Re: 6.11/regression/bisected - The commit c1385c1f0ba3 caused a new possible recursive locking detected warning at computer boot.

[Jonathan Cameron <Jonathan.Cameron@Huawei.com>]

On Fri, 26 Jul 2024 18:14:24 +0100
Jonathan Cameron <Jonathan.Cameron@Huawei.com> wrote:

> On Fri, 26 Jul 2024 18:26:01 +0200
> Thomas Gleixner <tglx@linutronix.de> wrote:
> 
> > On Thu, Jul 25 2024 at 18:13, Jonathan Cameron wrote:  
> > > On Tue, 23 Jul 2024 18:20:06 +0100
> > > Jonathan Cameron <Jonathan.Cameron@Huawei.com> wrote:
> > >    
> > >> > This is an interesting corner and perhaps reflects a flawed
> > >> > assumption we were making that for this path anything that can happen for an
> > >> > initially present CPU can also happen for a hotplugged one. On the hotplugged
> > >> > path the lock was always held and hence the static_key_enable() would
> > >> > have failed.    
> > 
> > No. The original code invoked this without cpus read locked via:
> > 
> > acpi_processor_driver.probe()
> >    __acpi_processor_start()
> >        ....
> > 
> > and the cpu hotplug callback finds it already set up, so it won't reach
> > the static_key_enable() anymore.
> >   
> > > One bit I need to check out tomorrow is to make sure this doesn't race with the
> > > workfn that is used to tear down the same static key on error.    
> > 
> > There is a simpler solution for that. See the uncompiled below.  
> 
> Thanks.  FWIW I got pretty much the same suggestion from Shameer this
> morning when he saw the workfn solution on list. Classic case of me
> missing the simple solution because I was down in the weeds.
> 
> I'm absolutely fine with this fix.
Hi Thomas,

I tested it on an emulated setup with your changes on top of
mainline as of today and the issue is resolved.

Would you mind posting a formal patch? Or I can do it on Monday if that's
easier for you.

Thanks

Jonathan

> 
> Mikhail, please could you test Thomas' proposal so we are absolutely sure
> nothing else is hiding.
> 
> Tglx's solution is much less likely to cause problems than what I proposed because
> it avoids changing the ordering.
> 
> Jonathan
> 
> 
> 
> 
> > 
> > Thanks,
> > 
> >         tglx
> > ---
> > diff --git a/arch/x86/kernel/cpu/aperfmperf.c b/arch/x86/kernel/cpu/aperfmperf.c
> > index b3fa61d45352..0b69bfbf345d 100644
> > --- a/arch/x86/kernel/cpu/aperfmperf.c
> > +++ b/arch/x86/kernel/cpu/aperfmperf.c
> > @@ -306,7 +306,7 @@ static void freq_invariance_enable(void)
> >  		WARN_ON_ONCE(1);
> >  		return;
> >  	}
> > -	static_branch_enable(&arch_scale_freq_key);
> > +	static_branch_enable_cpuslocked(&arch_scale_freq_key);
> >  	register_freq_invariance_syscore_ops();
> >  	pr_info("Estimated ratio of average max frequency by base frequency (times 1024): %llu\n", arch_max_freq_ratio);
> >  }
> > @@ -323,8 +323,10 @@ static void __init bp_init_freq_invariance(void)
> >  	if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL)
> >  		return;
> >  
> > -	if (intel_set_max_freq_ratio())
> > +	if (intel_set_max_freq_ratio()) {
> > +		guard(cpus_read_lock)();
> >  		freq_invariance_enable();
> > +	}
> >  }
> >  
> >  static void disable_freq_invariance_workfn(struct work_struct *work)
> > 
> >   
>