[PATCH] mm:vmscan: fix inaccurate reclaim during proactive reclaim

[PATCH] mm:vmscan: fix inaccurate reclaim during proactive reclaim

[Efly Young]

With commit f53af4285d77 ("mm: vmscan: fix extreme overreclaim
and swap floods"), proactive reclaim still seems inaccurate.

Our problematic scene also are almost anon pages. Request 1G
by writing memory.reclaim will reclaim 1.7G or other values
more than 1G by swapping.

This try to fix the inaccurate reclaim problem.

Signed-off-by: Efly Young <yangyifei03@kuaishou.com>
---
 mm/vmscan.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/mm/vmscan.c b/mm/vmscan.c
index 9c1c5e8b..2aea8d9 100644
--- a/mm/vmscan.c
+++ b/mm/vmscan.c
@@ -6208,7 +6208,7 @@ static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
 	unsigned long nr_to_scan;
 	enum lru_list lru;
 	unsigned long nr_reclaimed = 0;
-	unsigned long nr_to_reclaim = sc->nr_to_reclaim;
+	unsigned long nr_to_reclaim = (sc->nr_to_reclaim - sc->nr_reclaimed);
 	bool proportional_reclaim;
 	struct blk_plug plug;
 
-- 
1.8.3.1

Re: [PATCH] mm:vmscan: fix inaccurate reclaim during proactive reclaim

[Andrew Morton]

(cc hannes)

On Fri, 7 Jul 2023 18:32:26 +0800 Efly Young <yangyifei03@kuaishou.com> wrote:

> With commit f53af4285d77 ("mm: vmscan: fix extreme overreclaim
> and swap floods"), proactive reclaim still seems inaccurate.
> 
> Our problematic scene also are almost anon pages. Request 1G
> by writing memory.reclaim will reclaim 1.7G or other values
> more than 1G by swapping.
> 
> This try to fix the inaccurate reclaim problem.

It would be helpful to have some additional explanation of why you
believe the current code is incorrect?

> --- a/mm/vmscan.c
> +++ b/mm/vmscan.c
> @@ -6208,7 +6208,7 @@ static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
>  	unsigned long nr_to_scan;
>  	enum lru_list lru;
>  	unsigned long nr_reclaimed = 0;
> -	unsigned long nr_to_reclaim = sc->nr_to_reclaim;
> +	unsigned long nr_to_reclaim = (sc->nr_to_reclaim - sc->nr_reclaimed);
>  	bool proportional_reclaim;
>  	struct blk_plug plug;
>  

答复: ?????!?Re: [PATCH] mm:vmscan: fix inaccurate reclaim during proactive reclaim

[杨逸飞]

[-- Attachment #1: Type: text/plain, Size: 1934 bytes --]

Thank you for your reply. shrink_lruvec() is nested in deep loop. Reclaimer may have already reclaimed part of requested memory in one loop, but before adjust sc->nr_to_reclaim in outer loop, call shrink_lruvec() again will still follow the current sc->nr_to_reclaim to work. It will eventually lead to overreclaim. My problem case is easy to construct. Allocate lots of anonymous memory(e.g. 20G) in a memcg, then swapping by writing memory.reclaim and there is a certain probability of overreclaim.


________________________________
发件人: Andrew Morton <akpm@linux-foundation.org>
发送时间: 2023年7月8日 3:09
收件人: 杨逸飞
抄送: linux-mm@kvack.org; linux-kernel@vger.kernel.org; Johannes Weiner
主题: ?????!?Re: [PATCH] mm:vmscan: fix inaccurate reclaim during proactive reclaim

安全提示：此邮件来自公司外部。除非您确认发件人身份可信且邮件内容不含可疑信息，否则请勿回复或转发邮件、点击邮件链接或打开附件。


(cc hannes)

On Fri, 7 Jul 2023 18:32:26 +0800 Efly Young <yangyifei03@kuaishou.com> wrote:

> With commit f53af4285d77 ("mm: vmscan: fix extreme overreclaim
> and swap floods"), proactive reclaim still seems inaccurate.
>
> Our problematic scene also are almost anon pages. Request 1G
> by writing memory.reclaim will reclaim 1.7G or other values
> more than 1G by swapping.
>
> This try to fix the inaccurate reclaim problem.

