[PATCH net-next] udp_offload: Allow device GRO without checksum-complete

[PATCH net-next] udp_offload: Allow device GRO without checksum-complete

[Tom Herbert]

This patch adds a sysctl which allows GRO for a UDP offload protocol
to be performed in the device NAPI. This potentially is a performance
improvement if the savings of doing GRO in device NAPI outweighs the
cost of performing the checksum. Note that the performing the
checksum in device NAPI may negatively impact latency or throughput
of unrelated flows.

Performance results for VXLAN are below. Allowing GRO in device
NAPI does show performance improvement over doing GRO at the VXLAN
interface, however this performance is still less than what we see
with UDP checksums enabled (or getting checksum complete from the
device).

Test results: Running one netperf TCP_STREAM over VXLAN.

No UDP checksum, enable sysctl to allow GRO at device (this patch)
  TX CPU: 1.71
  RX CPU: 1.14
  6174 Mbps

UDP checksums and remote checksum offload enabled
  TX CPU: 1.97%
  RX CPU: 1.55%
  7527 Mbps

UDP checksums enabled
  TX CPU: 1.22%
  RX CPU: 1.86%
  6539 Mbps

No UDP checksums, GRO enabled on VXLAN interface
  TX CPU: 0.95%
  RX CPU: 1.78%
  4393 Mbps

No UDP checksum, GRO disabled VXLAN interface
  TX CPU: 1.31%
  RX CPU: 2.38%
  3613 Mbps

Signed-off-by: Tom Herbert <tom@herbertland.com>
---
 Documentation/networking/ip-sysctl.txt | 7 +++++++
 include/net/udp.h                      | 1 +
 net/ipv4/sysctl_net_ipv4.c             | 7 +++++++
 net/ipv4/udp.c                         | 3 +++
 net/ipv4/udp_offload.c                 | 7 ++++---
 5 files changed, 22 insertions(+), 3 deletions(-)

diff --git a/Documentation/networking/ip-sysctl.txt b/Documentation/networking/ip-sysctl.txt
index 46e88ed..d8563c08 100644
--- a/Documentation/networking/ip-sysctl.txt
+++ b/Documentation/networking/ip-sysctl.txt
@@ -711,6 +711,13 @@ udp_wmem_min - INTEGER
 	total pages of UDP sockets exceed udp_mem pressure. The unit is byte.
 	Default: 1 page
 
+udp_gro_nocsum_ok - BOOLEAN
+	If set, allow Generic Receive Offload (GRO) to be performed for UDP
+	offload protocols in the case that packets are being received
+	without an offloaded checksum. This implies that packets checksums
+	may be performed in the device NAPI routines which could negatively
+	impact unrelated flows.
+
 CIPSOv4 Variables:
 
 cipso_cache_enable - BOOLEAN
diff --git a/include/net/udp.h b/include/net/udp.h
index 6d4ed18..48eb6ae 100644
--- a/include/net/udp.h
+++ b/include/net/udp.h
@@ -103,6 +103,7 @@ extern atomic_long_t udp_memory_allocated;
 extern long sysctl_udp_mem[3];
 extern int sysctl_udp_rmem_min;
 extern int sysctl_udp_wmem_min;
+extern int sysctl_udp_gro_nocsum_ok;
 
 struct sk_buff;
 
diff --git a/net/ipv4/sysctl_net_ipv4.c b/net/ipv4/sysctl_net_ipv4.c
index 0330ab2..65fea78 100644
--- a/net/ipv4/sysctl_net_ipv4.c
+++ b/net/ipv4/sysctl_net_ipv4.c
@@ -766,6 +766,13 @@ static struct ctl_table ipv4_table[] = {
 		.proc_handler	= proc_dointvec_minmax,
 		.extra1		= &one
 	},
+	{
+		.procname	= "udp_gro_nocsum_ok",
+		.data		= &sysctl_udp_gro_nocsum_ok,
+		.maxlen		= sizeof(sysctl_udp_gro_nocsum_ok),
+		.mode		= 0644,
+		.proc_handler	= proc_dointvec_minmax,
+	},
 	{ }
 };
 
diff --git a/net/ipv4/udp.c b/net/ipv4/udp.c
index c0a15e7..1d91227 100644
--- a/net/ipv4/udp.c
+++ b/net/ipv4/udp.c
@@ -130,6 +130,9 @@ EXPORT_SYMBOL(sysctl_udp_wmem_min);
 atomic_long_t udp_memory_allocated;
 EXPORT_SYMBOL(udp_memory_allocated);
 
