[PATCH net-next v1 1/1] net: openvswitch: remove unnecessary vlan init in key_extract

[PATCH net-next v1 1/1] net: openvswitch: remove unnecessary vlan init in key_extract

[Eddy Tao]

Redefine clear_vlan to initialize one struct vlan_head
Define   clear_vlans to initialize key.eth.vlan and key.eth.cvlan
Calls the revised functions accurately

Reasoning:

For vlan packet, current code calls clear_vlan unnecessarily,
since parse_vlan sets key->eth.vlan and key->eth.cvlan correctly.
Only special case where return value <=0 needs inialization
certail key.eth.vlan or key.eth.cvlan specifically.

For none-vlan case, parse_vlan returns on the first parse_vlan_tag
which returns 0, in this case, calls clear_vlan

For MAC_PROTO_NONE, logic is intact after this revision

Signed-off-by: Eddy Tao <taoyuan_eddy@hotmail.com>
---
 net/openvswitch/flow.c | 24 ++++++++++++++++--------
 1 file changed, 16 insertions(+), 8 deletions(-)

diff --git a/net/openvswitch/flow.c b/net/openvswitch/flow.c
index e20d1a973417..30a90597cab6 100644
--- a/net/openvswitch/flow.c
+++ b/net/openvswitch/flow.c
@@ -480,12 +480,16 @@ static int parse_vlan_tag(struct sk_buff *skb, struct vlan_head *key_vh,
 	return 1;
 }
 
-static void clear_vlan(struct sw_flow_key *key)
+static inline void clear_vlan(struct vlan_head *vlan)
 {
-	key->eth.vlan.tci = 0;
-	key->eth.vlan.tpid = 0;
-	key->eth.cvlan.tci = 0;
-	key->eth.cvlan.tpid = 0;
+	vlan->tci = 0;
+	vlan->tpid = 0;
+}
+
+static inline void clear_vlans(struct sw_flow_key *key)
+{
+	clear_vlan(&key->eth.vlan);
+	clear_vlan(&key->eth.cvlan);
 }
 
 static int parse_vlan(struct sk_buff *skb, struct sw_flow_key *key)
@@ -498,14 +502,18 @@ static int parse_vlan(struct sk_buff *skb, struct sw_flow_key *key)
 	} else {
 		/* Parse outer vlan tag in the non-accelerated case. */
 		res = parse_vlan_tag(skb, &key->eth.vlan, true);
-		if (res <= 0)
+		if (res <= 0) {
+			clear_vlans(key);
 			return res;
+		}
 	}
 
 	/* Parse inner vlan tag. */
 	res = parse_vlan_tag(skb, &key->eth.cvlan, false);
-	if (res <= 0)
+	if (res <= 0) {
+		clear_vlan(&key->eth.cvlan);
 		return res;
+	}
 
 	return 0;
 }
@@ -918,8 +926,8 @@ static int key_extract(struct sk_buff *skb, struct sw_flow_key *key)
 	skb_reset_mac_header(skb);
 
 	/* Link layer. */
-	clear_vlan(key);
 	if (ovs_key_mac_proto(key) == MAC_PROTO_NONE) {
+		clear_vlans(key);
 		if (unlikely(eth_type_vlan(skb->protocol)))
 			return -EINVAL;
 
-- 
2.27.0

Re: [PATCH net-next v1 1/1] net: openvswitch: remove unnecessary vlan init in key_extract

[Simon Horman]

On Tue, Feb 07, 2023 at 12:31:33PM +0800, Eddy Tao wrote:
> Redefine clear_vlan to initialize one struct vlan_head
> Define   clear_vlans to initialize key.eth.vlan and key.eth.cvlan
> Calls the revised functions accurately
> 
> Reasoning:
> 
> For vlan packet, current code calls clear_vlan unnecessarily,
> since parse_vlan sets key->eth.vlan and key->eth.cvlan correctly.
> Only special case where return value <=0 needs inialization
> certail key.eth.vlan or key.eth.cvlan specifically.
> 
> For none-vlan case, parse_vlan returns on the first parse_vlan_tag
> which returns 0, in this case, calls clear_vlan
> 
> For MAC_PROTO_NONE, logic is intact after this revision
> 
> Signed-off-by: Eddy Tao <taoyuan_eddy@hotmail.com>

This seems like a complex, and perhaps error-prone, way to avoid
writing a few bytes. I do tend to think the extra code complexity
is not worth it without some performance justification.

