Re: [PATCH net-next] mld: add missing rtnl_lock() in do_ipv6_getsockopt()

On Tue, Mar 30, 2021 at 6:02 PM Taehee Yoo [off-list ref] wrote:

On 3/31/21 12:40 AM, Eric Dumazet wrote:
 > This seems a serious regression compared to old code (in net tree)
 >
 > Have you added RTNL requirement in all this code ?
 >
 > We would like to use RTNL only if strictly needed.

Yes, I agree with you.
This patchset actually relies on existed RTNL, which is
setsockopt_needs_rtnl().
And remained RTNL was replaced by mc_lock.
So, this patchset actually doesn't add new RTNL except in this case.

Fortunately, I think It can be replaced by RCU because,
1. ip6_mc_msfget() doesn't need the sleepable functions.
2. It is not the write critical section.
So, RCU can be used instead of RTNL for ip6_mc_msfget().
How do you think about it?

Yes please, do not add RTNL here if we can avoid it.

Otherwise some applications will slow down the whole stack, even with
different containers/netns.

(There is a single RTNL for the whole machine)