Re: [PATCH v4 resend 01/23] iidc: Introduce iidc.h

[Jason Gunthorpe <jgg@nvidia.com>]

On Mon, Apr 12, 2021 at 02:51:18PM +0000, Saleem, Shiraz wrote:

> > Where is ftype initialized?
> 
> Today it is just pf. But the upcoming Intel ethernet VF driver will
> set it to true.

Then it is dead code, don't send dead code to the kernel.

> > This cdev_info should just be a 'struct ice_pf *' and the "struct
> > iidc_core_dev_info" should be deleted entirely. You'll notice this
> > ends up looking nearly exactly like mlx5 does after this.
> 
> It was intentionally designed to be core device object carving out
> only pieces of the PF information required by the rdma driver. The
> next near-term PCI driver using IIDC can also this. Why expose the
> whole PF? This is a design choice no? Why do we need to follow mlx5?

When you get around to building your multi-driver API it should be
structured so it doesn't have a de-normalization of the data - don't
copy from one authoritative struct to some other struct just to get
some weird information hiding.

The PF driver should be a subclass if your "generic" driver and
directly embed some struct like this as the singular canonical source
of information, not be duplicated.

> I don't follow what the hackery is. Just because we use cdev_info in
> the .ops callbacks as opposed to ice_pf?

There are too many signs to ignore:
 - The obfuscated extensible structs being passed into ops that are
   only encoding a couple function call parameters
 - The ops that only have one implementation
 - The struct that is a complete copy of a different, but "internal",
   struct

You do stuff like this to make stable ABIs. This is forbidden by Linus
for in-kernel APIs, and it is not the kernel style in general to code
like this.

Jason