Re: [PATCH v7 02/10] mm/memcg: fold lru_lock in lock_page_lru

From: Konstantin Khlebnikov
Date: Mon Jan 13 2020 - 04:56:06 EST




On 13/01/2020 12.45, Alex Shi wrote:


å 2020/1/10 äå4:49, Konstantin Khlebnikov åé:
On 25/12/2019 12.04, Alex Shi wrote:
ÂFrom the commit_charge's explanations and mem_cgroup_commit_charge
comments, as well as call path when lrucare is ture, The lru_lock is
just to guard the task migration(which would be lead to move_account)
So it isn't needed when !PageLRU, and better be fold into PageLRU to
reduce lock contentions.

Signed-off-by: Alex Shi <alex.shi@xxxxxxxxxxxxxxxxx>
Cc: Johannes Weiner <hannes@xxxxxxxxxxx>
Cc: Michal Hocko <mhocko@xxxxxxxxxx>
Cc: Matthew Wilcox <willy@xxxxxxxxxxxxx>
Cc: Vladimir Davydov <vdavydov.dev@xxxxxxxxx>
Cc: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>
Cc: cgroups@xxxxxxxxxxxxxxx
Cc: linux-mm@xxxxxxxxx
Cc: linux-kernel@xxxxxxxxxxxxxxx
---
 mm/memcontrol.c | 9 ++++-----
 1 file changed, 4 insertions(+), 5 deletions(-)

diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index c5b5f74cfd4d..0ad10caabc3d 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c
@@ -2572,12 +2572,11 @@ static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
  static void lock_page_lru(struct page *page, int *isolated)
 {
-ÂÂÂ pg_data_t *pgdat = page_pgdat(page);
-
-ÂÂÂ spin_lock_irq(&pgdat->lru_lock);
ÂÂÂÂÂ if (PageLRU(page)) {
+ÂÂÂÂÂÂÂ pg_data_t *pgdat = page_pgdat(page);
ÂÂÂÂÂÂÂÂÂ struct lruvec *lruvec;
 + spin_lock_irq(&pgdat->lru_lock);

That's wrong. Here PageLRU must be checked again under lru_lock.
Hi, Konstantin,

For logical remain, we can get the lock and then release for !PageLRU.
but I still can figure out the problem scenario. Would like to give more hints?

That's trivial race: page could be isolated from lru between

if (PageLRU(page))
and
spin_lock_irq(&pgdat->lru_lock);





Also I don't like these functions:
- called lock/unlock but actually also isolates
- used just once
- pgdat evaluated twice

That's right. I will fold these functions into commit_charge.

Thanks
Alex