Re: [RFC PATCH net-next 11/11] net: enetc: add TX support for zero-copy XDP sockets

[Maciej Fijalkowski <maciej.fijalkowski@intel.com>]

On Wed, Feb 08, 2023 at 07:08:15PM +0200, Vladimir Oltean wrote:
> Hi Maciej,
> 
> On Wed, Feb 08, 2023 at 05:37:35PM +0100, Maciej Fijalkowski wrote:
> > On Mon, Feb 06, 2023 at 12:08:37PM +0200, Vladimir Oltean wrote:
> > 
> > Hey Vladimir,
> > 
> > > Schedule NAPI by hand from enetc_xsk_wakeup(), and send frames from the
> > > XSK TX queue from NAPI context. Add them to the completion queue from
> > > the enetc_clean_tx_ring() procedure which is common for all kinds of
> > > traffic.
> > > 
> > > We reuse one of the TX rings for XDP (XDP_TX/XDP_REDIRECT) for XSK as
> > > well. They are already cropped from the TX rings that the network stack
> > > can use when XDP is enabled (with or without AF_XDP).
> > > 
> > > As for XDP_REDIRECT calling enetc's ndo_xdp_xmit, I'm not sure if that
> > > can run simultaneously with enetc_poll() (on different CPUs, but towards
> > > the same TXQ). I guess it probably can, but idk what to do about it.
> > > The problem is that enetc_xdp_xmit() sends to
> > > priv->xdp_tx_ring[smp_processor_id()], while enetc_xsk_xmit() and XDP_TX
> > > send to priv->xdp_tx_ring[NAPI instance]. So when the NAPI instance runs
> > 
> > Why not use cpu id on the latter then?
> 
> Hmm, because I want the sendto() syscall to trigger wakeup of the NAPI
> that sends traffic to the proper queue_id, rather than to the queue_id
> affine to the CPU that the sendto() syscall was made?

Ok i was referring to the thing that using cpu id on both sides would
address concurrency. Regarding your need, what i did for ice was that i
assign xdp_ring to the rx_ring and within ndo_xsk_wakeup() i pick the
rx_ring based on queue_id that comes as an arg:

https://lore.kernel.org/bpf/20220822163257.2382487-3-anthony.l.nguyen@intel.com/