[PATCH] netdev: simple_tx_hash shouldn't hash inside fragments

[PATCH] netdev: simple_tx_hash shouldn't hash inside fragments

[Alexander Duyck]

Currently simple_tx_hash is hashing inside of udp fragments.  As a result
packets are getting getting sent to all queues when they shouldn't be.
This causes a serious performance regression which can be seen by sending
UDP frames larger than mtu on multiqueue devices.  This change will make
it so that fragments are hashed only as IP datagrams w/o any protocol
information.

Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
---

 net/core/dev.c |    6 ++++--
 1 files changed, 4 insertions(+), 2 deletions(-)

diff --git a/net/core/dev.c b/net/core/dev.c
index 60c51f7..5d9fa1e 100644
--- a/net/core/dev.c
+++ b/net/core/dev.c
@@ -122,6 +122,7 @@
 #include <linux/if_arp.h>
 #include <linux/if_vlan.h>
 #include <linux/ip.h>
+#include <net/ip.h>
 #include <linux/ipv6.h>
 #include <linux/in.h>
 #include <linux/jhash.h>
@@ -1667,7 +1668,7 @@ static u16 simple_tx_hash(struct net_device *dev, struct sk_buff *skb)
 {
 	u32 addr1, addr2, ports;
 	u32 hash, ihl;
-	u8 ip_proto;
+	u8 ip_proto = 0;
 
 	if (unlikely(!simple_tx_hashrnd_initialized)) {
 		get_random_bytes(&simple_tx_hashrnd, 4);
@@ -1676,7 +1677,8 @@ static u16 simple_tx_hash(struct net_device *dev, struct sk_buff *skb)
 
 	switch (skb->protocol) {
 	case __constant_htons(ETH_P_IP):
-		ip_proto = ip_hdr(skb)->protocol;
+		if (!(ip_hdr(skb)->frag_off & htons(IP_MF | IP_OFFSET)))
+			ip_proto = ip_hdr(skb)->protocol;
 		addr1 = ip_hdr(skb)->saddr;
 		addr2 = ip_hdr(skb)->daddr;
 		ihl = ip_hdr(skb)->ihl;

Re: [PATCH] netdev: simple_tx_hash shouldn't hash inside fragments

[David Miller]

From: Alexander Duyck <alexander.h.duyck@intel.com>
Date: Thu, 18 Sep 2008 17:43:37 -0700

> Currently simple_tx_hash is hashing inside of udp fragments.  As a result
> packets are getting getting sent to all queues when they shouldn't be.
> This causes a serious performance regression which can be seen by sending
> UDP frames larger than mtu on multiqueue devices.  This change will make
> it so that fragments are hashed only as IP datagrams w/o any protocol
> information.
> 
> Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>

As we discussed on IRC this fixes a pretty serious performance regression
compared to previous releases, so I added this to net-2.6 and will push
to Linus.

Thanks!