On Thu, Jul 06, 2017 at 09:53:46PM +0000, Casey Leedom wrote:
> | From: Jakub Kicinski <kubak...@wp.pl>
> | Sent: Thursday, July 6, 2017 12:02 PM
> |
> | IMHO if something gets replugged all the settings should be reset.
> | I feel that it's not entirely unlike replugging a USB adapter. Perhaps
> | we should introduce some (devlink) notifications for SFP module events
> | so userspace can apply whatever static user config it has?
>
> Absolutely a valid approach. As are all of the ones I outlined.
>
> But, and far more importantly, ideally _*ANY*_ such decision is made at an
> architectural level to apply to all Link Parameters and Vendor Products.
> The last thing a user wants to deal with is a hodge-podge of different
> policies for different adapters from different vendors.
Yes.
SFP needs to becomes a Linux device, similar to Copper PHYs are Linux
devices. With some core code which all drivers can use, implement
ethtool --dump-module-eeprom, report speeds to the MAC using
adjust_link, etc..
> how do users conceive of a "Port"?
For a user, it is something they configure via /etc/network/interfaces
and then use ifup/ifdown on.
> I.e. when a user requests that a particular
> Link Parameter be applied to a Port, are they thinking that it only applies
> to the current instantaneous combination of Adapter Transceiver Module Cage
> + Transceiver Module? Or do they conceptualize a "Port" as being a higher
> level entity?
>
> Or, let's make it Very Concrete with a specific example:
>
> 1. User applies some set of Link Parameters.
>
> 2. User attempts to bring Link up but it doesn't come up.
So these are effectively one step for the user, since the
configuration goes into /etc/network/interfaces, and it is only when
ifup is used is it applied. If the configuration is not valid, at this
point in time, i would expect ifup to give an error message.
> 3. User decides to try a different cable on the grounds that the first
> cable may be bad.
>
> 4. New cable is accidentally of a completely different type with completely
> different subsequent Physical Port Capabilities, not capable of supporting
> the user's selected Link Parameters.
And this is where it gets interesting, as you say. We are into a
hotplug model.
I think you also need to define 'cable' here. I assume you are not
talking about a piece of CAT 5 or glass fibre. You mean something
which is active. You are putting a different module into the SFP cage.
The extreme model would be, if you pull the module out, the whole
netdev is hot-unplugged. Plug a different modules in, the netdev is
hot-plugged. The user has to ifup it again, and would get an error
message if the configuration is invalid.
But i think this is too extreme.
I think the sfp device needs to give a hotplug event on unplug/plug.
A hot-unplug would result in an ifdown. And within the kernel, the
netdev is set down. If there is an "allow-hotplug" statement in
/etc/network/interfaces, on hot-plug, udev would try to ifup and get
an error and it will stay down. Without the "allow-hotplug" the
interface remains configured down until the user does an ifup and
would see an error message if the configuration is invalid.
Andrew
