Re: [PATCH 2/3] futex: Use RCU-based per-CPU reference counting instead of rcuref_t

[Hillf Danton <hdanton@sina.com>]

On Tue, 8 Jul 2025 11:16:26 +0200 Sebastian Andrzej Siewior wrote:
> On 2025-07-08 16:56:39 [+0800], Hillf Danton wrote:
> > On Mon,  7 Jul 2025 16:36:22 +0200 Sebastian Andrzej Siewior wrote:
> > > +static bool futex_ref_get(struct futex_private_hash *fph)
> > > +{
> > > +	struct mm_struct *mm = fph->mm;
> > > +
> > > +	guard(rcu)();
> > > +
> > Like regular refcount_t, it is buggy to touch fph if futex_atomic drops
> > to 0. And more important guard(rcu) does not prevent it from dropping to 0.
> 
> What is your intention with this? There is an inc-if-not-zero to ensure

I am just simply wondering why get and put do not work without the rcu guard?

> this does not happen. And it has to drop to zero in order to get
> replaced.
> 
> > > +	if (smp_load_acquire(&fph->state) == FR_PERCPU) {
> > > +		this_cpu_inc(*mm->futex_ref);
> > > +		return true;
> > > +	}
> > > +
> > > +	return atomic_long_inc_not_zero(&mm->futex_atomic);
> > > +}
> > > +
> > > +static bool futex_ref_put(struct futex_private_hash *fph)
> > > +{
> > > +	struct mm_struct *mm = fph->mm;
> > > +
> > > +	guard(rcu)();
> > > +
> > > +	if (smp_load_acquire(&fph->state) == FR_PERCPU) {
> > > +		this_cpu_dec(*mm->futex_ref);
> > > +		return false;
> > > +	}
> > > +
> > > +	return atomic_long_dec_and_test(&mm->futex_atomic);
> > > +}
> 
> Sebastian