linux-kernel - Re: [PATCH 1/3] mm/swap.c: pre-sort pages in pagevec for pagevec_lru_move

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <20201201081031.GQ17338@dhcp22.suse.cz>
Date:   Tue, 1 Dec 2020 09:10:31 +0100
From:   Michal Hocko <mhocko@...e.com>
To:     Alex Shi <alex.shi@...ux.alibaba.com>
Cc:     vbabka@...e.cz, Konstantin Khlebnikov <koct9i@...il.com>,
        Hugh Dickins <hughd@...gle.com>, Yu Zhao <yuzhao@...gle.com>,
        "Matthew Wilcox (Oracle)" <willy@...radead.org>,
        Andrew Morton <akpm@...ux-foundation.org>, linux-mm@...ck.org,
        linux-kernel@...r.kernel.org
Subject: Re: [PATCH 1/3] mm/swap.c: pre-sort pages in pagevec for
 pagevec_lru_move_fn

On Tue 01-12-20 16:02:13, Alex Shi wrote:
> Pages in pagevec may have different lruvec, so we have to do relock in
> function pagevec_lru_move_fn(), but a relock may cause current cpu wait
> for long time on the same lock for spinlock fairness reason.
> 
> Before per memcg lru_lock, we have to bear the relock since the spinlock
> is the only way to serialize page's memcg/lruvec. Now TestClearPageLRU
> could be used to isolate pages exculsively, and stable the page's
> lruvec/memcg. So it gives us a chance to sort the page's lruvec before
> moving action in pagevec_lru_move_fn. Then we don't suffer from the
> spinlock's fairness wait.

Do you have any data to show any improvements from this?

> Signed-off-by: Alex Shi <alex.shi@...ux.alibaba.com>
> Cc: Konstantin Khlebnikov <koct9i@...il.com>
> Cc: Hugh Dickins <hughd@...gle.com>
> Cc: Yu Zhao <yuzhao@...gle.com>
> Cc: Michal Hocko <mhocko@...e.com>
> Cc: Matthew Wilcox (Oracle) <willy@...radead.org>
> Cc: Andrew Morton <akpm@...ux-foundation.org>
> Cc: linux-mm@...ck.org
> Cc: linux-kernel@...r.kernel.org
> ---
>  mm/swap.c | 92 +++++++++++++++++++++++++++++++++++++++++++++++--------
>  1 file changed, 79 insertions(+), 13 deletions(-)
> 
> diff --git a/mm/swap.c b/mm/swap.c
> index 490553f3f9ef..17d8990e5ca7 100644
> --- a/mm/swap.c
> +++ b/mm/swap.c
> @@ -201,29 +201,95 @@ int get_kernel_page(unsigned long start, int write, struct page **pages)
>  }
>  EXPORT_SYMBOL_GPL(get_kernel_page);
>  
> +/* Pratt's gaps for shell sort, https://en.wikipedia.org/wiki/Shellsort */
> +static int gaps[] = { 6, 4, 3, 2, 1, 0};
> +
> +/* Shell sort pagevec[] on page's lruvec.*/
> +static void shell_sort(struct pagevec *pvec, unsigned long *lvaddr)
> +{
> +	int g, i, j, n = pagevec_count(pvec);
> +
> +	for (g=0; gaps[g] > 0 && gaps[g] <= n/2; g++) {
> +		int gap = gaps[g];
> +
> +		for (i = gap; i < n; i++) {
> +			unsigned long tmp = lvaddr[i];
> +			struct page *page = pvec->pages[i];
> +
> +			for (j = i - gap; j >= 0 && lvaddr[j] > tmp; j -= gap) {
> +				lvaddr[j + gap] = lvaddr[j];
> +				pvec->pages[j + gap] = pvec->pages[j];
> +			}
> +			lvaddr[j + gap] = tmp;
> +			pvec->pages[j + gap] = page;
> +		}
> +	}
> +}
> +
> +/* Get lru bit cleared page and their lruvec address, release the others */
> +void sort_isopv(struct pagevec *pvec, struct pagevec *isopv,
> +		unsigned long *lvaddr)
> +{
> +	int i, j;
> +	struct pagevec busypv;
> +
> +	pagevec_init(&busypv);
> +
> +	for (i = 0, j = 0; i < pagevec_count(pvec); i++) {
> +		struct page *page = pvec->pages[i];
> +
> +		pvec->pages[i] = NULL;
> +
> +		/* block memcg migration during page moving between lru */
> +		if (!TestClearPageLRU(page)) {
> +			pagevec_add(&busypv, page);
> +			continue;
> +		}
> +		lvaddr[j++] = (unsigned long)
> +				mem_cgroup_page_lruvec(page, page_pgdat(page));
> +		pagevec_add(isopv, page);
> +	}
> +	pagevec_reinit(pvec);
> +	if (pagevec_count(&busypv))
> +		release_pages(busypv.pages, busypv.nr);
> +
> +	shell_sort(isopv, lvaddr);
> +}
> +
>  static void pagevec_lru_move_fn(struct pagevec *pvec,
>  	void (*move_fn)(struct page *page, struct lruvec *lruvec))
>  {
> -	int i;
> +	int i, n;
>  	struct lruvec *lruvec = NULL;
>  	unsigned long flags = 0;
> +	unsigned long lvaddr[PAGEVEC_SIZE];
> +	struct pagevec isopv;
>  
> -	for (i = 0; i < pagevec_count(pvec); i++) {
> -		struct page *page = pvec->pages[i];
> +	pagevec_init(&isopv);
>  
> -		/* block memcg migration during page moving between lru */
> -		if (!TestClearPageLRU(page))
> -			continue;
> +	sort_isopv(pvec, &isopv, lvaddr);
>  
> -		lruvec = relock_page_lruvec_irqsave(page, lruvec, &flags);
> -		(*move_fn)(page, lruvec);
> +	n = pagevec_count(&isopv);
> +	if (!n)
> +		return;
>  
> -		SetPageLRU(page);
> +	lruvec = (struct lruvec *)lvaddr[0];
> +	spin_lock_irqsave(&lruvec->lru_lock, flags);
> +
> +	for (i = 0; i < n; i++) {
> +		/* lock new lruvec if lruvec changes, we have sorted them */
> +		if (lruvec != (struct lruvec *)lvaddr[i]) {
> +			spin_unlock_irqrestore(&lruvec->lru_lock, flags);
> +			lruvec = (struct lruvec *)lvaddr[i];
> +			spin_lock_irqsave(&lruvec->lru_lock, flags);
> +		}
> +
> +		(*move_fn)(isopv.pages[i], lruvec);
> +
> +		SetPageLRU(isopv.pages[i]);
>  	}
> -	if (lruvec)
> -		unlock_page_lruvec_irqrestore(lruvec, flags);
> -	release_pages(pvec->pages, pvec->nr);
> -	pagevec_reinit(pvec);
> +	spin_unlock_irqrestore(&lruvec->lru_lock, flags);
> +	release_pages(isopv.pages, isopv.nr);
>  }
>  
>  static void pagevec_move_tail_fn(struct page *page, struct lruvec *lruvec)
> -- 
> 2.29.GIT
> 

-- 
Michal Hocko
SUSE Labs