Re: [PATCH v2 6/8] mm/lru: remove rcu_read_lock to fix performance regression

From: Matthew Wilcox
Date: Tue Nov 12 2019 - 09:38:46 EST


On Tue, Nov 12, 2019 at 10:06:26PM +0800, Alex Shi wrote:
> Intel 0day report there are performance regression on this patchset.
> The detailed info points to rcu_read_lock + PROVE_LOCKING which causes
> queued_spin_lock_slowpath waiting too long time to get lock.
> Remove rcu_read_lock is safe here since we had a spinlock hold.

Argh. You have not sent these patches in a properly reviewable form!
I wasted all that time reviewing the earlier patch in this series only to
find out that you changed it here. FIX THE PATCH, don't send a fix-patch
on top of it!

> Reported-by: kbuild test robot <lkp@xxxxxxxxx>
> Signed-off-by: Alex Shi <alex.shi@xxxxxxxxxxxxxxxxx>
> Cc: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>
> Cc: Johannes Weiner <hannes@xxxxxxxxxxx>
> Cc: Roman Gushchin <guro@xxxxxx>
> Cc: Shakeel Butt <shakeelb@xxxxxxxxxx>
> Cc: Chris Down <chris@xxxxxxxxxxxxxx>
> Cc: Tejun Heo <tj@xxxxxxxxxx>
> Cc: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
> Cc: linux-mm@xxxxxxxxx
> Cc: linux-kernel@xxxxxxxxxxxxxxx
> ---
> include/linux/memcontrol.h | 29 ++++++++++++-----------------
> 1 file changed, 12 insertions(+), 17 deletions(-)
>
> diff --git a/include/linux/memcontrol.h b/include/linux/memcontrol.h
> index 2421b720d272..f869897a68f0 100644
> --- a/include/linux/memcontrol.h
> +++ b/include/linux/memcontrol.h
> @@ -1307,20 +1307,18 @@ static inline struct lruvec *relock_page_lruvec_irq(struct page *page,
> struct pglist_data *pgdat = page_pgdat(page);
> struct lruvec *lruvec;
>
> - rcu_read_lock();
> + if (!locked_lruvec)
> + goto lock;
> +
> lruvec = mem_cgroup_page_lruvec(page, pgdat);
>
> - if (locked_lruvec == lruvec) {
> - rcu_read_unlock();
> + if (locked_lruvec == lruvec)
> return lruvec;
> - }
> - rcu_read_unlock();
>
> - if (locked_lruvec)
> - spin_unlock_irq(&locked_lruvec->lru_lock);
> + spin_unlock_irq(&locked_lruvec->lru_lock);
>
> +lock:
> lruvec = lock_page_lruvec_irq(page, pgdat);
> -
> return lruvec;
> }
>
> @@ -1331,21 +1329,18 @@ static inline struct lruvec *relock_page_lruvec_irqsave(struct page *page,
> struct pglist_data *pgdat = page_pgdat(page);
> struct lruvec *lruvec;
>
> - rcu_read_lock();
> + if (!locked_lruvec)
> + goto lock;
> +
> lruvec = mem_cgroup_page_lruvec(page, pgdat);
>
> - if (locked_lruvec == lruvec) {
> - rcu_read_unlock();
> + if (locked_lruvec == lruvec)
> return lruvec;
> - }
> - rcu_read_unlock();
>
> - if (locked_lruvec)
> - spin_unlock_irqrestore(&locked_lruvec->lru_lock,
> - locked_lruvec->flags);
> + spin_unlock_irqrestore(&locked_lruvec->lru_lock, locked_lruvec->flags);
>
> +lock:
> lruvec = lock_page_lruvec_irqsave(page, pgdat);
> -
> return lruvec;
> }
>
> --
> 1.8.3.1
>
>