Re: [PATCH net-next RFC v3 01/14] devlink: Add reload action option to devlink reload command

[Moshe Shemesh]

On 9/4/2020 10:56 PM, Jakub Kicinski wrote:
> External email: Use caution opening links or attachments
>
>
> On Fri, 4 Sep 2020 11:04:50 +0200 Jiri Pirko wrote:
> > Thu, Sep 03, 2020 at 09:47:19PM CEST, kuba@xxxxxxxxxx wrote:
> > > On Thu, 3 Sep 2020 07:57:29 +0200 Jiri Pirko wrote:
> > > > Wed, Sep 02, 2020 at 05:30:25PM CEST, kuba@xxxxxxxxxx wrote:
> > > > > On Wed, 2 Sep 2020 11:46:27 +0200 Jiri Pirko wrote:
> > > > > > > ? Do we need such change there too or keep it as is, each action by itself
> > > > > > > and return what was performed ?
> > > > > > Well, I don't know. User asks for X, X should be performed, not Y or Z.
> > > > > > So perhaps the return value is not needed.
> > > > > > Just driver advertizes it supports X, Y, Z and the users says:
> > > > > > 1) do X, driver does X
> > > > > > 2) do Y, driver does Y
> > > > > > 3) do Z, driver does Z
> > > > > > [
> > > > > > I think this kindof circles back to the original proposal...
> > > > > Why? User does not care if you activate new devlink params when
> > > > > activating new firmware. Trust me. So why make the user figure out
> > > > > which of all possible reset option they should select? If there is
> > > > > a legitimate use case to limit what is reset - it should be handled
> > > > > by a separate negative attribute, like --live which says don't reset
> > > > > anything.
> > > > I see. Okay. Could you please sum-up the interface as you propose it?
> > > What I proposed on v1, pass requested actions as a bitfield, driver may
> > > perform more actions, we can return performed actions in the response.
> > Okay. So for example for mlxsw, user might say:
> > 1) I want driver reinit
> >      kernel reports: fw reset and driver reinit was done
> > 2) I want fw reset
> >      kernel reports: fw reset and driver reinit was done
> > 3) I want fw reset and driver reinit
> >      kernel reports: fw reset and driver reinit was done
> Yup.
>
> > > Then separate attribute to carry constraints for the request, like
> > > --live.
> > Hmm, this is a bit unclear how it is supposed to work. The constraints
> > apply for all? I mean, the actions are requested by a bitfield.
> > So the user can say:
> > I want fw reset and driver reinit --live. "--live" applies to both fw
> > reset and driver reinit? That is odd.
> The way I was thinking about it - the constraint expresses what sort of
> downtime the user can accept. So yes, it'd apply to all. If any of the
> reset actions does not meet the constraint then error should be
> returned.
>
> In that sense I don't like --live because it doesn't really say much.
> AFAIU it means 1) no link flap; 2) < 2 sec datapath downtime; 3) no
> configuration is lost in kernel or device (including netdev config,
> link config, flow rules, counters etc.). I was hoping at least the
> documentation in patch 14 would be more precise.


Actually, while writing "no-reset" or "live-patching" I meant also no 
downtime at all and nothing resets (config, rules ... anything), that 
fits mlx5 live-patching.

However, to make it more generic,  I can allow few seconds downtime and 
add similar constrains as you mentioned here to "no-reset". I will add 
that to the documentation patch.

> I think you're saying that it's strange to express that as a constraint
> because internally it maps to one of two fw reset types. And there is
> only one driver reinit procedure. But I don't think that the
> distinction of reset types is valuable to the user. What matters is if
> application SLAs will be met or not.
>
> I assume that deeper/longer reset is always less risky and would be
> preferred unless constraint is specified.
>
> > > I'd think the supported actions in devlink_ops would be fine as a
> > > bitfield, too. Combinations are often hard to capture in static data.