evie/android_kernel_oneplus_msm8998
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

net: rmnet_data: Call the GRO handler for TCP packets only

Browse source 
While running a UDP test using gro handler, there was a slight
increase in MIPS observed. This is because all UDP packets passed
through napi_gro_receive need to be checked if it is an encapsulated
UDP packet which could actually be coalesced and processed further.

As of now, clients of rmnet_data do not support this and we can
save MIPS on this by passing only TCP packets through the GRO
handler. All non TCP packets will directly be passed directly to
the network stack. This also helps us to avoid an atomic operation
unnecessarily when GRO path is not exercised.

There is 1% savings in MIPS observed when using a single stream
UDP DL connection at 300Mbps.

Change-Id: Ie601000de21afeacfca93f23117aeb0f7cefda98
Signed-off-by: Subash Abhinov Kasiviswanathan <subashab@codeaurora.org>
This commit is contained in:
Subash Abhinov Kasiviswanathan 2015-08-14 18:08:30 -06:00 • committed by David Keitel
parent 747f0aec1b
commit 2b455e304e
1 changed files with 51 additions and 7 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

58
net/rmnet_data/rmnet_data_handlers.c
View file

@ -20,6 +20,8 @@
#include <linux/rmnet_data.h>
#include <linux/net_map.h>
#include <linux/netdev_features.h>
#include <linux/ip.h>
#include <linux/ipv6.h>
#include "rmnet_data_private.h"
#include "rmnet_data_config.h"
#include "rmnet_data_vnd.h"
@ -43,6 +45,12 @@ module_param(dump_pkt_tx, uint, S_IRUGO | S_IWUSR);
MODULE_PARM_DESC(dump_pkt_tx, "Dump packets exiting egress handler");
#endif /* CONFIG_RMNET_DATA_DEBUG_PKT */
#define RMNET_DATA_IP_VERSION_4 0x40
#define RMNET_DATA_IP_VERSION_6 0x60
#define RMNET_DATA_GRO_RCV_FAIL 0
#define RMNET_DATA_GRO_RCV_PASS 1
/* ***************** Helper Functions *************************************** */
/**
@ -57,10 +65,10 @@ MODULE_PARM_DESC(dump_pkt_tx, "Dump packets exiting egress handler");
static inline void __rmnet_data_set_skb_proto(struct sk_buff *skb)
{
switch (skb->data[0] & 0xF0) {
case 0x40: /* IPv4 */
case RMNET_DATA_IP_VERSION_4:
skb->protocol = htons(ETH_P_IP);
break;
case 0x60: /* IPv6 */
case RMNET_DATA_IP_VERSION_6:
skb->protocol = htons(ETH_P_IPV6);
break;
default:
@ -177,6 +185,37 @@ static void rmnet_reset_mac_header(struct sk_buff *skb)
}
#endif /*NET_SKBUFF_DATA_USES_OFFSET*/
/**
* rmnet_check_skb_can_gro() - Check is skb can be passed through GRO handler
*
* Determines whether to pass the skb to the GRO handler napi_gro_receive() or
* handle normally by passing to netif_receive_skb().
*
* Warning:
* This assumes that only TCP packets can be coalesced by the GRO handler which
* is not true in general. We lose the ability to use GRO for cases like UDP
* encapsulation protocols.
*
* Return:
* - RMNET_DATA_GRO_RCV_FAIL if packet is sent to netif_receive_skb()
* - RMNET_DATA_GRO_RCV_PASS if packet is sent to napi_gro_receive()
*/
static int rmnet_check_skb_can_gro(struct sk_buff *skb)
{
switch (skb->data[0] & 0xF0) {
case RMNET_DATA_IP_VERSION_4:
if (ip_hdr(skb)->protocol == IPPROTO_TCP)
return RMNET_DATA_GRO_RCV_PASS;
break;
case RMNET_DATA_IP_VERSION_6:
if (ipv6_hdr(skb)->nexthdr == IPPROTO_TCP)
return RMNET_DATA_GRO_RCV_PASS;
/* Fall through */
}
return RMNET_DATA_GRO_RCV_FAIL;
}
/**
* __rmnet_deliver_skb() - Deliver skb
*
@ -211,11 +250,16 @@ static rx_handler_result_t __rmnet_deliver_skb(struct sk_buff *skb,
case RX_HANDLER_PASS:
skb->pkt_type = PACKET_HOST;
rmnet_reset_mac_header(skb);
if (skb->dev->features & NETIF_F_GRO) {
napi = rmnet_vnd_get_napi(skb->dev);
napi_schedule(napi);
gro_res = napi_gro_receive(napi, skb);
trace_rmnet_gro_downlink(gro_res);
if (rmnet_check_skb_can_gro(skb)) {
if (skb->dev->features & NETIF_F_GRO) {
napi = rmnet_vnd_get_napi(skb->dev);
napi_schedule(napi);
gro_res = napi_gro_receive(napi, skb);
trace_rmnet_gro_downlink(gro_res);
} else {
netif_receive_skb(skb);
}
} else {
netif_receive_skb(skb);
}