Re: [RFC net-next 14/15] devlink: add by-instance dump infra

[Jakub Kicinski <kuba@kernel.org>]

On Thu, 15 Dec 2022 10:11:03 +0100 Jiri Pirko wrote:
> Instead of having this extra list of ops struct, woudn't it make sence
> to rather implement this dumpit_one infra directly as a part of generic
> netlink code?

I was wondering about that, but none of the ideas were sufficiently
neat to implement :( There's a lot of improvements that can be done
in the core, starting with making more of the info structures shared
between do and dump in genl :( 

> Something like:
> 
>  	{
>  		.cmd = DEVLINK_CMD_RATE_GET,
>  		.doit = devlink_nl_cmd_rate_get_doit,
> 		.dumpit_one = devlink_nl_cmd_rate_get_dumpit_one,
> 		.dumpit_one_walk = devlink_nl_dumpit_one_walk,
>  		.internal_flags = DEVLINK_NL_FLAG_NEED_RATE,
>  		/* can be retrieved by unprivileged users */
>  	},

Growing the struct ops (especially the one called _small_) may be 
a hard sale for a single user. For split ops, it's a different story,
because we can possibly have a flag that changes the interpretation
of the union. Maybe.

I'd love to have a way of breaking down the ops so that we can factor
out the filling of the message (the code that is shared between doit
and dump). Just for the walk I don't think it's worth it.

I went in the same direction as ethtool because if over time we arrive
at a similar structure we can use that as a corner stone.

All in all, I think this patch is a reasonable step forward. 
But definitely agree that the genl infra is still painfully basic.