Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > linux.kernel > #1586888 > unrolled thread

[PATCH v2 2/2] mm/cgroup: delay soft limit data allocation

Back to article view | Back to linux.kernel

This discussion starts older than the indexed window; earlier articles aren't shown. The article labeled Started by below is the oldest one visible, not the original post.

Contents

  [PATCH v2 2/2] mm/cgroup: delay soft limit data allocation Laurent Dufour <ldufour@linux.vnet.ibm.com> - 2017-02-23 14:40 +0100
    Re: [PATCH v2 2/2] mm/cgroup: delay soft limit data allocation Michal Hocko <mhocko@kernel.org> - 2017-02-23 16:40 +0100
      Re: [PATCH v2 2/2] mm/cgroup: delay soft limit data allocation Johannes Weiner <hannes@cmpxchg.org> - 2017-02-23 20:10 +0100

#1586888 — [PATCH v2 2/2] mm/cgroup: delay soft limit data allocation

Until a soft limit is set to a cgroup, the soft limit data are useless
so delay this allocation when a limit is set.

Suggested-by: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Laurent Dufour <ldufour@linux.vnet.ibm.com>
---
 mm/memcontrol.c | 67 ++++++++++++++++++++++++++++++++++++++++++++-------------
 1 file changed, 52 insertions(+), 15 deletions(-)

diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index a9f10fde44a6..c639c898809d 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c
@@ -142,7 +142,7 @@ struct mem_cgroup_tree {
 	struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
 };
 
-static struct mem_cgroup_tree soft_limit_tree __read_mostly;
+static struct mem_cgroup_tree *soft_limit_tree __read_mostly;
 
 /* for OOM */
 struct mem_cgroup_eventfd_list {
@@ -381,10 +381,52 @@ mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
 	return memcg->nodeinfo[nid];
 }
 
+static bool soft_limit_initialize(void)
+{
+	static DEFINE_MUTEX(soft_limit_mutex);
+	struct mem_cgroup_tree *tree;
+	bool ret = true;
+	int node;
+
+	mutex_lock(&soft_limit_mutex);
+	if (soft_limit_tree)
+		goto bail;
+
+	tree = kmalloc(sizeof(*soft_limit_tree), GFP_KERNEL);
+	if (!tree) {
+		ret = false;
+		goto bail;
+	}
+	for_each_node(node) {
+		struct mem_cgroup_tree_per_node *rtpn;
+
+		rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
+				    node_online(node) ? node : NUMA_NO_NODE);
+		if (!rtpn)
+			goto cleanup;
+
+		rtpn->rb_root = RB_ROOT;
+		spin_lock_init(&rtpn->lock);
+		tree->rb_tree_per_node[node] = rtpn;
+	}
+	WRITE_ONCE(soft_limit_tree, tree);
+bail:
+	mutex_unlock(&soft_limit_mutex);
+	return ret;
+cleanup:
+	for_each_node(node)
+		kfree(tree->rb_tree_per_node[node]);
+	kfree(tree);
+	ret = false;
+	goto bail;
+}
+
 static struct mem_cgroup_tree_per_node *
 soft_limit_tree_node(int nid)
 {
-	return soft_limit_tree.rb_tree_per_node[nid];
+	if (!soft_limit_tree)
+		return NULL;
+	return soft_limit_tree->rb_tree_per_node[nid];
 }
 
 static struct mem_cgroup_tree_per_node *
@@ -392,7 +434,9 @@ soft_limit_tree_from_page(struct page *page)
 {
 	int nid = page_to_nid(page);
 
-	return soft_limit_tree.rb_tree_per_node[nid];
+	if (!soft_limit_tree)
+		return NULL;
+	return soft_limit_tree->rb_tree_per_node[nid];
 }
 
 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
@@ -3003,6 +3047,10 @@ static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
 		}
 		break;
 	case RES_SOFT_LIMIT:
+		if (!soft_limit_initialize()) {
+			ret = -ENOMEM;
+			break;
+		}
 		memcg->soft_limit = nr_pages;
 		ret = 0;
 		break;
@@ -5777,7 +5825,7 @@ __setup("cgroup.memory=", cgroup_memory);
  */
 static int __init mem_cgroup_init(void)
 {
-	int cpu, node;
+	int cpu;
 
 #ifndef CONFIG_SLOB
 	/*
@@ -5797,17 +5845,6 @@ static int __init mem_cgroup_init(void)
 		INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
 			  drain_local_stock);
 
-	for_each_node(node) {
-		struct mem_cgroup_tree_per_node *rtpn;
-
-		rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
-				    node_online(node) ? node : NUMA_NO_NODE);
-
-		rtpn->rb_root = RB_ROOT;
-		spin_lock_init(&rtpn->lock);
-		soft_limit_tree.rb_tree_per_node[node] = rtpn;
-	}
-
 	return 0;
 }
 subsys_initcall(mem_cgroup_init);
-- 
2.7.4

[toc] | [next] | [standalone]

#1586961

On Thu 23-02-17 14:36:39, Laurent Dufour wrote:
> Until a soft limit is set to a cgroup, the soft limit data are useless
> so delay this allocation when a limit is set.

Hmm, I am still undecided whether this is actually worth it. On one hand
distribution kernels tend to have quite large NUMA_SHIFT (e.g. SLES has
NUMA_SHIFT=10 and then we will save 8kB+12kB which is not hell of a lot
but always good if we can save that, especially for a rarely used
feature. The code grown on the other hand (it was in __init section
previously) which is a minus, on the other hand.

What do you think Johannes?

This would be a useful info in the changelog, btw.

