Re: WARNING: at net/core/dev.c:2161 net_rx_action()

[David Miller <davem@davemloft.net>]

From: Ingo Molnar <mingo@elte.hu>
Date: Tue, 16 Oct 2007 00:03:57 +0200

> * David Miller <davem@davemloft.net> wrote:
> 
> > From: Ingo Molnar <mingo@elte.hu>
> > Date: Mon, 15 Oct 2007 13:24:30 +0200
> > 
> > > got this warning with Linus' latest -git tree:
> > > 
> > >   WARNING: at net/core/dev.c:2161 net_rx_action()
> > >    [<80564db4>] net_rx_action+0xce/0x186
> > >    [<8011ba98>] __do_softirq+0x6c/0xcf
> > >    [<8011bb2d>] do_softirq+0x32/0x36
> > >    [<8011bcae>] irq_exit+0x35/0x40
> > >    [<80104fdb>] do_IRQ+0x5c/0x71
> > >    [<801048cd>] do_nmi+0x8f/0x238
> > >    [<801033a3>] common_interrupt+0x23/0x30
> > >    =======================
> > 
> > This is a driver bug, the work "budget" passed into a driver's 
> > ->poll() handler should never be exceeded.  That's what this warning 
> > assertion is checking.
> > 
> > What ethernet card is in your system and what driver is being used to 
> > drive it?
> 
> it's forcedeth.
> 
> i've checked nv_napi_poll(), and i dont see how it could return larger 
> than 'limit' number of packets.
> 
> it could return packets == limit though:
> 
>                 pkts = nv_rx_process_optimized(dev, budget);
> ...
> 
>         if (pkts < budget) {
>                 /* re-enable receive interrupts */
>                 spin_lock_irqsave(&np->lock, flags);
> 
>                 __netif_rx_complete(dev, napi);
> 
> ...
>         return pkts;
> 
> shouldnt that be "pkts <= budget"? But even that shouldnt cause a larger 
> than limit return. Weird.

No, not in this case.  The driver must only netif_rx_complete()
if it consumed strictly less than "budget" worth of work.

> there are two networking cards in the system, the other one is a:
> 
>   eth1: RealTek RTL8139 at 0xf080e000, 00:c0:df:03:68:5d, IRQ 11
>   eth1:  Identified 8139 chip type 'RTL-8139B'
> 
> but this one should be inactive (not plugged into the network). Should i 
> try to get a debug print out of the actual 'weight' and 'work' integers, 
> and of the n->poll function address?

That might help.

I don't see any possible nv_rx_process{,_optimized}() can return "work
> budget" either.  But I do notice these loops unconditionally execute
at least once, perhaps budget is being passed erroneously in as zero?