Re: [PATCH v2] perf/core: Fix group leader use-after-free after sibling detach

[Aditya Chillara <aditya.chillara@oss.qualcomm.com>]

On 8/7/2026 1:20 AM, Peter Zijlstra wrote:
> 
> Finally got to look at this. In principle this seems okay, but while
> staring at it, I had a few questions, see below.
> 
> On Tue, Jun 30, 2026 at 12:42:12AM +0530, Aditya Chillara wrote:
>> ---
>>  kernel/events/core.c | 62 +++++++++++++++++++++++++++++++++++-----------------
>>  1 file changed, 42 insertions(+), 20 deletions(-)
>>
>> diff --git a/kernel/events/core.c b/kernel/events/core.c
>> index 954c36e28101..744643ada948 100644
>> --- a/kernel/events/core.c
>> +++ b/kernel/events/core.c
>> @@ -2253,6 +2253,8 @@ static void put_event(struct perf_event *event);
>>  static void __event_disable(struct perf_event *event,
>>  			    struct perf_event_context *ctx,
>>  			    enum perf_event_state state);
>> +static void event_sched_out(struct perf_event *event,
>> +			    struct perf_event_context *ctx);
>>  
>>  static void perf_put_aux_event(struct perf_event *event)
>>  {
>> @@ -2343,6 +2345,44 @@ static inline struct list_head *get_event_list(struct perf_event *event)
>>  				    &event->pmu_ctx->flexible_active;
>>  }
>>  
>> +/* @sibling must already be unlinked from its old leader's sibling_list. */
>> +static void perf_promote_sibling_to_leader(struct perf_event *sibling,
>> +					   struct perf_event_context *ctx,
>> +					   int group_caps)
>> +{
>> +	/*
>> +	 * Events that have PERF_EV_CAP_SIBLING require being part of
>> +	 * a group and cannot exist on their own, schedule them out
>> +	 * and move them into the ERROR state. Also see
>> +	 * _perf_event_enable(), it will not be able to recover this
>> +	 * ERROR state.
>> +	 */
>> +	if (sibling->event_caps & PERF_EV_CAP_SIBLING) {
>> +		event_sched_out(sibling, ctx);
>> +
>> +		/*
>> +		 * The guards keep this correct even when @sibling is already
>> +		 * disabled (see __perf_remove_from_context()).
>> +		 */
>> +		if (sibling->state > PERF_EVENT_STATE_OFF)
>> +			perf_cgroup_event_disable(sibling, ctx);
>> +		if (sibling->state > PERF_EVENT_STATE_ERROR)
>> +			perf_event_set_state(sibling, PERF_EVENT_STATE_ERROR);
>> +	}
> 
> The below code used __event_disable(); and this change is not
> mentioned in the Changelog. Why was this changed?

I should've mentioned this in the Changelog, I'll update it.

In case of __perf_remove_from_context(event, ...) where the event is part of a group
but is not the leader; by the time we get to perf_group_detach():

1. perf_cgroup_event_disable(event, ctx) would have been called
2. The sibling's state may have been set to EXIT/REVOKED/DEAD if not OFF

So we cannot use __event_disable() because it would unconditionally call
1. perf_cgroup_event_disable() that would decrement nr_cgroups twice
2. perf_event_set_state(event, PERF_EVENT_STATE_ERROR) that could bring back
   the state from EXIT/REVOKED/DEAD/OFF to ERROR.

In previous code, in case of __perf_remove_from_context(event, ...) where the event
is a leader; __event_disable() was called for siblings unconditionally but I do not
see a reason why the added guards don't hold.

Should we move the guards to __event_disable() instead? That would simplify the code:

static void __event_disable(struct perf_event *event,
			    struct perf_event_context *ctx,
			    enum perf_event_state state)
{
	event_sched_out(event, ctx);
	if (event->state > PERF_EVENT_STATE_OFF)
		perf_cgroup_event_disable(event, ctx);
	perf_event_set_state(event, min(event->state, state));
}

Also, event_sched_out() in case of pending_disable calls perf_cgroup_event_disable()
and sets the state to OFF, but __event_disable() immediately calls
perf_cgroup_event_disable() again.

> 
>> +
>> +	sibling->group_leader = sibling;
>> +	sibling->group_caps = group_caps;
>> +
>> +	if (sibling->attach_state & PERF_ATTACH_CONTEXT) {
>> +		add_event_to_groups(sibling, ctx);
>> +
>> +		if (sibling->state == PERF_EVENT_STATE_ACTIVE)
>> +			list_add_tail(&sibling->active_list, get_event_list(sibling));
>> +	}
>> +
>> +	perf_event__header_size(sibling);
>> +}
>> +
>>  static void perf_group_detach(struct perf_event *event)
>>  {
>>  	struct perf_event *leader = event->group_leader;
>> @@ -2368,6 +2408,7 @@ static void perf_group_detach(struct perf_event *event)
>>  		list_del_init(&event->sibling_list);
>>  		event->group_leader->nr_siblings--;
>>  		event->group_leader->group_generation++;
> 
> Here we can do 's/event->group_//'

I'll update this.

> 
> Also, this case 'leader != event' we remove one sibling from a group and
> decrement leader->nr_siblings...
> 
>> +		perf_promote_sibling_to_leader(event, ctx, event->event_caps);
>>  		goto out;
>>  	}
>>  
>> @@ -2377,29 +2418,10 @@ static void perf_group_detach(struct perf_event *event)
>>  	 * to whatever list we are on.
>>  	 */
>>  	list_for_each_entry_safe(sibling, tmp, &event->sibling_list, sibling_list) {
>> -
>> -		/*
>> -		 * Events that have PERF_EV_CAP_SIBLING require being part of
>> -		 * a group and cannot exist on their own, schedule them out
>> -		 * and move them into the ERROR state. Also see
>> -		 * _perf_event_enable(), it will not be able to recover this
>> -		 * ERROR state.
>> -		 */
>> -		if (sibling->event_caps & PERF_EV_CAP_SIBLING)
>> -			__event_disable(sibling, ctx, PERF_EVENT_STATE_ERROR);
>> -
>> -		sibling->group_leader = sibling;
>>  		list_del_init(&sibling->sibling_list);
>>  
>>  		/* Inherit group flags from the previous leader */
>> -		sibling->group_caps = event->group_caps;
>> -
>> -		if (sibling->attach_state & PERF_ATTACH_CONTEXT) {
>> -			add_event_to_groups(sibling, event->ctx);
>> -
>> -			if (sibling->state == PERF_EVENT_STATE_ACTIVE)
>> -				list_add_tail(&sibling->active_list, get_event_list(sibling));
>> -		}
>> +		perf_promote_sibling_to_leader(sibling, ctx, event->group_caps);
>>  
>>  		WARN_ON_ONCE(sibling->ctx != event->ctx);
>>  	}
> 
> This is the case 'leader == event' (per not being the other case), and
> this we remove all siblings, but then do not set leader->nr_siblings =
> 0, should we ?

Yes, I'll add this.

Thank you,
Aditya