[PATCH v6 23/29] context-tracking: Introduce work deferral infrastructure

[Frederic Weisbecker]

Le Wed, Oct 29, 2025 at 11:09:50AM +0100, Valentin Schneider a écrit :
> On 28/10/25 15:00, Frederic Weisbecker wrote:
> > Le Fri, Oct 10, 2025 at 05:38:33PM +0200, Valentin Schneider a écrit :
> >> +	old = atomic_read(&ct->state);
> >> +
> >> +	/*
> >> +	 * The work bit must only be set if the target CPU is not executing
> >> +	 * in kernelspace.
> >> +	 * CT_RCU_WATCHING is used as a proxy for that - if the bit is set, we
> >> +	 * know for sure the CPU is executing in the kernel whether that be in
> >> +	 * NMI, IRQ or process context.
> >> +	 * Set CT_RCU_WATCHING here and let the cmpxchg do the check for us;
> >> +	 * the state could change between the atomic_read() and the cmpxchg().
> >> +	 */
> >> +	old |= CT_RCU_WATCHING;
> >
> > Most of the time, the task should be either idle or in userspace. I'm still not
> > sure why you start with a bet that the CPU is in the kernel with RCU watching.
> >
> 
> Right I think I got that the wrong way around when I switched to using
> CT_RCU_WATCHING vs CT_STATE_KERNEL. That wants to be
> 
>   old &= ~CT_RCU_WATCHING;
> 
> i.e. bet the CPU is NOHZ-idle, if it's not the cmpxchg fails and we don't
> store the work bit.

Right.

> 
> >> +	/*
> >> +	 * Try setting the work until either
> >> +	 * - the target CPU has entered kernelspace
> >> +	 * - the work has been set
> >> +	 */
> >> +	do {
> >> +		ret = atomic_try_cmpxchg(&ct->state, &old, old | (work << CT_WORK_START));
> >> +	} while (!ret && !(old & CT_RCU_WATCHING));
> >
> > So this applies blindly to idle as well, right? It should work but note that
> > idle entry code before RCU watches is also fragile.
> >
> 
> Yeah I remember losing some hair trying to grok the idle entry situation;
> we could keep this purely NOHZ_FULL and have the deferral condition be:
> 
>   (ct->state & CT_STATE_USER) && !(ct->state & CT_RCU_WATCHING)

Well, after all what works for NOHZ_FULL should also work for idle. It's
preceded by entry code as well (or rather __cpuidle).

Thanks.

-- 
Frederic Weisbecker
SUSE Labs