Re: [PATCH v5 net-next 6/6] net: ethernet: ti: cpsw: add XDP support

[Jesper Dangaard Brouer]

On Sun, 30 Jun 2019 20:23:48 +0300 Ivan Khoronzhuk <ivan.khoronzhuk@xxxxxxxxxx> wrote:

> Add XDP support based on rx page_pool allocator, one frame per page.
> Page pool allocator is used with assumption that only one rx_handler
> is running simultaneously. DMA map/unmap is reused from page pool
> despite there is no need to map whole page.
> 
> Due to specific of cpsw, the same TX/RX handler can be used by 2
> network devices, so special fields in buffer are added to identify
> an interface the frame is destined to. Thus XDP works for both
> interfaces, that allows to test xdp redirect between two interfaces
> easily. Aslo, each rx queue have own page pools, but common for both
> netdevs.

Looking at the details what happen when a single RX-queue can receive
into multiple net_device'es.  I realize that this driver will
violate/kill some of the "hidden"/implicit RX-bulking that the
XDP_REDIRECT code depend on for performance.

Specifically, it violate this assumption:
 https://github.com/torvalds/linux/blob/v5.2-rc7/kernel/bpf/devmap.c#L324-L329

	/* Ingress dev_rx will be the same for all xdp_frame's in
	 * bulk_queue, because bq stored per-CPU and must be flushed
	 * from net_device drivers NAPI func end.
	 */
	if (!bq->dev_rx)
		bq->dev_rx = dev_rx;

This drivers "NAPI func end", can have received into multiple
net_devices, before it's NAPI cycle ends.  Thus, violating this code
assumption.

Knowing all xdp_frame's in the bulk queue is from the same net_device,
can be used to further optimize XDP.  E.g. the dev->netdev_ops->ndo_xdp_xmit()
call don't take fully advantage of this, yet.  If we merge this driver,
it will block optimizations in this area.

NACK

-- 
Best regards,
  Jesper Dangaard Brouer
  MSc.CS, Principal Kernel Engineer at Red Hat
  LinkedIn: http://www.linkedin.com/in/brouer