2020-03-18 03:03:11

by Yang Shi

Subject: [v4 PATCH 2/2] mm: swap: use smp_mb__after_atomic() to order LRU bit set

Memory barrier is needed after setting LRU bit, but smp_mb() is too
strong. Some architectures, i.e. x86, imply memory barrier with atomic
operations, so replacing it with smp_mb__after_atomic() sounds better,
which is nop on strong ordered machines, and full memory barriers on
others. With this change the vm-scalability cases would perform better
on x86, I saw total 6% improvement with this patch and previous inline
fix.

The test data (lru-file-readtwice throughput) against v5.6-rc4:
mainline w/ inline fix w/ both (adding this)
150MB 154MB 159MB

Fixes: 9c4e6b1a7027 ("mm, mlock, vmscan: no more skipping pagevecs")
Acked-by: Vlastimil Babka <[email protected]>
Reviewed-and-Tested-by: Shakeel Butt <[email protected]>
Signed-off-by: Yang Shi <[email protected]>
---
v3: * Added Shakeel's review-and-test signature.
v2: * Solved the comment from Vlastimil and added his Ack.

mm/swap.c | 4 ++--
1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/mm/swap.c b/mm/swap.c
index cf39d24..74ea08a 100644
--- a/mm/swap.c
+++ b/mm/swap.c
@@ -931,7 +931,6 @@ static void __pagevec_lru_add_fn(struct page *page, struct lruvec *lruvec,

VM_BUG_ON_PAGE(PageLRU(page), page);

- SetPageLRU(page);
/*
* Page becomes evictable in two ways:
* 1) Within LRU lock [munlock_vma_page() and __munlock_pagevec()].
@@ -958,7 +957,8 @@ static void __pagevec_lru_add_fn(struct page *page, struct lruvec *lruvec,
* looking at the same page) and the evictable page will be stranded
* in an unevictable LRU.
*/
- smp_mb();
+ SetPageLRU(page);
+ smp_mb__after_atomic();

if (page_evictable(page)) {
lru = page_lru(page);
--
1.8.3.1

2020-03-19 14:06:35

by Johannes Weiner

Subject: Re: [v4 PATCH 2/2] mm: swap: use smp_mb__after_atomic() to order LRU bit set

On Wed, Mar 18, 2020 at 11:02:21AM +0800, Yang Shi wrote:
> Memory barrier is needed after setting LRU bit, but smp_mb() is too
> strong. Some architectures, i.e. x86, imply memory barrier with atomic
> operations, so replacing it with smp_mb__after_atomic() sounds better,
> which is nop on strong ordered machines, and full memory barriers on
> others. With this change the vm-scalability cases would perform better
> on x86, I saw total 6% improvement with this patch and previous inline
> fix.
>
> The test data (lru-file-readtwice throughput) against v5.6-rc4:
> mainline w/ inline fix w/ both (adding this)
> 150MB 154MB 159MB
>
> Fixes: 9c4e6b1a7027 ("mm, mlock, vmscan: no more skipping pagevecs")
> Acked-by: Vlastimil Babka <[email protected]>
> Reviewed-and-Tested-by: Shakeel Butt <[email protected]>
> Signed-off-by: Yang Shi <[email protected]>

Acked-by: Johannes Weiner <[email protected]>