In snmp_mibII module assume that virtual interfaces can work at maximum speed available in the system

[ae]

Summary

bsnmpd follows RFC 2863 and provides 64-bit counters only when interface's ifi_baudrate is big enough.
Virtual interfaces like gif(4), gre(4), enc(4), lo(4), etc., usually have zero baudrate, and bsnmpd doesn't provide 64-bit counters for them.
This patch changes this - if in a system are present some HighSpeed interfaces, it is assumed, that virtual interfaces also capable to work at such speed.

[bz]

Hmm I tried to add syrinx but she's not on Phab it seems.

Other question: why do we not set baudrate on the virtual interfaces?

[ae]

In D15112#318277, @bz wrote:

Other question: why do we not set baudrate on the virtual interfaces?

I think first of it is historically, then it is because we don't know which value to set :)
Also there is related problem with LACP lagg: https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=193953

[ae]

Maybe better solution is just provide 64-bit counters for all interfaces? It seems net-snmpd does so.

[harti]

In D15112#318283, @ae wrote:

Maybe better solution is just provide 64-bit counters for all interfaces? It seems net-snmpd does so.

I did not do this in the past because it was not possible to add or increment 64-bit counters atomically. Did this change? Instead of providing bad values I decided to follow the words of the RFC.

[glebius]

I did not do this in the past because it was not possible to add or increment 64-bit counters atomically. Did this change? Instead of providing bad values I decided to follow the words of the RFC.

Today kernel achieves that. The getifaddrs API guarantees to report counter values to userland atomically.

[harti]

How does it do that? The problem is whether a 64-bit increment or add is atomic with regard to a read from another CPU.

[eugen_grosbein.net]

In D15112#323137, @harti wrote:

How does it do that? The problem is whether a 64-bit increment or add is atomic with regard to a read from another CPU.

This is described in counter(9) manual page.

[harti]

Ah, thanks. This looks fine. Then we really could go for 64-bit counters for all interfaces.