Re: [PATCH] sh_eth: NAPI requires netif_receive_skb()

[Eric Dumazet <eric.dumazet@gmail.com>]

On Thu, 2013-09-26 at 18:12 +0200, Guennadi Liakhovetski wrote:
> Hi
> 
> On Wed, 4 Sep 2013, David Miller wrote:
> 
> > From: Sergei Shtylyov <sergei.shtylyov@cogentembedded.com>
> > Date: Tue, 3 Sep 2013 03:03:10 +0400
> > 
> > > Driver supporting NAPI should use NAPI-specific function for receiving packets,
> > > so netif_rx() should be changed to netif_receive_skb().
> > > 
> > > Signed-off-by: Sergei Shtylyov <sergei.shtylyov@cogentembedded.com>
> 
> This patch breaks NFS boot on Armadillo800eva for me. Network 
> communication slows down to a crawl with
> 
> net eth0: Receive FIFO Overflow
> nfs: server 192.168.x.y not responding, still trying
> 
> With this patch reverted (e.g. in today's Linus tree snapshot) boot is 
> restored.

RX_RING_SIZE is 64

This driver refills the Rx ring buffers only _after_ the loop to drain
ready buffers. This was OK with the previous behavior (netif_rx() is
damn fast)

With this low amount of buffers, underrun can happen with the new code,
as netif_receive_skb() adds delay during the drain.

Most likely driver needs to refill buffers one by one (or small batches)
instead of in one go after the drain.