Messages in this thread | | | Date | Thu, 23 Nov 2023 18:49:49 -0300 | From | Arnaldo Carvalho de Melo <> | Subject | Re: [RFC PATCH v1] perf parse-events: Make legacy events lower priority than sysfs/json |
| |
Em Thu, Nov 23, 2023 at 07:18:57AM -0800, Ian Rogers escreveu: > On Thu, Nov 23, 2023 at 6:37 AM Mark Rutland <mark.rutland@arm.com> wrote: > > On Wed, Nov 22, 2023 at 08:29:22PM -0800, Ian Rogers wrote: > > > This is a large behavioral change: > > > 1) the scope of the change means it should bake on linux-next and I > > > don't believe should be a 6.7-rc fix.
> > I'm happy for this to bake, but I do think it needs to be backported for the > > sake of users, especially given that it *restores* the old behaviour.
> It is going to change the behavior for a far larger set of users. I'm > also concerned that:
> ``` > $ perf list > ... > cpu-cycles OR cpu/cpu-cycles/ [Kernel PMU event] > ... > ```
> implies that cpu/cpu-cycles/ is a synonym for cpu-cycles, which is no > longer true (or pick another event from sysfs whose name is the same > as a legacy event). I'm not sure what a fix in perf list for this > would look like.
It is a mess, indeed, cpu-cycles should be equivalent to cpu/cpu-cycles/, map to the same HW PMU counter.
But by now I think we need to just reword the output of perf list to not equate those events, and instead provide a more informative output, if that is at all possible.
Something like:
Legacy events:
cpu-cycles: some nice explanation about the intent for this one (Ingo's probably).
And then:
PMU events: (and this is even less clear :-()
cpu/cpu-cycles/: other nice explanation about the intent for this one, if different, on this arch, for the "legacy" cpu-cycles one.
The original intent, that I called "Ingo's" was to try to somehow generalize some concepts (CPU cycles, for instante) so that we could get a rough idea that would allow us to somehow compare results using the tools over different architectures (and micro-arches, etc).
I think that with these generic names we're now in damage control mode: what matters is to keep what people got used to to continue to hold. And that is what I think is the main argument here.
And I really think we should just head people like Hector (perf power users) and provide a way to aggregate "cycles" over all cores, probably not in the default output, but in a really simple way, as he seems to want that and I would'n t be surprised that that may be something useful after all 8-).
So back to checking why this patch isn't working in some of the container arches I test this whole shebang, sigh.
- Arnaldo
| |