Re: [tip: perf/urgent] perf: Fix perf_remove_from_context() event scheduling

[Peter Zijlstra]

On Sat, Oct 24, 2026 at 11:34:03AM -0000, tip-bot2 for Peter Zijlstra wrote:
> The following commit has been merged into the perf/urgent branch of tip:
> 
> Commit-ID:     d095321349222961813333277881b3c74f10b17b
> Gitweb:        https://git.kernel.org/tip/d095321349222961813333277881b3c74f10b17b
> Author:        Peter Zijlstra <peterz@xxxxxxxxxxxxx>
> AuthorDate:    Wed, 14 Oct 2026 14:45:51 +02:00
> Committer:     Peter Zijlstra <peterz@xxxxxxxxxxxxx>
> CommitterDate: Fri, 23 Oct 2026 19:09:52 +02:00
> 
> perf: Fix perf_remove_from_context() event scheduling
> 
> Q8 noted that it is possible to get into an inconsistent state when
> using perf_remove_from_context() on a group leader event.
> 
> Specifically, when the leader gets scheduled out, we should (when not
> destroying the grouping) schedule out the complete group, similar to
> __perf_event_disable().
> 

I noticed that I forgot the Fixes line :-/ This is an ancient oversight
afaict and needs backporting to all still maintained kernels.

Fixes: b04243ef7006 ("perf: Complete software pmu grouping")

> Reported-by: Q8 <cipher.gnep@xxxxxxxxx>
> Tested-by: Q8 <cipher.gnep@xxxxxxxxx>
> Signed-off-by: Peter Zijlstra (Intel) <peterz@xxxxxxxxxxxxx>
> ---
>  kernel/events/core.c |  9 +++++++++
>  1 file changed, 9 insertions(+)
> 
> diff --git a/kernel/events/core.c b/kernel/events/core.c
> index e248f34..f36a3bb 100644
> --- a/kernel/events/core.c
> +++ b/kernel/events/core.c
> @@ -2590,6 +2590,7 @@ __perf_remove_from_context(struct perf_event *event,
>  	enum perf_event_state state = PERF_EVENT_STATE_OFF;
>  	unsigned long flags = (unsigned long)info;
>  
> +	perf_pmu_disable(pmu_ctx->pmu);
>  	ctx_time_update(cpuctx, ctx);
>  
>  	/*
> @@ -2603,6 +2604,13 @@ __perf_remove_from_context(struct perf_event *event,
>  	if (flags & DETACH_DEAD)
>  		state = PERF_EVENT_STATE_DEAD;
>  
> +	/*
> +	 * When disabling a group leader, the whole group becomes ineligible
> +	 * to run, so schedule out the full group.
> +	 */
> +	if (!(flags & DETACH_GROUP) && event == event->group_leader)
> +		group_sched_out(event, ctx);
> +
>  	__event_disable(event, ctx, state);
>  
>  	if (flags & DETACH_GROUP)
> @@ -2632,6 +2640,7 @@ __perf_remove_from_context(struct perf_event *event,
>  			cpuctx->task_ctx = NULL;
>  		}
>  	}
> +	perf_pmu_enable(pmu_ctx->pmu);
>  }
>  
>  /*