Jun