bge(4) watchdog timeout with APIC 5.4-prerel setup with 64bit PCI cards

[Pawel Malachowski]

Hello,

I have serious problems with 3Com 996(B-T?) bge(4) NICs in some kind of setups.
It fails with ``kernel: bge2: watchdog timeout -- resetting'' and disconnets
from network (it takes from 20 seconds to few minutes, very annoying).


Details:

Machine A: 5.3-RELEASE, GENERIC, ACPI disabled, IBM, Pentium 4, 32bit PCI,
GIANT LOCK.
. bge on board BCM5702
. 3Com bge NICs BCM5701
Cards connected to 100Mbps switch.
Machine A works just fine and routes about 20-50k pps (spending about 33%
of CPU in interrupts).


Machine B: 5.4-PRERELEASE, GENERIC, IBM (x225?), dual Xeon, 64bit PCI.
HTT disabled, ACPI disabled, GIANT LOCK.
. bge0 on board BCM5703
. two 3Com bge{1,2} NICs BCM5701 (identical as in A, but in 64bit PCI slots)
Cards connected to 100Mbps switch.
This machine works just fine only when onboard bge0 NIC is being used.
If I move vlans wih real traffic (only about 5-10Mbit/s) to bge1 or bge2,
it will fail with watchdog timeout few times per day (sometimes few times
per hour).

Cables were replaced, switch was changed.
SMP was disabled.
System was updated from 5.3-RELEASE to latest RELENG_5.

Nothing helps!

I'm currently running UP GENERIC kernel with APIC disabled, seems to work
fine (1 hour).


-- 
Paweł Małachowski