Re: [PATCH v2] perf/core: restore __perf_remove_from_context when DETACH_EXIT not set

[Peter Zijlstra]

On Tue, Jun 03, 2025 at 04:33:04PM +0800, David Wang wrote:
> commit a3c3c66670ce ("perf/core: Fix child_total_time_enabled accounting
> bug at task exit") made changes to __perf_remove_from_context() to
> coordinate its changes with perf_event_exit_event(), but the change are
> unconditional, it impacts callpaths to __perf_remove_from_context()
> other than from perf_event_exit_event(). One of the impact is to cgroup,
> which is not properly handled and would cause kernel panic with high
> probalibity during reboot on some system[1].

Sorry, but no. This does not describe the problem adequately. I would
have to go read your [1] to figure out what is actually broken.

That is, having read the above, I'm still clueless as to what the actual
problem is.

> To confine the side effects, make the changes to
> __perf_remove_from_context() conditional, restore to its previous state
> except when DETACH_EXIT is set.
> 
> Closes: https://lore.kernel.org/lkml/20250601173603.3920-1-00107082@xxxxxxx/ [1]
> Fixes: a3c3c66670ce ("perf/core: Fix child_total_time_enabled accounting bug at task exit")
> Signed-off-by: David Wang <00107082@xxxxxxx>
> ---
> Changes:
> Address yeoreum.yun@xxxxxxx's concern about missing cgroup event.
> ---
>  kernel/events/core.c | 11 ++++++-----
>  1 file changed, 6 insertions(+), 5 deletions(-)
> 
> diff --git a/kernel/events/core.c b/kernel/events/core.c
> index 95e703891b24..e2c0f34b0789 100644
> --- a/kernel/events/core.c
> +++ b/kernel/events/core.c
> @@ -2466,7 +2466,7 @@ __perf_remove_from_context(struct perf_event *event,
>  			   void *info)
>  {
>  	struct perf_event_pmu_context *pmu_ctx = event->pmu_ctx;
> -	enum perf_event_state state = PERF_EVENT_STATE_OFF;
> +	enum perf_event_state exit_state = PERF_EVENT_STATE_EXIT;
>  	unsigned long flags = (unsigned long)info;
>  
>  	ctx_time_update(cpuctx, ctx);
> @@ -2475,19 +2475,20 @@ __perf_remove_from_context(struct perf_event *event,
>  	 * Ensure event_sched_out() switches to OFF, at the very least
>  	 * this avoids raising perf_pending_task() at this time.
>  	 */
> -	if (flags & DETACH_EXIT)
> -		state = PERF_EVENT_STATE_EXIT;
>  	if (flags & DETACH_DEAD) {
>  		event->pending_disable = 1;
> -		state = PERF_EVENT_STATE_DEAD;
> +		exit_state = PERF_EVENT_STATE_DEAD;
>  	}
>  	event_sched_out(event, ctx);
> -	perf_event_set_state(event, min(event->state, state));
>  	if (flags & DETACH_GROUP)
>  		perf_group_detach(event);
>  	if (flags & DETACH_CHILD)
>  		perf_child_detach(event);
>  	list_del_event(event, ctx);
> +	if (flags & DETACH_EXIT)
> +		perf_event_set_state(event, min(event->state, exit_state));
> +	if (flags & DETACH_DEAD)
> +		event->state = PERF_EVENT_STATE_DEAD;

Urgh, no. Trying to reverse engineer the above, the intent appears to be
to not set OFF.

This can be achieved by doing:

-       enum perf_event_state state = PERF_EVENT_STATE_OFF;
+       enum perf_event_state state = event->state;

No other changes required. You also move the location of
perf_event_set_state(), but it is entirely unclear to me if that is
actually needed.

Worse, you split the means of setting state -- that is entirely uncalled
for.