[RESEND [PATCH] 2/2] mm/mmap.c: unlink vma before rb_erase

From: Wei Yang <richardw.yang@linux.intel.com>
To: akpm@linux-foundation.org, vbabka@suse.cz,
	kirill.shutemov@linux.intel.com, yang.shi@linux.alibaba.com
Cc: linux-mm@kvack.org, linux-kernel@vger.kernel.org,
	Wei Yang <richardw.yang@linux.intel.com>
Subject: [RESEND [PATCH] 2/2] mm/mmap.c: unlink vma before rb_erase
Date: Wed, 28 Aug 2019 14:06:14 +0800	[thread overview]
Message-ID: <20190828060614.19535-3-richardw.yang@linux.intel.com> (raw)
In-Reply-To: <20190828060614.19535-1-richardw.yang@linux.intel.com>

Current sequence to remove a vma is:

  vma_rb_erase_ignore()
  __vma_unlink_list()
  vma_gap_update()

This may do some extra subtree_gap propagation due the vma is unlink
from list after rb_erase.

For example, we have a tree:

                    a
                    [0x9000, 0x10000]
                /            \
            b                 c
            [0x8000, 0x9000]  [0x10000, 0x11000]
         /
        d
        [0x6000, 0x7000]

The gap for each node is:

  a's gap = 0x6000
  b's gap = 0x6000
  c's gap = 0x0
  d's gap = 0x6000

Now we want to remove node d. Since we don't unlink d from link when
doing rb_erase, b's gap would still be computed to 0x1000. This leads to
the vma_gap_update() after list unlink would recompute b and a's gap.

For this case, by unlink the list before rb_erase, we would have one
time less of vma_compute_subtree_gap.

Signed-off-by: Wei Yang <richardw.yang@linux.intel.com>
---
 mm/mmap.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/mm/mmap.c b/mm/mmap.c
index 672ad7dc6b3c..907939690a30 100644
--- a/mm/mmap.c
+++ b/mm/mmap.c
@@ -678,8 +678,8 @@ static __always_inline void __vma_unlink_common(struct mm_struct *mm,
 						struct vm_area_struct *vma,
 						struct vm_area_struct *ignore)
 {
-	vma_rb_erase_ignore(vma, &mm->mm_rb, ignore);
 	__vma_unlink_list(mm, vma);
+	vma_rb_erase_ignore(vma, &mm->mm_rb, ignore);
 	/* Kill the cache */
 	vmacache_invalidate(mm);
 }
-- 
2.17.1

