Re: [PATCH] perf cgroups: Don't rotate events for cgroups unnecessarily

[Ian Rogers]

On Fri, Jun 14, 2019 at 2:43 PM Liang, Kan <kan.liang@xxxxxxxxxxxxxxx> wrote:
>
>
>
> On 6/14/2019 3:10 PM, Stephane Eranian wrote:
> > On Thu, Jun 13, 2019 at 9:13 AM Liang, Kan <kan.liang@xxxxxxxxxxxxxxx> wrote:
> >>
> >>
> >>
> >> On 6/1/2019 4:27 AM, Ian Rogers wrote:
> >>> Currently perf_rotate_context assumes that if the context's nr_events !=
> >>> nr_active a rotation is necessary for perf event multiplexing. With
> >>> cgroups, nr_events is the total count of events for all cgroups and
> >>> nr_active will not include events in a cgroup other than the current
> >>> task's. This makes rotation appear necessary for cgroups when it is not.
> >>>
> >>> Add a perf_event_context flag that is set when rotation is necessary.
> >>> Clear the flag during sched_out and set it when a flexible sched_in
> >>> fails due to resources.
> >>>
> >>> Signed-off-by: Ian Rogers <irogers@xxxxxxxxxx>
> >>> ---
> >>>    include/linux/perf_event.h |  5 +++++
> >>>    kernel/events/core.c       | 42 +++++++++++++++++++++++---------------
> >>>    2 files changed, 30 insertions(+), 17 deletions(-)
> >>>
> >>> diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h
> >>> index 15a82ff0aefe..7ab6c251aa3d 100644
> >>> --- a/include/linux/perf_event.h
> >>> +++ b/include/linux/perf_event.h
> >>> @@ -747,6 +747,11 @@ struct perf_event_context {
> >>>        int                             nr_stat;
> >>>        int                             nr_freq;
> >>>        int                             rotate_disable;
> >>> +     /*
> >>> +      * Set when nr_events != nr_active, except tolerant to events not
> >>> +      * needing to be active due to scheduling constraints, such as cgroups.
> >>> +      */
> >>> +     int                             rotate_necessary;
> >>
> >> It looks like the rotate_necessary is only useful for cgroup and cpuctx.
> >> Why not move it to struct perf_cpu_context and under #ifdef
> >> CONFIG_CGROUP_PERF?
> >> And rename it cgrp_rotate_necessary?
> >>
> > I am not sure I see the point here. What I'd like to see is a uniform
> > signal for rotation needed in per-task, per-cpu or per-cgroup modes > Ian's patch does that. It does make it a lot more efficient in cgroup
> > mode, by avoiding unnecessary rotations, and does not alter/improve
> > on any of the other two modes.
>
> I just thought if it is only used by cgroup, it may be better to move it
> under #ifdef CONFIG_CGROUP_PERF. For users who don't care about cgroup,
> it may save some space.
> But if it's designed as a uniform signal, this is OK.
>
> Thanks,
> Kan

Thanks Kan,
I think you have a wider valid point that by specializing to the CPU
or task context the code could avoid work. Similarly, adding the state
to perf_event_context is limited overhead for non-cgroup builds that
could be optimized away by going back to "nr_active == nr_events". I
agree with Stephane that we should try to keep the code uniform for
the task and CPU case for consistency - and this is the state in the
patch.

Are there any blockers to getting this merged?

