"Luigi Genoni" <luigi.genoni@xxxxxxxxxxx> writes:expect.
> (e-mail resent because not delivered using my other e-mail account)
>
> Hi,
> this night a linux server 8 dual core CPU Optern 2600Mhz crashed just
> after giving this message
>
> Jan 22 04:48:28 frey kernel: do_IRQ: 1.98 No irq handler for vector
Ok. This indicates that the hardware is doing something we didn't
We don't know which irq the hardware was trying to deliver when itnet
sent vector 0x98 to cpu 1.
> I have no other logs, and I eventually lost the OOPS since I have no
> console setled up.AMD
If you had an oops it may have meant the above message was a secondary
symptom. Groan. If it stayed up long enough to give an OOPS then
there is a chance the above message appearing only once had nothing
to do with the actual crash.
How long had the system been up?
> As I said sistem is running linux 2.6.19 compiled with gcc 4.1.1 for
> Opteron (attached see .config), no kernel preemption excepted the BKLbe
> preemption. glibc 2.4.
>
> System has 16 GB RAM and 8 dual core Opteron 2600Mhz.
>
> I am running irqbalance 0.55.
>
> any hints on what has happened?
Three guesses.
- A race triggered by irq migration (but I would expect more people to
yelling). The code path where that message comes from is new in 2.6.19so
it may not have had all of the bugs found yet :(4000 8000 ; do
- A weird hardware or BIOS setup.
- A secondary symptom triggered by some other bug.
If this winds up being reproducible we should be able to track it down.
If not this may end up in the files of crap something bad happened that
we don't understand.
The one condition I know how to test for (if you are willing) is an
irq migration race. Simply by triggering irq migration much more often,
and thus increasing our chances of hitting a problem.
Stopping irqbalance and running something like:
for irq in 0 24 28 29 44 45 60 68 ; do
while :; do
for mask in 1 2 4 8 10 20 40 80 100 200 400 800 1000 2000
echo mask > /proc/irq/$irq/smp_affinity1
sleep 1
done
done &
done
Should force every irq to migrate once a second, and removing the sleep
is even harsher, although we max at one irq migration by irq received.irq
If some variation of the above loop does not trigger the do_IRQ ??? No
handler for vector message chances are it isn't a race in irq migration.right
If we can rule out the race scenario it will at least put us in the
direction for guessing what went wrong with your box.
Eric