It would be helpful to have some additional explanation of why you
believe the current code is incorrect?

> --- a/mm/vmscan.c
> +++ b/mm/vmscan.c
> @@ -6208,7 +6208,7 @@ static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
>       unsigned long nr_to_scan;
>       enum lru_list lru;
>       unsigned long nr_reclaimed = 0;
> -     unsigned long nr_to_reclaim = sc->nr_to_reclaim;
> +     unsigned long nr_to_reclaim = (sc->nr_to_reclaim - sc->nr_reclaimed);
>       bool proportional_reclaim;
>       struct blk_plug plug;
>


[-- Attachment #2: Type: text/html, Size: 3366 bytes --]

Re: [PATCH] mm:vmscan: fix inaccurate reclaim during proactive reclaim

[Johannes Weiner]

On Fri, Jul 07, 2023 at 06:32:26PM +0800, Efly Young wrote:
> With commit f53af4285d77 ("mm: vmscan: fix extreme overreclaim
> and swap floods"), proactive reclaim still seems inaccurate.
> 
> Our problematic scene also are almost anon pages. Request 1G
> by writing memory.reclaim will reclaim 1.7G or other values
> more than 1G by swapping.
> 
> This try to fix the inaccurate reclaim problem.

I can see how this happens. Direct and kswapd reclaim have much
smaller nr_to_reclaim targets, so it's less noticable when we loop a
few times. Proactive reclaim can come in with a rather large value.

What does the reproducer setup look like? Are you calling reclaim on a
higher level cgroup with several children? Or is the looping coming
from having multiple zones alone?

> Signed-off-by: Efly Young <yangyifei03@kuaishou.com>
> ---
>  mm/vmscan.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/mm/vmscan.c b/mm/vmscan.c
> index 9c1c5e8b..2aea8d9 100644
> --- a/mm/vmscan.c
> +++ b/mm/vmscan.c
> @@ -6208,7 +6208,7 @@ static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
>  	unsigned long nr_to_scan;
>  	enum lru_list lru;
>  	unsigned long nr_reclaimed = 0;
> -	unsigned long nr_to_reclaim = sc->nr_to_reclaim;
> +	unsigned long nr_to_reclaim = (sc->nr_to_reclaim - sc->nr_reclaimed);

This can underflow. shrink_list() eats SWAP_CLUSTER_MAX batches out of
lru_pages >> priority, and only checks reclaimed > to_reclaim
after. This will then disable the bailout mechanism entirely.

In general, I'm not sure this is the best spot to fix the problem:

- During reclaim/compaction, should_continue_reclaim() may decide that
  more reclaim is required before compaction can proceed. But the
  second cycle might not do anything now, since you remember the work
  done by the previous one.

- shrink_node_memcgs() might do the full batch against the first
  cgroup and not touch the second one anymore. This will result in
  super lopsided behavior when you target a tree of multiple groups.

There might be other spots that break, I haven't checked.

You could go through them one by one, of course. But the truth is,
larger reclaim targets are the rare exception. Trying to support them
at the risk of breaking all other reclaim users seems ill-advised.

A better approach might be to just say: "don't call reclaim with large
numbers". Have proactive reclaim code handle the batching into smaller
chunks:

diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index e8ca4bdcb03c..4b016806dcc7 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c
@@ -6696,7 +6696,7 @@ static ssize_t memory_reclaim(struct kernfs_open_file *of, char *buf,
 			lru_add_drain_all();
 
 		reclaimed = try_to_free_mem_cgroup_pages(memcg,
-						nr_to_reclaim - nr_reclaimed,
+						min(nr_to_reclaim - nr_reclaimed, SWAP_CLUSTER_MAX),
 						GFP_KERNEL, reclaim_options);
 
 		if (!reclaimed && !nr_retries--)

