Re: [PATCH v2 46/46] mm/filemap: Add FGP_STABLE

On Wed, Jun 23, 2021 at 01:43:30PM +0200, Christoph Hellwig wrote:

On Tue, Jun 22, 2021 at 01:15:51PM +0100, Matthew Wilcox (Oracle) wrote:

quoted

Allow filemap_get_folio() to wait for writeback to complete (if the
filesystem wants that behaviour).  This is the folio equivalent of
grab_cache_page_write_begin(), which is moved into the folio-compat
file as a reminder to migrate all the code using it.  This paves the
way for getting rid of AOP_FLAG_NOFS once grab_cache_page_write_begin()
is removed.

We actually should kill FGP_NOFS as well by switching everything over
to memalloc_nofs_{save, restore} eventually, given how error prone
all these manual flags settings are.

Well, yes, but it's been four years and we still have over 1100 uses of
GFP_NOFS.  Until someone takes on that Augean Stables, we're going to need
FGP_NOFS.  I added that context to the readahead path in f2c817bed58d,
but of course that doesn't let me remove any uses of GFP_NOFS.

quoted

diff --git a/mm/folio-compat.c b/mm/folio-compat.c
index 78365eaee7d3..206bedd621d0 100644
--- a/mm/folio-compat.c
+++ b/mm/folio-compat.c

@@ -115,6 +115,7 @@ void lru_cache_add(struct page *page)
 }
 EXPORT_SYMBOL(lru_cache_add);
 
+noinline
 struct page *pagecache_get_page(struct address_space *mapping, pgoff_t index,
 		int fgp_flags, gfp_t gfp)

How did that sneak in here?

Without it, pagecache_get_page() gets inlined by
grab_cache_page_write_begin() which is just too much code.