On Mon, May 04, 2015 at 07:54:59PM +0900, Minchan Kim wrote:

So, I guess once below compiler optimization happens in __page_set_anon_rmap,
it could be corrupt in page_refernced.

__page_set_anon_rmap:
        page->mapping = (struct address_space *) anon_vma;
        page->mapping = (struct address_space *)((void *)page_mapping + PAGE_MAPPING_ANON);

Because page_referenced checks it with PageAnon which has no memory barrier.
So if above compiler optimization happens, page_referenced can pass the anon
page in rmap_walk_file, not ramp_walk_anon. It's my theory. :)

FWIW

If such splits were possible, we would have bugs all over the kernel
IMO. An example is do_wp_page() vs shrink_active_list(). In do_wp_page()
we can call page_move_anon_rmap(), which sets page->mapping in exactly
the same fashion as above-mentioned __page_set_anon_rmap():

	anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
	page->mapping = (struct address_space *) anon_vma;

The page in question may be on an LRU list, because nowhere in
do_wp_page() we remove it from the list, neither do we take any LRU
related locks. The page is locked, that's true, but shrink_active_list()
calls page_referenced() on an unlocked page, so according to your logic
they can race with the latter receiving a page with page->mapping equal
to anon_vma w/o PAGE_MAPPING_ANON bit set:

CPU0				CPU1
----				----
do_wp_page			shrink_active_list
 lock_page			 page_referenced
				  PageAnon->yes, so skip trylock_page
 page_move_anon_rmap
  page->mapping = anon_vma
				  rmap_walk
				   PageAnon->no
				   rmap_walk_file
				    BUG
  page->mapping = page->mapping+PAGE_MAPPING_ANON

However, this does not happen.

Thanks,
Vladimir