[PATCH 0/3] Add support for rx flow hash configuration in a network device

[PATCH 0/3] Add support for rx flow hash configuration in a network device

[Santwona Behera]

Added new interface to ethtool to configure receive network flow 
classification. This patch adds support for flow distribution using 
hashing where the device can be configured with the fields in the 
receive packet to hash onto one of the multiple Rx rings.

This feature has 3 related patches:

1. An ethtool application patch.
2. A kernel patch for the ethtool interface.
3. A kernel patch for the niu driver that implements the ethtool config 
ioctl functions. NOTE: The niu driver already has rx flow hashing 
implemented. With these 3 patches, this feature is now configurable 
using ethtool.

Thanks!

--santwona

Re: [PATCH 0/3] Add support for rx flow hash configuration in a network device

[Andi Kleen]

Santwona Behera <Santwona.Behera@Sun.COM> writes:

> Added new interface to ethtool to configure receive network flow
> classification. This patch adds support for flow distribution using
> hashing where the device can be configured with the fields in the
> receive packet to hash onto one of the multiple Rx rings.
>
> This feature has 3 related patches:
>
> 1. An ethtool application patch.
> 2. A kernel patch for the ethtool interface.
> 3. A kernel patch for the niu driver that implements the ethtool
> config ioctl functions. NOTE: The niu driver already has rx flow
> hashing implemented. With these 3 patches, this feature is now
> configurable using ethtool.

Are you sure ethtool is the right interface for this?

Me thinks the kernel should just do the right thing automatically
for this, by distributing the flows over CPUs.

-Andi

Re: [PATCH 0/3] Add support for rx flow hash configuration in a network device

[David Miller]

From: Andi Kleen <andi@firstfloor.org>
Date: Sun, 29 Jun 2008 14:17:56 +0200

> Are you sure ethtool is the right interface for this?

Yes.

> Me thinks the kernel should just do the right thing automatically
> for this, by distributing the flows over CPUs.

How the work, and what kind of work, should be distributed across the
machine is the user's business.

Or maybe we should just get rid of all of those NUMA knobs we have
too?

We've discussed this before and Jeff Garzik and I already told
everyone that we want to make this configurable and ethtool is
the way to do it no matter how adhoc and device specific the
interfaces might end up being.

Please don't send submitters mixed messages like this after we've
told them to do things this way, it only leads to discouragement
for new contributors.

Re: [PATCH 0/3] Add support for rx flow hash configuration in a network device

[Andi Kleen]

> Please don't send submitters mixed messages like this after we've

Sorry must have missed the official bulletin on the latest party line.

> told them to do things this way, it only leads to discouragement
> for new contributors.

I suspect "adhoc and device specific interfaces" like you are
advocating will finally lead to discouragement of end users. 
But it is your choice.

-Andi

Re: [PATCH 0/3] Add support for rx flow hash configuration in a network device

[Roland Dreier]

 > We've discussed this before and Jeff Garzik and I already told
 > everyone that we want to make this configurable and ethtool is
 > the way to do it no matter how adhoc and device specific the
 > interfaces might end up being.

I think it was a few weeks ago that Jeff said:

  I know user interfaces are annoying because you have to think about 
  chips other than your own, but that's life.  Other hardware vendors have 
  to do it too.

  Letting each driver have a different user interface is /unfriendly/ to 
  both developers users.

I guess as long as you stick an "ethtool" in front of the different
interfaces it all becomes OK.

 - R.

Re: [PATCH 0/3] Add support for rx flow hash configuration in a network device

[David Miller]

From: Roland Dreier <rdreier@cisco.com>
Date: Mon, 30 Jun 2008 20:33:45 -0700

>  > We've discussed this before and Jeff Garzik and I already told
>  > everyone that we want to make this configurable and ethtool is
>  > the way to do it no matter how adhoc and device specific the
>  > interfaces might end up being.
> 
> I think it was a few weeks ago that Jeff said:
> 
>   I know user interfaces are annoying because you have to think about 
>   chips other than your own, but that's life.  Other hardware vendors have 
>   to do it too.
> 
>   Letting each driver have a different user interface is /unfriendly/ to 
>   both developers users.
> 
> I guess as long as you stick an "ethtool" in front of the different
> interfaces it all becomes OK.

This interface we're talking about is in fact not all that device
specific, and it's extensible so that we can describe different types
of flow hashing types if so needed to support other chips.

Currently Sun's Neptune is the only one that offers so many options,
and to be honest I expect other implementations to be subsets of what
Neptune can do.

Re: [PATCH 0/3] Add support for rx flow hash configuration in a network device

[David Miller]

From: Santwona Behera <Santwona.Behera@Sun.COM>
Date: Fri, 27 Jun 2008 17:06:16 -0700

> This feature has 3 related patches:
> 
> 1. An ethtool application patch.
> 2. A kernel patch for the ethtool interface.
> 3. A kernel patch for the niu driver that implements the ethtool config 
> ioctl functions. NOTE: The niu driver already has rx flow hashing 
> implemented. With these 3 patches, this feature is now configurable 
> using ethtool.

Thanks for doing this work Santwona.

I've applied patch 2 and 3 to my net-next-2.6 kernel tree.

I spoke with Jeff Garzik, and he will add your first patch
to the ethtool tree when he gets a chance.

Thanks again!