Re: [PATCH v1] memcg: add charging of already allocated slab objects

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Mon, Aug 26, 2024 at 04:29:08PM -0700, Shakeel Butt wrote:
> At the moment, the slab objects are charged to the memcg at the
> allocation time. However there are cases where slab objects are
> allocated at the time where the right target memcg to charge it to is
> not known. One such case is the network sockets for the incoming
> connection which are allocated in the softirq context.
> 
> Couple hundred thousand connections are very normal on large loaded
> server and almost all of those sockets underlying those connections get
> allocated in the softirq context and thus not charged to any memcg.
> However later at the accept() time we know the right target memcg to
> charge. Let's add new API to charge already allocated objects, so we can
> have better accounting of the memory usage.
> 
> To measure the performance impact of this change, tcp_crr is used from
> the neper [1] performance suite. Basically it is a network ping pong
> test with new connection for each ping pong.
> 
> The server and the client are run inside 3 level of cgroup hierarchy
> using the following commands:
> 
> Server:
>  $ tcp_crr -6
> 
> Client:
>  $ tcp_crr -6 -c -H ${server_ip}
> 
> If the client and server run on different machines with 50 GBPS NIC,
> there is no visible impact of the change.
> 
> For the same machine experiment with v6.11-rc5 as base.
> 
>           base (throughput)     with-patch
> tcp_crr   14545 (+- 80)         14463 (+- 56)
> 
> It seems like the performance impact is within the noise.
> 
> Link: https://github.com/google/neper [1]
> Signed-off-by: Shakeel Butt <shakeel.butt@xxxxxxxxx>

Hi Shakeel,

I like the idea and performance numbers look good. However some comments on
the implementation:

> ---
> 
> Changes since the RFC:
> - Added check for already charged slab objects.
> - Added performance results from neper's tcp_crr
> 
>  include/linux/slab.h            |  1 +
>  mm/slub.c                       | 54 +++++++++++++++++++++++++++++++++
>  net/ipv4/inet_connection_sock.c |  5 +--
>  3 files changed, 58 insertions(+), 2 deletions(-)
> 
> diff --git a/include/linux/slab.h b/include/linux/slab.h
> index eb2bf4629157..05cfab107c72 100644
> --- a/include/linux/slab.h
> +++ b/include/linux/slab.h
> @@ -547,6 +547,7 @@ void *kmem_cache_alloc_lru_noprof(struct kmem_cache *s, struct list_lru *lru,
>  			    gfp_t gfpflags) __assume_slab_alignment __malloc;
>  #define kmem_cache_alloc_lru(...)	alloc_hooks(kmem_cache_alloc_lru_noprof(__VA_ARGS__))
>  
> +bool kmem_cache_charge(void *objp, gfp_t gfpflags);
>  void kmem_cache_free(struct kmem_cache *s, void *objp);
>  
>  kmem_buckets *kmem_buckets_create(const char *name, slab_flags_t flags,
> diff --git a/mm/slub.c b/mm/slub.c
> index c9d8a2497fd6..580683597b5c 100644
> --- a/mm/slub.c
> +++ b/mm/slub.c
> @@ -2185,6 +2185,16 @@ void memcg_slab_free_hook(struct kmem_cache *s, struct slab *slab, void **p,
>  
>  	__memcg_slab_free_hook(s, slab, p, objects, obj_exts);
>  }
> +
> +static __fastpath_inline
> +bool memcg_slab_post_charge(struct kmem_cache *s, void *p, gfp_t flags)
> +{
> +	if (likely(!memcg_kmem_online()))
> +		return true;

We do have this check in kmem_cache_charge(), why do we need to check it again?

