=> > So you're saying that you can reproduce this crash with acenic?
=>
=> Yes, I found capital logical bug there. Evictor on smp can corrupt queues.
=> Shame on me, I _knew_ that this pattern is wrong, socket hash tables need
=> the same, but they are right. So, fix is to work like sockets.
=>
=> Patch is appended. Please, test.
=>
=> Alexey
The patch appears to fix the problems I was seeing.
I've been blasting UDP streams at the smp box for
an hour or so. It used to die in a few seconds.
I'm also not seeing the RX-ERR increment at all. I was
often not able to get an skb. Definitely a BIG
improvement from my perspective - Thanks!
I'll be doing some more performance characterization
shortly. At first glance, I think the udp receive rate
is down quite a bit.
rcc2 33% netstat -i
Kernel Interface table
Iface MTU Met RX-OK RX-ERR RX-DRP RX-OVR TX-OK TX-ERR TX-DRP TX-OVR Flg
eth0 1500 0 4834 0 0 0 607 0 0 0 BRU
lo 16192 0 46 0 0 0 46 0 0 0 LRU
myri0 9000 0 133058589 0 0 0 2867 0 0 0 BRU
|