[NET]: Speed up __alloc_skb()

From: Benjamin LaHaise <bcrl@kvack.org>

In __alloc_skb(), the use of skb_shinfo() which casts a u8 * to the 
shared info structure results in gcc being forced to do a reload of the 
pointer since it has no information on possible aliasing.  Fix this by 
using a pointer to refer to skb_shared_info.

By initializing skb_shared_info sequentially, the write combining buffers 
can reduce the number of memory transactions to a single write.  Reorder 
the initialization in __alloc_skb() to match the structure definition.  
There is also an alignment issue on 64 bit systems with skb_shared_info 
by converting nr_frags to a short everything packs up nicely.

Also, pass the slab cache pointer according to the fclone flag instead 
of using two almost identical function calls.

This raises bw_unix performance up to a peak of 707KB/s when combined 
with the spinlock patch.  It should help other networking protocols, too.
Signed-off-by: David S. Miller <davem@davemloft.net>

include/linux/skbuff.h

@@ -133,7 +133,7 @@ struct skb_frag_struct {
  */
 struct skb_shared_info {
 	atomic_t	dataref;
-	unsigned int	nr_frags;
+	unsigned short	nr_frags;
 	unsigned short	tso_size;
 	unsigned short	tso_segs;
 	unsigned short  ufo_size;

net/core/skbuff.c

@@ -135,17 +135,13 @@ void skb_under_panic(struct sk_buff *skb, int sz, void *here)
 struct sk_buff *__alloc_skb(unsigned int size, gfp_t gfp_mask,
 			    int fclone)
 {
+	struct skb_shared_info *shinfo;
 	struct sk_buff *skb;
 	u8 *data;
 
 	/* Get the HEAD */
-	if (fclone)
-		skb = kmem_cache_alloc(skbuff_fclone_cache,
-				       gfp_mask & ~__GFP_DMA);
-	else
-		skb = kmem_cache_alloc(skbuff_head_cache,
-				       gfp_mask & ~__GFP_DMA);
-
+	skb = kmem_cache_alloc(fclone ? skbuff_fclone_cache : skbuff_head_cache,
+				gfp_mask & ~__GFP_DMA);
 	if (!skb)
 		goto out;
 
@@ -162,6 +158,16 @@ struct sk_buff *__alloc_skb(unsigned int size, gfp_t gfp_mask,
 	skb->data = data;
 	skb->tail = data;
 	skb->end  = data + size;
+	/* make sure we initialize shinfo sequentially */
+	shinfo = skb_shinfo(skb);
+	atomic_set(&shinfo->dataref, 1);
+	shinfo->nr_frags  = 0;
+	shinfo->tso_size = 0;
+	shinfo->tso_segs = 0;
+	shinfo->ufo_size = 0;
+	shinfo->ip6_frag_id = 0;
+	shinfo->frag_list = NULL;
+
 	if (fclone) {
 		struct sk_buff *child = skb + 1;
 		atomic_t *fclone_ref = (atomic_t *) (child + 1);
@@ -171,13 +177,6 @@ struct sk_buff *__alloc_skb(unsigned int size, gfp_t gfp_mask,
 
 		child->fclone = SKB_FCLONE_UNAVAILABLE;
 	}
-	atomic_set(&(skb_shinfo(skb)->dataref), 1);
-	skb_shinfo(skb)->nr_frags  = 0;
-	skb_shinfo(skb)->tso_size = 0;
-	skb_shinfo(skb)->tso_segs = 0;
-	skb_shinfo(skb)->frag_list = NULL;
-	skb_shinfo(skb)->ufo_size = 0;
-	skb_shinfo(skb)->ip6_frag_id = 0;
 out:
 	return skb;
 nodata: