[PATCH] mm/rmap: optimize folio_move_anon_rmap()

[PATCH] mm/rmap: optimize folio_move_anon_rmap()

[Chen Taotao]

When a folio belongs exclusively to one process after a COW event,
folio_move_anon_rmap() always moves the folio into the anon_vma
belongs only to this process.

However, if the folio already belongs to the anon_vma of the this
process, we don't need to move it again. In this case, we first
check if the folio already belongs to the anna_vma of the this
process, and only move it if it does not.

The above changes may improve the performance of vm faults in some
scenarios, because the performance loss caused by WRITE_ONCE() is
much more than the performance loss caused by add a judgment.

Signed-off-by: Chen Taotao <chentt10@chinatelecom.cn>
---
 mm/rmap.c | 11 +++++++----
 1 file changed, 7 insertions(+), 4 deletions(-)

diff --git a/mm/rmap.c b/mm/rmap.c
index 3746a5531018..5720cb78162d 100644
--- a/mm/rmap.c
+++ b/mm/rmap.c
@@ -1216,12 +1216,15 @@ void folio_move_anon_rmap(struct folio *folio, struct vm_area_struct *vma)
 	VM_BUG_ON_VMA(!anon_vma, vma);
 
 	anon_vma += 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.
+	 * If anon_vma != folio->mapping ensure that anon_vma and the
+	 * PAGE_MAPPING_ANON bit are writtensimultaneously, so a concurrent
+	 * reader (eg folio_referenced()'s folio_test_anon()) will not see
+	 * one without the other.
 	 */
-	WRITE_ONCE(folio->mapping, anon_vma);
+	if (anon_vma != folio->mapping)
+		WRITE_ONCE(folio->mapping, anon_vma);
 }
 
 /**
-- 
2.27.0

Re: [PATCH] mm/rmap: optimize folio_move_anon_rmap()

[David Hildenbrand]

On 12.05.24 14:35, Chen Taotao wrote:
> When a folio belongs exclusively to one process after a COW event,
> folio_move_anon_rmap() always moves the folio into the anon_vma
> belongs only to this process.
> 
> However, if the folio already belongs to the anon_vma of the this
> process, we don't need to move it again. In this case, we first
> check if the folio already belongs to the anna_vma of the this
> process, and only move it if it does not.
> 
> The above changes may improve the performance of vm faults in some
> scenarios, because the performance loss caused by WRITE_ONCE() is
> much more than the performance loss caused by add a judgment.

Please proof that by real numbers. I don't think it will make a real 
difference, and we likely don't want that change.


-- 
Cheers,

David / dhildenb

Re: [PATCH] mm/rmap: optimize folio_move_anon_rmap()

[Anshuman Khandual]

On 5/13/24 21:07, David Hildenbrand wrote:
> On 12.05.24 14:35, Chen Taotao wrote:
>> When a folio belongs exclusively to one process after a COW event,
>> folio_move_anon_rmap() always moves the folio into the anon_vma
>> belongs only to this process.
>>
>> However, if the folio already belongs to the anon_vma of the this
>> process, we don't need to move it again. In this case, we first
>> check if the folio already belongs to the anna_vma of the this
>> process, and only move it if it does not.
>>
>> The above changes may improve the performance of vm faults in some
>> scenarios, because the performance loss caused by WRITE_ONCE() is
>> much more than the performance loss caused by add a judgment.
> 
> Please proof that by real numbers. I don't think it will make a real difference, and we likely don't want that change.

Agreed, only scenarios when pre-condition check makes any sense is
if subsequent actions are expensive such as writing into registers
etc. But in this case both 'if' and 'WRITE_ONCE' statements are of
comparable cost, hence adding additional conditional check is only
going to increase the cost on average i.e when both gets executed.