[PATCH-mm v2] mm/list_lru: Optimize memcg_reparent_list_lru_node()

[PATCH-mm v2] mm/list_lru: Optimize memcg_reparent_list_lru_node()

[Waiman Long]

Since commit 2c80cd57c743 ("mm/list_lru.c: fix list_lru_count_node()
to be race free"), we are tracking the total number of lru
entries in a list_lru_node in its nr_items field.  In the case of
memcg_reparent_list_lru_node(), there is nothing to be done if nr_items
is 0.  We don't even need to take the nlru->lock as no new lru entry
could be added by a racing list_lru_add() to the draining src_idx memcg
at this point.

Signed-off-by: Waiman Long <longman@redhat.com>
---
 mm/list_lru.c | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/mm/list_lru.c b/mm/list_lru.c
index ba76428ceece..c669d87001a6 100644
--- a/mm/list_lru.c
+++ b/mm/list_lru.c
@@ -394,6 +394,12 @@ static void memcg_reparent_list_lru_node(struct list_lru *lru, int nid,
 	int dst_idx = dst_memcg->kmemcg_id;
 	struct list_lru_one *src, *dst;
 
+	/*
+	 * If there is no lru entry in this nlru, we can skip it immediately.
+	 */
+	if (!READ_ONCE(nlru->nr_items))
+		return;
+
 	/*
 	 * Since list_lru_{add,del} may be called under an IRQ-safe lock,
 	 * we have to use IRQ-safe primitives here to avoid deadlock.
-- 
2.27.0

Re: [PATCH-mm v2] mm/list_lru: Optimize memcg_reparent_list_lru_node()

[Roman Gushchin]

On Tue, Mar 08, 2022 at 08:18:24PM -0500, Waiman Long wrote:
> Since commit 2c80cd57c743 ("mm/list_lru.c: fix list_lru_count_node()
> to be race free"), we are tracking the total number of lru
> entries in a list_lru_node in its nr_items field.  In the case of
> memcg_reparent_list_lru_node(), there is nothing to be done if nr_items
> is 0.  We don't even need to take the nlru->lock as no new lru entry
> could be added by a racing list_lru_add() to the draining src_idx memcg
> at this point.
> 
> Signed-off-by: Waiman Long <longman@redhat.com>
> ---
>  mm/list_lru.c | 6 ++++++
>  1 file changed, 6 insertions(+)
> 
> diff --git a/mm/list_lru.c b/mm/list_lru.c
> index ba76428ceece..c669d87001a6 100644
> --- a/mm/list_lru.c
> +++ b/mm/list_lru.c
> @@ -394,6 +394,12 @@ static void memcg_reparent_list_lru_node(struct list_lru *lru, int nid,
>  	int dst_idx = dst_memcg->kmemcg_id;
>  	struct list_lru_one *src, *dst;
>  
> +	/*
> +	 * If there is no lru entry in this nlru, we can skip it immediately.
> +	 */
> +	if (!READ_ONCE(nlru->nr_items))
> +		return;

This is a per-node counter, not a per-memcg, right?
If so, do we optimize for the case when all lru items belong to one node and
others are empty?

Re: [PATCH-mm v2] mm/list_lru: Optimize memcg_reparent_list_lru_node()

[Waiman Long]

