Re: Problem with Netpoll based netdumping and NAPI

Re: Problem with Netpoll based netdumping and NAPI

[Matt Mackall]

[changed cc: from linux-net to netdev]

On Mon, Apr 19, 2004 at 06:21:48PM +0530, Hariprasad Nellitheertha wrote:
> Hi All,
> 
> I am facing a problem while trying to network dump using LKCD. My 
> debugging so far indicates that this is due to both NAPI and NETPOLL 
> being enabled.
> 
> I am using LKCD on the 2.6.5 kernel and both the client and server are 
> i386 boxes. The dumping machine has an e100 card. I have built the kernel
> with both CONFIG_E100_NAPI and CONFIG_NET_POLL_CONTROLLER (and the other
> netpoll related options) selected.
> 
> LKCD uses netpoll for its network dump implementation. The problem we see
> is that the network dump driver does not receive any packet from the 
> card driver and hence dumping fails. In e100_intr(), we call 
> netif_rx_schedule() if we are using the NAPI feature. netif_rx_schedule, 
> in turn, ends up adding the processing of this packet to the NET_RX_SOFTIRQ 
> softirq.

Netpoll should be manually calling the NAPI poll function like this 
after calling the interrupt handler (in netpoll_poll()):

      /* If scheduling is stopped, tickle NAPI bits */
         if(trapped && np->dev->poll &&
            test_bit(__LINK_STATE_RX_SCHED, &np->dev->state))
                 np->dev->poll(np->dev, &budget);

Please ensure that LKCD is calling netpoll_set_trap(1) which tells it
that packet scheduling is stopped.

I've tested this path primarily with tg3 and kgdb-over-ethernet, but
it should be functionally quite similar to e100 and lkcd.

-- 
Matt Mackall : http://www.selenic.com : Linux development and consulting

Re: Problem with Netpoll based netdumping and NAPI

[Hariprasad Nellitheertha]

Hi Matt,

On Mon, Apr 19, 2004 at 12:42:54PM -0500, Matt Mackall wrote:
> [changed cc: from linux-net to netdev]
> 
> On Mon, Apr 19, 2004 at 06:21:48PM +0530, Hariprasad Nellitheertha wrote:
> > Hi All,
> > 
> > I am facing a problem while trying to network dump using LKCD. My 
> > debugging so far indicates that this is due to both NAPI and NETPOLL 
> > being enabled.
> > 
> > I am using LKCD on the 2.6.5 kernel and both the client and server are 
> > i386 boxes. The dumping machine has an e100 card. I have built the kernel
> > with both CONFIG_E100_NAPI and CONFIG_NET_POLL_CONTROLLER (and the other
> > netpoll related options) selected.
> > 
> > LKCD uses netpoll for its network dump implementation. The problem we see
> > is that the network dump driver does not receive any packet from the 
> > card driver and hence dumping fails. In e100_intr(), we call 
> > netif_rx_schedule() if we are using the NAPI feature. netif_rx_schedule, 
> > in turn, ends up adding the processing of this packet to the NET_RX_SOFTIRQ 
> > softirq.
> 
> Netpoll should be manually calling the NAPI poll function like this 
> after calling the interrupt handler (in netpoll_poll()):
> 
>       /* If scheduling is stopped, tickle NAPI bits */
>          if(trapped && np->dev->poll &&
>             test_bit(__LINK_STATE_RX_SCHED, &np->dev->state))
>                  np->dev->poll(np->dev, &budget);
> 
> Please ensure that LKCD is calling netpoll_set_trap(1) which tells it
> that packet scheduling is stopped.

This was indeed the problem. We were not calling netpoll_set_trap in LKCD.
Adding this fixed the problem. Thanks so much for your help with this.

Regards, Hari

> 
> I've tested this path primarily with tg3 and kgdb-over-ethernet, but
> it should be functionally quite similar to e100 and lkcd.
> 
> -- 
> Matt Mackall : http://www.selenic.com : Linux development and consulting

-- 
Hariprasad Nellitheertha
Linux Technology Center
India Software Labs
IBM India, Bangalore