| Message ID | 20240829175339.2424521-1-shakeel.butt@linux.dev (mailing list archive) |
|---|---|
| State | Superseded |
| Delegated to: | Netdev Maintainers |
| Headers | show |
| Series | [v3] memcg: add charging of already allocated slab objects | expand |
| Context | Check | Description |
|---|---|---|
| netdev/tree_selection | success | Not a local patch |
On 8/29/24 19:53, 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@linux.dev> Thanks, pushed to slab/for-next for test coverage, hopefully net people will ack. Also one thing: We should add some kernel doc for this, no? Explaining when people are supposed to use this, that objects from KMALLOC_NORMAL will be ignored, and what the return value means (including where it's faked to be true). > +bool kmem_cache_charge(void *objp, gfp_t gfpflags) > +{ > + if (!memcg_kmem_online()) > + return true; > + > + return memcg_slab_post_charge(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); > > release_sock(newsk); > }
On Fri, Aug 30, 2024 at 11:07:32AM GMT, Vlastimil Babka wrote: > On 8/29/24 19:53, 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@linux.dev> > > Thanks, pushed to slab/for-next for test coverage, hopefully net people will > ack. > > Also one thing: > > We should add some kernel doc for this, no? Explaining when people are > supposed to use this, that objects from KMALLOC_NORMAL will be ignored, and > what the return value means (including where it's faked to be true). > Yes this makes sense. I will add this info similar to the kmalloc() have. Should I send a v4 with this details?
On 8/30/24 21:44, Shakeel Butt wrote: >> >> Also one thing: >> >> We should add some kernel doc for this, no? Explaining when people are >> supposed to use this, that objects from KMALLOC_NORMAL will be ignored, and >> what the return value means (including where it's faked to be true). >> > > Yes this makes sense. I will add this info similar to the kmalloc() > have. Should I send a v4 with this details? Yeah please, will replace v3 with that. Thanks
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/slab.h b/mm/slab.h index dcdb56b8e7f5..9f907e930609 100644 --- a/mm/slab.h +++ b/mm/slab.h @@ -443,6 +443,13 @@ static inline bool is_kmalloc_cache(struct kmem_cache *s) return (s->flags & SLAB_KMALLOC); } +static inline bool is_kmalloc_normal(struct kmem_cache *s) +{ + if (!is_kmalloc_cache(s)) + return false; + return !(s->flags & (SLAB_CACHE_DMA|SLAB_ACCOUNT|SLAB_RECLAIM_ACCOUNT)); +} + /* Legal flag mask for kmem_cache_create(), for various configurations */ #define SLAB_CORE_FLAGS (SLAB_HWCACHE_ALIGN | SLAB_CACHE_DMA | \ SLAB_CACHE_DMA32 | SLAB_PANIC | \ diff --git a/mm/slub.c b/mm/slub.c index c9d8a2497fd6..3f2a89f7a23a 100644 --- a/mm/slub.c +++ b/mm/slub.c @@ -2185,6 +2185,41 @@ 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(void *p, gfp_t flags) +{ + struct slabobj_ext *slab_exts; + struct kmem_cache *s; + struct folio *folio; + struct slab *slab; + unsigned long off; + + folio = virt_to_folio(p); + if (!folio_test_slab(folio)) { + return folio_memcg_kmem(folio) || + (__memcg_kmem_charge_page(folio_page(folio, 0), flags, + folio_order(folio)) == 0); + } + + slab = folio_slab(folio); + s = slab->slab_cache; + + /* Ignore KMALLOC_NORMAL cache to avoid circular dependency. */ + if (is_kmalloc_normal(s)) + return true; + + /* Ignore already charged objects. */ + slab_exts = slab_obj_exts(slab); + if (slab_exts) { + off = obj_to_index(s, slab, p); + if (unlikely(slab_exts[off].objcg)) + return true; + } + + 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 +2233,11 @@ 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(void *p, gfp_t flags) +{ + return true; +} #endif /* CONFIG_MEMCG */ /* @@ -4062,6 +4102,15 @@ void *kmem_cache_alloc_lru_noprof(struct kmem_cache *s, struct list_lru *lru, } EXPORT_SYMBOL(kmem_cache_alloc_lru_noprof); +bool kmem_cache_charge(void *objp, gfp_t gfpflags) +{ + if (!memcg_kmem_online()) + return true; + + return memcg_slab_post_charge(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); release_sock(newsk); }
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@linux.dev> --- v2: https://lore.kernel.org/all/20240827235228.1591842-1-shakeel.butt@linux.dev/ Change since v2: - Add handling of already charged large kmalloc objects. - Move the normal kmalloc cache check into a function. v1: https://lore.kernel.org/all/20240826232908.4076417-1-shakeel.butt@linux.dev/ Changes since v1: - Correctly handle large allocations which bypass slab - Rearrange code to avoid compilation errors for !CONFIG_MEMCG builds RFC: https://lore.kernel.org/all/20240824010139.1293051-1-shakeel.butt@linux.dev/ 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/slab.h | 7 +++++ mm/slub.c | 49 +++++++++++++++++++++++++++++++++ net/ipv4/inet_connection_sock.c | 5 ++-- 4 files changed, 60 insertions(+), 2 deletions(-)