Re: [tip:perf/core] [perf] da916e96e2: BUG:KASAN:null-ptr-deref_in_put_event

[James Clark]

On 15/04/2025 2:14 pm, Peter Zijlstra wrote:
> On Tue, Apr 15, 2025 at 12:08:40PM +0200, Peter Zijlstra wrote:
> > On Tue, Apr 15, 2025 at 10:14:05AM +0100, James Clark wrote:
> > > On 15/04/2025 5:46 am, Oliver Sang wrote:
> >
> > > > yes, below patch fixes the issues we observed for da916e96e2. thanks
> > > >
> > > > Tested-by: kernel test robot <oliver.sang@xxxxxxxxx>
> > >
> > > Also fixes the same issues we were seeing:
> > >
> > > Tested-by: James Clark <james.clark@xxxxxxxxxx>
> >
> > Excellent, thank you both! Now I gotta go write me a Changelog :-)
>
> Hmm, so while writing Changelog, I noticed something else was off. The
> case where event->parent was set to EVENT_TOMBSTONE now didn't have a
> put_event(parent) anymore. So that needs to be put back in as well.
>
> Frederic, afaict this should still be okay, since if we're detached,
> then nothing will try and access event->parent in the free path.
>
> Also, nothing in perf_pending_task() will try and access either
> event->parent or event->pmu.
>
> ---
> Subject: perf: Fix event->parent life-time issue
> From: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
> Date: Tue Apr 15 12:12:52 CEST 2025
>
> Due to an oversight in merging da916e96e2de ("perf: Make
> perf_pmu_unregister() useable") on top of 56799bc03565 ("perf: Fix
> hang while freeing sigtrap event"), it is now possible to hit
> put_event(EVENT_TOMBSTONE), which makes the computer sad.
>
> This also means that for the event->parent == EVENT_TOMBSTONE, the
> put_event() matching inherit_event() has gone missing.
>
> Previously this was done in perf_event_release_kernel() after calling
> perf_remove_from_context(), but with it delegated to put_event(), this
> case is now entirely missed, leading to leaks.
>
> Fixes: da916e96e2de ("perf: Make perf_pmu_unregister() useable")
> Signed-off-by: Peter Zijlstra (Intel) <peterz@xxxxxxxxxxxxx>
> ---
>   kernel/events/core.c |    3 ++-
>   1 file changed, 2 insertions(+), 1 deletion(-)
>
> --- a/kernel/events/core.c
> +++ b/kernel/events/core.c
> @@ -2343,6 +2343,7 @@ static void perf_child_detach(struct per
>   	 * not being a child event. See for example unaccount_event().
>   	 */
>   	event->parent = EVENT_TOMBSTONE;
> +	put_event(parent_event);
>   }
>   
>   static bool is_orphaned_event(struct perf_event *event)
> @@ -5688,7 +5689,7 @@ static void put_event(struct perf_event
>   	_free_event(event);
>   
>   	/* Matches the refcount bump in inherit_event() */
> -	if (parent)
> +	if (parent && parent != EVENT_TOMBSTONE)
>   		put_event(parent);
>   }
>   

Hi Peter,

Unrelated to the pointer deref issue, I'm also seeing perf stat not 
working due to this commit. And that's both with and without this fixup:

 -> perf stat -- true

 Performance counter stats for 'true':

     <not counted> msec task-clock 

     <not counted>      context-switches 

     <not counted>      cpu-migrations 

     <not counted>      page-faults 

     <not counted>      armv8_cortex_a53/instructions/ 

     <not counted>      armv8_cortex_a57/instructions/ 

     <not counted>      armv8_cortex_a53/cycles/ 

     <not counted>      armv8_cortex_a57/cycles/ 

     <not counted>      armv8_cortex_a53/branches/ 

     <not counted>      armv8_cortex_a53/branch-misses/ 

     <not counted>      armv8_cortex_a57/branch-misses/ 


       0.074139992 seconds time elapsed

       0.000000000 seconds user
       0.054797000 seconds sys

Didn't look into it more other than bisecting it to this commit, but I 
can dig more unless the issue is obvious. This is on Arm big.LITTLE, 
although I didn't test it elsewhere so I'm not sure if that's relevant 
or not.

Thanks
James