On Tue, 9 May 2006, Lee Revell wrote:
On Tue, 2006-05-09 at 15:45 -0700, Kjetil S. Matheussen wrote:
Hmm, it sounds like a solution could be to separate timers that just
wake up a process from ones that do actual work and run them in separate
kernel threads.
I don't understand why you want that. To me (which
knows about nothing about how the kernel works), the solution to
the problem is crystal clear: The softirq timer needs to have the highest
priority, and the only thing the sofirq timer threads does is to
make sure threads that should be woken up are put into some kernel
schedule queue somehow. I don't understand why the softirq timer threads
should cause any latency problems, what does it do except scheduling
waiting threads?
Well as you have seen the softirq timer thread apparently does route
cache flushing and a bunch of other things we don't want it to do. I
was under the impression these were already done by separate threads.
I think we are in agreement.
Good. :-)
I have to double check how it works in -rt
- I've been running mainline lately which still has a single ksoftirqd
process.
I think the ksoftirqd processes have the same problem. At least for
das_watchdog, it must be set to 99 for the watchdog to work. There is
even code to check that it/they have priority 99.
