Re: [PATCH] perf/core: Fix cgroup events tracking

[Ravi Bangoria <ravi.bangoria@amd.com>]

On 06-Dec-22 8:20 AM, Chengming Zhou wrote:
> We encounter perf warnings when using cgroup events like:
> ```
> cd /sys/fs/cgroup
> mkdir test
> perf stat -e cycles -a -G test
> ```
> 
> WARNING: CPU: 0 PID: 690 at kernel/events/core.c:849 perf_cgroup_switch+0xb2/0xc0
> [   91.393417] Call Trace:
> [   91.393772]  <TASK>
> [   91.394080]  __schedule+0x4ae/0x9f0
> [   91.394535]  ? _raw_spin_unlock_irqrestore+0x23/0x40
> [   91.395145]  ? __cond_resched+0x18/0x20
> [   91.395622]  preempt_schedule_common+0x2d/0x70
> [   91.396163]  __cond_resched+0x18/0x20
> [   91.396621]  wait_for_completion+0x2f/0x160
> [   91.397137]  ? cpu_stop_queue_work+0x9e/0x130
> [   91.397665]  affine_move_task+0x18a/0x4f0

nit: These timestamps can be removed in commit log.

> 
> WARNING: CPU: 0 PID: 690 at kernel/events/core.c:829 ctx_sched_in+0x1cf/0x1e0
> [   91.430151] Call Trace:
> [   91.430490]  <TASK>
> [   91.430793]  ? ctx_sched_out+0xb7/0x1b0
> [   91.431274]  perf_cgroup_switch+0x88/0xc0
> [   91.431778]  __schedule+0x4ae/0x9f0
> [   91.432215]  ? _raw_spin_unlock_irqrestore+0x23/0x40
> [   91.432825]  ? __cond_resched+0x18/0x20
> [   91.433299]  preempt_schedule_common+0x2d/0x70
> [   91.433839]  __cond_resched+0x18/0x20
> [   91.434298]  wait_for_completion+0x2f/0x160
> [   91.434808]  ? cpu_stop_queue_work+0x9e/0x130
> [   91.435334]  affine_move_task+0x18a/0x4f0
> 
> The above two warnings are not complete here since I remove other
> unimportant information. The problem is caused by the perf cgroup
> events tracking:
> 
> CPU0					CPU1
> perf_event_open()
>   perf_event_alloc()
>     account_event()
>       account_event_cpu()
>         atomic_inc(perf_cgroup_events)
> 					__perf_event_task_sched_out()
> 					  if (atomic_read(perf_cgroup_events))
> 					    perf_cgroup_switch()
> 					      // kernel/events/core.c:849
> 					      WARN_ON_ONCE(cpuctx->ctx.nr_cgroups == 0)
> 					      if (READ_ONCE(cpuctx->cgrp) == cgrp) // false
> 					        return
> 					      perf_ctx_lock()
> 					      ctx_sched_out()
> 					      cpuctx->cgrp = cgrp
> 					      ctx_sched_in()
> 					        perf_cgroup_set_timestamp()
> 					          // kernel/events/core.c:829
> 					          WARN_ON_ONCE(!ctx->nr_cgroups)
> 					      perf_ctx_unlock()
>   perf_install_in_context()
>     add_event_to_ctx()
>       list_add_event()
>         perf_cgroup_event_enable()
>           ctx->nr_cgroups++
>           cpuctx->cgrp = X

IIUC, since it's a cgroup event, perf_install_in_context() will do:
cpu_function_call(cpu, __perf_install_in_context, event). And thus,
callchain starting with add_event_to_ctx() will be executed on CPU1,
not on CPU0.

> We can see from above that we wrongly use percpu atomic perf_cgroup_events
> to check if we need to perf_cgroup_switch(), which should only be used
> when we know this CPU has cgroup events enabled.
> 
> The commit bd2756811766 ("perf: Rewrite core context handling") change
> to have only one context per-CPU, so we can just use cpuctx->cgrp to
> check if this CPU has cgroup events enabled.
> 
> So percpu atomic perf_cgroup_events is not needed.
> 
> Signed-off-by: Chengming Zhou <zhouchengming@bytedance.com>

Fixes: bd2756811766 ("perf: Rewrite core context handling")

Otherwise looks good.
Tested-by: Ravi Bangoria <ravi.bangoria@amd.com>

Thanks,
Ravi