On Fri, Feb 25, 2022 at 01:31:09AM +0000, Matthew Wilcox wrote:
> On Mon, Feb 14, 2022 at 08:00:11PM +0000, Matthew Wilcox (Oracle) wrote:
> > folio_mapped() is expensive because it has to check each page's mapcount
> > field. A cheaper check is whether there are any extra references to
> > the page, other than the one we own and the ones held by the page cache.
> > The call to remove_mapping() will fail in any case if it cannot freeze
> > the refcount, but failing here avoids cycling the i_pages spinlock.
>
> This is the patch that's causing ltp's readahead02 test to break.
> Haven't dug into why yet, but it happens without large folios, so
> I got something wrong.
This fixes it:
+++ b/mm/truncate.c
@@ -288,7 +288,8 @@ int invalidate_inode_page(struct page *page)
if (folio_test_dirty(folio) || folio_test_writeback(folio))
return 0;
/* The refcount will be elevated if any page in the folio is mapped */
- if (folio_ref_count(folio) > folio_nr_pages(folio) + 1)
+ if (folio_ref_count(folio) >
+ folio_nr_pages(folio) + 1 + folio_has_private(folio))
return 0;
if (folio_has_private(folio) && !filemap_release_folio(folio, 0))
return 0;
Too late for today's -next, but I'll push it out tomorrow.
