Re: [PATCH RFC] remove jump_label optimization for perf sched events

[Gleb Natapov <gleb@redhat.com>]

On Thu, Nov 17, 2011 at 01:49:19PM +0100, Peter Zijlstra wrote:
> On Thu, 2011-11-17 at 14:30 +0200, Gleb Natapov wrote:
> > jump_lable patching is very expensive operation that involves pausing all
> > cpus. The patching of perf_sched_events jump_label is easily controllable
> > from userspace by unprivileged user. When user runs loop like this
> > "while true; do perf stat -e cycles true; done" the performance of my
> > test application that just increments a counter for one second drops by
> > 4%. This is on a 16 cpu box with my test application using only one of
> > them. An impact on a real server doing real work will be much worse.
> > Performance of KVM PMU drops nearly 50% due to jump_lable for "perf
> > record" since KVM PMU implementation creates and destroys perf event
> > frequently.
> 
> Ideally we'd fix text_poke to not use stop_machine() we know how to, but
> we haven't had the green light from Intel/AMD yet.
> 
> Rostedt was going to implement it anyway and see if anything breaks.
> 
Hmm interesting.

> Also, virt might be able to pull something smart on text_poke() dunno.
> 
The problem with virt is not text_poke() in a guest, but the one in a
host. The guest I am testing with has only one cpu. Basically creating
fist perf event/destroying last perf event is very expensive currently
and when "perf record" is running in a guest this happens a lot in a
host.

> That said, I'd much rather throttle this particular jump label than
> remove it altogether, some people really don't like all this scheduler
> hot path crap.
What about moving perf_event_task_sched() to sched_(in|out)_preempt_notifiers?
preempt notifiers checking is already on the scheduler hot path, so no
additional overhead for perf case.

--
			Gleb.