__page_cache_release() needs to recheck the page count inside the LRU
lock, because shrink_cache() may have found the page on the LRU and
incremented its refcount again.

Which is carefully documented over __pagevec_release().  Duh.