+int sysctl_udp_gro_nocsum_ok;
+EXPORT_SYMBOL(sysctl_udp_gro_nocsum_ok);
+
 #define MAX_UDP_PORTS 65536
 #define PORTS_PER_CHAIN (MAX_UDP_PORTS / UDP_HTABLE_SIZE_MIN)
 
diff --git a/net/ipv4/udp_offload.c b/net/ipv4/udp_offload.c
index f938616..1666f44 100644
--- a/net/ipv4/udp_offload.c
+++ b/net/ipv4/udp_offload.c
@@ -300,9 +300,10 @@ struct sk_buff **udp_gro_receive(struct sk_buff **head, struct sk_buff *skb,
 	int flush = 1;
 
 	if (NAPI_GRO_CB(skb)->udp_mark ||
-	    (skb->ip_summed != CHECKSUM_PARTIAL &&
-	     NAPI_GRO_CB(skb)->csum_cnt == 0 &&
-	     !NAPI_GRO_CB(skb)->csum_valid))
+	    ((skb->ip_summed != CHECKSUM_PARTIAL &&
+	      NAPI_GRO_CB(skb)->csum_cnt == 0 &&
+	      !NAPI_GRO_CB(skb)->csum_valid) &&
+	      !sysctl_udp_gro_nocsum_ok))
 		goto out;
 
 	/* mark that this skb passed once through the udp gro layer */
-- 
1.8.1

Re: [PATCH net-next] udp_offload: Allow device GRO without checksum-complete

[Eric Dumazet]

On Mon, 2015-08-24 at 12:34 -0700, Tom Herbert wrote:

> diff --git a/net/ipv4/udp.c b/net/ipv4/udp.c
> index c0a15e7..1d91227 100644
> --- a/net/ipv4/udp.c
> +++ b/net/ipv4/udp.c
> @@ -130,6 +130,9 @@ EXPORT_SYMBOL(sysctl_udp_wmem_min);
>  atomic_long_t udp_memory_allocated;
>  EXPORT_SYMBOL(udp_memory_allocated);
>  
> +int sysctl_udp_gro_nocsum_ok;
> +EXPORT_SYMBOL(sysctl_udp_gro_nocsum_ok);
> +

1) Why is this exported ?

2) I do not believe it is specific to UDP path.

  We could have the same sysctl for GRE or IPIP or XXX encaps ?

Re: [PATCH net-next] udp_offload: Allow device GRO without checksum-complete

[Ramu Ramamurthy]

