[PATCH net-next] udp: Avoid post-GRO UDP checksum recalculation

[PATCH net-next] udp: Avoid post-GRO UDP checksum recalculation

[Sean Tranchetti]

Currently, when resegmenting an unexpected UDP GRO packet, the full UDP
checksum will be calculated for every new SKB created by skb_segment()
because the netdev features passed in by udp_rcv_segment() lack any
information about checksum offload capabilities.

We have no need to perform this calculation again, as
  1) The GRO implementation guarantees that any packets making it to the
     udp_rcv_segment() function had correct checksums, and, more
     importantly,
  2) Upon the successful return of udp_rcv_segment(), we immediately pull
     the UDP header off and either queue the segment to the socket or
     hand it off to a new protocol handler. In either case, the checksum
     is not needed.

Fixes: cf329aa42b66 ("udp: cope with UDP GRO packet misdirection")
Cc: Paolo Abeni <pabeni@redhat.com>
Cc: Subash Abhinov Kasiviswanathan <subashab@codeaurora.org>
Signed-off-by: Sean Tranchetti <stranche@codeaurora.org>
---
 include/net/udp.h | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/include/net/udp.h b/include/net/udp.h
index d8ce937..6164d5c 100644
--- a/include/net/udp.h
+++ b/include/net/udp.h
@@ -472,11 +472,15 @@ static inline struct sk_buff *udp_rcv_segment(struct sock *sk,
 					      struct sk_buff *skb, bool ipv4)
 {
 	struct sk_buff *segs;
+	netdev_features_t features = NETIF_F_SG;
+
+	/* Avoid csum recalculation in skb_segment() */
+	features |= NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM;
 
 	/* the GSO CB lays after the UDP one, no need to save and restore any
 	 * CB fragment
 	 */
-	segs = __skb_gso_segment(skb, NETIF_F_SG, false);
+	segs = __skb_gso_segment(skb, features, false);
 	if (unlikely(IS_ERR_OR_NULL(segs))) {
 		int segs_nr = skb_shinfo(skb)->gso_segs;
 
-- 
1.9.1

Re: [PATCH net-next] udp: Avoid post-GRO UDP checksum recalculation

[Paolo Abeni]

On Thu, 2019-05-23 at 13:36 -0600, Sean Tranchetti wrote:
> Currently, when resegmenting an unexpected UDP GRO packet, the full UDP
> checksum will be calculated for every new SKB created by skb_segment()
> because the netdev features passed in by udp_rcv_segment() lack any
> information about checksum offload capabilities.
> 
> We have no need to perform this calculation again, as
>   1) The GRO implementation guarantees that any packets making it to the
>      udp_rcv_segment() function had correct checksums, and, more
>      importantly,
>   2) Upon the successful return of udp_rcv_segment(), we immediately pull
>      the UDP header off and either queue the segment to the socket or
>      hand it off to a new protocol handler. In either case, the checksum
>      is not needed.

I *think* there is a possible, even if unlikely, exception to the
above: if userspace has set the IP_CHECKSUM sockopt, recvmsg can later
try to access skb csum.

I think that setting NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM only if
!inet_get_convert_csum() would address the above,

Other than that LGTM, thanks for catching this!

Paolo

p.s. I suspect that with this patch GRO + resegmentation is notably
faster than the plain unaggregated path, do you have by chance any
related datapoint?

Re: [PATCH net-next] udp: Avoid post-GRO UDP checksum recalculation

[David Miller]

From: Sean Tranchetti <stranche@codeaurora.org>
Date: Thu, 23 May 2019 13:36:17 -0600

> @@ -472,11 +472,15 @@ static inline struct sk_buff *udp_rcv_segment(struct sock *sk,
>  					      struct sk_buff *skb, bool ipv4)
>  {
>  	struct sk_buff *segs;
> +	netdev_features_t features = NETIF_F_SG;

When you respin this based upon Paolo's feedback, please reverse christmas
tree these local variables.

Thank you.