Merge branch 'net-core-use-a-dedicated-kmem_cache-for-skb-head-allocs'

Eric Dumazet says:

====================
net: core: use a dedicated kmem_cache for skb head allocs

Our profile data show that using kmalloc(non_const_size)/kfree(ptr)
has a certain cost, because kfree(ptr) has to pull a 'struct page'
in cpu caches.

Using a dedicated kmem_cache for TCP skb->head allocations makes
a difference, both in cpu cycles and memory savings.

This kmem_cache could also be used for GRO skb allocations,
this is left as a future exercise.
====================

Link: https://lore.kernel.org/r/20230206173103.2617121-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

1 files changed, 82 insertions, 33 deletions

diff --git a/net/core/skbuff.c b/net/core/skbuff.c
index 624e9e4ec116..bdb1e015e32b 100644
--- a/net/core/skbuff.c
+++ b/net/core/skbuff.c
@@ -89,6 +89,34 @@ 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
+
+/* skb_small_head_cache and related code is only supported
+ * for CONFIG_SLAB and CONFIG_SLUB.
+ * As soon as SLOB is removed from the kernel, we can clean up this.
+ */
+#if !defined(CONFIG_SLOB)
+# define HAVE_SKB_SMALL_HEAD_CACHE 1
+#endif
+
+#ifdef HAVE_SKB_SMALL_HEAD_CACHE
+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.
+ * This should ensure that SKB_SMALL_HEAD_HEADROOM is a unique
+ * size, and we can differentiate heads from skb_small_head_cache
+ * vs system slabs by looking at their size (skb_end_offset()).
+ */
+#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)
+#endif /* HAVE_SKB_SMALL_HEAD_CACHE */
+
 int sysctl_max_skb_frags __read_mostly = MAX_SKB_FRAGS;
 EXPORT_SYMBOL(sysctl_max_skb_frags);
 
@@ -478,17 +506,37 @@ EXPORT_SYMBOL(napi_build_skb);
  * may be used. Otherwise, the packet data may be discarded until enough
  * memory is free
  */
-static void *kmalloc_reserve(size_t size, gfp_t flags, int node,
+static void *kmalloc_reserve(unsigned int *size, gfp_t flags, int node,
 			     bool *pfmemalloc)
 {
-	void *obj;
 	bool ret_pfmemalloc = false;
+	unsigned int obj_size;
+	void *obj;
 
+	obj_size = SKB_HEAD_ALIGN(*size);
+#ifdef HAVE_SKB_SMALL_HEAD_CACHE
+	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;
+		}
+	}
+#endif
+	*size = obj_size = kmalloc_size_roundup(obj_size);
 	/*
 	 * Try a regular allocation, when that fails and we're not entitled
 	 * to the reserves, fail.
 	 */
-	obj = kmalloc_node_track_caller(size,
+	obj = kmalloc_node_track_caller(obj_size,
 					flags | __GFP_NOMEMALLOC | __GFP_NOWARN,
 					node);
 	if (obj || !(gfp_pfmemalloc_allowed(flags)))
@@ -496,7 +544,7 @@ static void *kmalloc_reserve(size_t size, gfp_t flags, int node,
 
 	/* Try again but now we are using pfmemalloc reserves */
 	ret_pfmemalloc = true;
-	obj = kmalloc_node_track_caller(size, flags, node);
+	obj = kmalloc_node_track_caller(obj_size, flags, node);
 
 out:
 	if (pfmemalloc)
@@ -533,7 +581,6 @@ struct sk_buff *__alloc_skb(unsigned int size, gfp_t gfp_mask,
 {
 	struct kmem_cache *cache;
 	struct sk_buff *skb;
-	unsigned int osize;
 	bool pfmemalloc;
 	u8 *data;
 
@@ -558,18 +605,14 @@ struct sk_buff *__alloc_skb(unsigned int size, gfp_t gfp_mask,
 	 * aligned memory blocks, unless SLUB/SLAB debug is enabled.
 	 * Both skb->head and skb_shared_info are cache line aligned.
 	 */
-	size = SKB_DATA_ALIGN(size);
-	size += SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
-	osize = kmalloc_size_roundup(size);
-	data = kmalloc_reserve(osize, gfp_mask, node, &pfmemalloc);
+	data = kmalloc_reserve(&size, gfp_mask, node, &pfmemalloc);
 	if (unlikely(!data))
 		goto nodata;
 	/* kmalloc_size_roundup() might give us more room than requested.
 	 * Put skb_shared_info exactly at the end of allocated zone,
 	 * to allow max possible filling before reallocation.
 	 */
-	size = SKB_WITH_OVERHEAD(osize);
-	prefetchw(data + size);
+	prefetchw(data + SKB_WITH_OVERHEAD(size));
 
