[Bug 19619] Re: Interface statistics wrong

Alexandre Otto Strube surak at surak.eti.br
Wed Oct 11 08:55:21 UTC 2006

The upstream bug was rejected as being duplicate, and the original one
was closed, because it's not a bug from gnome-nettool, but from the
kernel. Quoting the mailing list:

>> Surely I'm not the only person to have encountered this issue by now; When
>> the kernel records the tx/rx packet counts and their respective sizes,
>> they rollover very quickly on 32-bit hardware. To put this in better
>> perspective - After roughly 4 Gb worth of network traffic, the Kernel
>> tends to restart at 0 for the rx/tx stats in net_device_stat.

>> Could/should it be changed such that net_device_stat uses larger types for
>> these fields? 4 Gb is not a whole lot of traffic, its discouraging to
>> check ifconfig only a few hours later to find out the kernel has
>> rolled-over several times since you last checked. 32-bit servers with
>> large volumes of traffic are simply swamped with rollovers, rendering any
>> practical use for net_device_stat's rx/tx byte count out of the question.

>> I'm not capable of patching this myself, so it's more along the
>> lines of a comment/suggestion... <smirks>

>> Any insight would be appreciated,

> Lots of replies, eh?  or maybe some private ones?

> This has been brought up a few times previously, so you could
> check the archives to see what some other answers are.

> One point is that an SNMP agent could do this for you.
> Are you using one?

> Linux net device counters don't have to track RFCs, but another
> observation is that RFCs require 64-bit ("high capacity")
> counters in cases where a 32-bit counter might overflow within
> 1 hour (IIRC).

> And last, to try to help you understand, one of the reasons
> that it hasn't already been done is that 64-bit
> arithmetic on a 32-bit processor requires using a lock
> to make it atomic, and that's expensive.

Interface statistics wrong

More information about the desktop-bugs mailing list