Re: [PATCH 00/14]: Killing qdisc->ops->requeue().

[Jarek Poplawski <jarkao2@gmail.com>]

On Tue, Oct 14, 2008 at 01:37:14PM -0700, Alexander Duyck wrote:
> Jarek Poplawski wrote:
>> On Tue, Oct 14, 2008 at 09:41:30AM -0700, Alexander Duyck wrote:
>> ...
>>> I think if anything it seems like you guys actually found the cpu
>>> performance issue a while back in the fact that the dev_requeue_skb was
>>> calling __netif_schedule when requeuing on a stopped queue.  That is the
>>> one piece I would say needs to be changed so that you only call
>>> __netif_schedule if the skb is not going to a stopped queue.
>>
>> BTW, since one of the other "dreams" of killing requeuing failed, I
>> think this proposal is worth checking, but David was concerned about
>> some issues with buggy drivers after __netif_schedule() removing.
>>
>> Jarek P.
>
> I would say not to remove it.  Just add a check to verify that the queue  
> the packet is bound for is not stopped prior to calling it.  If the  
> queue is stopped it should be rescheduled by the wake_queue call when  
> the device has cleared the queue.

Sure, I meant "after __netif_schedule() call removing". Anyway, there
were a discussion on something like this:

http://marc.info/?l=linux-netdev&m=122197536025956&w=2

> Unfortunately I am currently working on other projects so I don't really  
>  have the time to create and test a patch for this.  Hopefully my input  
> has proven useful.

Without this __netif_schedule() call we should expect at least higher
cpu use here while packets are xmitted to a stopped queue compared to
older kernels or the current one (especially with multi qdisc/class/
filter configs), so first we should know this requeuing from the top is
really necessary, and I didn't see anything convincing about this yet.

Thanks,
Jarek P.