Re: [PATCH] Re: softirq buggy

[Andrea Arcangeli <andrea@suse.de>]

[-- Attachment #1: Type: text/plain, Size: 2730 bytes --]

On Mon, Apr 09, 2001 at 05:26:27PM +0200, Manfred Spraul wrote:
> The return path to user space checks for pending softirqs. A delay of

And it breaks the loop too if new softirq events become pending again in
background.

> 1/HZ is only possible if the cpu loops in kernel space without returning
> to user space - and the functions that can loop check

It is also possible when new events are posted and I think it makes
sense to scale the softirq load with the scheduler when it is flooding.

Theoretically one could move the _whole_ softirq load into the ksoftirqd, but
that would increase the latency too much and I think it is better to use it
only as a fallback when we have to giveup but we still would like to keep
processing the softirq load so we let the scheduler to choose if we still can
do that or if we should giveup on the softirq.

> Is a full thread really necessary? Just setting 'need_resched' should be
> enough, schedule() checks for pending softirqs.

If you abuse need_resched then you can starve userspace again, if you are ok
to starve userspace indefinitely then it is more efficient to keep looping
forever into do_softirq as far as new events are posted in background instead
of exiting do_softirq and waiting the scheduler to kickin again.

> And do you have a rough idea how often that new thread is scheduled
> under load?

The scheduling is not as heavy as with tasks, it's a kernel thread
so the tlb isn't touched. However yes it will generate some overhead
with schedule() compared to just waiting the 1/HZ but letting the scheduler to
understand when the softirq should keep running instead of another task is
supposed to be a feature. I run a netpipe run with an alpha SMP as receiver
with the ksoftirqd patch and then without and the numbers didn't changed at all
even if the ksofitrqd was often running (1/2% of the load of the machine).

> Btw, you don't schedule the ksoftirqd thread if do_softirq() returns
> from the 'if(in_interrupt())' check.

That's not necessary and it's intentional, such check will be passed in the
last do_softirq executed before returning to userspace or kernel normal
context, the reason of such check is only to avoid recursing too much on the
stack during nested irqs.

> I assume that this is the most common case of delayed softirq
> processing:
> 
> ; in process context
> spin_lock_bh();
> ; hw interrupt arrives
> ; do_softirq returns immediately
> spin_unlock_bh();

This is yet another case and it's handled before returning to userspace so the
latency should still be pretty small (and there would be no singificant
advantage and almost certainly only a performance drop in waking up ksoftirqd
from the `do_softirq returns immediatly' line).

Andrea

[-- Attachment #2: ksofitrqd.png --]
[-- Type: image/png, Size: 3940 bytes --]