Ingo Molnar wrote:
>>below i've also attached a softirq.c patch against 2.6.8-rc2 that does
>>unconditional deferring. (this patch is of course not intended to be
>>merged upstream as-is, since normally we want to process softirqs
>>right after the irq context.)
> 
> 
> i've got a more complete patch against vanilla 2.6.8-rc2:
> 
>  http://redhat.com/~mingo/voluntary-preempt/defer-softirqs-2.6.8-rc2-A2
> 
> which introduces the following tunable:
> 
>     /proc/sys/kernel/defer_softirqs  (default: 0)
> 
> this, if enabled, causes all softirqs to be processed within ksoftirqd,
> and it also breaks out of the softirq loop if preemption of ksoftirqd
> has been triggered by a higher-prio task.
> 
> I've also added this additional break-out to the -H6 patch of
> voluntary-preempt:
> 
>   http://redhat.com/~mingo/voluntary-preempt/voluntary-preempt-2.6.8-rc2-H6
> 
> it's enabled by default.
> 

What do you think about deferring softirqs just while in critical
sections?

I'm not sure how well this works, and it is CONFIG_PREEMPT only
but in theory it should prevent unbounded softirqs while under
locks without taking the performance hit of doing the context
switch.