Re: [patch] softirq-2.4.5-A1

[Andrea Arcangeli <andrea@suse.de>]

On Sun, May 27, 2001 at 09:08:51PM +0200, Ingo Molnar wrote:
> i took at look at your ksoftirq stuff yesterday, and i think it's
> completely unnecessery and adds serious overhead to softirq handling. The
> whole point of softirqs is to have maximum scalability and no
> serialization. Why did you add ksoftirqd, would you mind explaining it?

The only case ksoftirqd runs is when the stock kernel does the wrong
thing and potentially delays the softirq of 1/HZ. Nothing else.

When current kernel does the right thing ksoftirq cannot generate any
scalability problem and furthmore ksoftirqd is a per-cpu thing so if you
face a scalability problem with it that simply means you need to fix the
scheduler because then it means you would face a scalability issue as
well every time a tux thread calls schedule().

90% of the time ksoftirqd will never run, when it runs it means you want
to pay for a scheduler run to get it running. The price of the scheduler
is just the price for the logic that balance the softirq load in a fair
manner and without buggy latencies.

Andrea