On 16-08-2010 18:08, Eric Blake wrote:
On 08/16/2010 03:54 AM, Soren Hansen wrote:
09:41:01.134: debug : umlMonitorCommand:698 : Run command 'config con0' 09:41:01.134: debug : umlMonitorCommand:733 : res.error: 6, res.extra: 0, res.length: 4096, res.data: 09:41:01.134: debug : umlMonitorCommand:736 : nbytes: 0 09:41:01.134: debug : umlMonitorCommand:738 : res.error: 0, res.extra: 0, res.length: 4, res.data: pts This one's a 2.6.34.1 kernel. The one where I didn't see the problem is a 2.6.32.something-or-the-other kernel. Mindboggling. Oh my. This really does look like a kernel bug. Can you confirm it with an strace?
Annoyingly, no. It doesn't happen in the primary libvirt thread, so I have to pass -f to strace, and uml doesn't let you PTRACE it, so I can't trigger the problem.
Have you reported this regression to the right kernel folks?
It doesn't happen in 2.6.35, so I seems isolated to that particular kernel. It's really very, very odd.
I guess it would help if we could write a simpler test program to isolate whether this recvfrom bug exists in a bare minimum number of syscalls. Meanwhile, I have no idea how to work around a buggy recvfrom that doesn't return the correct number of bytes.
No, you're right. Whatever we do to work around this is going to suck somehow. -- Soren Hansen Ubuntu Developer http://www.ubuntu.com/