All cases that I know of so far, involve a vnode that's being flushed out
to disk after going "inactive". The process in state `D' is waiting
in vflushbuf(). Obviously, there's a race somewhere for the buffers that
were (once) associated with the vnode in question.

-pk