Re: [PATCH RFC v2 0/3] mm: Properly document tail pages for a folio

[Matthew Wilcox <willy@infradead.org>]

On Mon, Aug 14, 2023 at 04:21:55PM -0400, Peter Xu wrote:
> On Mon, Aug 14, 2023 at 08:58:44PM +0100, Matthew Wilcox wrote:
> > On Mon, Aug 14, 2023 at 02:44:08PM -0400, Peter Xu wrote:
> > 
> > Look, this is all still too complicated.  And you're trying to make
> > something better that I'm trying to make disappear.  I'd really rather
> > you spent your time worrying about making userfaultfd use folios
> > than faffing with this.
> 
> I saw that internally some of uffd already start to use folio, while I
> don't think the syscall part needs changing yet - the ranged API should
> work for folio when it comes, and other than that folio should be hidden
> and transparent, afaiu.
> 
> Do you mean when large folios can land on anon/shmem we can start to
> allocate large folios there for uffd operations?  Or something else?

Hm, I thought there were some parts that still needed to be converted.
But I don't see anything obvious right now.

> > @@ -360,6 +363,7 @@ struct folio {
> >  			unsigned long _head_2a;
> >  	/* public: */
> >  			struct list_head _deferred_list;
> > +			/* three more words available here */
> 
> .. not really three more words here but 2 for 32 bits and 1 for 64 bits.
> In my patch 3 I used "8 bytes free" so it's applicable to both.

I always forget about THP_SWAP using tail->private.  That actually needs
to be asserted by the compiler, not just documented.  Something along
these lines.

diff --git a/include/linux/mm_types.h b/include/linux/mm_types.h
index 659c7b84726c..3880b3f2e321 100644
--- a/include/linux/mm_types.h
+++ b/include/linux/mm_types.h
@@ -340,8 +340,11 @@ struct folio {
 			atomic_t _pincount;
 #ifdef CONFIG_64BIT
 			unsigned int _folio_nr_pages;
-#endif
+			/* 4 byte gap here */
 	/* private: the union with struct page is transitional */
+			/* Fix THP_SWAP to not use tail->private */
+			unsigned long _private_1;
+#endif
 		};
 		struct page __page_1;
 	};
@@ -362,6 +365,9 @@ struct folio {
 	/* public: */
 			struct list_head _deferred_list;
 	/* private: the union with struct page is transitional */
+			unsigned long _avail_2a;
+			/* Fix THP_SWAP to not use tail->private */
+			unsigned long _private_2a;
 		};
 		struct page __page_2;
 	};
@@ -386,12 +392,18 @@ FOLIO_MATCH(memcg_data, memcg_data);
 			offsetof(struct page, pg) + sizeof(struct page))
 FOLIO_MATCH(flags, _flags_1);
 FOLIO_MATCH(compound_head, _head_1);
+#ifdef CONFIG_64BIT
+FOLIO_MATCH(private, _private_1);
+#endif
 #undef FOLIO_MATCH
 #define FOLIO_MATCH(pg, fl)						\
 	static_assert(offsetof(struct folio, fl) ==			\
 			offsetof(struct page, pg) + 2 * sizeof(struct page))
 FOLIO_MATCH(flags, _flags_2);
 FOLIO_MATCH(compound_head, _head_2);
+FOLIO_MATCH(flags, _flags_2a);
+FOLIO_MATCH(compound_head, _head_2a);
+FOLIO_MATCH(private, _private_2a);
 #undef FOLIO_MATCH
 
 /*

This is against the patchset I just posted which frees up a word in the
first tail page.