Re: [PATCH v2 2/4] percpu_stats: Enable 64-bit counts in 32-bit architectures

[Waiman Long]

On 04/08/2016 12:47 PM, Tejun Heo wrote:
> Hello, Waiman.
>
> On Fri, Apr 08, 2016 at 12:16:20PM -0400, Waiman Long wrote:
> > +/**
> > + * __percpu_stats_add - add given count to percpu value
> > + * @pcs : Pointer to percpu_stats structure
> > + * @stat: The statistics count that needs to be updated
> > + * @cnt:  The value to be added to the statistics count
> > + */
> > +void __percpu_stats_add(struct percpu_stats *pcs, int stat, int cnt)
> > +{
> > +	/*
> > +	 * u64_stats_update_begin/u64_stats_update_end alone are not safe
> > +	 * against recursive add on the same CPU caused by interrupt.
> > +	 * So we need to set the PCPU_STAT_INTSAFE flag if this is required.
> > +	 */
> > +	if (IS_STATS64(pcs)) {
> > +		uint64_t *pstats64;
> > +		unsigned long flags;
> > +
> > +		pstats64 = get_cpu_ptr(pcs->stats64);
> > +		if (pcs->flags&  PCPU_STAT_INTSAFE)
> > +			local_irq_save(flags);
> > +
> > +		u64_stats_update_begin(&pcs->sync);
> > +		pstats64[stat] += cnt;
> > +		u64_stats_update_end(&pcs->sync);
> > +
> > +		if (pcs->flags&  PCPU_STAT_INTSAFE)
> > +			local_irq_restore(flags);
> > +
> > +		put_cpu_ptr(pcs->stats64);
> > +	}
> > +}
> Heh, that's a handful, and, right, u64_stats needs separate irq
> protection.  I'm not sure.  If we have to do the above, it's likely
> that it'll perform worse than percpu_counter on 32bits.  On 64bits,
> percpu_counter would incur extra preempt_disable/enable() operations
> but that comes from it not using this_cpu_add_return().  I wonder
> whether it'd be better to either use percpu_counter instead or if
> necessary extend it to handle multiple counters.  What do you think?
>
> Thanks.

Yes, I think it will be more efficient to use percpu_counter in this 
case. The preempt_disable/enable() calls are pretty cheap. Once in a 
while, you need to take the lock and update the global count. How about 
I change the 2nd patch to use percpu_counter internally when 64-bit 
counts are needed in 32-bit archs, but use the regular percpu counts on 
64-bit archs? If you are OK with that, I can update the patch accordingly.

Cheers,
Longman