Re: [PATCH v2 2/6] perf/core: Introduce percpu perf_cgroup

[Namhyung Kim <namhyung@kernel.org>]

On Tue, Mar 22, 2022 at 5:10 AM Chengming Zhou
<zhouchengming@bytedance.com> wrote:
>
> Although we don't have incosistency problem any more, we can
> have other problem like:
>
> CPU1                                    CPU2
> (in context_switch)                     (attach running task)
>                                         prev->cgroups = cgrp2
> perf_cgroup_sched_switch(prev, next)
>         cgrp2 == cgrp2 is True
>
> If perf_cgroup of prev task changes from cgrp1 to cgrp2,
> perf_cgroup_sched_switch() will skip perf_cgroup_switch(),
> so the CPU would still schedule the cgrp1 events, but we should
> schedule the cgrp2 events.

Ah ok, now I see the problem in changing prev->cgroup too.

>
> The reason of this problem is that we shouldn't use the changeable
> prev->cgroups to decide whether skip perf_cgroup_switch().
>
> This patch introduces a percpu perf_cgroup to cache the perf_cgroup
> that scheduled in cpuctxes, which later used to compare with the
> perf_cgroup of next task to decide whether skip perf_cgroup_switch().
>
> Since the perf_cgroup_switch() can be called after the context switch,
> the cgroup events might be scheduled already. So we put the comparison
> of perf_cgroups in perf_cgroup_switch(), and delete the unused function
> perf_cgroup_sched_switch().
>
> We must clear the percpu perf_cgroup cache when the last cgroup event
> disabled.
>
> Fixes: a8d757ef076f ("perf events: Fix slow and broken cgroup context switch code")
> Signed-off-by: Chengming Zhou <zhouchengming@bytedance.com>
> ---
>  kernel/events/core.c | 63 ++++++++++++++++----------------------------
>  1 file changed, 22 insertions(+), 41 deletions(-)
>
> diff --git a/kernel/events/core.c b/kernel/events/core.c
> index 8b5cf2aedfe6..848a3bfa9513 100644
> --- a/kernel/events/core.c
> +++ b/kernel/events/core.c
> @@ -826,6 +826,7 @@ perf_cgroup_set_timestamp(struct task_struct *task,
>         }
>  }
>
> +static DEFINE_PER_CPU(struct perf_cgroup *, cpu_perf_cgroup);
>  static DEFINE_PER_CPU(struct list_head, cgrp_cpuctx_list);
>
>  /*
> @@ -833,6 +834,7 @@ static DEFINE_PER_CPU(struct list_head, cgrp_cpuctx_list);
>   */
>  static void perf_cgroup_switch(struct task_struct *task)
>  {
> +       struct perf_cgroup *cgrp;
>         struct perf_cpu_context *cpuctx, *tmp;
>         struct list_head *list;
>         unsigned long flags;
> @@ -843,11 +845,21 @@ static void perf_cgroup_switch(struct task_struct *task)
>          */
>         local_irq_save(flags);
>
> +       cgrp = perf_cgroup_from_task(task, NULL);
> +       if (cgrp == __this_cpu_read(cpu_perf_cgroup))
> +               goto out;
> +
> +       __this_cpu_write(cpu_perf_cgroup, cgrp);
> +
>         list = this_cpu_ptr(&cgrp_cpuctx_list);
>         list_for_each_entry_safe(cpuctx, tmp, list, cgrp_cpuctx_entry) {
>                 WARN_ON_ONCE(cpuctx->ctx.nr_cgroups == 0);
>
>                 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
> +
> +               if (cpuctx->cgrp == cgrp)

Missing perf_ctx_unlock().

Thanks,
Namhyung

> +                       continue;
> +
>                 perf_pmu_disable(cpuctx->ctx.pmu);
>
>                 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
> @@ -855,14 +867,11 @@ static void perf_cgroup_switch(struct task_struct *task)
>                  * must not be done before ctxswout due
>                  * to event_filter_match() in event_sched_out()
>                  */
> -               cpuctx->cgrp = perf_cgroup_from_task(task,
> -                                                    &cpuctx->ctx);
> +               cpuctx->cgrp = cgrp;
>                 /*
>                  * set cgrp before ctxsw in to allow
>                  * event_filter_match() to not have to pass
>                  * task around
> -                * we pass the cpuctx->ctx to perf_cgroup_from_task()
> -                * because cgroup events are only per-cpu
>                  */
>                 cpu_ctx_sched_in(cpuctx, EVENT_ALL, task);
>
> @@ -870,35 +879,10 @@ static void perf_cgroup_switch(struct task_struct *task)
>                 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
>         }
>
> +out:
>         local_irq_restore(flags);
>  }