On 2015-08-24 12:34, Tom Herbert wrote:
> This patch adds a sysctl which allows GRO for a UDP offload protocol
> to be performed in the device NAPI. This potentially is a performance
> improvement if the savings of doing GRO in device NAPI outweighs the
> cost of performing the checksum. Note that the performing the
> checksum in device NAPI may negatively impact latency or throughput
> of unrelated flows.
> 
> Performance results for VXLAN are below. Allowing GRO in device
> NAPI does show performance improvement over doing GRO at the VXLAN
> interface, however this performance is still less than what we see
> with UDP checksums enabled (or getting checksum complete from the
> device).
> 
> Test results: Running one netperf TCP_STREAM over VXLAN.
> 
> No UDP checksum, enable sysctl to allow GRO at device (this patch)
>   TX CPU: 1.71
>   RX CPU: 1.14
>   6174 Mbps
> 
> UDP checksums and remote checksum offload enabled
>   TX CPU: 1.97%
>   RX CPU: 1.55%
>   7527 Mbps
> 
> UDP checksums enabled
>   TX CPU: 1.22%
>   RX CPU: 1.86%
>   6539 Mbps
> 
> No UDP checksums, GRO enabled on VXLAN interface
>   TX CPU: 0.95%
>   RX CPU: 1.78%
>   4393 Mbps
> 
> No UDP checksum, GRO disabled VXLAN interface
>   TX CPU: 1.31%
>   RX CPU: 2.38%
>   3613 Mbps
> 
> Signed-off-by: Tom Herbert <tom@herbertland.com>
> ---
>  Documentation/networking/ip-sysctl.txt | 7 +++++++
>  include/net/udp.h                      | 1 +
>  net/ipv4/sysctl_net_ipv4.c             | 7 +++++++
>  net/ipv4/udp.c                         | 3 +++
>  net/ipv4/udp_offload.c                 | 7 ++++---
>  5 files changed, 22 insertions(+), 3 deletions(-)
> 
> diff --git a/Documentation/networking/ip-sysctl.txt
> b/Documentation/networking/ip-sysctl.txt
> index 46e88ed..d8563c08 100644
> --- a/Documentation/networking/ip-sysctl.txt
> +++ b/Documentation/networking/ip-sysctl.txt
> @@ -711,6 +711,13 @@ udp_wmem_min - INTEGER
>  	total pages of UDP sockets exceed udp_mem pressure. The unit is byte.
>  	Default: 1 page
> 
> +udp_gro_nocsum_ok - BOOLEAN
> +	If set, allow Generic Receive Offload (GRO) to be performed for UDP
> +	offload protocols in the case that packets are being received
> +	without an offloaded checksum. This implies that packets checksums
> +	may be performed in the device NAPI routines which could negatively
> +	impact unrelated flows.
> +
>  CIPSOv4 Variables:
> 
>  cipso_cache_enable - BOOLEAN
> diff --git a/include/net/udp.h b/include/net/udp.h
> index 6d4ed18..48eb6ae 100644
> --- a/include/net/udp.h
> +++ b/include/net/udp.h
> @@ -103,6 +103,7 @@ extern atomic_long_t udp_memory_allocated;
>  extern long sysctl_udp_mem[3];
>  extern int sysctl_udp_rmem_min;
>  extern int sysctl_udp_wmem_min;
> +extern int sysctl_udp_gro_nocsum_ok;
> 
>  struct sk_buff;
> 
> diff --git a/net/ipv4/sysctl_net_ipv4.c b/net/ipv4/sysctl_net_ipv4.c
> index 0330ab2..65fea78 100644
> --- a/net/ipv4/sysctl_net_ipv4.c
> +++ b/net/ipv4/sysctl_net_ipv4.c
> @@ -766,6 +766,13 @@ static struct ctl_table ipv4_table[] = {
>  		.proc_handler	= proc_dointvec_minmax,
>  		.extra1		= &one
>  	},
> +	{
> +		.procname	= "udp_gro_nocsum_ok",
> +		.data		= &sysctl_udp_gro_nocsum_ok,
> +		.maxlen		= sizeof(sysctl_udp_gro_nocsum_ok),
> +		.mode		= 0644,
> +		.proc_handler	= proc_dointvec_minmax,
> +	},
>  	{ }
>  };
> 
> diff --git a/net/ipv4/udp.c b/net/ipv4/udp.c
> index c0a15e7..1d91227 100644
> --- a/net/ipv4/udp.c
> +++ b/net/ipv4/udp.c
> @@ -130,6 +130,9 @@ EXPORT_SYMBOL(sysctl_udp_wmem_min);
>  atomic_long_t udp_memory_allocated;
>  EXPORT_SYMBOL(udp_memory_allocated);
> 
> +int sysctl_udp_gro_nocsum_ok;
> +EXPORT_SYMBOL(sysctl_udp_gro_nocsum_ok);
> +
>  #define MAX_UDP_PORTS 65536
>  #define PORTS_PER_CHAIN (MAX_UDP_PORTS / UDP_HTABLE_SIZE_MIN)
> 
> diff --git a/net/ipv4/udp_offload.c b/net/ipv4/udp_offload.c
> index f938616..1666f44 100644
> --- a/net/ipv4/udp_offload.c
> +++ b/net/ipv4/udp_offload.c
> @@ -300,9 +300,10 @@ struct sk_buff **udp_gro_receive(struct sk_buff
> **head, struct sk_buff *skb,
>  	int flush = 1;
> 
>  	if (NAPI_GRO_CB(skb)->udp_mark ||
> -	    (skb->ip_summed != CHECKSUM_PARTIAL &&
> -	     NAPI_GRO_CB(skb)->csum_cnt == 0 &&
> -	     !NAPI_GRO_CB(skb)->csum_valid))
> +	    ((skb->ip_summed != CHECKSUM_PARTIAL &&
> +	      NAPI_GRO_CB(skb)->csum_cnt == 0 &&
> +	      !NAPI_GRO_CB(skb)->csum_valid) &&
> +	      !sysctl_udp_gro_nocsum_ok))
>  		goto out;
> 
>  	/* mark that this skb passed once through the udp gro layer */

Thanks for making this configurable, It would help with 10G adapters 
including ( intel 82599es , intel br kx4 dual-port)