Thanks,
Ian


> >
> >> Thanks,
> >> Kan
> >>
> >>>        refcount_t                      refcount;
> >>>        struct task_struct              *task;
> >>>
> >>> diff --git a/kernel/events/core.c b/kernel/events/core.c
> >>> index abbd4b3b96c2..41ae424b9f1d 100644
> >>> --- a/kernel/events/core.c
> >>> +++ b/kernel/events/core.c
> >>> @@ -2952,6 +2952,12 @@ static void ctx_sched_out(struct perf_event_context *ctx,
> >>>        if (!ctx->nr_active || !(is_active & EVENT_ALL))
> >>>                return;
> >>>
> >>> +     /*
> >>> +      * If we had been multiplexing, no rotations are necessary now no events
> >>> +      * are active.
> >>> +      */
> >>> +     ctx->rotate_necessary = 0;
> >>> +
> >>>        perf_pmu_disable(ctx->pmu);
> >>>        if (is_active & EVENT_PINNED) {
> >>>                list_for_each_entry_safe(event, tmp, &ctx->pinned_active, active_list)
> >>> @@ -3325,6 +3331,15 @@ static int flexible_sched_in(struct perf_event *event, void *data)
> >>>                        sid->can_add_hw = 0;
> >>>        }
> >>>
> >>> +     /*
> >>> +      * If the group wasn't scheduled then set that multiplexing is necessary
> >>> +      * for the context. Note, this won't be set if the event wasn't
> >>> +      * scheduled due to event_filter_match failing due to the earlier
> >>> +      * return.
> >>> +      */
> >>> +     if (event->state == PERF_EVENT_STATE_INACTIVE)
> >>> +             sid->ctx->rotate_necessary = 1;
> >>> +
> >>>        return 0;
> >>>    }
> >>>
> >>> @@ -3690,24 +3705,17 @@ ctx_first_active(struct perf_event_context *ctx)
> >>>    static bool perf_rotate_context(struct perf_cpu_context *cpuctx)
> >>>    {
> >>>        struct perf_event *cpu_event = NULL, *task_event = NULL;
> >>> -     bool cpu_rotate = false, task_rotate = false;
> >>> -     struct perf_event_context *ctx = NULL;
> >>> +     struct perf_event_context *task_ctx = NULL;
> >>> +     int cpu_rotate, task_rotate;
> >>>
> >>>        /*
> >>>         * Since we run this from IRQ context, nobody can install new
> >>>         * events, thus the event count values are stable.
> >>>         */
> >>>
> >>> -     if (cpuctx->ctx.nr_events) {
> >>> -             if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
> >>> -                     cpu_rotate = true;
> >>> -     }
> >>> -
> >>> -     ctx = cpuctx->task_ctx;
> >>> -     if (ctx && ctx->nr_events) {
> >>> -             if (ctx->nr_events != ctx->nr_active)
> >>> -                     task_rotate = true;
> >>> -     }
> >>> +     cpu_rotate = cpuctx->ctx.rotate_necessary;
> >>> +     task_ctx = cpuctx->task_ctx;
> >>> +     task_rotate = task_ctx ? task_ctx->rotate_necessary : 0;
> >>>
> >>>        if (!(cpu_rotate || task_rotate))
> >>>                return false;
> >>> @@ -3716,7 +3724,7 @@ static bool perf_rotate_context(struct perf_cpu_context *cpuctx)
> >>>        perf_pmu_disable(cpuctx->ctx.pmu);
> >>>
> >>>        if (task_rotate)
> >>> -             task_event = ctx_first_active(ctx);
> >>> +             task_event = ctx_first_active(task_ctx);
> >>>        if (cpu_rotate)
> >>>                cpu_event = ctx_first_active(&cpuctx->ctx);
> >>>
> >>> @@ -3724,17 +3732,17 @@ static bool perf_rotate_context(struct perf_cpu_context *cpuctx)
> >>>         * As per the order given at ctx_resched() first 'pop' task flexible
> >>>         * and then, if needed CPU flexible.
> >>>         */
> >>> -     if (task_event || (ctx && cpu_event))
> >>> -             ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE);
> >>> +     if (task_event || (task_ctx && cpu_event))
> >>> +             ctx_sched_out(task_ctx, cpuctx, EVENT_FLEXIBLE);
> >>>        if (cpu_event)
> >>>                cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
> >>>
> >>>        if (task_event)
> >>> -             rotate_ctx(ctx, task_event);
> >>> +             rotate_ctx(task_ctx, task_event);
> >>>        if (cpu_event)
> >>>                rotate_ctx(&cpuctx->ctx, cpu_event);
> >>>
> >>> -     perf_event_sched_in(cpuctx, ctx, current);
> >>> +     perf_event_sched_in(cpuctx, task_ctx, current);
> >>>
> >>>        perf_pmu_enable(cpuctx->ctx.pmu);
> >>>        perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
> >>>