On 3/8/22 21:13, Roman Gushchin wrote:
> On Tue, Mar 08, 2022 at 08:18:24PM -0500, Waiman Long wrote:
>> Since commit 2c80cd57c743 ("mm/list_lru.c: fix list_lru_count_node()
>> to be race free"), we are tracking the total number of lru
>> entries in a list_lru_node in its nr_items field.  In the case of
>> memcg_reparent_list_lru_node(), there is nothing to be done if nr_items
>> is 0.  We don't even need to take the nlru->lock as no new lru entry
>> could be added by a racing list_lru_add() to the draining src_idx memcg
>> at this point.
>>
>> Signed-off-by: Waiman Long <longman@redhat.com>
>> ---
>>   mm/list_lru.c | 6 ++++++
>>   1 file changed, 6 insertions(+)
>>
>> diff --git a/mm/list_lru.c b/mm/list_lru.c
>> index ba76428ceece..c669d87001a6 100644
>> --- a/mm/list_lru.c
>> +++ b/mm/list_lru.c
>> @@ -394,6 +394,12 @@ static void memcg_reparent_list_lru_node(struct list_lru *lru, int nid,
>>   	int dst_idx = dst_memcg->kmemcg_id;
>>   	struct list_lru_one *src, *dst;
>>   
>> +	/*
>> +	 * If there is no lru entry in this nlru, we can skip it immediately.
>> +	 */
>> +	if (!READ_ONCE(nlru->nr_items))
>> +		return;
> This is a per-node counter, not a per-memcg, right?
Right. list_lru_node is a per-node structure inside list_lru.
> If so, do we optimize for the case when all lru items belong to one node and
> others are empty?

That is actually the case that I am trying to optimize for.

If a system has many containers. It is also likely each container may 
mount one or more container specific filesystems. Since a container 
likely use just a few cpus, it is highly that only the list_lru_node 
that contains those cpus will be utilized while the rests may be empty.

I got the idea of doing this patch when I was looking at a crash dump 
related to the list_lru code. That particular crash dump has more than 
13k list_lru's and thousands of mount points. I had notice even if 
nr_items of a list_lru_node is 0, it still tries to transfer lru entries 
from source idx to dest idx. Without doing an lock/unlock and loading a 
cacheline from the memcg_lrus, it can save some time. That can be 
substantial saving if we are talking about thousands of list_lru's.

Cheers,
Longman

Re: [PATCH-mm v2] mm/list_lru: Optimize memcg_reparent_list_lru_node()

[Roman Gushchin]

On Tue, Mar 08, 2022 at 10:12:48PM -0500, Waiman Long wrote:
> On 3/8/22 21:13, Roman Gushchin wrote:
> > On Tue, Mar 08, 2022 at 08:18:24PM -0500, Waiman Long wrote:
> > > Since commit 2c80cd57c743 ("mm/list_lru.c: fix list_lru_count_node()
> > > to be race free"), we are tracking the total number of lru
> > > entries in a list_lru_node in its nr_items field.  In the case of
> > > memcg_reparent_list_lru_node(), there is nothing to be done if nr_items
> > > is 0.  We don't even need to take the nlru->lock as no new lru entry
> > > could be added by a racing list_lru_add() to the draining src_idx memcg
> > > at this point.
> > > 
> > > Signed-off-by: Waiman Long <longman@redhat.com>
> > > ---
> > >   mm/list_lru.c | 6 ++++++
> > >   1 file changed, 6 insertions(+)
> > > 
> > > diff --git a/mm/list_lru.c b/mm/list_lru.c
> > > index ba76428ceece..c669d87001a6 100644
> > > --- a/mm/list_lru.c
> > > +++ b/mm/list_lru.c
> > > @@ -394,6 +394,12 @@ static void memcg_reparent_list_lru_node(struct list_lru *lru, int nid,
> > >   	int dst_idx = dst_memcg->kmemcg_id;
> > >   	struct list_lru_one *src, *dst;
> > > +	/*
> > > +	 * If there is no lru entry in this nlru, we can skip it immediately.
> > > +	 */
> > > +	if (!READ_ONCE(nlru->nr_items))
> > > +		return;
> > This is a per-node counter, not a per-memcg, right?
> Right. list_lru_node is a per-node structure inside list_lru.
> > If so, do we optimize for the case when all lru items belong to one node and
> > others are empty?
> 
> That is actually the case that I am trying to optimize for.
> 
> If a system has many containers. It is also likely each container may mount
> one or more container specific filesystems. Since a container likely use
> just a few cpus, it is highly that only the list_lru_node that contains
> those cpus will be utilized while the rests may be empty.
> 
> I got the idea of doing this patch when I was looking at a crash dump
> related to the list_lru code. That particular crash dump has more than 13k
> list_lru's and thousands of mount points. I had notice even if nr_items of a
> list_lru_node is 0, it still tries to transfer lru entries from source idx
> to dest idx. Without doing an lock/unlock and loading a cacheline from the
> memcg_lrus, it can save some time. That can be substantial saving if we are
> talking about thousands of list_lru's.

