Re: [PATCH 01/12] mm/ksm: catch tail page abnormal in page_stable_node

From: David Hildenbrand
Date: Wed Mar 20 2024 - 05:29:48 EST


On 20.03.24 10:05, Alex Shi wrote:


On 3/18/24 8:25 PM, David Hildenbrand wrote:
On 18.03.24 13:14, alexs@xxxxxxxxxx wrote:
From: Alex Shi <alexs@xxxxxxxxxx>

commit 19138349ed59 ("mm/migrate: Add folio_migrate_flags()") change the
meaning of func page_stable_node() to check the compound head for tail
'page' instead of tail page self.
But seems both semantics are same at results, the func always return NULL
  for tail page. So adding a bug monitor here in case of abnormal.

Signed-off-by: Alex Shi <alexs@xxxxxxxxxx>
Cc: Izik Eidus <izik.eidus@xxxxxxxxxxxxxxxxxx>
Cc: Matthew Wilcox <willy@xxxxxxxxxxxxx>
Cc: Andrea Arcangeli <aarcange@xxxxxxxxxx>
Cc: Hugh Dickins <hughd@xxxxxxxxxx>
Cc: Chris Wright <chrisw@xxxxxxxxxxxx>
To: linux-kernel@xxxxxxxxxxxxxxx
To: linux-mm@xxxxxxxxx
To: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>
---
  mm/ksm.c | 7 ++++++-
  1 file changed, 6 insertions(+), 1 deletion(-)

diff --git a/mm/ksm.c b/mm/ksm.c
index 8c001819cf10..3ff469961927 100644
--- a/mm/ksm.c
+++ b/mm/ksm.c
@@ -1091,7 +1091,12 @@ static inline struct ksm_stable_node *folio_stable_node(struct folio *folio)
    static inline struct ksm_stable_node *page_stable_node(struct page *page)
  {
-    return folio_stable_node(page_folio(page));
+    struct ksm_stable_node *node;
+
+    node = folio_stable_node(page_folio(page));
+    VM_BUG_ON_PAGE(PageTail(page) && node, page);

I don't really understand why we would want this.

Only KSM folios can have a node in the stable tree. KSM folios cannot be large folios. At that is precisely what folio_stable_node() checks.

If we'd have a large folio identify as a KSM folio we'd be in much bigger trouble.


Besides, I'm sure you read "22) Do not crash the kernel" in Documentation/process/coding-style.rst


Hi David,

Thanks for comments!
Forgive my stupidity, I understand KSM stable tree has no compound pages, but when searching a tail page in ksm_do_scan(), why we couldn't be in a race, that another VM doing THP collapse on the same contents pages, while the 3rd vm is doing hugepage spliting?

We always call cmp_and_merge_page() while holding a reference on the page.

There, we call page_stable_node() directly and via stable_tree_search()->page_stable_node() on that page.

When stable_tree_search() returns a kpage, we also hold a reference to that kpage. So calling page_stable_node() on the kpage behaves the same.

As we are holding page references, pages cannot be split/merged and we should not see any races in page_stable_node().

Am I missing something?

Note that your change would also not help here: if it would be racy, you'd also not reliably catch any tail pages.

But it should not be racy unless I am missing something.

--
Cheers,

David / dhildenb