> Suggested-by: Michal Hocko <mhocko@kernel.org>
> Signed-off-by: Laurent Dufour <ldufour@linux.vnet.ibm.com>

The patch looks good to me so feel free to add
Reviewed-by: Michal Hocko <mhocko@suse.com>

> ---
>  mm/memcontrol.c | 67 ++++++++++++++++++++++++++++++++++++++++++++-------------
>  1 file changed, 52 insertions(+), 15 deletions(-)
> 
> diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> index a9f10fde44a6..c639c898809d 100644
> --- a/mm/memcontrol.c
> +++ b/mm/memcontrol.c
> @@ -142,7 +142,7 @@ struct mem_cgroup_tree {
>  	struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
>  };
>  
> -static struct mem_cgroup_tree soft_limit_tree __read_mostly;
> +static struct mem_cgroup_tree *soft_limit_tree __read_mostly;
>  
>  /* for OOM */
>  struct mem_cgroup_eventfd_list {
> @@ -381,10 +381,52 @@ mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
>  	return memcg->nodeinfo[nid];
>  }
>  
> +static bool soft_limit_initialize(void)
> +{
> +	static DEFINE_MUTEX(soft_limit_mutex);
> +	struct mem_cgroup_tree *tree;
> +	bool ret = true;
> +	int node;
> +
> +	mutex_lock(&soft_limit_mutex);
> +	if (soft_limit_tree)
> +		goto bail;
> +
> +	tree = kmalloc(sizeof(*soft_limit_tree), GFP_KERNEL);
> +	if (!tree) {
> +		ret = false;
> +		goto bail;
> +	}
> +	for_each_node(node) {
> +		struct mem_cgroup_tree_per_node *rtpn;
> +
> +		rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
> +				    node_online(node) ? node : NUMA_NO_NODE);
> +		if (!rtpn)
> +			goto cleanup;
> +
> +		rtpn->rb_root = RB_ROOT;
> +		spin_lock_init(&rtpn->lock);
> +		tree->rb_tree_per_node[node] = rtpn;
> +	}
> +	WRITE_ONCE(soft_limit_tree, tree);
> +bail:
> +	mutex_unlock(&soft_limit_mutex);
> +	return ret;
> +cleanup:
> +	for_each_node(node)
> +		kfree(tree->rb_tree_per_node[node]);
> +	kfree(tree);
> +	ret = false;
> +	goto bail;
> +}
> +
>  static struct mem_cgroup_tree_per_node *
>  soft_limit_tree_node(int nid)
>  {
> -	return soft_limit_tree.rb_tree_per_node[nid];
> +	if (!soft_limit_tree)
> +		return NULL;
> +	return soft_limit_tree->rb_tree_per_node[nid];
>  }
>  
>  static struct mem_cgroup_tree_per_node *
> @@ -392,7 +434,9 @@ soft_limit_tree_from_page(struct page *page)
>  {
>  	int nid = page_to_nid(page);
>  
> -	return soft_limit_tree.rb_tree_per_node[nid];
> +	if (!soft_limit_tree)
> +		return NULL;
> +	return soft_limit_tree->rb_tree_per_node[nid];
>  }
>  
>  static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
> @@ -3003,6 +3047,10 @@ static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
>  		}
>  		break;
>  	case RES_SOFT_LIMIT:
> +		if (!soft_limit_initialize()) {
> +			ret = -ENOMEM;
> +			break;
> +		}
>  		memcg->soft_limit = nr_pages;
>  		ret = 0;
>  		break;
> @@ -5777,7 +5825,7 @@ __setup("cgroup.memory=", cgroup_memory);
>   */
>  static int __init mem_cgroup_init(void)
>  {
> -	int cpu, node;
> +	int cpu;
>  
>  #ifndef CONFIG_SLOB
>  	/*
> @@ -5797,17 +5845,6 @@ static int __init mem_cgroup_init(void)
>  		INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
>  			  drain_local_stock);
>  
> -	for_each_node(node) {
> -		struct mem_cgroup_tree_per_node *rtpn;
> -
> -		rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
> -				    node_online(node) ? node : NUMA_NO_NODE);
> -
> -		rtpn->rb_root = RB_ROOT;
> -		spin_lock_init(&rtpn->lock);
> -		soft_limit_tree.rb_tree_per_node[node] = rtpn;
> -	}
> -
>  	return 0;
>  }
>  subsys_initcall(mem_cgroup_init);
> -- 
> 2.7.4

-- 
Michal Hocko
SUSE Labs

[toc] | [prev] | [next] | [standalone]

#1587092

On Thu, Feb 23, 2017 at 04:31:07PM +0100, Michal Hocko wrote:
> On Thu 23-02-17 14:36:39, Laurent Dufour wrote:
> > Until a soft limit is set to a cgroup, the soft limit data are useless
> > so delay this allocation when a limit is set.
> 
> Hmm, I am still undecided whether this is actually worth it. On one hand
> distribution kernels tend to have quite large NUMA_SHIFT (e.g. SLES has
> NUMA_SHIFT=10 and then we will save 8kB+12kB which is not hell of a lot
> but always good if we can save that, especially for a rarely used
> feature. The code grown on the other hand (it was in __init section
> previously) which is a minus, on the other hand.
> 
> What do you think Johannes?

Hohumm, saving 5 pages on a NUMA machine vs. the additional complexity
and the increased risk of memory problems when somebody sets up a soft
limit after some uptime... I don't think I can give a strong yes or no
on this one, so inertia wins for me; I'd just leave it alone.

[toc] | [prev] | [standalone]

Back to top | Article view | linux.kernel

csiph-web