Offloading features in VLAN interfaces

Offloading features in VLAN interfaces

[Olivier Crameri]

Hi,

after running some experiments, we realised that using VLAN support in
Linux caused some overhead. It turned out to be that the offloading
features of our NIC (tso, sg, checksum offloading) were not beeing
used.

After digging a little bit in the code, we realised that the VLAN code
did not set the features parameter of the net_device structure for the
virtual interface. Attached is a patch with a suggested solution to
this problem.

Please let us know if this is an acceptable patch.

Thanks,
Olivier Crameri, HP Labs

Re: Offloading features in VLAN interfaces

[Olivier Crameri]

[-- Attachment #1: Type: text/plain, Size: 698 bytes --]

Same thing but with the patch this time.

Sorry about that,
Olivier Crameri

On 9/29/06, Olivier Crameri <olivier.crameri@gmail.com> wrote:
> Hi,
>
> after running some experiments, we realised that using VLAN support in
> Linux caused some overhead. It turned out to be that the offloading
> features of our NIC (tso, sg, checksum offloading) were not beeing
> used.
>
> After digging a little bit in the code, we realised that the VLAN code
> did not set the features parameter of the net_device structure for the
> virtual interface. Attached is a patch with a suggested solution to
> this problem.
>
> Please let us know if this is an acceptable patch.
>
> Thanks,
> Olivier Crameri, HP Labs
>

[-- Attachment #2: vlan.patch --]
[-- Type: application/octet-stream, Size: 1780 bytes --]

--- net/8021q/vlan.c.orig	2006-09-29 16:20:01.000000000 -0700
+++ net/8021q/vlan.c	2006-09-29 17:16:31.000000000 -0700
@@ -10,7 +10,9 @@
  *              Fix for packet capture - Nick Eggleston <nick@dccinc.com>;
  *		Add HW acceleration hooks - David S. Miller <davem@redhat.com>;
  *		Correct all the locking - David S. Miller <davem@redhat.com>;
- *		Use hash table for VLAN groups - David S. Miller <davem@redhat.com>
+ *		Use hash table for VLAN groups - David S. Miller <davem@redhat.com>;
+ *		Advertise real device's features - Olivier Crameri, Hewlett-Packard 
+ *		<olivier.crameri@epfl.ch>
  *
  *		This program is free software; you can redistribute it and/or
  *		modify it under the terms of the GNU General Public License
@@ -512,6 +514,15 @@ static struct net_device *register_vlan_
 	}
 	new_dev->hard_header_parse = real_dev->hard_header_parse;
 
+	/* Advertise the same offloading features than the real device */
+	new_dev->features = real_dev->features & 
+				(NETIF_F_IP_CSUM |
+				 NETIF_F_NO_CSUM |
+				 NETIF_F_HW_CSUM |
+				 NETIF_F_SG |
+				 NETIF_F_FRAGLIST |
+				 NETIF_F_TSO );
+	
 	VLAN_DEV_INFO(new_dev)->vlan_id = VLAN_ID; /* 1 through VLAN_VID_MASK */
 	VLAN_DEV_INFO(new_dev)->real_dev = real_dev;
 	VLAN_DEV_INFO(new_dev)->dent = NULL;
@@ -663,6 +674,23 @@ static int vlan_device_event(struct noti
 				break;
 		}
 		break;
+	
+	case NETDEV_FEAT_CHANGE:
+		/* Propagate feature changes to VLAN devices */
+		for (i = 0; i < VLAN_GROUP_ARRAY_LEN; i++) {
+			vlandev = grp->vlan_devices[i];
+			if (!vlandev)
+				continue;
+			vlandev->features = dev->features & ( 
+					NETIF_F_IP_CSUM |
+				 	NETIF_F_NO_CSUM |
+				 	NETIF_F_HW_CSUM |
+				 	NETIF_F_SG |
+				 	NETIF_F_FRAGLIST |
+				 	NETIF_F_TSO );
+		}
+		break;
+
 	};
 
 out:

Re: Offloading features in VLAN interfaces

[Stephen Hemminger]

On Fri, 29 Sep 2006 17:25:35 -0700
"Olivier Crameri" <olivier.crameri@gmail.com> wrote:

> Same thing but with the patch this time.
> 
> Sorry about that,
> Olivier Crameri
> 
> On 9/29/06, Olivier Crameri <olivier.crameri@gmail.com> wrote:
> > Hi,
> >
> > after running some experiments, we realised that using VLAN support in
> > Linux caused some overhead. It turned out to be that the offloading
> > features of our NIC (tso, sg, checksum offloading) were not beeing
> > used.
> >
> > After digging a little bit in the code, we realised that the VLAN code
> > did not set the features parameter of the net_device structure for the
> > virtual interface. Attached is a patch with a suggested solution to
> > this problem.
> >
> > Please let us know if this is an acceptable patch.
> >
> > Thanks,
> > Olivier Crameri, HP Labs
> >

Patch needs to include "Developer's Certificate of Origin"
see Documentation/SubmittingPatches section 11.

Patch should not be sent as an attachment (if at all possible).

Skip the marque comment. It is standard practice now to use the revision
control system change log for attribution.

Please handle all the current offload features including GSO.

Re: Offloading features in VLAN interfaces

[Ben Greear]

Olivier Crameri wrote:
> Same thing but with the patch this time.
Since the VLAN device's features may also change in the handler, 
shouldn't we check and generate a feature-change
event for the VLAN device(s) as well?

Ben


-- 
Ben Greear <greearb@candelatech.com> 
Candela Technologies Inc  http://www.candelatech.com