Chad Reese wrote:
I assume this is happening under high network load. Some early Octeon
ethernet drivers had a problem where they could starve the rest of the
system processing incoming packets. The message you are getting is the
kernel warning you that userspace hasn't been given any processing time.
What is probably happening is the cavium_ethernet driver is spending all
its time in a receive tasklet getting packets and then dropping them.
This has been fixed in a later cavium SDK, but it looks like you are
running an ancient kernel. A newer kernel is available on the cavium
support site.
It seems unlikely it's due to cpu load. It's only receiving around
30000 packets/sec total, and before the problem shows up "top" shows
some cores at around 25% and others totally idle. The only symptom of
the problem is that the system simply stops responding.
You wouldn't happen to know which versions fixed the problem you
described above, would you? Upgrading the kernel isn't an option, but
it wouldn't be the first time we've had to backport things.
The cavium ethernet driver doesn't use the standard NAPI interface since
it doesn't support multicore receive for a single port.
Ah...I was wondering about that.
Chris
|