Re: [RFC net-next v4 2/2] net/mlx5e: Add per queue netdev-genl stats

[Tariq Toukan]

On 07/06/2024 3:19, Jakub Kicinski wrote:
> On Thu, 6 Jun 2024 14:54:40 -0700 Joe Damato wrote:
> > > > Compare the values in /proc/net/dev match the output of cli for the same
> > > > device, even while the device is down.
> > > >
> > > > Note that while the device is down, per queue stats output nothing
> > > > (because the device is down there are no queues):
> > >
> > > This part is not true anymore.
> >
> > It is true with this patch applied and running the command below.
> > Maybe I should have been more explicit that using cli.py outputs []
> > when scope = queue, which could be an internal cli.py thing, but
> > this is definitely true with this patch.
> >
> > Did you test it and get different results?
>
> To avoid drivers having their own interpretations what "closed" means,
> core hides all queues in closed state:
>
> https://elixir.bootlin.com/linux/v6.10-rc1/source/net/core/netdev-genl.c#L582

Oh, so the kernel doesn't even call the driver's 
mlx5e_get_queue_stats_rx/tx callbacks if interface is down. Although our 
driver can easily satisfy the query and provide the stats.

I think the kernel here makes some design assumption about the stats, 
and enforces it on all vendor drivers.
I don't think it's a matter of "closed channel" interpretation, it's 
more about persistent stats.
IMO the kernel should be generic enough to let both designs (persistent 
and non-persistent stats) integrate naturally with this new queue 
netdev-genl stats feature.

> > > PTP RQ index is naively assigned to zero:
> > > rq->ix           = MLX5E_PTP_CHANNEL_IX;
> > >
> > > but this isn't to be used as the stats index.
> > > Today, the PTP-RQ has no matcing rxq in the kernel level.
> > > i.e. turning PTP-RQ on won't add a kernel-level RXQ to the
> > > real_num_rx_queues.
> > > Maybe we better do.
> > > If not, and the current state is kept, the best we can do is let the PTP-RQ
> > > naively contribute its queue-stat to channel 0.
> >
> > OK, it sounds like the easiest thing to do is just count PTP as
> > channel 0, so if i == 0, I'll in the PTP stats.
> >
> > But please see below regarding testing whether or not PTP is
> > actually enabled or not.
>
> If we can I think we should avoid making queue 0 too special.
> If someone configures steering and only expects certain packets on
> queue 0 - getting PTP counted there will be a surprise.
> I vote to always count it towards base.

+1, let's count PTP RX in the base, especially that it has no matching 
kernel-level rxq.

Another option is to add one more kernel rxq for it (i.e. set 
real_num_rx_queues to num_channels + 1). But, that would be a bigger 
change, we can keep it for a followup discussion.