Solving Ethernet watchdog timer deadlocks

Question

I have a debian linux box (Debian Squeeze) that deadlocks every few hours if I run a python script that sniffs an interface...

The stack trace is attached to the bottom of this question. Essentially, I have a Broadcom ethernet interface (bnx2 driver) that seems to die when I start a sniffing session and then it tries to transmit a frame out the same interface.

From what I can tell, a kernel watchdog timer is tripping...

NETDEV WATCHDOG: eth3 (bnx2): transmit queue 0 timed out

I think there is a way to control watchdog timers with ioctl (ref: EmbeddedFreak: How to use linux watchdog).

Questions (Original):

How can I find which watchdog timer(s) is controlling eth3? Bonus points if you can tell me how to change the timer or even disable the watchdog...

Questions (Revised):

How can I prevent the ethernet watchdog timer from causing problems?

---

Stack trace

Apr 30 08:38:44 Hotcoffee kernel: [275460.837147] ------------[ cut here ]------------
Apr 30 08:38:44 Hotcoffee kernel: [275460.837166] WARNING: at /build/buildd-linux-2.6_2.6.32-41squeeze2-amd64-NDo8b7/linux-2.6-2.6.32/debian/build/source_amd64_none/net/sched/sch_generic.c:261 dev_watchdog+0xe2/0x194()
Apr 30 08:38:44 Hotcoffee kernel: [275460.837169] Hardware name: PowerEdge R710
Apr 30 08:38:44 Hotcoffee kernel: [275460.837171] NETDEV WATCHDOG: eth3 (bnx2): transmit queue 0 timed out
Apr 30 08:38:44 Hotcoffee kernel: [275460.837172] Modules linked in: 8021q garp stp parport_pc ppdev lp parport pci_stub vboxpci vboxnetadp vboxnetflt vboxdrv ext2 loop psmouse power_meter button dcdbas evdev pcspkr processor serio_raw ext4 mbcache jbd2 crc16 sg sr_mod cdrom ses ata_generic sd_mod usbhid hid crc_t10dif enclosure uhci_hcd ehci_hcd megaraid_sas ata_piix thermal libata usbcore nls_base scsi_mod bnx2 thermal_sys [last unloaded: scsi_wait_scan]
Apr 30 08:38:44 Hotcoffee kernel: [275460.837202] Pid: 0, comm: swapper Not tainted 2.6.32-5-amd64 #1
Apr 30 08:38:44 Hotcoffee kernel: [275460.837204] Call Trace:
Apr 30 08:38:44 Hotcoffee kernel: [275460.837206]  <IRQ>  [<ffffffff81263086>] ? dev_watchdog+0xe2/0x194
Apr 30 08:38:44 Hotcoffee kernel: [275460.837211]  [<ffffffff81263086>] ? dev_watchdog+0xe2/0x194
Apr 30 08:38:44 Hotcoffee kernel: [275460.837217]  [<ffffffff8104df9c>] ? warn_slowpath_common+0x77/0xa3
Apr 30 08:38:44 Hotcoffee kernel: [275460.837220]  [<ffffffff81262fa4>] ? dev_watchdog+0x0/0x194
Apr 30 08:38:44 Hotcoffee kernel: [275460.837223]  [<ffffffff8104e024>] ? warn_slowpath_fmt+0x51/0x59
Apr 30 08:38:44 Hotcoffee kernel: [275460.837228]  [<ffffffff8104a4ba>] ? try_to_wake_up+0x289/0x29b
Apr 30 08:38:44 Hotcoffee kernel: [275460.837231]  [<ffffffff81262f78>] ? netif_tx_lock+0x3d/0x69
Apr 30 08:38:44 Hotcoffee kernel: [275460.837237]  [<ffffffff8124dda3>] ? netdev_drivername+0x3b/0x40
Apr 30 08:38:44 Hotcoffee kernel: [275460.837240]  [<ffffffff81263086>] ? dev_watchdog+0xe2/0x194
Apr 30 08:38:44 Hotcoffee kernel: [275460.837242]  [<ffffffff8103fa2a>] ? __wake_up+0x30/0x44
Apr 30 08:38:44 Hotcoffee kernel: [275460.837249]  [<ffffffff8105a71b>] ? run_timer_softirq+0x1c9/0x268
Apr 30 08:38:44 Hotcoffee kernel: [275460.837252]  [<ffffffff81053dc7>] ? __do_softirq+0xdd/0x1a6
Apr 30 08:38:44 Hotcoffee kernel: [275460.837257]  [<ffffffff8102462a>] ? lapic_next_event+0x18/0x1d
Apr 30 08:38:44 Hotcoffee kernel: [275460.837262]  [<ffffffff81011cac>] ? call_softirq+0x1c/0x30
Apr 30 08:38:44 Hotcoffee kernel: [275460.837265]  [<ffffffff8101322b>] ? do_softirq+0x3f/0x7c
Apr 30 08:38:44 Hotcoffee kernel: [275460.837267]  [<ffffffff81053c37>] ? irq_exit+0x36/0x76
Apr 30 08:38:44 Hotcoffee kernel: [275460.837270]  [<ffffffff810250f8>] ? smp_apic_timer_interrupt+0x87/0x95
Apr 30 08:38:44 Hotcoffee kernel: [275460.837273]  [<ffffffff81011673>] ? apic_timer_interrupt+0x13/0x20
Apr 30 08:38:44 Hotcoffee kernel: [275460.837274]  <EOI>  [<ffffffffa01bc509>] ? acpi_idle_enter_bm+0x27d/0x2af [processor]
Apr 30 08:38:44 Hotcoffee kernel: [275460.837283]  [<ffffffffa01bc502>] ? acpi_idle_enter_bm+0x276/0x2af [processor]
Apr 30 08:38:44 Hotcoffee kernel: [275460.837289]  [<ffffffff8123a0ba>] ? cpuidle_idle_call+0x94/0xee
Apr 30 08:38:44 Hotcoffee kernel: [275460.837293]  [<ffffffff8100fe97>] ? cpu_idle+0xa2/0xda
Apr 30 08:38:44 Hotcoffee kernel: [275460.837297]  [<ffffffff8151c140>] ? early_idt_handler+0x0/0x71
Apr 30 08:38:44 Hotcoffee kernel: [275460.837301]  [<ffffffff8151ccdd>] ? start_kernel+0x3dc/0x3e8
Apr 30 08:38:44 Hotcoffee kernel: [275460.837304]  [<ffffffff8151c3b7>] ? x86_64_start_kernel+0xf9/0x106
Apr 30 08:38:44 Hotcoffee kernel: [275460.837306] ---[ end trace 92c65e52c9e327ec ]---

