Re: [tip: perf/core] perf: Simplify perf_event_free_task() wait

[Peter Zijlstra]

On Wed, Apr 09, 2025 at 03:01:12PM +0200, Frederic Weisbecker wrote:

> > diff --git a/kernel/events/core.c b/kernel/events/core.c
> > index 3c92b75..fa6dab0 100644
> > --- a/kernel/events/core.c
> > +++ b/kernel/events/core.c
> > @@ -1270,6 +1270,9 @@ static void put_ctx(struct perf_event_context *ctx)
> >  		if (ctx->task && ctx->task != TASK_TOMBSTONE)
> >  			put_task_struct(ctx->task);
> >  		call_rcu(&ctx->rcu_head, free_ctx);
> > +	} else if (ctx->task == TASK_TOMBSTONE) {
> > +		smp_mb(); /* pairs with wait_var_event() */
> > +		wake_up_var(&ctx->refcount);
> 
> So there are three situations:
> 
> * If perf_event_free_task() has removed all the children from the parent list
>   before perf_event_release_kernel() got a chance to even iterate them, then
>   it's all good as there is no get_ctx() pending.
> 
> * If perf_event_release_kernel() iterates a child event, but it gets freed
>   meanwhile by perf_event_free_task() while the mutexes are temporarily
>   unlocked, it's all good because while locking again the ctx mutex,
>   perf_event_release_kernel() observes TASK_TOMBSTONE.
> 
> * But if perf_event_release_kernel() frees the child event before
>   perf_event_free_task() got a chance, we may face this scenario:
> 
>     perf_event_release_kernel()                                  perf_event_free_task()
>     --------------------------                                   ------------------------
>     mutex_lock(&event->child_mutex)
>     get_ctx(child->ctx)
>     mutex_unlock(&event->child_mutex)
> 
>     mutex_lock(ctx->mutex)
>     mutex_lock(&event->child_mutex)
>     perf_remove_from_context(child)
>     mutex_unlock(&event->child_mutex)
>     mutex_unlock(ctx->mutex)
> 
>                                                                  // This lock acquires ctx->refcount == 2
>                                                                  // visibility
>                                                                  mutex_lock(ctx->mutex)
>                                                                  ctx->task = TASK_TOMBSTONE
>                                                                  mutex_unlock(ctx->mutex)
>                                                                  
>                                                                  wait_var_event()
>                                                                      // enters prepare_to_wait() since
>                                                                      // ctx->refcount == 2
>                                                                      // is guaranteed to be seen
>                                                                      set_current_state(TASK_INTERRUPTIBLE)
>                                                                      smp_mb()
>                                                                      if (ctx->refcount != 1)
>                                                                          schedule()
>     put_ctx()
>        // NOT fully ordered! Only RELEASE semantics
>        refcount_dec_and_test()
>            atomic_fetch_sub_release()
>        // So TASK_TOMBSTONE is not guaranteed to be seen
>        if (ctx->task == TASK_TOMBSTONE)
>            wake_up_var()
> 
> Basically it's a broken store buffer:
> 
>     perf_event_release_kernel()                                  perf_event_free_task()
>     --------------------------                                   ------------------------
>     ctx->task = TASK_TOMBSTONE                                   smp_store_release(&ctx->refcount, ctx->refcount - 1)
>     smp_mb()
>     READ_ONCE(ctx->refcount)                                     READ_ONCE(ctx->task)
> 
> 
> So you need this:
> 
> diff --git a/kernel/events/core.c b/kernel/events/core.c
> index fa6dab08be47..c4fbbe25361a 100644
> --- a/kernel/events/core.c
> +++ b/kernel/events/core.c
> @@ -1270,9 +1270,10 @@ static void put_ctx(struct perf_event_context *ctx)
>  		if (ctx->task && ctx->task != TASK_TOMBSTONE)
>  			put_task_struct(ctx->task);
>  		call_rcu(&ctx->rcu_head, free_ctx);
> -	} else if (ctx->task == TASK_TOMBSTONE) {
> +	} else {
>  		smp_mb(); /* pairs with wait_var_event() */
> -		wake_up_var(&ctx->refcount);
> +		if (ctx->task == TASK_TOMBSTONE)
> +			wake_up_var(&ctx->refcount);
>  	}
>  }

Very good, thanks!

I'll make that smp_mb__after_atomic() instead, but yes, this barrier
needs to move before the loading of ctx->task.

I'll transform this into a patch and stuff on top.