[Devel] [PATCH vz7] mm/slab: use monotonic serial number for naming per memcg caches

[Kirill Tkhai]

On 12.10.2018 18:54, Konstantin Khorenko wrote:
> Currently, we use mem_cgroup->kmemcg_id to guarantee kmem_cache->name
> uniqueness. Unfortunately kmemcg_id release happens on cgroup offline
> when kmem caches are  still alive and their sysfs entries are not removed.
> 
> Thus it's possible to start/suspend/resume/stop a Container and get into
> situation when mem cgroup is offline (thus mem_cgroup::kmemcg_id is
> released), but not freed and corresponding slab caches are alive along
> with their sysfs entries like:
>   /sys/kernel/slab/filp/cgroup/filp($KMEMCG_ID:$MEMCG_NAME)
> 
> If we start a Container after that, new memory cgroup is created and
> gets same kmemcg_id, new kmem_cache is created and sysfs_slab_add()
> fails due to duplicate name provided which triggers a warning:
> 
>   WARNING: CPU: 0 PID: 31346 at lib/kobject.c:239
>     kobject_add_internal+0x6ef/0x910
>   kobject_add_internal failed for kmalloc-512(3:100) with -EEXIST, don't
>     try to register things with the same name in the same directory.
> 
> In older kernels we use css_id(mem_cgroup_css(memcg)), but now cssid is
> dropped.
> In mainstream monotonic counter css->serial_nr is used instead,
> so let's add monotonic serial_nr as well, but as nobody needs it at the
> moment except memcg, let it be part of struct mem_cgroup.

Looks good. Please, look at small nits below.

> https://jira.sw.ru/browse/PSBM-89047
> 
> Signed-off-by: Konstantin Khorenko <khorenko at virtuozzo.com>
> ---
>  include/linux/memcontrol.h | 11 +++++++++++
>  mm/memcontrol.c            | 18 ++++++++++++++++++
>  mm/slab_common.c           | 12 ++++++++++--
>  3 files changed, 39 insertions(+), 2 deletions(-)
> 
> diff --git a/include/linux/memcontrol.h b/include/linux/memcontrol.h
> index 640a5802e398..56eb202ac654 100644
> --- a/include/linux/memcontrol.h
> +++ b/include/linux/memcontrol.h
> @@ -549,6 +549,8 @@ bool __memcg_kmem_newpage_charge(struct page *page, gfp_t gfp, int order);
>  void __memcg_kmem_uncharge_pages(struct page *page, int order);
>  
>  int memcg_cache_id(struct mem_cgroup *memcg);
> +void memcg_set_serial_nr(struct mem_cgroup *memcg, u64 val);
> +u64 memcg_get_serial_nr(struct mem_cgroup *memcg);
>  
>  struct kmem_cache *
>  __memcg_kmem_get_cache(struct kmem_cache *cachep, gfp_t gfp);
> @@ -685,6 +687,15 @@ static inline int memcg_cache_id(struct mem_cgroup *memcg)
>  	return -1;
>  }
>  
> +static inline void memcg_set_serial_nr(struct mem_cgroup *memcg, u64 val)
> +{
> +}
> +
> +static inline u64 memcg_get_serial_nr(struct mem_cgroup *memcg)
> +{
> +	return 0;
> +}
> +
>  static inline void memcg_get_cache_ids(void)
>  {
>  }
> diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> index e57263fea566..2061d5b116bf 100644
> --- a/mm/memcontrol.c
> +++ b/mm/memcontrol.c
> @@ -404,6 +404,14 @@ struct mem_cgroup {
>  #if defined(CONFIG_MEMCG_KMEM)
>          /* Index in the kmem_cache->memcg_params.memcg_caches array */
>  	int kmemcg_id;
> +	/*
> +	 * Monotonically increasing unique serial number which defines a
> +	 * uniform order among all mem cgroups.
> +	 * It's used to construct unique kmem_cache name which should remain
> +	 * unique after mem cgroup offline when kmemcg_id is released.
> +	 */
> +	u64 serial_nr;
> +
>  	/* List of memcgs sharing the same kmemcg_id */
>  	struct list_head kmemcg_sharers;
>  #endif
> @@ -3432,6 +3440,16 @@ int memcg_cache_id(struct mem_cgroup *memcg)
>  	return memcg ? memcg->kmemcg_id : -1;
>  }
>  
> +void memcg_set_serial_nr(struct mem_cgroup *memcg, u64 val)
> +{
> +	memcg->serial_nr = val;
> +}
> +
> +u64 memcg_get_serial_nr(struct mem_cgroup *memcg)
> +{
> +	return memcg->serial_nr;
> +}

Do we really need this struct mem_cgroup::serial_nr and the get/set primitives?
Are we going to use them for something else in the future?

> +
>  static int memcg_alloc_cache_id(void)
>  {
>  	int id, size;
> diff --git a/mm/slab_common.c b/mm/slab_common.c
> index 1b2a7d208596..24abd2e88efe 100644
> --- a/mm/slab_common.c
> +++ b/mm/slab_common.c
> @@ -421,6 +421,7 @@ static void do_kmem_cache_release(struct list_head *release,
>  void memcg_create_kmem_cache(struct mem_cgroup *memcg,
>  			     struct kmem_cache *root_cache)
>  {
> +	static atomic64_t serial_nr_cursor = ATOMIC64_INIT(0);
>  	static char memcg_name_buf[NAME_MAX + 1]; /* protected by slab_mutex */
>  	struct memcg_cache_array *arr;
>  	struct kmem_cache *s = NULL;
> @@ -453,8 +454,15 @@ void memcg_create_kmem_cache(struct mem_cgroup *memcg,
>  	strlcpy(memcg_name_buf, cgroup_name(mem_cgroup_css(memcg)->cgroup),
>  		NAME_MAX + 1);
>  	rcu_read_unlock();
> -	cache_name = kasprintf(GFP_KERNEL, "%s(%d:%s)", root_cache->name,
> -			       idx, memcg_name_buf);
> +
> +	/*
> +	 * Assign a monotonically increasing serial number.
> +	 * It guarantees kmem_cache name uniqueness even if a mem cgroup goes
> +	 * offline, releases kmemcg_id and new mem cgroup gets same kmemcg_id.
> +	 */
> +	memcg_set_serial_nr(memcg, atomic64_inc_return(&serial_nr_cursor));

Here we are already held slab_mutex, so it looks like atomic is overkill.
Simple u64 should be enough.

> +	cache_name = kasprintf(GFP_KERNEL, "%s(%llu:%s)", root_cache->name,
> +			       memcg_get_serial_nr(memcg), memcg_name_buf);
>  	if (!cache_name)
>  		goto out_unlock;

Kirill