Re: [PATCH v2 4/5] dts: add ability to start/stop testpmd ports

[Jeremy Spewock <jspewock@iol.unh.edu>]

On Fri, Sep 6, 2024 at 12:46 PM Luca Vizzarro <Luca.Vizzarro@arm.com> wrote:
>
> On 23/08/2024 13:16, Juraj Linkeš wrote:
> > As Jeremy mentioned, adding the verify argument may be worthwhile, but
> > maybe only if we actually identify a usecase where we wouldn't want to
> > do the verification.
>
> Yeah, as I pointed out, it feels unlikely to pretend that they are
> started (or stopped). I think that could cause my unpredicted behaviour
> and confusion. But also as said already, it can always be added on demand.
>

I agree that we can save this as something to add as we need it and
that there could be some confusion to pretending the ports are in a
certain state. Since we actually save the state in testpmd, we
definitely want that to be accurate. The main thing I was thinking
about this for however was less about the state tracking and more
about what the expectation is for developers when they specifically
specify they do not want a method to be verified. There aren't many
cases where you wouldn't want to verify the outcome of a command, but
I would imagine the two reasons you would do so is either you don't
care about the result, or you specifically don't want to be disruptive
to the rest of the run if this command fails (like if the framework
has a testpmd shell running for some kind of process but doesn't want
to end the run if it fails). When we enforce the verification of some
part of the process like stopping the port in this case, that
assumption about not verifying changes. The second reason I mentioned
is why I believe capabilities checking specifies no-verify when it
updates the MTU, for example. I imagine we don't want the test run to
end if a testpmd method fails its check with capabilities, we would
want the capability to just be unsupported. Of course the developer
could catch the exception themselves, but in that case I think it
would make more sense for catching the exception to be the primary way
to avoid verification and we could just remove the parameter all
together so there isn't this edge-case for avoiding verification with
decorated methods. I guess even the static parameter that specifies
whether or not to verify doesn't really address this problem either,
haha. Maybe it would even make more sense for the decorator to assume
that it shouldn't throw an exception if the ports state changing
failed (but still avoid updating the port state) and let the verifying
process be completely down to if the method being decorated succeeded
or not with the ports in whatever state they were at the time since, I
would assume, if the ports fail to update their state the method will
also fail anyway.

All-in-all however, there likely won't be many cases where the ports
actually fail to stop in the first place, and not verifying in general
is uncommon. If what I mentioned above makes sense and we're fine with
the slight change in thinking, I think it is fine to merge this and
just keep in the back of our minds that this is the case, and maybe
either document somewhere that you have to catch the exception with
decorated methods, or just remove the verify parameter from decorated
methods all together and enforce that they must be verified. As
mentioned, it's not like we are forcing ourselves to stick with it,
adding the parameter is pretty easy to do on demand.

> > I also have two other points:
> > 1. Do we want a decorator that does both - starting and stopping? Is the
> > idea to have all methods that require a certain state to be decorated
> > and in that case we don't need this? E.g. if there are multiple
> > configuration steps, only the first one stops the ports (if started) and
> > we start the ports only when a method needs that (such as starting pkt
> > fwd)? I think this workflow should be documented in the class docstring
> > (the important part being that starting and stopping of ports is done
> > automatically).
>
> The way I envisioned these decorators is by using a lazy approach. I
> don't think it may be right to eagerly restart ports after stopping
> them, because maybe we want to run another command after that will stop
> them again. So only start and stop as needed. Ideally every port that
> requires a specific state of the ports need to be decorated. I went
> through all the methods in the class to check which would and which
> wouldn't, and it seems alright like this.
>
> > 2. Do we want decorators that start/stop just one port? I guess this is
> > basically useless with the above workflow.
>
> Would it actually matter? We are not really using the other ports in
> parallel here I believe. May bring unnecessary complexity. I am thinking
> that maybe if needed it could be added later.