64 bit packet counters

[Jesper Skriver]

On Sat, Nov 08, 2003 at 10:13:48AM +0100, Harti Brandt wrote:

> On Fri, 7 Nov 2003, Alex Hoff wrote:
>
> AH>Hi,
> AH>
> AH>We are attempting to implement the IF-MIB, which requires the
> AH>use of 64 bit packet counters and the differentiation between
> AH>multicast and broadcast pkts. Since changing the if_data (by
> AH>adding new counters and changing the existing to u_int64) is a bad
> AH>idea, does anyone have any good ideas on how to do this? I was
> AH>thinking of tacking on a new struct (lets call it ifx_data) on at
> AH>the end of the current if_net struct with the appropriate counters
> AH>(i/opacket, i/obyte, i/obcast, i/omcast). Apart from having to do
> AH>a little double counting is there any obvious pitfals with this
> AH>approach? Does anyone have an better ideas? Is there currently any
> AH>plans to update the network stack to handle this properly?
>
> You may lookup the discussions in the mailing lists. As far as
> I remember the problem with 64 bit counting was that this needs
> locks because not on all architectures you have atomic 64bit add
> operations. A simple method that does not involve kernel changes (and
> that I plan to implement in my snmp daemon) is to periodically monitor
> the counters (depending on the interface speed) and detect wraps in
> the daemon.

What is done in other places is to have normal 32 bit counters that are
incremented per packet, and have a background task/thread to update the
64 bit counters from the 32 bit counters.

That way, we avoid the locking issue per packet.

/Jesper

-- 
Jesper Skriver, jesper(at)skriver(dot)dk  -  CCIE #5456

One Unix to rule them all, One Resolver to find them,
One IP to bring them all and in the zone to bind them.