Re: [PATCH v2] l2tp: use per-cpu variables for u64_stats updates

[Rick Jones <rick.jones2@hp.com>]

On 06/27/2012 01:39 PM, Eric Dumazet wrote:
> On Wed, 2012-06-27 at 13:21 -0700, Rick Jones wrote:
>
>> It is a question of the speed of the communications more than the
>> bitness of the processor no?
>
> Why ?

The desire to have a greater than 32 bit counter stems from how quickly 
it will wrap, and how quickly it will wrap depends on the speed of 
communication.

> In 2012 or 2013, 64bits kernels are the norm, and 32bit the exception.

In servers and laptops I would be inclined to agree.  Elsewhere I am not 
so sure.   And while I don't know of its status, there is at least one 
hit on an RFC where, back in 2001, when a great many things were 32 bit, 
various 64-bit counters were added for L2TP Domain statistics:

http://tools.ietf.org/html/draft-ietf-l2tpext-l2tp-mib-03

(I think that became http://tools.ietf.org/html/rfc3371 in 2002)

So, it is possible people were simply painting a bikeshed, but 10 years 
ago, when the prevalence of 64 bit processors was not nearly so great, 
folks involved in L2TP MIB definitions found it worthwhile to make 
various counters 64-bit.

> Should we add complex code to l2tp just for being able to run it on
> 32bit kernels with 64bit stats ?
>
> Considering this code is buggy at the v1 & v2, I am really wondering.
>
> All sane SNMP applications are ready to cope with 32bits counters
> wrapping.

Once (maybe twice?) but no more than that.

> Machines that could wrap the 32bit counter several times per second are
> probably running on 64bit kernels.

I don't think it requires wrapping a 32 bit counter several times a 
second to warrant a > 32 bit counter.

> Also percpu stats are overkill unless a device is really meant to be
> used in // by many cpus.

I take it "//" is used as a shorthand for parallel?

rick