Re: [PATCH v5 net-next 02/12] net-shapers: implement NL get operation

[Jakub Kicinski <kuba@kernel.org>]

On Fri, 30 Aug 2024 17:43:08 +0200 Paolo Abeni wrote:
> Please allow me to put a few high level questions together, to both 
> underline them as most critical, and keep the thread focused.
> 
> On 8/30/24 03:20, Jakub Kicinski wrote:
>  > This 'binding' has the same meaning as 'binding' in TCP ZC? :(  
> 
> I hope we can agree that good naming is difficult. I thought we agreed 
> on such naming in the past week’s discussion. The term 'binding' is 
> already used in the networking stack in many places to identify 
> different things (i.e. device tree, socket, netfilter.. ). The name 
> prefix avoids any ambiguity and I think this a good name, but if you 
> have any better suggestions, this change should be trivial.

Ack. Maybe we can cut down the number of ambiguous nouns elsewhere:

maybe call net_shaper_info -> net_shaper ?

maybe net_shaper_data -> net_shaper_hierarchy ?

>  > I've been wondering if we shouldn't move this lock
>  > directly into net_device and combine it with the RSS lock.
>  > Create a "per-netdev" lock, instead of having multiple disparate
>  > mutexes which are hard to allocate?  
> 
> The above looks like a quite unrelated refactor and one I think it will 
> not be worthy. The complexity of locking code in this series is very 
> limited, and self-encapsulated. Different locks for different things 
> increases scalability. Possibly we will not see much contention on the 
> same device, but some years ago we did not think there would be much 
> contention on RTNL...

We need to do this, anyway. Let me do it myself, then.

> Additionally, if we use a per _network device_ lock, future expansion of 
> the core to support devlink objects will be more difficult.

You parse out the binding you can store a pointer to the right mutex.

> [about separate handle from shaper_info arguments]
>  > Wouldn't it be convenient to store the handle in the "info"
>  > object? AFAIU the handle is forever for an info, so no risk of it
>  > being out of sync…  
> 
> Was that way a couple of iterations ago. Jiri explicitly asked for the 
> separation, I asked for confirmation and nobody objected.

Could you link to that? I must have not read it.
You can keep it wrapped in a struct *_handle, that's fine.
But it can live inside the shaper object.

> Which if the 2 options is acceptable from both of you?
> 
> [about queue limit and channel reconf]
>  > we probably want to trim the queue shapers on channel reconfig,
>  > then, too? :(  
> 
> what about exposing to the drivers an helper alike:
> 
> 	net_shaper_notify_delete(binding, handle);
> 
> that tells the core the shaper at the given handle just went away in the 
> H/W? The driver will call it in the queue deletion helper, and such 
> helper could be later on used more generically, i.e. for vf/devlink port 
> deletion.

We can either prevent disabling queues which have shapers attached, 
or auto-removing the shapers. No preference on that. But put the
callback in the core, please, netif_set_real_num_rx_queues() ?
Why not?

>  > It's not just for introspection, it's also for the core to do
>  > error checking.  
> 
> Actually, in the previous discussions it was never mentioned to use 
> capabilities to fully centralize the error checking.
> 
> This really looks like another feature, and can easily be added in a 
> second time (say, a follow-up series), with no functionality loss.
> 
> I (or anybody else) can’t keep adding new features at every iteration. 
> At some point we need to draw a line, and we should agree that the scope 
> of this activity has already expanded a lot in the past year. I would 
> like to draw such a line here.

I can help you. Just tell me which parts you want me to take care of.