Re: [PATCH net-next] rtnetlink: Honour NLM_F_ECHO flag in rtnl_{new, set}link

[Jakub Kicinski <kuba@kernel.org>]

Let me clarify one more time in case Hangbin is waiting for 
the discussion to resolve...

On Fri, 23 Sep 2022 10:43:53 +0200 Nicolas Dichtel wrote:
> Le 22/09/2022 à 16:51, Guillaume Nault a écrit :
> > I just don't see any way to pass a handle back to user space at the
> > moment. The echo mechanism did that and was generic to all netlink
> > families (as long as nlmsg_notify() was called with the right
> > parameters).

In NEWLINK, right? In NEWLINK there is no way to pass it back 
at the moment. A newly added command can just respond with the handle
always. The problem with NEWLINK is that it _used to_ not respond so 
we can't make it start responding because it will confuse existing user
space.

At the protocol level NEW is no different than GET, whether it sends 
a response back is decided by whoever implements the command.

So yes, for NEWLINK we need a way to inform the kernel that user space
wants a reply. It can be via ECHO, it could be via a new attr.

What I'm trying to argue about is *not* whether NEWLINK should support
ECHO but whether requiring ECHO to get a response for newly added
CREATE / NEW commands is a good idea. I think it is not, and new
commands should just always respond with the handle.

My main concern with using ECHO is that it breaks the one-to-one
relationship between a request and a response. There may be multiple
notifications generated due to a command, and if we want to retain 
the "ECHO will loop back to you all resulting notifications" semantics,
which I think we should, then there can be multiple "responses".

This also has implications for the command IDs used in notifications.
A lot of modern genl families use different IDs for notifications to
make it easily distinguishable from responses.


I guess tl;dr is Hangbin should go forward with the v2, and I should
document the expectations clearly..