[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>
diff --git a/include/linux/skbuff.h b/include/linux/skbuff.h
index 9716771..483cfc4 100644
--- a/include/linux/skbuff.h
+++ b/include/linux/skbuff.h
@@ -133,7 +133,7 @@
*/
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;