Answer 1

I have read a similar story from GeNUA. Their workaround was to restart the network driver (OpenBSD). On Linux this would translate to: ifdown eth3 && rmmod bnx2 && modprobe bnx2 && ifup eth3.

The core problem was an internal coding problem with pointers on a PAE system in conjunction with the broadcom-driver.

Answer 2

Commenting out my code that called ethtool to modify the NIC buffers stopped watchdog timers from tripping on the bnx2 card.

I still want to find an answer to the question about watchdog timers, but I will ask another question

def _linux_buffer_alloc(iface=None, rx_ring_buffers=768,
    netdev_max_backlog=30000):

    default_rx = 255
    default_rx_jumbo = 0
    default_netdev_max_backlog = 1000
    ## Set linux rx ring buffers (to prevent tcpdump 'dropped by intf' msg)
## FIXME: removing for now due to systematic deadlocks with the bnx2 driver
#    sample: ethtool -G eth3 rx 768
#    cmd = 'ethtool -G %s rx %s' % (iface, rx_ring_buffers)
#    p = Popen(cmd.split(' '), stdout=PIPE)
#    p.communicate(); time.sleep(0.15)
#    sample: ethtool -G eth3 rx-jumbo 0
#    cmd = 'ethtool -G %s rx-jumbo %s' % (iface, default_rx_jumbo)
#    p = Popen(cmd.split(' '), stdout=PIPE)
#    p.communicate(); time.sleep(0.15)
## /FIXME