[net-next v19 4/7] net: mtip: Add net_device_ops functions to the L2 switch driver

[Jakub Kicinski]

On Thu, 11 Sep 2025 23:55:47 +0200 Łukasz Majewski wrote:
> > > Ok. No adjustments needed then. Good :)    
> > 
> > No, you were talking about build_skb() which is Rx.
> > This is the patch that adds Tx. Tx is wrong.  
> 
> The same approach is taken in fec_main.c (@ fec_enet_txq_submit_skb()
> function).

FWIW I'm 99% sure we were once investigating a bug in FEC related to
modifying timestamped packets, leading to crashes. Maybe there is more.

> > > could be replaced just with mtip_switch_tx(napi->dev);
> > > as TX via napi->dev shall be forward to both ports if required.
> > > 
> > > I will check if this can be done in such a way.    
> > 
> > Not napi->dev. You have to attribute sent packets to the right netdev.  
> 
> And then we do have some issue to solve. To be more specific -
> fec_main.c to avoid starvation just from fec_enet_rx_napi() calls
> fec_enet_tx() with only one net device (which it supports).
> 
> I wanted to mimic such behaviour with L2 switch driver (at
> mtip_rx_napi()), but then the question - which network device (from
> available two) shall be assigned?
> 
> The net device passed to mtip_switch_tx() is only relevant for
> "housekeeping/statistical data" as in fact we just provide another
> descriptor to the HW to be sent.
> 
> Maybe I shall extract the net device pointer from the skb structure?

Exactly :)

> > > You mean a separate SW queues for each devices? This is not
> > > supported in the MTIP L2 switch driver. Maybe such high level SW
> > > queues management is available in the upper layers?    
> > 
> > Not possible, each netdev has it's own private qdisc tree.  
> 
> Please correct me if I'm wrong, but aren't packets from those queues
> end up with calling ->ndo_start_xmit() function?

Right. I think I'm lost, why does this matter?

> > I think I explained this enough times. Next version is v20.
> > If it's not significantly better than this one, I'm going to have 
> > to ask you to stop posting this driver.  
> 
> I don't know how to reply to this comment, really. 
> 
> I've spent many hours of my spare time to upstream this driver.
> I'm just disappointed (and maybe I will not say more because of high
> level of my frustration).

I believe mlxsw has fewer DMA queues than ports. But TBH I'm not sure
how they handle the congestion. In your case since you only have two
ports (at most) I think you can trivially just always stop and start
both.