On 3/28/19 1:09 PM, Jakub Kicinski wrote:

On Thu, 28 Mar 2019 17:34:39 +0100, Jiri Pirko wrote:

quoted

Thu, Mar 28, 2019 at 10:53:47AM CET, mkubecek@suse.cz wrote:

quoted

On Thu, Mar 28, 2019 at 10:21:26AM +0100, Jiri Pirko wrote:  

quoted

Wed, Mar 27, 2019 at 11:25:54PM CET, mkubecek@suse.cz wrote:  

quoted

I'm all for implementing new features which are are related to physical
device (ASIC) rather than network interface only in devlink (at the
level of kernel-userspace interface). But for features already provided
by ethtool (userspace utility) I can't help seeing the state of devlink
support in NIC drivers as a serious blocker.  

What I'm thinking about at for some time now would be en explicit
default devlink and devlink_port registration for every netdev for
drivers that does not support devlink themselves. I need to think about
it more, but it seems doable. Then we can hang appropriate things there
and make the ethtoolnl/devlink separation clear. I believe we need to do
it.  

That sounds great, such "generic devlink" implementation would be a way
around. Kernel could then emulate features which are not implemented by
an actual devlink handler (i.e. "generic devlink" is used or particular
handler is missing) by falling back to ethtool_ops handler so that
userspace could rely on devlink API for things like device information,
various dumps, flashing etc. without losing anything.  

Yep. Plan to do that next week.

The problem with the main part of dev info - fw_version - is that it is
often overloaded in drivers and becomes impossible to parse for users.
I'd rather we didn't dump that nasty chaos in devlink info and let it
die with ethtool IOCTL.

Flashing can also be handled at user space tool level.

Yes, I really like how sfc does it which is to expose MTD devices and
certainly have a custom flashing tool, that should be the way to go IMHO.
-- 
Florian