Multiple TX queue implementation and Batching

Multiple TX queue implementation and Batching

[Krishna Kumar2]

Hi Dave,

Would it be possible to get a sneak preview of your multiple tx queue
implementation (along with a complete re-write of the entire tx path)?

I would like to see if batching would make sense with the new design,
or becomes redundant (or something to be done entirely at the drivers).

Thanks,

- KK

Re: Multiple TX queue implementation and Batching

[David Miller]

From: Krishna Kumar2 <krkumar2@in.ibm.com>
Date: Thu, 19 Jun 2008 13:17:53 +0530

> Would it be possible to get a sneak preview of your multiple tx queue
> implementation (along with a complete re-write of the entire tx path)?
> 
> I would like to see if batching would make sense with the new design,
> or becomes redundant (or something to be done entirely at the drivers).

Right now I'm just hacking drivers in preparation for this, so
that the transmit queueing is seperated from the location
in the driver where this actually occurs.

I'll post a set of tg3 driver patches once I'm done debugging
things.

Re: Multiple TX queue implementation and Batching

[Krishna Kumar2]

Hi Dave,

David Miller <davem@davemloft.net> wrote on 06/19/2008 02:40:39 PM:

> Right now I'm just hacking drivers in preparation for this, so
> that the transmit queueing is seperated from the location
> in the driver where this actually occurs.
>
> I'll post a set of tg3 driver patches once I'm done debugging
> things.

What about core networking changes? I think you were planning to
remove a lot of stuff (device busy checks, requeue's, etc?) and
also implement lockless code (sch_generic.c at the very least).
Are those ready for taking an initial look?

Thanks,

- KK

Re: Multiple TX queue implementation and Batching

[David Miller]

From: Krishna Kumar2 <krkumar2@in.ibm.com>
Date: Thu, 19 Jun 2008 14:47:52 +0530

> What about core networking changes? I think you were planning to
> remove a lot of stuff (device busy checks, requeue's, etc?) and
> also implement lockless code (sch_generic.c at the very least).
> Are those ready for taking an initial look?

We can't do that completely until every driver has it's TX path
converted, which isn't likely to happen so we'll have a compat
layer for a long time.

I haven't touched one line of generic code yet.  It isn't necessary
to test out all of these ideas.

Re: Multiple TX queue implementation and Batching

[David Miller]

Also, please don't be so impatient.

If I had testable code ready to publish, I would publish it.  Every
reply I send here is time away from debugging and hacking on my
changes.