Re:Re:[PATCH] mm:vmscan: fix inaccurate reclaim during proactive reclaim

[Efly Young]

>> With commit f53af4285d77 ("mm: vmscan: fix extreme overreclaim
>> and swap floods"), proactive reclaim still seems inaccurate.
>> 
>> Our problematic scene also are almost anon pages. Request 1G
>> by writing memory.reclaim will reclaim 1.7G or other values
>> more than 1G by swapping.
>> 
>> This try to fix the inaccurate reclaim problem.
>
> I can see how this happens. Direct and kswapd reclaim have much
> smaller nr_to_reclaim targets, so it's less noticable when we loop a
> few times. Proactive reclaim can come in with a rather large value.
> 
> What does the reproducer setup look like? Are you calling reclaim on a
> higher level cgroup with several children? Or is the looping coming
> from having multiple zones alone?

Thank you for your comment. The process in a leaf cgroup without children
just malloc 20G anonymous memory and sleep, then calling reclaim in the
leaf cgroup. Before commit f53af4285d77 ("mm: vmscan: fix extreme
overreclaim and swap floods"), reclaimer may reclaim many times the amount
of request. Now it should eventually reclaim in [request, 2 * request).

>> Signed-off-by: Efly Young <yangyifei03@kuaishou.com>
>> ---
>>  mm/vmscan.c | 2 +-
>>  1 file changed, 1 insertion(+), 1 deletion(-)
>> 
>> diff --git a/mm/vmscan.c b/mm/vmscan.c
>> index 9c1c5e8b..2aea8d9 100644
>> --- a/mm/vmscan.c
>> +++ b/mm/vmscan.c
>> @@ -6208,7 +6208,7 @@ static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
>>  	unsigned long nr_to_scan;
>>  	enum lru_list lru;
>>  	unsigned long nr_reclaimed = 0;
>> -	unsigned long nr_to_reclaim = sc->nr_to_reclaim;
>> +	unsigned long nr_to_reclaim = (sc->nr_to_reclaim - sc->nr_reclaimed);
>
> This can underflow. shrink_list() eats SWAP_CLUSTER_MAX batches out of
> lru_pages >> priority, and only checks reclaimed > to_reclaim
> after. This will then disable the bailout mechanism entirely.
> 
> In general, I'm not sure this is the best spot to fix the problem:
> 
> - During reclaim/compaction, should_continue_reclaim() may decide that
>   more reclaim is required before compaction can proceed. But the
>   second cycle might not do anything now, since you remember the work
>   done by the previous one.
> 
> - shrink_node_memcgs() might do the full batch against the first
>   cgroup and not touch the second one anymore. This will result in
>   super lopsided behavior when you target a tree of multiple groups.
> 
> There might be other spots that break, I haven't checked.
> 
> You could go through them one by one, of course. But the truth is,
> larger reclaim targets are the rare exception. Trying to support them
> at the risk of breaking all other reclaim users seems ill-advised.

I agree with your view. These explanations are more considerate. Thank
you again for helping me out.

> A better approach might be to just say: "don't call reclaim with large
> numbers". Have proactive reclaim code handle the batching into smaller
> chunks:
> 
> diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> index e8ca4bdcb03c..4b016806dcc7 100644
> --- a/mm/memcontrol.c
> +++ b/mm/memcontrol.c
> @@ -6696,7 +6696,7 @@ static ssize_t memory_reclaim(struct kernfs_open_file *of, char *buf,
>  			lru_add_drain_all();
>  
>  		reclaimed = try_to_free_mem_cgroup_pages(memcg,
> -						nr_to_reclaim - nr_reclaimed,
> +						min(nr_to_reclaim - nr_reclaimed, SWAP_CLUSTER_MAX),
>  						GFP_KERNEL, reclaim_options);
>  
>  		if (!reclaimed && !nr_retries--)

May be this way could solve the inaccurate proactive reclaim
problem without breaking the original balance. But may be less
efficient than before?