[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Thu, 2 Feb 2023 18:58:01 +0000
From: Eric Dumazet <edumazet@...gle.com>
To: "David S . Miller" <davem@...emloft.net>,
Jakub Kicinski <kuba@...nel.org>,
Paolo Abeni <pabeni@...hat.com>
Cc: netdev@...r.kernel.org, eric.dumazet@...il.com,
Alexander Duyck <alexanderduyck@...com>,
Soheil Hassas Yeganeh <soheil@...gle.com>,
Eric Dumazet <edumazet@...gle.com>
Subject: [PATCH net-next 4/4] net: add dedicated kmem_cache for typical/small skb->head
Recent removal of ksize() in alloc_skb() increased
performance because we no longer read
the associated struct page.
We have an equivalent cost at kfree_skb() time.
kfree(skb->head) has to access a struct page,
often cold in cpu caches to get the owning
struct kmem_cache.
Considering that many allocations are small,
we can have our own kmem_cache to avoid the cache line miss.
This also saves memory because these small heads
are no longer padded to 1024 bytes.
CONFIG_SLUB=y
$ grep skbuff_small_head /proc/slabinfo
skbuff_small_head 2907 2907 640 51 8 : tunables 0 0 0 : slabdata 57 57 0
CONFIG_SLAB=y
$ grep skbuff_small_head /proc/slabinfo
skbuff_small_head 607 624 640 6 1 : tunables 54 27 8 : slabdata 104 104 5
Note: after Kees Cook patches and this one, we might
be able to revert commit
dbae2b062824 ("net: skb: introduce and use a single page frag cache")
because GRO_MAX_HEAD is also small.
Signed-off-by: Eric Dumazet <edumazet@...gle.com>
Cc: Paolo Abeni <pabeni@...hat.com>
---
net/core/skbuff.c | 52 ++++++++++++++++++++++++++++++++++++++++++-----
1 file changed, 47 insertions(+), 5 deletions(-)
diff --git a/net/core/skbuff.c b/net/core/skbuff.c
index ae0b2aa1f01e8060cc4fe69137e9bd98e44280cc..3e540b4924701cc57b6fbd1b668bab3b652ee94c 100644
--- a/net/core/skbuff.c
+++ b/net/core/skbuff.c
@@ -89,6 +89,19 @@ static struct kmem_cache *skbuff_fclone_cache __ro_after_init;
#ifdef CONFIG_SKB_EXTENSIONS
static struct kmem_cache *skbuff_ext_cache __ro_after_init;
#endif
+static struct kmem_cache *skb_small_head_cache __ro_after_init;
+
+#define SKB_SMALL_HEAD_SIZE SKB_HEAD_ALIGN(MAX_TCP_HEADER)
+
+/* We want SKB_SMALL_HEAD_CACHE_SIZE to not be a power of two. */
+#define SKB_SMALL_HEAD_CACHE_SIZE \
+ (is_power_of_2(SKB_SMALL_HEAD_SIZE) ? \
+ (SKB_SMALL_HEAD_SIZE + L1_CACHE_BYTES) : \
+ SKB_SMALL_HEAD_SIZE)
+
+#define SKB_SMALL_HEAD_HEADROOM \
+ SKB_WITH_OVERHEAD(SKB_SMALL_HEAD_CACHE_SIZE)
+
int sysctl_max_skb_frags __read_mostly = MAX_SKB_FRAGS;
EXPORT_SYMBOL(sysctl_max_skb_frags);
@@ -486,6 +499,21 @@ static void *kmalloc_reserve(unsigned int *size, gfp_t flags, int node,
void *obj;
obj_size = SKB_HEAD_ALIGN(*size);
+ if (obj_size <= SKB_SMALL_HEAD_CACHE_SIZE &&
+ !(flags & KMALLOC_NOT_NORMAL_BITS)) {
+
+ /* skb_small_head_cache has non power of two size,
+ * likely forcing SLUB to use order-3 pages.
+ * We deliberately attempt a NOMEMALLOC allocation only.
+ */
+ obj = kmem_cache_alloc_node(skb_small_head_cache,
+ flags | __GFP_NOMEMALLOC | __GFP_NOWARN,
+ node);
+ if (obj) {
+ *size = SKB_SMALL_HEAD_CACHE_SIZE;
+ goto out;
+ }
+ }
*size = obj_size = kmalloc_size_roundup(obj_size);
/*
* Try a regular allocation, when that fails and we're not entitled
@@ -805,6 +833,14 @@ static bool skb_pp_recycle(struct sk_buff *skb, void *data)
return page_pool_return_skb_page(virt_to_page(data));
}
+static void skb_kfree_head(void *head, unsigned int end_offset)
+{
+ if (end_offset == SKB_SMALL_HEAD_HEADROOM)
+ kmem_cache_free(skb_small_head_cache, head);
+ else
+ kfree(head);
+}
+
static void skb_free_head(struct sk_buff *skb)
{
unsigned char *head = skb->head;
@@ -814,7 +850,7 @@ static void skb_free_head(struct sk_buff *skb)
return;
skb_free_frag(head);
} else {
- kfree(head);
+ skb_kfree_head(head, skb_end_offset(skb));
}
}
@@ -1995,7 +2031,7 @@ int pskb_expand_head(struct sk_buff *skb, int nhead, int ntail,
return 0;
nofrags:
- kfree(data);
+ skb_kfree_head(data, size);
nodata:
return -ENOMEM;
}
@@ -4633,6 +4669,12 @@ void __init skb_init(void)
0,
SLAB_HWCACHE_ALIGN|SLAB_PANIC,
NULL);
+ skb_small_head_cache = kmem_cache_create("skbuff_small_head",
+ SKB_SMALL_HEAD_CACHE_SIZE,
+ 0,
+ SLAB_HWCACHE_ALIGN | SLAB_PANIC,
+ NULL);
+
skb_extensions_init();
}
@@ -6297,7 +6339,7 @@ static int pskb_carve_inside_header(struct sk_buff *skb, const u32 off,
if (skb_cloned(skb)) {
/* drop the old head gracefully */
if (skb_orphan_frags(skb, gfp_mask)) {
- kfree(data);
+ skb_kfree_head(data, size);
return -ENOMEM;
}
for (i = 0; i < skb_shinfo(skb)->nr_frags; i++)
@@ -6405,7 +6447,7 @@ static int pskb_carve_inside_nonlinear(struct sk_buff *skb, const u32 off,
memcpy((struct skb_shared_info *)(data + size),
skb_shinfo(skb), offsetof(struct skb_shared_info, frags[0]));
if (skb_orphan_frags(skb, gfp_mask)) {
- kfree(data);
+ skb_kfree_head(data, size);
return -ENOMEM;
}
shinfo = (struct skb_shared_info *)(data + size);
@@ -6441,7 +6483,7 @@ static int pskb_carve_inside_nonlinear(struct sk_buff *skb, const u32 off,
/* skb_frag_unref() is not needed here as shinfo->nr_frags = 0. */
if (skb_has_frag_list(skb))
kfree_skb_list(skb_shinfo(skb)->frag_list);
- kfree(data);
+ skb_kfree_head(data, size);
return -ENOMEM;
}
skb_release_data(skb, SKB_CONSUMED);
--
2.39.1.456.gfc5497dd1b-goog
Powered by blists - more mailing lists