Messages in this thread | | | Date | Thu, 17 Nov 2011 16:12:06 +0200 | From | Avi Kivity <> | Subject | Re: [PATCH RFC] remove jump_label optimization for perf sched events |
| |
On 11/17/2011 03:47 PM, Peter Zijlstra wrote: > On Thu, 2011-11-17 at 15:24 +0200, Avi Kivity wrote: > > On 11/17/2011 03:10 PM, Peter Zijlstra wrote: > > > On Thu, 2011-11-17 at 15:00 +0200, Gleb Natapov wrote: > > > > > > > > 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. > > > > > > Same problem really, some people complain about the overhead of preempt > > > notifiers, also not all kernels have those in. > > > > We could combine the two, sort-circuit preempt notifiers with jump > > labels if empty && not much activity on them. > > Jump-labels are still more efficient, also I don't much like preempt > notifiers. > > > > Futhermore I loathe notifier lists because they obscure wtf is done. > > > > That's life in a general purpose kernel, if everyone gets their hook in > > to keep their code clean, the scheduler will bloat. > > Uhm, no. The bloat isn't different, the only difference is you can > actually see it. So I very much prefer direct hooks. > > > An advantage of preempt notifiers is that you can make the perf code > > modular. > > Yeah, and you know I loathe modules even more.
Is there something you like?
-- error compiling committee.c: too many arguments to function
| |