It's better disable load balance for receiving packet when CONFIG_RPS is enabled.
If not, octeon-ethernet driver select CPU and then the rps select again CPU.
It can be ipi interrupts overhead and packet reordering could be possible.
Signed-off-by: Eunbong Song <eunb.song@xxxxxxxxxxx>
---
drivers/staging/octeon/ethernet-rx.c | 2 ++
1 files changed, 2 insertions(+), 0 deletions(-)
diff --git a/drivers/staging/octeon/ethernet-rx.c b/drivers/staging/octeon/ethernet-rx.c
index b2b6c3c..44e372f 100644
--- a/drivers/staging/octeon/ethernet-rx.c
+++ b/drivers/staging/octeon/ethernet-rx.c
@@ -286,6 +286,7 @@ static int cvm_oct_napi_poll(struct napi_struct *napi, int budget)
did_work_request = 1;
}
+#ifndef CONFIG_RPS
if (rx_count == 0) {
/*
* First time through, see if there is enough
@@ -300,6 +301,7 @@ static int cvm_oct_napi_poll(struct napi_struct *napi, int budget)
if (backlog > budget * cores_in_use && napi != NULL)
cvm_oct_enable_one_cpu();
}
+#endif
rx_count++;
skb_in_hw = USE_SKBUFFS_IN_HW && work->word2.s.bufs == 1;
--
1.7.0.1
