Re: Are jump labels broken on 6.11-rc1?

[Peter Zijlstra <peterz@infradead.org>]

On Wed, Aug 07, 2024 at 04:03:12PM +0200, Thomas Gleixner wrote:

> > +	if (static_key_dec(key, true)) // dec-not-one
> 
> Eeew.

:-) I knew you'd hate on that

> Something like the below?
> 
> Thanks,
> 
>         tglx
> ---
> @@ -250,49 +250,71 @@ void static_key_disable(struct static_ke
>  }
>  EXPORT_SYMBOL_GPL(static_key_disable);
>  
> -static bool static_key_slow_try_dec(struct static_key *key)
> +static bool static_key_dec(struct static_key *key, bool dec_not_one)
>  {
> +	int v = atomic_read(&key->enabled);
>  
>  	do {
>  		/*
> +		 * Warn about the '-1' case; since that means a decrement is
> +		 * concurrent with a first (0->1) increment. IOW people are
> +		 * trying to disable something that wasn't yet fully enabled.
> +		 * This suggests an ordering problem on the user side.
> +		 *
> +		 * Warn about the '0' case; simple underflow.
>  		 */
> +		if (WARN_ON_ONCE(v <= 0))
> +			return v;
> +
> +		if (dec_not_one && v == 1)
> +			return v;
> +
>  	} while (!likely(atomic_try_cmpxchg(&key->enabled, &v, v - 1)));
>  
> +	return v;
> +}
> +
> +/*
> + * Fastpath: Decrement if the reference count is greater than one
> + *
> + * Returns false, if the reference count is 1 or -1 to force the caller
> + * into the slowpath.
> + *
> + * The -1 case is to handle a decrement during a concurrent first enable,
> + * which sets the count to -1 in static_key_slow_inc_cpuslocked(). As the
> + * slow path is serialized the caller will observe 1 once it acquired the
> + * jump_label_mutex, so the slow path can succeed.
> + */
> +static bool static_key_dec_not_one(struct static_key *key)
> +{
> +	int v = static_key_dec(key, true);
> +
> +	return v != 1 && v != -1;

	if (v < 0)
		return false;

	/*
	 * Notably, 0 (underflow) returns true such that it bails out
	 * without doing anything.
	 */
	return v != 1;

Perhaps?

> +}
> +
> +/*
> + * Slowpath: Decrement and test whether the refcount hit 0.
> + *
> + * Returns true if the refcount hit zero, i.e. the previous value was one.
> + */
> +static bool static_key_dec_and_test(struct static_key *key)
> +{
> +	int v = static_key_dec(key, false);
> +
> +	lockdep_assert_held(&jump_label_mutex);
> +	return v == 1;
>  }

But yeah, this is nicer!