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

[Ben Greear <greearb@candelatech.com>]

On 06/27/2012 01:50 PM, Stephen Hemminger wrote:
> On Wed, 27 Jun 2012 22:39:01 +0200
> Eric Dumazet <eric.dumazet@gmail.com> wrote:
>
>> All sane SNMP applications are ready to cope with 32bits counters
>> wrapping.
>
> Actually that statement depends on the data rate. SNMP daemons work
> by polling at periodic intervals. The limit for detecting roll over depends
> on the rate and the interval. I believe the ubiquitous net-snmp code uses
> something a 30 second polling interval for lots of it's caches. This means
> it rolls over too fast at 10G. Polling faster can help but net-snmp is
> a pig about updates.
>
> I just realized the whole x32 (running 32 bit apps on 64 bit kernel) is broken
> for things like /proc/net/dev where 64 bit kernel will give 64 bit values and
> the 32 bit app (like net-snmp) is expecting unsigned long (32 bits).

It's worse than that:  Even on 64-bit kernels, counters that are returned by
netlink and /proc/net/dev as 64-bit may still wrap themselves at 32-bit
intervals.

I found that I just have to be very paranoid, and assume that if a '64-bit'
number wraps its high 32-bits between polls then it is really just a 32-bit
number and do that wrap properly (and poll more often).

Thanks,
Ben

-- 
Ben Greear <greearb@candelatech.com>
Candela Technologies Inc  http://www.candelatech.com