Re: [PATCH next] mm/swap.c: reduce lock contention in lru_cache_add

[Yu Zhao <yuzhao@google.com>]

On Fri, Nov 20, 2020 at 04:27:27PM +0800, Alex Shi wrote:
> The current relock logical will change lru_lock when found a new
> lruvec, so if 2 memcgs are reading file or alloc page at same time,
> they could hold the lru_lock alternately, and wait for each other for
> fairness attribute of ticket spin lock.
> 
> This patch will sort that all lru_locks and only hold them once in
> above scenario. That could reduce fairness waiting for lock reget.
> Than, vm-scalability/case-lru-file-readtwice could get ~5% performance
> gain on my 2P*20core*HT machine.
> 
> Suggested-by: Konstantin Khlebnikov <koct9i@gmail.com>
> Signed-off-by: Alex Shi <alex.shi@linux.alibaba.com>
> Cc: Konstantin Khlebnikov <koct9i@gmail.com>
> Cc: Andrew Morton <akpm@linux-foundation.org>
> Cc: Hugh Dickins <hughd@google.com>
> Cc: Yu Zhao <yuzhao@google.com>
> Cc: Michal Hocko <mhocko@suse.com>
> Cc: linux-mm@kvack.org
> Cc: linux-kernel@vger.kernel.org
> ---
>  mm/swap.c | 57 +++++++++++++++++++++++++++++++++++++++++++++++--------
>  1 file changed, 49 insertions(+), 8 deletions(-)
> 
> diff --git a/mm/swap.c b/mm/swap.c
> index 490553f3f9ef..c787b38bf9c0 100644
> --- a/mm/swap.c
> +++ b/mm/swap.c
> @@ -1009,24 +1009,65 @@ static void __pagevec_lru_add_fn(struct page *page, struct lruvec *lruvec)
>  	trace_mm_lru_insertion(page, lru);
>  }
>  
> +struct lruvecs {
> +	struct list_head lists[PAGEVEC_SIZE];
> +	struct lruvec *vecs[PAGEVEC_SIZE];
> +};
> +
> +/* Sort pvec pages on their lruvec */
> +int sort_page_lruvec(struct lruvecs *lruvecs, struct pagevec *pvec)
> +{
> +	int i, j, nr_lruvec;
> +	struct page *page;
> +	struct lruvec *lruvec = NULL;
> +
> +	lruvecs->vecs[0] = NULL;
> +	for (i = nr_lruvec = 0; i < pagevec_count(pvec); i++) {
> +		page = pvec->pages[i];
> +		lruvec = mem_cgroup_page_lruvec(page, page_pgdat(page));
> +
> +		/* Try to find a same lruvec */
> +		for (j = 0; j <= nr_lruvec; j++)
> +			if (lruvec == lruvecs->vecs[j])
> +				break;
> +
> +		/* A new lruvec */
> +		if (j > nr_lruvec) {
> +			INIT_LIST_HEAD(&lruvecs->lists[nr_lruvec]);
> +			lruvecs->vecs[nr_lruvec] = lruvec;
> +			j = nr_lruvec++;
> +			lruvecs->vecs[nr_lruvec] = 0;
> +		}
> +
> +		list_add_tail(&page->lru, &lruvecs->lists[j]);
> +	}
> +
> +	return nr_lruvec;
> +}
> +
>  /*
>   * Add the passed pages to the LRU, then drop the caller's refcount
>   * on them.  Reinitialises the caller's pagevec.
>   */
>  void __pagevec_lru_add(struct pagevec *pvec)
>  {
> -	int i;
> -	struct lruvec *lruvec = NULL;
> +	int i, nr_lruvec;
>  	unsigned long flags = 0;
> +	struct page *page;
> +	struct lruvecs lruvecs;
>  
> -	for (i = 0; i < pagevec_count(pvec); i++) {
> -		struct page *page = pvec->pages[i];
> +	nr_lruvec = sort_page_lruvec(&lruvecs, pvec);

Simply looping pvec multiple times (15 at most) for different lruvecs
would be better because 1) it requires no extra data structures and
therefore has better cache locality (theoretically faster) 2) it only
loops once when !CONFIG_MEMCG and !CONFIG_NUMA and therefore has no
impact on Android and Chrome OS.

> -		lruvec = relock_page_lruvec_irqsave(page, lruvec, &flags);
> -		__pagevec_lru_add_fn(page, lruvec);
> +	for (i = 0; i < nr_lruvec; i++) {
> +		spin_lock_irqsave(&lruvecs.vecs[i]->lru_lock, flags);
> +		while (!list_empty(&lruvecs.lists[i])) {
> +			page = lru_to_page(&lruvecs.lists[i]);
> +			list_del(&page->lru);
> +			__pagevec_lru_add_fn(page, lruvecs.vecs[i]);
> +		}
> +		spin_unlock_irqrestore(&lruvecs.vecs[i]->lru_lock, flags);
>  	}
> -	if (lruvec)
> -		unlock_page_lruvec_irqrestore(lruvec, flags);
> +
>  	release_pages(pvec->pages, pvec->nr);
>  	pagevec_reinit(pvec);
>  }
> -- 
> 2.29.GIT
>