Re: [PATCH] tracing: Remove pointless memory barriers

[David Laight <david.laight.linux@gmail.com>]

On Wed, 9 Jul 2025 11:08:27 -0400
Steven Rostedt <rostedt@goodmis.org> wrote:

> On Tue, 8 Jul 2025 09:42:19 +0200
> Nam Cao <namcao@linutronix.de> wrote:
> 
> > So yes, smp_rmb() is only useful inbetween reads, and smp_wmb() is
> > only userful inbetween writes.  
> 
> Hmm, I wonder if barriers isn't needed but atomic values are?
> 
> That is, it looks like rv_monitoring_on() is looking to read the
> current state, where as turn_monitoring_on/off() changes the state.
> 
> Perhaps instead of barriers, it should use atomics?
> 
>  bool rv_monitoring_on(void)
>  {
> 	return atomic_read(&monitoring_on);
>  }
>  
>  static void turn_monitoring_off(void)
>  {
> 	atomic_set(&monitoring_on, 0);
>  }
>  
> 
> Doesn't atomic make sure the values are seen when they are changed?

No.
It normally just ensures the read/write aren't 'torn'.
Atomics are used for read-modify-writes to ensure two cpu don't
do read-read-modify-modify-write-write losing one of the changes.
(They can need special instructions for read and write - but normally don't.)
So here just the same as the volatile accesses READ_ONCE() and WRITE_ONCE().

	David


> 
> As this code is more about looking at state and not ordering, and I
> think that's what atomics are about.
> 
> -- Steve
>