 	/*
 	 * Only clear those fields we need to clear, not those that we will
@@ -577,7 +620,7 @@ struct sk_buff *__alloc_skb(unsigned int size, gfp_t gfp_mask,
 	 * the tail pointer in struct sk_buff!
 	 */
 	memset(skb, 0, offsetof(struct sk_buff, tail));
-	__build_skb_around(skb, data, osize);
+	__build_skb_around(skb, data, size);
 	skb->pfmemalloc = pfmemalloc;
 
 	if (flags & SKB_ALLOC_FCLONE) {
@@ -632,8 +675,7 @@ struct sk_buff *__netdev_alloc_skb(struct net_device *dev, unsigned int len,
 		goto skb_success;
 	}
 
-	len += SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
-	len = SKB_DATA_ALIGN(len);
+	len = SKB_HEAD_ALIGN(len);
 
 	if (sk_memalloc_socks())
 		gfp_mask |= __GFP_MEMALLOC;
@@ -732,8 +774,7 @@ struct sk_buff *__napi_alloc_skb(struct napi_struct *napi, unsigned int len,
 		data = page_frag_alloc_1k(&nc->page_small, gfp_mask);
 		pfmemalloc = NAPI_SMALL_PAGE_PFMEMALLOC(nc->page_small);
 	} else {
-		len += SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
-		len = SKB_DATA_ALIGN(len);
+		len = SKB_HEAD_ALIGN(len);
 
 		data = page_frag_alloc(&nc->page, len, gfp_mask);
 		pfmemalloc = nc->page.pfmemalloc;
@@ -809,6 +850,16 @@ 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)
+{
+#ifdef HAVE_SKB_SMALL_HEAD_CACHE
+	if (end_offset == SKB_SMALL_HEAD_HEADROOM)
+		kmem_cache_free(skb_small_head_cache, head);
+	else
+#endif
+		kfree(head);
+}
+
 static void skb_free_head(struct sk_buff *skb)
 {
 	unsigned char *head = skb->head;
@@ -818,7 +869,7 @@ 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));
 	}
 }
 
@@ -1938,10 +1989,7 @@ int pskb_expand_head(struct sk_buff *skb, int nhead, int ntail,
 	if (skb_pfmemalloc(skb))
 		gfp_mask |= __GFP_MEMALLOC;
 
-	size = SKB_DATA_ALIGN(size);
-	size += SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
-	size = kmalloc_size_roundup(size);
-	data = kmalloc_reserve(size, gfp_mask, NUMA_NO_NODE, NULL);
+	data = kmalloc_reserve(&size, gfp_mask, NUMA_NO_NODE, NULL);
 	if (!data)
 		goto nodata;
 	size = SKB_WITH_OVERHEAD(size);
@@ -2004,7 +2052,7 @@ 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;
 }
@@ -4641,6 +4689,13 @@ void __init skb_init(void)
 						0,
 						SLAB_HWCACHE_ALIGN|SLAB_PANIC,
 						NULL);
+#ifdef HAVE_SKB_SMALL_HEAD_CACHE
+	skb_small_head_cache = kmem_cache_create("skbuff_small_head",
+						SKB_SMALL_HEAD_CACHE_SIZE,
+						0,
+						SLAB_HWCACHE_ALIGN | SLAB_PANIC,
+						NULL);
+#endif
 	skb_extensions_init();
 }
 
@@ -6289,10 +6344,7 @@ static int pskb_carve_inside_header(struct sk_buff *skb, const u32 off,
 	if (skb_pfmemalloc(skb))
 		gfp_mask |= __GFP_MEMALLOC;
 
-	size = SKB_DATA_ALIGN(size);
-	size += SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
-	size = kmalloc_size_roundup(size);
-	data = kmalloc_reserve(size, gfp_mask, NUMA_NO_NODE, NULL);
+	data = kmalloc_reserve(&size, gfp_mask, NUMA_NO_NODE, NULL);
 	if (!data)
 		return -ENOMEM;
 	size = SKB_WITH_OVERHEAD(size);
@@ -6308,7 +6360,7 @@ 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++)
@@ -6408,10 +6460,7 @@ static int pskb_carve_inside_nonlinear(struct sk_buff *skb, const u32 off,
 	if (skb_pfmemalloc(skb))
 		gfp_mask |= __GFP_MEMALLOC;
 
-	size = SKB_DATA_ALIGN(size);
-	size += SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
-	size = kmalloc_size_roundup(size);
-	data = kmalloc_reserve(size, gfp_mask, NUMA_NO_NODE, NULL);
+	data = kmalloc_reserve(&size, gfp_mask, NUMA_NO_NODE, NULL);
 	if (!data)
 		return -ENOMEM;
 	size = SKB_WITH_OVERHEAD(size);
@@ -6419,7 +6468,7 @@ 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);
@@ -6455,7 +6504,7 @@ 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);