Oops

I set my backup script off again and it went fine until it tried to pull data from curator, which is the firewall and one of the recently upgraded systems, at which point the data only trickled across at (later experiment showed) about 15Kbyte/sec.

After investigating a bit, I ruled out everything up to and including curator's ssh (since it was blocking in writes and consuming next to no CPU, and worked fine over loopback). I initially suspected an IRQ conflict - there are still ISA devices in this machine - but fiddling with jumpers, swapping cards out and disabling the onboard serial and parallel made no difference.

#chiark made several helpful remarks but to no avail, until I remembered that (i) eth0 and eth1 had swapped meaning during the upgrade (which might have been reversible but I decided to cope in software) (ii) I'd not told the traffic shaper about the change. So the kernel was clamping data going to the internal network to 128Kbit/second. (It's supposed to do this to data going to the internet, as the cable modem, or possibly NTL's router, is broken.)

Fix this and it all started behaving again.