Hello, I have the same opinion. Silent saturation (with log). The goal is to have the same behavior on all BGP implementation. Alexandre
On Dec 2, 2020, at 6:45 PM, Daniel Suchy <danny@danysek.cz> wrote:
Hello, commercial BGP implementations (Juniper, Nokia, Cisco) usually implements silent saturation based on your point (2) in such cases, Quagga does this too (see route_value_adjust function [1]; which handles these overflows). It's not limited to locpref, also other attibutes can be modified by addition/subtraction.
I suggest to follow this practice also in Bird code, as this behavior is commonly expected here.
Length of relevant attributes are defined in RFC and I don't expect any move from 32b to 64b integers here - as this will break compability between speakers.
- Daniel
[1] https://github.com/Quagga/quagga/blob/88d6516676cbcefb6ecdc1828cf59ba3a6e5fe...
On 12/1/20 5:51 PM, Ondrej Zajicek wrote:
There are at least three reasonable behaviors for fixed-range arithmetics: 1) silent overflow (current behavior) 2) silent saturation (as you suggested) 3) explicit (logged) error Not really sure which is better. (1) is likely least astonishment (for fixed-range), (2) is most useful when overflows are intentional, (3) is most useful when overflows are unintentional. I have minor preference for (3), but no strong preference either way to change anything. Also note that in cases when fixed-range attributes have lower range than 32bits (i.e. 16bit preference and 24bit ospf_metrics), we generate explicit error on overflow on assignment. Or perhaps we should just move to arbitrary-range integers .. many new RFCs already use 64bit attributes anyways.