Re: [PATCH net-next 4/5] devlink: embed driver's priv data callback param into devlink_resource

[Jiri Pirko <jiri@resnulli.us>]

Mon, Aug 12, 2024 at 01:50:06PM CEST, przemyslaw.kitszel@intel.com wrote:
>On 8/9/24 13:02, Jiri Pirko wrote:
>> Fri, Aug 09, 2024 at 04:41:50AM CEST, kuba@kernel.org wrote:
>> > On Wed, 7 Aug 2024 08:49:57 +0200 Jiri Pirko wrote:
>> > > > 	lockdep_assert_held(&devlink->lock);
>> > > > 
>> > > > 	resource = devlink_resource_find(devlink, NULL, resource_id);
>> > > > -	if (WARN_ON(!resource))
>> > > > +	if (WARN_ON(!resource || occ_priv_size > resource->priv_size))
>> > > 
>> > > Very odd. You allocate a mem in devl_resource_register() and here you
>> > > copy data to it. Why the void pointer is not enough for you? You can
>> > > easily alloc struct in the driver and pass a pointer to it.
>> > > 
>> > > This is quite weird. Please don't.
>> > 
>> > The patch is a bit of a half measure, true.
>
>Another option to suit my wants would be to just pass resource_id to the
>callbacks, would you accept that?

Why, the callback is registered for particular resource. Passing ID is
just redundant.


>
>> > 
>> > Could you shed more light on the design choices for the resource API,
>> > tho? Why the tying of objects by driver-defined IDs? It looks like
>> 
>> The ids are exposed all the way down to the user. They are the same
>> across the reboots and allow user to use the same scripts. Similar to
>> port index for example.
>> 
>> 
>> > the callback for getting resources occupancy is "added" later once
>> > the resource is registered? Is this some legacy of the old locking
>> > scheme? It's quite unusual.
>
>I did such review last month, many decisions really bother me :F, esp:
>- whole thing is about limiting resources, driver asks HW for occupancy.

Can you elaborate what's exactly wrong with that?


>
>Some minor things:
>- resizing request validation: parent asks children for permission;
>- the function to commit the size after the reload is named
>  devl_resource_size_get().
>
>From the user perspective, I'm going to add a setter, that will be
>another mode of operation (if compared to the first thing on my complain
>list):
>+ there is a limit that is constant, and driver/user allocates resource
>  from such pool.
>
>> 
>> It's been some while since I reviewed this, but afaik the reason is that
>> the occupancy was not possible to obtain during reload, yet the resource
>> itself stayed during reload. This is now not a problem, since
>> devlink->lock protects it. I don't see why occupancy getter cannot be
>> put during resource register, you are correct.
>> 
>I could add that to my todo list

Cool.