Re: [GIT PULL] hotfixes for 6.3-rc1

["Huang, Ying" <ying.huang@intel.com>]

Vlastimil Babka <vbabka@suse.cz> writes:

> On 3/6/23 02:25, Huang, Ying wrote:
>> Hi, Linus,
>> 
>> Linus Torvalds <torvalds@linux-foundation.org> writes:
>> 
>>> On Sat, Mar 4, 2023 at 3:21 PM Andrew Morton <akpm@linux-foundation.org> wrote:
>>>>
>>>> Ah. Ying did it this way:
>>>
>>> Yeah, I saw that patch flying past, but I actually think that it only
>>> silences the warning almost by mistake. There's nothing fundamental in
>>> there that a compiler wouldn't just follow across two assignments, and
>>> it just happens to now not trigger any more.
>>>
>>> Assigning to a union entry is a more fundamental operation in that
>>> respect. Not that the compiler still doesn't see that it's assigning a
>>> value that in the end is not really type compatible, so a different
>>> version of gcc could still warn, but at that point I feel like it's
>>> more of an actual compiler bug than just "oh, the compiler didn't
>>> happen to follow the cast through a temporary".
>> 
>> Yes.  Your fix is much better.  This can be used for
>> __page_set_anon_rmap() family too to make the code look better?
>
> Those are trickier due to the PAGE_MAPPING_ANON tagging bit which your
> code doesn't need to handle because you can simply store an untagged
> anon_vma pointer. Otherwise we could have the union at the struct page
> level already (but probably not at this point as IIRC Matthew is
> planning to have completely separate types for anon and file folios).
>
> So right now we have e.g. folio_get_anon_vma() using unsigned long as
> the intermediate variable, page_move_anon_rmap() using a void *
> variable, __page_set_anon_rmap() casting through a (void *) ... Is there
> a single recommended way for "tagged pointers" that we could unify that to?

Ah, you are right.  We need to deal with tagging bit and maybe struct
movable_operations *.  I tried to write the below debug patch (only
build test it).  The code adds 1 or 2 lines for each function.  But to
be honest, the original force casting method appears more natural to me.

Best Regards,
Huang, Ying

---------------------------8<------------------------------------
From 68a0f54921beca8aeaa8fe78867e62b5a66658b8 Mon Sep 17 00:00:00 2001
From: Huang Ying <ying.huang@intel.com>
Date: Thu, 9 Mar 2023 15:29:58 +0800
Subject: [PATCH] dbg mapping_ptr

---
 mm/rmap.c | 28 ++++++++++++++++++++--------
 1 file changed, 20 insertions(+), 8 deletions(-)

diff --git a/mm/rmap.c b/mm/rmap.c
index 8632e02661ac..50ee208baff9 100644
--- a/mm/rmap.c
+++ b/mm/rmap.c
@@ -466,6 +466,13 @@ void __init anon_vma_init(void)
 			SLAB_PANIC|SLAB_ACCOUNT);
 }
 
+union mapping_ptr {
+	struct address_space *mapping;
+	unsigned long tag;
+	struct anon_vma *anon_vma;
+	struct movable_operations *mops;
+};
+
 /*
  * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
  *
@@ -493,16 +500,17 @@ void __init anon_vma_init(void)
 struct anon_vma *folio_get_anon_vma(struct folio *folio)
 {
 	struct anon_vma *anon_vma = NULL;
-	unsigned long anon_mapping;
+	union mapping_ptr mptr;
 
 	rcu_read_lock();
-	anon_mapping = (unsigned long)READ_ONCE(folio->mapping);
-	if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
+	mptr.mapping = READ_ONCE(folio->mapping);
+	if ((mptr.tag & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
 		goto out;
 	if (!folio_mapped(folio))
 		goto out;
 
-	anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
+	mptr.tag &= ~PAGE_MAPPING_FLAGS;
+	anon_vma = mptr.anon_vma;
 	if (!atomic_inc_not_zero(&anon_vma->refcount)) {
 		anon_vma = NULL;
 		goto out;
@@ -1115,18 +1123,20 @@ int folio_total_mapcount(struct folio *folio)
 void page_move_anon_rmap(struct page *page, struct vm_area_struct *vma)
 {
 	void *anon_vma = vma->anon_vma;
+	union mapping_ptr mptr;
 	struct folio *folio = page_folio(page);
 
 	VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
 	VM_BUG_ON_VMA(!anon_vma, vma);
 
-	anon_vma += PAGE_MAPPING_ANON;
+	mptr.anon_vma = anon_vma;
+	mptr.tag |= PAGE_MAPPING_ANON;
 	/*
 	 * Ensure that anon_vma and the PAGE_MAPPING_ANON bit are written
 	 * simultaneously, so a concurrent reader (eg folio_referenced()'s
 	 * folio_test_anon()) will not see one without the other.
 	 */
-	WRITE_ONCE(folio->mapping, anon_vma);
+	WRITE_ONCE(folio->mapping, mptr.mapping);
 	SetPageAnonExclusive(page);
 }
 
@@ -1142,6 +1152,7 @@ static void __page_set_anon_rmap(struct folio *folio, struct page *page,
 	struct vm_area_struct *vma, unsigned long address, int exclusive)
 {
 	struct anon_vma *anon_vma = vma->anon_vma;
+	union mapping_ptr mptr;
 
 	BUG_ON(!anon_vma);
 
@@ -1162,8 +1173,9 @@ static void __page_set_anon_rmap(struct folio *folio, struct page *page,
 	 * the PAGE_MAPPING_ANON type identifier, otherwise the rmap code
 	 * could mistake the mapping for a struct address_space and crash.
 	 */
-	anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
-	WRITE_ONCE(folio->mapping, (struct address_space *) anon_vma);
+	mptr.anon_vma = anon_vma;
+	mptr.tag |= PAGE_MAPPING_ANON;
+	WRITE_ONCE(folio->mapping, mptr.mapping);
 	folio->index = linear_page_index(vma, address);
 out:
 	if (exclusive)
-- 
2.39.2