OTOH, I think that perhaps a better question might be: do the bytes need to
be cleared under any circumstances? I suspect key is discarded when an
error occurs.

> ---
>  net/openvswitch/flow.c | 24 ++++++++++++++++--------
>  1 file changed, 16 insertions(+), 8 deletions(-)
> 
> diff --git a/net/openvswitch/flow.c b/net/openvswitch/flow.c
> index e20d1a973417..30a90597cab6 100644
> --- a/net/openvswitch/flow.c
> +++ b/net/openvswitch/flow.c
> @@ -480,12 +480,16 @@ static int parse_vlan_tag(struct sk_buff *skb, struct vlan_head *key_vh,
>  	return 1;
>  }
>  
> -static void clear_vlan(struct sw_flow_key *key)
> +static inline void clear_vlan(struct vlan_head *vlan)
>  {
> -	key->eth.vlan.tci = 0;
> -	key->eth.vlan.tpid = 0;
> -	key->eth.cvlan.tci = 0;
> -	key->eth.cvlan.tpid = 0;
> +	vlan->tci = 0;
> +	vlan->tpid = 0;
> +}
> +
> +static inline void clear_vlans(struct sw_flow_key *key)
> +{
> +	clear_vlan(&key->eth.vlan);
> +	clear_vlan(&key->eth.cvlan);
>  }

This is a nice cleanup, IMHO :)
>  
>  static int parse_vlan(struct sk_buff *skb, struct sw_flow_key *key)
> @@ -498,14 +502,18 @@ static int parse_vlan(struct sk_buff *skb, struct sw_flow_key *key)
>  	} else {
>  		/* Parse outer vlan tag in the non-accelerated case. */
>  		res = parse_vlan_tag(skb, &key->eth.vlan, true);
> -		if (res <= 0)
> +		if (res <= 0) {
> +			clear_vlans(key);

I think this makes more sense in the caller.

>  			return res;
> +		}
>  	}
>  
>  	/* Parse inner vlan tag. */
>  	res = parse_vlan_tag(skb, &key->eth.cvlan, false);
> -	if (res <= 0)
> +	if (res <= 0) {
> +		clear_vlan(&key->eth.cvlan);
>  		return res;
> +	}
>  
>  	return 0;
>  }
> @@ -918,8 +926,8 @@ static int key_extract(struct sk_buff *skb, struct sw_flow_key *key)
>  	skb_reset_mac_header(skb);
>  
>  	/* Link layer. */
> -	clear_vlan(key);
>  	if (ovs_key_mac_proto(key) == MAC_PROTO_NONE) {
> +		clear_vlans(key);
>  		if (unlikely(eth_type_vlan(skb->protocol)))
>  			return -EINVAL;

I think you missed the following case further down:

		if (unlikely(key->eth.type == htons(0)))
			return -ENOMEM;

Re: [PATCH net-next v1 1/1] net: openvswitch: remove unnecessary vlan init in key_extract

[Eddy Tao]

Hi, Simon:

     Thanks for looking into this. by looking into the compiled 
instructions, i am convinced the patch is logically correct but not help 
in performance, and it adds complexity,

Below function is actually optimized to one movq instruction with no 
special -O settings on my centos8, which seems good enough for me

static void clear_vlan(struct sw_flow_key *key)
{
     key->eth.vlan.tci = 0;
     key->eth.vlan.tpid = 0;
     key->eth.cvlan.tci = 0;
     key->eth.cvlan.tpid = 0;
}

optimized to movq   $0x0,0x158(%rbx)

I am withdrawing this patch

About your comment below, since it is after the parse_vlan handling, the 
code looks fine and does not affect the logic of the patch

 >>I think you missed the following case further down: >>>>if 
(unlikely(key->eth.type == htons(0))) Have a great day

eddy