Re: [PATCH net-next v11 3/6] net: add a generation counter for dev->mc changes

[Yuyang Huang]

> To clarify: every I mentioned in my previous email was about possible
> follow-ups - not to be included here.

Understood, I will send them as separate patches.

> My understanding is that there is sort of a pending/implicit question
> from Jakub on v9: why are you doing this? is does it solve some
> specific scaling problem or is 'nl is nicer" thing?

I think the motivation question was on the v5 thread, and I replied
there as well.

My original motivation is to simplify some userspace hardware offload
code I am working on, and to finish moving "ip maddr" to netlink.

RTM_GETMULTICAST already reports IPv4 and IPv6 group memberships with
IFA_MC_USERS. dev->mc is the only part of the multicast state still
available only from /proc/net/dev_mcast, so every consumer that needs
the full picture keeps two paths: netlink for L3 and procfs text
parsing for L2. The offload code above is one such consumer. iproute2
is another: "ip maddr show" already gets IPv4/IPv6 over netlink but
still parses procfs for its link-layer entries. The iproute2 change to
drop that is ready to follow this series.

Also, compared to procfs, the dump provides a nicer API: it supports
per-device filtering, other netns through IFA_TARGET_NETNSID, and
reports concurrent changes with NLM_F_DUMP_INTR.

Thanks,

Yuyang