Re: [EXT] Re: [PATCH net-next v3 2/4] octeon_ep: PF-VF mailbox version support

[Leon Romanovsky]

On Mon, Dec 11, 2023 at 02:51:19PM +0000, Shinas Rasheed wrote:
> > > > This architecture design is unlikely to work in the real world unless
> > > > you control both PF and VF environment. Mostly PF is running some old
> > > > legacy distribution while VFs run more modern OS and this check will
> > > > prevent to run new driver in VF.
> > > >
> > > > Thanks
> > >
> > > Thanks for the review. This version validation only concerns regarding the
> > >control net API layer (which is used to communicate with
> > > the firmware). In the case you have described, this instead enables new VF
> > >drivers to atleast work atop legacy PF drivers (note legacy here still
> > > refers to PF drivers which support this backward compatibility), although
> > >they might not be able to use the latest control net functionalities that they
> > > have been enabled for.
> > 
> > The question what will be in X years from now, when you will have v100?
> > Will you fallback to v0 for backward compatibility?
> > 
> > >
> > > In the absence of such a backward compatibility, VF drivers would issue
> > >control net requests which PF drivers wouldn't know, only leading to logs of
> > > incompatibility errors and erroneous usage.
> > >
> > > Also again please note that this version compatibility only concerns the
> > > control net infrastructure and API (the control plane).
> > 
> > It doesn't matter, even in best scenario, you can't guarantee that code in VM
> > actually
> > implements version Y fully and will need to check correctness per-command
> > anyway.
> > 
> > Thanks
> 
> I'm afraid as to what else can be an alternative? The control net commands have to be decoded and passed by the PF driver for the VFs,
> as only the PFs have access to talk to firmware directly. The VF drivers do not have an alternative way to query control net APIs, and may fail 
> if the control net APIs they have are not even recognized by the PF to decode them.
> 
> Either VF commands which the PF can't support can be blocked at the source (by the equivalent PF-VF backward compatibility which will exist in VF drivers)
> by this negotiation, or we have to let commands come through and fail them, leading to just redundancy in terms of running code. I don't see how this negotiation in
> any way 'limit' the VF drivers.
> 
> As you said, in essence the VF drivers will have to fallback to v0 for backward compatibility if the native host uses some old OS having older PF drivers. If not, the 
> commands would come and fail anyways at the PF. Either way, it's an error case and this negotiation is just to decide if we are going to allow letting such commands in.

I don't know what netdev maintainers will do with this code, I just
pointed to this architecture/HW troublesome design.

Thanks

> 
> Thanks for your time