diff mbox series

[1/3] mm/swap.c: pre-sort pages in pagevec for pagevec_lru_move_fn

Message ID 1606809735-43300-1-git-send-email-alex.shi@linux.alibaba.com (mailing list archive)
State New, archived
Headers show
Series [1/3] mm/swap.c: pre-sort pages in pagevec for pagevec_lru_move_fn | expand

Commit Message

Alex Shi Dec. 1, 2020, 8:02 a.m. UTC 
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.

Signed-off-by: Alex Shi <alex.shi@linux.alibaba.com>
Cc: Konstantin Khlebnikov <koct9i@gmail.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Yu Zhao <yuzhao@google.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: linux-mm@kvack.org
Cc: linux-kernel@vger.kernel.org
---
 mm/swap.c | 92 +++++++++++++++++++++++++++++++++++++++++++++++--------
 1 file changed, 79 insertions(+), 13 deletions(-)

Comments

Michal Hocko Dec. 1, 2020, 8:10 a.m. UTC | #1 
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@linux.alibaba.com>
> Cc: Konstantin Khlebnikov <koct9i@gmail.com>
> Cc: Hugh Dickins <hughd@google.com>
> Cc: Yu Zhao <yuzhao@google.com>
> Cc: Michal Hocko <mhocko@suse.com>
> Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
> Cc: Andrew Morton <akpm@linux-foundation.org>
> Cc: linux-mm@kvack.org
> Cc: linux-kernel@vger.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
>
Alex Shi Dec. 1, 2020, 8:20 a.m. UTC | #2 
在 2020/12/1 下午4:10, Michal Hocko 写道:
> 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?
> 

Hi Michal,

Thanks for quick response.

Not yet. I am running for data. but according to the lru_add result, there
should be a big gain for multiple memcgs scenario.

Also I don't except a quick accept, just send out the idea for comments 
when the thread is still warm. :)

Thanks
Alex
diff mbox series

Patch

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)