RFC softirq: Per vector threading

So this is a first shot to implement what Linus suggested. To summarize: when a softirq vector is stormed and needs more time than what IRQ tail can offer, the whole softirq processing is offloaded to ksoftirqd. But this has an impact on other softirq vectors that are then subject to scheduler latencies.

So the softirqs time limits is now per vector and only the vectors that get stormed are offloaded to a thread (workqueue).

This is in a very Proof of concept state. It doesn't even boot successfully once in a while. So I'll do more debugging tomorrow (today in fact) but you get the big picture.

It probably won't come free given the clock reads around softirq callbacks.

git://git.kernel.org/pub/scm/linux/kernel/git/frederic/linux-dynticks.git
softirq/poc

HEAD: 0e982634115283710d0801048e5a316def26f31d

Thanks,
Frederic
---

Frederic Weisbecker (2):
softirq: Account time and iteration stats per vector
softirq: Per vector thread deferment

 kernel/softirq.c | 123 +++++++++++++++++++++++++++++++++++++++++++++++++++----
 1 file changed, 114 insertions(+), 9 deletions(-)