The real device over which the rmnet devices are installed also
aggregate multiple IP packets and sends them as a single large
aggregate frame to the hardware. This causes degraded throughput
for TCP TX due to bufferbloat.

To overcome this problem, pacing shift value of 8 is set using the
sk_pacing_shift_update() helper. This value was determined based
on experiments with a single stream TCP TX using iperf for a
duration of 30s.

Pacing shift | Observed data rate (Mbps)
          10 | 9
           9 | 140
           8 | 146 (Max link rate)

Signed-off-by: Subash Abhinov Kasiviswanathan <subas...@codeaurora.org>
---
 drivers/net/ethernet/qualcomm/rmnet/rmnet_handlers.c | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/drivers/net/ethernet/qualcomm/rmnet/rmnet_handlers.c 
b/drivers/net/ethernet/qualcomm/rmnet/rmnet_handlers.c
index 8e1f43a..8f8c4f2 100644
--- a/drivers/net/ethernet/qualcomm/rmnet/rmnet_handlers.c
+++ b/drivers/net/ethernet/qualcomm/rmnet/rmnet_handlers.c
@@ -16,6 +16,7 @@
 #include <linux/netdevice.h>
 #include <linux/netdev_features.h>
 #include <linux/if_arp.h>
+#include <net/sock.h>
 #include "rmnet_private.h"
 #include "rmnet_config.h"
 #include "rmnet_vnd.h"
@@ -204,6 +205,8 @@ void rmnet_egress_handler(struct sk_buff *skb)
        struct rmnet_priv *priv;
        u8 mux_id;
 
+       sk_pacing_shift_update(skb->sk, 8);
+
        orig_dev = skb->dev;
        priv = netdev_priv(orig_dev);
        skb->dev = priv->real_dev;
-- 
1.9.1

Reply via email to