Re: [PATCH] ifb: add multi-queue support

[Eric Dumazet <eric.dumazet@gmail.com>]

Changli Gao a écrit :
> On Fri, Nov 13, 2009 at 5:57 PM, Jarek Poplawski <jarkao2@gmail.com> wrote:
>> I'm not against your solution at all. It only needs more proof...
> 
> I know. :)
> 
>> You seem to forget the main networking paths now are just softirq, and
>> it's probably for some reason. If kernel threads are good enough, it
>> seems we should do more such changes.
>>
> 
> I find there is still a softirqd kernel thread for each online CPU,
> and these threads will be waked up if there are more SoftIRQs need to
> be done after restarting the SoftIRQ processing many times, to keep
> the whole system responsible. Did I miss sth.? And at the other side,
> real time branch just wants to make all the activities based on kernel
> threads, even ISR.
> 
> 


You seem to focus on stress loads, where we enter ksoftirqd more to get
more throughput (and cpu fairness), at latency expense.

But in normal situations, ksoftirq is not started at all.