Re: page cache: Store only head pages in i_pages

[Matthew Wilcox <willy@infradead.org>]

On Fri, Mar 29, 2019 at 05:25:34PM -0400, Qian Cai wrote:
> On Fri, 2019-03-29 at 12:59 -0700, Matthew Wilcox wrote:
> > Oh ... is it a race?
> > 
> > so CPU A does:
> > 
> > page = find_get_page(swap_address_space(entry), offset)
> >         page = find_subpage(page, offset);
> > trylock_page(page);
> > 
> > while CPU B does:
> > 
> > xa_lock_irq(&address_space->i_pages);
> > __delete_from_swap_cache(page, entry);
> >         xas_store(&xas, NULL);
> >         ClearPageSwapCache(page);
> > xa_unlock_irq(&address_space->i_pages);
> > 
> > and if the ClearPageSwapCache happens between the xas_load() and the
> > find_subpage(), we're stuffed.  CPU A has a reference to the page, but
> > not a lock, and find_get_page is running under RCU.
> > 
> > I suppose we could fix this by taking the i_pages xa_lock around the
> > call to find_get_pages().  If indeed, that's what this problem is.
> > Want to try this patch?
> 
> Confirmed. Well spotted!

Excellent!  I'm not comfortable with the rule that you have to be holding
the i_pages lock in order to call find_get_page() on a swap address_space.
How does this look to the various smart people who know far more about the
MM than I do?

The idea is to ensure that if this race does happen, the page will be
handled the same way as a pagecache page.  If __delete_from_swap_cache()
can be called while the page is still part of a VMA, then this patch
will break page_to_pgoff().  But I don't think that can happen ... ?

(also, is this the right memory barrier to use to ensure that the old
value of page->index cannot be observed if the PageSwapCache bit is
obseved as having been cleared?)

diff --git a/mm/swap_state.c b/mm/swap_state.c
index 2e15cc335966..a715efcf0991 100644
--- a/mm/swap_state.c
+++ b/mm/swap_state.c
@@ -165,13 +165,16 @@ void __delete_from_swap_cache(struct page *page, swp_entry_t entry)
 	VM_BUG_ON_PAGE(!PageSwapCache(page), page);
 	VM_BUG_ON_PAGE(PageWriteback(page), page);
 
+	page->index = idx;
+	smp_mb__before_atomic();
+	ClearPageSwapCache(page);
+
 	for (i = 0; i < nr; i++) {
 		void *entry = xas_store(&xas, NULL);
 		VM_BUG_ON_PAGE(entry != page, entry);
 		set_page_private(page + i, 0);
 		xas_next(&xas);
 	}
-	ClearPageSwapCache(page);
 	address_space->nrpages -= nr;
 	__mod_node_page_state(page_pgdat(page), NR_FILE_PAGES, -nr);
 	ADD_CACHE_INFO(del_total, nr);