Re: [RFC 1/3] mm/ksm: add anonymous check in find_mergeable_vma

From: David Hildenbrand
Date: Fri Jun 07 2024 - 06:13:30 EST


On 07.06.24 11:33, Alex Shi wrote:


On 6/5/24 5:56 PM, David Hildenbrand wrote:
On 05.06.24 11:53, alexs@xxxxxxxxxx wrote:
From: "Alex Shi (tencent)" <alexs@xxxxxxxxxx>

We do vma_set_anonyous in do_mmap(), and then vma_is_anonymous()
checking workable, use it as a extra check since ksm only care anonymous
pages.

Signed-off-by: Alex Shi (tencent) <alexs@xxxxxxxxxx>
---
  mm/ksm.c | 3 ++-
  1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/mm/ksm.c b/mm/ksm.c
index f5138f43f0d2..088bce39cd33 100644
--- a/mm/ksm.c
+++ b/mm/ksm.c
@@ -742,7 +742,8 @@ static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm,
      if (ksm_test_exit(mm))
          return NULL;
      vma = vma_lookup(mm, addr);
-    if (!vma || !(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
+    if (!vma || !(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma ||
+            !vma_is_anonymous(vma))

Doesn't KSM also apply to COW'ed pages in !anon mappings? At least that's what I recall.
I didn't a evidence for this. :(

In write_protect_page(), "PageAnonExclusive(&folio->page);" has a "VM_BUG_ON_PGFLAGS(!PageAnon(page), page);"
So is this hints the vma also need to be anonymous one?


vma_is_anonymous(vma) is restricted to anonymous folios and the shared zeropage, but other VMAs (MAP_PRIVATE file-backed VMAs) can contain anonymous folios as well.

--
Cheers,

David / dhildenb