We use a percpu variable named flow_hash_info, which holds 12 bytes.
It is currently marked as ____cacheline_aligned, which makes linker
skip space to properly align this variable.
Before :
c065cc90 D per_cpu__softnet_data
c065cd00 d per_cpu__flow_tables
<Here, hole of 124 bytes>
c065cd80 d per_cpu__flow_hash_info
<Here, hole of 116 bytes>
c065ce00 d per_cpu__flow_flush_tasklets
c065ce14 d per_cpu__rt_cache_stat
This alignement is quite unproductive, and removing it reduces the
size of percpu data (by 240 bytes on my x86 machine), and improves
performance (flow_tables & flow_hash_info can share a single cache
line)
After patch :
c065cc04 D per_cpu__softnet_data
c065cc4c d per_cpu__flow_tables
c065cc50 d per_cpu__flow_hash_info
c065cc5c d per_cpu__flow_flush_tasklets
c065cc70 d per_cpu__rt_cache_stat
Signed-off-by: Eric Dumazet <[EMAIL PROTECTED]>
diff --git a/net/core/flow.c b/net/core/flow.c
index a618f89..2a735ba 100644
--- a/net/core/flow.c
+++ b/net/core/flow.c
@@ -52,7 +52,7 @@ struct flow_percpu_info {
int hash_rnd_recalc;
u32 hash_rnd;
int count;
-} ____cacheline_aligned;
+};
static DEFINE_PER_CPU(struct flow_percpu_info, flow_hash_info) = { 0 };
#define flow_hash_rnd_recalc(cpu) \
