Re: [RFC PATCH 12/23] thermal: intel: hfi: Convert table_lock to use flags-handling variants

[Peter Zijlstra]

On Tue, Sep 27, 2022 at 01:34:07PM +0200, Peter Zijlstra wrote:
> On Fri, Sep 09, 2022 at 04:11:54PM -0700, Ricardo Neri wrote:
> 
> > --- a/drivers/thermal/intel/intel_hfi.c
> > +++ b/drivers/thermal/intel/intel_hfi.c
> > @@ -175,9 +175,10 @@ static struct workqueue_struct *hfi_updates_wq;
> >  static void get_hfi_caps(struct hfi_instance *hfi_instance,
> >  			 struct thermal_genl_cpu_caps *cpu_caps)
> >  {
> > +	unsigned long flags;
> >  	int cpu, i = 0;
> >  
> > -	raw_spin_lock_irq(&hfi_instance->table_lock);
> > +	raw_spin_lock_irqsave(&hfi_instance->table_lock, flags);
> >  	for_each_cpu(cpu, hfi_instance->cpus) {
> >  		struct hfi_cpu_data *caps;
> >  		s16 index;
> 
> ^^^^ Anti-pattern alert!
> 
> Now your IRQ latency depends on nr_cpus -- which is a fair fail. The
> existing code is already pretty crap in that it has the preemption
> latency depend on nr_cpus.
> 
> While I'm here looking at the HFI stuff, did they fix that HFI interrupt
> broadcast mess already? Sending an interrupt to *all* CPUs is quite
> insane.

Anyway; given the existence of this here loop; why not have something
like:

DEFINE_PER_CPU(int, hfi_ipc_class);

	class = // extract from HFI mess
	WRITE_ONCE(per_cpu(hfi_ipc_class, cpu), class);

And then have the tick use this_cpu_read(hfi_ipc_class)? No extra
locking required.