On 11/10/2019 10.22, Wei Yang wrote:
In function __anon_vma_prepare(), we will try to find anon_vma if it is
possible to reuse it. While on fork, the logic is different.
Since commit 5beb49305251 ("mm: change anon_vma linking to fix
multi-process server scalability issue"), function anon_vma_clone()
tries to allocate new anon_vma for child process. But the logic here
will allocate a new anon_vma for each vma, even in parent this vma
is mergeable and share the same anon_vma with its sibling. This may do
better for scalability issue, while it is not necessary to do so
especially after interval tree is used.
Commit 7a3ef208e662 ("mm: prevent endless growth of anon_vma hierarchy")
tries to reuse some anon_vma by counting child anon_vma and attached
vmas. While for those mergeable anon_vmas, we can just reuse it and not
necessary to go through the logic.
After this change, kernel build test reduces 20% anon_vma allocation.
Do the same kernel build test, it shows run time in sys reduced 11.6%.
Origin:
real 2m50.467s
user 17m52.002s
sys 1m51.953s
real 2m48.662s
user 17m55.464s
sys 1m50.553s
real 2m51.143s
user 17m59.687s
sys 1m53.600s
Patched:
real 2m39.933s
user 17m1.835s
sys 1m38.802s
real 2m39.321s
user 17m1.634s
sys 1m39.206s
real 2m39.575s
user 17m1.420s
sys 1m38.845s
Signed-off-by: Wei Yang <richardw.yang@xxxxxxxxxxxxxxx>
Acked-by: Konstantin Khlebnikov <khlebnikov@xxxxxxxxxxxxxx>
---
mm/rmap.c | 13 +++++++++++++
1 file changed, 13 insertions(+)
diff --git a/mm/rmap.c b/mm/rmap.c
index c34414567474..2c13e2bfd393 100644
--- a/mm/rmap.c
+++ b/mm/rmap.c
@@ -268,6 +268,19 @@ int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
{
struct anon_vma_chain *avc, *pavc;
struct anon_vma *root = NULL;
+ struct vm_area_struct *prev = dst->vm_prev, *pprev = src->vm_prev;
+
+ /*
+ * If parent share anon_vma with its vm_prev, keep this sharing in in
+ * child.
+ *
+ * 1. Parent has vm_prev, which implies we have vm_prev.
+ * 2. Parent and its vm_prev have the same anon_vma.
+ */
+ if (!dst->anon_vma && src->anon_vma &&
+ pprev && pprev->anon_vma == src->anon_vma)
+ dst->anon_vma = prev->anon_vma;
+
I believe that in present code "prev" cannot be NULL if !dst->anon_vma && src->anon_vma is true.
It would be safer to check this explicitly.
list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
struct anon_vma *anon_vma;
