netscale.patch

Author: Wangyang Guo  Guo, Wangyang <wangyang.guo@intel.com>


The dst_entry struct has a reference count that gets touched
quite a bit in scenarios where many connections happen
from/to the same IP. This dirty cache line is shared
with other members that are read (mostly) that are 
used quite a bit (for example, ltwstate)

In addition, this struct is embedded in other structs
and the __refcnt shows up as falls sharing even in that 
context.

An example workload is the phoronix pts/memcached benchmark
(in 1:100 mode) that exagerates the problem in its setup.

Without the patch below the workload gets a score of
771377, while the patch below improves this to
1027113.. a 30%+ gain.

It's likely possible to reorder some fields in the struct
to reduce the size of the needed padding, but this is
the simplest solution.

Signed-off-by: Arjan van de Ven <arjan@linux.intel.com>


--- a/include/net/dst.h	2022-12-11 22:15:18.000000000 +0000
+++ b/include/net/dst.h	2023-02-13 14:48:45.498505188 +0000
@@ -66,6 +66,7 @@
 	 */
 #ifdef CONFIG_64BIT
 	atomic_t		__refcnt;	/* 64-bit offset 64 */
+	int			__pad2[15];
 #endif
 	int			__use;
 	unsigned long		lastuse;