Cool! Makes total sense to me. Thanks for the explanation!
Would you mind to add this text to the commit log?

Reviewed-by: Roman Gushchin <roman.gushchin@linux.dev>

Thanks!

Re: [PATCH-mm v2] mm/list_lru: Optimize memcg_reparent_list_lru_node()

[Waiman Long]

On 3/8/22 23:46, Roman Gushchin wrote:
> On Tue, Mar 08, 2022 at 10:12:48PM -0500, Waiman Long wrote:
>> On 3/8/22 21:13, Roman Gushchin wrote:
>>> On Tue, Mar 08, 2022 at 08:18:24PM -0500, Waiman Long wrote:
>>>> Since commit 2c80cd57c743 ("mm/list_lru.c: fix list_lru_count_node()
>>>> to be race free"), we are tracking the total number of lru
>>>> entries in a list_lru_node in its nr_items field.  In the case of
>>>> memcg_reparent_list_lru_node(), there is nothing to be done if nr_items
>>>> is 0.  We don't even need to take the nlru->lock as no new lru entry
>>>> could be added by a racing list_lru_add() to the draining src_idx memcg
>>>> at this point.
>>>>
>>>> Signed-off-by: Waiman Long <longman@redhat.com>
>>>> ---
>>>>    mm/list_lru.c | 6 ++++++
>>>>    1 file changed, 6 insertions(+)
>>>>
>>>> diff --git a/mm/list_lru.c b/mm/list_lru.c
>>>> index ba76428ceece..c669d87001a6 100644
>>>> --- a/mm/list_lru.c
>>>> +++ b/mm/list_lru.c
>>>> @@ -394,6 +394,12 @@ static void memcg_reparent_list_lru_node(struct list_lru *lru, int nid,
>>>>    	int dst_idx = dst_memcg->kmemcg_id;
>>>>    	struct list_lru_one *src, *dst;
>>>> +	/*
>>>> +	 * If there is no lru entry in this nlru, we can skip it immediately.
>>>> +	 */
>>>> +	if (!READ_ONCE(nlru->nr_items))
>>>> +		return;
>>> This is a per-node counter, not a per-memcg, right?
>> Right. list_lru_node is a per-node structure inside list_lru.
>>> If so, do we optimize for the case when all lru items belong to one node and
>>> others are empty?
>> That is actually the case that I am trying to optimize for.
>>
>> If a system has many containers. It is also likely each container may mount
>> one or more container specific filesystems. Since a container likely use
>> just a few cpus, it is highly that only the list_lru_node that contains
>> those cpus will be utilized while the rests may be empty.
>>
>> I got the idea of doing this patch when I was looking at a crash dump
>> related to the list_lru code. That particular crash dump has more than 13k
>> list_lru's and thousands of mount points. I had notice even if nr_items of a
>> list_lru_node is 0, it still tries to transfer lru entries from source idx
>> to dest idx. Without doing an lock/unlock and loading a cacheline from the
>> memcg_lrus, it can save some time. That can be substantial saving if we are
>> talking about thousands of list_lru's.
> Cool! Makes total sense to me. Thanks for the explanation!
> Would you mind to add this text to the commit log?
>
> Reviewed-by: Roman Gushchin <roman.gushchin@linux.dev>

Sure. I will add some of it into the commit log.

Cheers,
Longman