Re: [PATCH v2 1/3] sched: Add helper kstat_cpu_softirqs_sum()

[Paul E. McKenney]

On Fri, Oct 28, 2022 at 10:38:15AM +0800, Leizhen (ThunderTown) wrote:
> 
> 
> On 2022/10/28 3:04, Elliott, Robert (Servers) wrote:
> > 
> >> Similar to kstat_cpu_irqs_sum(), it counts the sum of all software
> >> interrupts on a specified CPU.
> >>
> >> diff --git a/include/linux/kernel_stat.h b/include/linux/kernel_stat.h
> >> @@ -67,6 +67,17 @@ static inline unsigned int kstat_softirqs_cpu(unsigned int irq, int cpu)
> >>         return kstat_cpu(cpu).softirqs[irq];
> >>  }
> >>
> >> +static inline unsigned int kstat_cpu_softirqs_sum(int cpu)
> >> +{
> >> +	int i;
> >> +	unsigned int sum = 0;
> >> +
> >> +	for (i = 0; i < NR_SOFTIRQS; i++)
> >> +		sum += kstat_softirqs_cpu(i, cpu);
> >> +
> >> +	return sum;
> >> +}
> > 
> > In the function upon which this is based:
> > 
> > struct kernel_stat {
> >         unsigned long irqs_sum;
> >         unsigned int softirqs[NR_SOFTIRQS];
> > };
> > 
> > static inline unsigned int kstat_cpu_irqs_sum(unsigned int cpu)
> > {
> >         return kstat_cpu(cpu).irqs_sum;
> > }
> > 
> > kstat_cpu_irqs_sum returns an unsigned long as an unsigned int, which
> > could cause large values to be truncated. Should that return
> > unsigned long? The only existing caller is fs/proc/stat.c which
> 
> This should be a mistake on:
> commit f2c66cd8eeddedb4 ("/proc/stat: scalability of irq num per cpu")
> 
> I'll correct it to "unsigned long" in the next version. Thanks.
> 
> > puts it into a u64:
> >         u64 sum = 0;
> >         ...
> >         sum             += kstat_cpu_irqs_sum(i);
> > 
> > The softirqs field is an unsigned int, so the new function doesn't have
> > this inconsistency.
> 
> OK.
> 
> To be honest, I did the math. CONFIG_HZ=250
> 2^32 / 250 / 3600 / 24 / 365 = 0.545 < 1 year

For this to be a problem, our RCU CPU stall warning would have to be
for a months-long grace period, even on systems with HZ=1000.  In almost
all cases, the system would have OOMed long before then.

> So, in theory, for those 32-bit processors, we should use "unsigned long long".
> Of course, from a programming point of view, 64-bit consists of two 32-bits,
> and there is an atomicity problem. I think that's probably why members of
> struct kernel_stat don't use u64.
> 
> However, it seems that the type of member softirqs can currently be changed to
> unsigned long. So, at least on a 64-bit processor, it won't have a count
> overflow problem.

An unsigned long should suffice.  ;-)

							Thanx, Paul