Re: [PATCH v2 01/28] coresight: Change device mode to atomic type

[Yeoreum Yun <yeoreum.yun@arm.com>]

Hi,
>
> On Wed, Jul 02, 2025 at 10:49:01AM +0100, Yeoreum Yun wrote:
> > Hi Leo,
> >
> > >  {
> > > -	return local_cmpxchg(&csdev->mode, CS_MODE_DISABLED, new_mode) ==
> > > -	       CS_MODE_DISABLED;
> > > +	int curr = CS_MODE_DISABLED;
> > > +
> > > +	return atomic_try_cmpxchg_acquire(&csdev->mode, &curr, new_mode);
> > >  }
> >
> > Just question. why is acquire symentic enough in here?
>
> My understanding is that acquire semantics ensure ordering between
> cmpxchg_acquire() and all memory accesses that follow it. However, it
> does not guarantee that memory accesses appearing before the acquire
> are ordered as well.
>
> This is exactly what we want in the driver. We must ensure to first grab
> an active device mode, then it is safe to proceed later operations (e.g.
> set configurations in driver data and access registers).
>
> > before this change, local_cmpxchg seems to use full_fenced.
>
> Not really. Arm64 has atomic instruction for cmpxchg, it does not use
> full_fenced. It should run into the path of arch_cmpxchg().

No, It local_cmpxchg backend with atomic_cmpxchg() and it seems include
full fenced.

But, you've explained the what i concerned with below diagram clear

CPU0                                          CPU1
// store csmode in store buffer
atomic_cmpxchg_acquired(csmode)
                                             // couldn't see the contents in store buffer
                                             atomic_read_acquire(csmode);

                                             Send SMP call
              ,-------------------------------/
             v
atomic_cmpxchg_acquired(csmode)  ==> Fail to grab another mode

However, I think it would be good to change the atomic_try_cmpxchg()
instead of atomic_cmpxchg_acquired() to prevent redundnat IPI trigger for
observation failure.

Thanks.

--
Sincerely,
Yeoreum Yun