> +
> +	return __memcg_slab_post_alloc_hook(s, NULL, flags, 1, &p);
> +}
> +
>  #else /* CONFIG_MEMCG */
>  static inline bool memcg_slab_post_alloc_hook(struct kmem_cache *s,
>  					      struct list_lru *lru,
> @@ -2198,6 +2208,13 @@ static inline void memcg_slab_free_hook(struct kmem_cache *s, struct slab *slab,
>  					void **p, int objects)
>  {
>  }
> +
> +static inline bool memcg_slab_post_charge(struct kmem_cache *s,
> +					  void *p,
> +					  gfp_t flags)
> +{
> +	return true;
> +}
>  #endif /* CONFIG_MEMCG */
>  
>  /*
> @@ -4062,6 +4079,43 @@ void *kmem_cache_alloc_lru_noprof(struct kmem_cache *s, struct list_lru *lru,
>  }
>  EXPORT_SYMBOL(kmem_cache_alloc_lru_noprof);
>  
> +#define KMALLOC_TYPE (SLAB_KMALLOC | SLAB_CACHE_DMA | \
> +		      SLAB_ACCOUNT | SLAB_RECLAIM_ACCOUNT)
> +
> +bool kmem_cache_charge(void *objp, gfp_t gfpflags)
> +{
> +	struct slabobj_ext *slab_exts;
> +	struct kmem_cache *s;
> +	struct folio *folio;
> +	struct slab *slab;
> +	unsigned long off;
> +
> +	if (!memcg_kmem_online())
> +		return true;
> +
> +	folio = virt_to_folio(objp);
> +	if (unlikely(!folio_test_slab(folio)))
> +		return false;

Does it handle the case of a too-big-to-be-a-slab-object allocation?
I think it's better to handle it properly. Also, why return false here?

> +
> +	slab = folio_slab(folio);
> +	s = slab->slab_cache;
> +
> +	/* Ignore KMALLOC_NORMAL cache to avoid circular dependency. */
> +	if ((s->flags & KMALLOC_TYPE) == SLAB_KMALLOC)
> +		return true;

And true here? It seems to be a bit inconsistent.
Also, if we have this check here, it means your function won't handle kmallocs
at all? Because !KMALLOC_NORMAL allocations won't get here.

> +
> +	/* Ignore already charged objects. */
> +	slab_exts = slab_obj_exts(slab);
> +	if (slab_exts) {
> +		off = obj_to_index(s, slab, objp);
> +		if (unlikely(slab_exts[off].objcg))
> +			return true;
> +	}
> +
> +	return memcg_slab_post_charge(s, objp, gfpflags);
> +}
> +EXPORT_SYMBOL(kmem_cache_charge);
> +
>  /**
>   * kmem_cache_alloc_node - Allocate an object on the specified node
>   * @s: The cache to allocate from.
> diff --git a/net/ipv4/inet_connection_sock.c b/net/ipv4/inet_connection_sock.c
> index 64d07b842e73..3c13ca8c11fb 100644
> --- a/net/ipv4/inet_connection_sock.c
> +++ b/net/ipv4/inet_connection_sock.c
> @@ -715,6 +715,7 @@ struct sock *inet_csk_accept(struct sock *sk, struct proto_accept_arg *arg)
>  	release_sock(sk);
>  	if (newsk && mem_cgroup_sockets_enabled) {
>  		int amt = 0;
> +		gfp_t gfp = GFP_KERNEL | __GFP_NOFAIL;
>  
>  		/* atomically get the memory usage, set and charge the
>  		 * newsk->sk_memcg.
> @@ -731,8 +732,8 @@ struct sock *inet_csk_accept(struct sock *sk, struct proto_accept_arg *arg)
>  		}
>  
>  		if (amt)
> -			mem_cgroup_charge_skmem(newsk->sk_memcg, amt,
> -						GFP_KERNEL | __GFP_NOFAIL);
> +			mem_cgroup_charge_skmem(newsk->sk_memcg, amt, gfp);
> +		kmem_cache_charge(newsk, gfp);

Wait, so we assume that newsk->sk_memcg === current memcg? Or we're ok with them being
different?

Thanks!




[Index of Archives]     [Linux ARM Kernel]     [Linux ARM]     [Linux Omap]     [Fedora ARM]     [IETF Annouce]     [Security]     [Bugtraq]     [Linux OMAP]     [Linux MIPS]     [eCos]     [Asterisk Internet PBX]     [Linux API]     [Monitors]

  Powered by Linux