On 16/02/16 20:26, roopa wrote:
On 2/16/16, 7:41 AM, David Miller wrote:
Statistics not provided via netlink are useless in real installations.
In fact I would say to forego the proc interface entirely, it's a second
class citizen for statistics gathering and has a non-triviel per-device
cost for instantiation.
+1
I agree with the cost too.
Robert, I was thinking of responding to your series to add netlink stats for
AF_MPLS in
rtnl_af_ops (similar to IFLA_INET6_STATS). But, soon realized that it is
currently used by ipv6 alone
and it also ended up with a skip filter (RTEXT_FILTER_SKIP_STATS). So,
extending that interface for
mpls is not the right thing to do either.
ipv4 doesn't have per-interface stats, so the fact that it doesn't use
fill_link_af to export stats doesn't really add much to the argument.
The real issue with the IFLA_INET6_STATS mechanism is that they are
included in netlink broadcasts, not just netlink unicasts and there is
no way of filtering for broadcasts at the moment.
There is work being done for a separate netlink infrastructure for stats.
I would wait for that infrastructure to be ready to add mpls stats. It should
be available soon.
Great, any details on what it would look like? Can I assist in any way
in the development?
In the mean time, I'll rebase and resubmit the ip ttl propagation patch
separately.
Thanks,
Rob
