Re: [PATCH 1/2] tracing/perf: Prevent double unregister of perf probes

[Aditya Chillara <quic_achillar@quicinc.com>]

On 7/9/2025 7:53 PM, Steven Rostedt wrote:
> On Wed, 9 Jul 2025 11:11:09 +0530
> Aditya Chillara <quic_achillar@quicinc.com> wrote:
> 
>> Double perf_trace_event_unreg is allowed causing perf_refcount to go
>> negative. total_ref_count also goes negative because the return value
>> of tracepoint_probe_unregister is ignored.
>>
>> Once total_ref_count is negative, the next call to perf_trace_event_reg
>> will register perf_probe but will not allocate perf_trace_buf and sets
>> it to NULL instead.
>>
>> The subsequent trace_##call() will mem abort in perf_trace_buf_alloc
>> because memset will be called on the NULL perf_trace_buf.
>>
>> Gracefully handle the error in perf_trace_event_unreg to prevent
>> double unregister.
>>
>> Signed-off-by: Aditya Chillara <quic_achillar@quicinc.com>
>> ---
>>  kernel/trace/trace_event_perf.c | 8 ++++++--
>>  kernel/trace/trace_events.c     | 3 +--
>>  2 files changed, 7 insertions(+), 4 deletions(-)
>>
>> diff --git a/kernel/trace/trace_event_perf.c b/kernel/trace/trace_event_perf.c
>> index 61e3a2620fa3c9417ac23cf5a18aeb86e7393dcc..247db88accd88eb0acf3692ea593d576519ce8b1 100644
>> --- a/kernel/trace/trace_event_perf.c
>> +++ b/kernel/trace/trace_event_perf.c
>> @@ -154,12 +154,16 @@ static int perf_trace_event_reg(struct trace_event_call *tp_event,
>>  static void perf_trace_event_unreg(struct perf_event *p_event)
>>  {
>>  	struct trace_event_call *tp_event = p_event->tp_event;
>> -	int i;
>> +	int i, ret;
>>  
>>  	if (--tp_event->perf_refcount > 0)
>>  		return;
>>  
>> -	tp_event->class->reg(tp_event, TRACE_REG_PERF_UNREGISTER, NULL);
>> +	ret = tp_event->class->reg(tp_event, TRACE_REG_PERF_UNREGISTER, NULL);
> 
> The only time unreg() fails is when it doesn't find a tracepoint to
> unregister.
> 
> There should be no reason to check the return value of unregister if
> you have your accounting correct. Thus I think you are fixing a symptom
> of a bug elsewhere.

The exact problem was introduced by:
https://github.com/torvalds/linux/commit/7ef5aa081f989ecfecc1df02068a80aebbd3ec31
(perf/core: Simplify the perf_event_alloc() error path)
where __free_event was calling event->destroy() even though it would
have been called by perf_try_init_event in case it failed.

This was fixed by:
https://github.com/torvalds/linux/commit/da02f54e81db2f7bf6af9d1d0cfc5b41ec6d0dcb
(perf/core: Clean up perf_try_init_event())

This patch prevents from crashing even if that happens, and there
will be a warning anyway to notice the double unregister.

> 
> -- Steve
> 
> 
>> +	if (ret) {
>> +		++tp_event->perf_refcount;
>> +		return;
>> +	}
>>  

Best Regards,
Aditya