bizarre nfe(4) problem

[Don Lewis]

On 11 Aug, Pyun YongHyeon wrote:
> On Sat, Aug 11, 2007 at 06:22:22PM +0900, To Don Lewis wrote:
>  > On Fri, Aug 10, 2007 at 10:42:19PM -0700, Don Lewis wrote:
>  >  > I've a rather strange nfe(4) problem that appears to be repeatable.  I
>  >  > recently started running -CURRENT on a older socket 754 motherboard with
>  >  > the nForce3 chipset.  Initially, I was running an SMP kernel, but I had
>  >  > problems with sporadic "nfe0: watchdog timeout (missed Tx interrupts) --
>  >  > recovering" problems that would intermittently cause the system to lose
>  >  > network connectivity which it would recover from.  The kernel was very
>  >  > similar to GENERIC, with just the addition of "options DEBUG_VFS_LOCKS"
>  >  > and the replacement of atapicd with atapicam.
>  >  > 
>  >  > The nfe0 problem totally went away when I removed "options SMP" and
>  >  > "device apic" from the kernel configuration, except under the following
>  >  > very specific circumstances:
>  >  > 
>  >  > 	A vncserver session using the GNOME desktop was started on the
>  >  > 	system.
>  >  > 	
>  >  > 	There was no keyboard or mouse activity on the console for an
>  >  > 	extended period of time, allowing the GNOME screen saver to kick
>  >  > 	in and lock the screen.
>  >  > 
>  >  > The system would run fine in this state for many hours, and would accept
>  >  > incoming SMTP connections, etc.
>  >  > 
>  >  > 	A remote vncclient makes a connection to the vncserver session
>  >  > 	and the password was entired on the client.
>  >  > 
>  >  > At this point the nfe0 interface would appear to go deaf.  This might
>  >  > happen before or slightly after the password dialog box appeared for the
>  >  > vnc session.  For a short while, the system would be able to transmit
>  >  > TCP packets, ntp queries, etc., but it would not respond to any incoming
>  >  > packets (ping, TCP connection requests, etc.). Eventually, the ARP cache
>  >  > would time out and the only packets being transmitted would be ARP
>  >  > requests and the occasional UDP broadcast from the samba server running
>  >  > on the machine.
>  >  > 
>  >  > Pressing any key on the (PS/2) keyboard would instantly bring the
>  >  > network interface back to life.  Examination of /var/log/messages showed
>  >  > lots of "nfe0: watchdog timeout" messages for the entire time that nfe0
>  >  > was not listening to the network.
>  >  > 
>  >  > I've had this problem happen twice.  Both times were after an extended
>  >  > period of console inactivity.   An incoming vnc connection is not
>  >  > sufficient to trigger the problem if the console was recently active,
>  >  > and even waiting for the GNOME screensaver to put the monitor in DPMS
>  >  > power save mode before initiating the vnc connection does not appear to
>  >  > be sufficient to trigger the problem.
>  >  > 
>  >  > I believe that nfe0 was sharing an interrupt with one of the USB ports
>  >  > when the kernel was compiled with "device apic", but it is not sharing
>  >  > an interrupt without "device apic".
>  >  > 
>  >  > Any thoughts on how to debug this problem?
>  >  > 
>  >  > 
>  >  > # vmstat -i
>  >  > interrupt                          total       rate
>  >  > irq0: clk                       41903449       1000
>  >  > irq1: atkbd0                       39034          0
>  >  > irq3: ohci0                            5          0
>  >  > irq7: ppc0                             2          0
>  >  > irq8: rtc                        5362802        127
>  >  > irq9: ohci1 ahc0+                1963559         46
>  >  > irq10: nfe0+                      225593          5
>  >               ^^
>  > You have nfe0+ which indicates vmstat had run out of room to
>  > display somthing. I'm not sure but it's still sharing interrupt
>  > with other device?
> 
> It seems the interrupt is shared with atapci1.

Which is unused ...