Bug 16314
Summary: | Erroneous idle times for processes. | ||
---|---|---|---|
Product: | Timers | Reporter: | overlordq |
Component: | Other | Assignee: | Jeremy Fitzhardinge (jeremy) |
Status: | RESOLVED CODE_FIX | ||
Severity: | normal | CC: | jeremy |
Priority: | P1 | ||
Hardware: | All | ||
OS: | Linux | ||
Kernel Version: | 2.6.34 | Subsystem: | |
Regression: | Yes | Bisected commit-id: |
Description
overlordq
2010-06-28 22:15:52 UTC
So what was the last kernel version that worked properly for you? # umount /foot/mouth Seems I can get 2.6.31.13 to work, but 2.6.34 is a no go. Jeremy: Any thoughts on this one? That's interesting. I'm not sure what timebase uptime is computed with. It sounds like it may be using something like the scheduler clock, which under Xen does not include time stolen by other domains. How busy is the domain in question, and how busy is the overall system? Hm, but according to /proc/stat, there's no stolen time on any cpu (second last column, I think), so that suggests the rest of the system is very idle. Has the uptime computation changed between .31 and .34? The domain is effectively idle, very low traffic, not sure on the overall system, I'd have to check. The problem shows up more the longer the box is up. Like right now only root 13 16857174 0.0 0 0 ? S Jun29 21067849:29 [events/2] out of the kernel threads is showing up wrong, other process tend toward this the longer its up. This weekend I'll double check .31 to see if leaving it up longer manifests problems and try to narrow down a specific revision when I have time to recompile a lot of kernels. Recent mainline and stable kernels have a fix to convert Xen's sched_clock to use plain wallclock time. Do they still exhibit this problem? Thanks, J I'll test a new stock kernel and see if the problem still occurs. In the mean time I have been using a proposed patch[1] from the xen-devel lists which also mitigates the problem. 1 - http://lists.xensource.com/archives/html/xen-devel/2010-07/msg00738.html OK, that fixes the symptom that the upstream patch fixes properly. 2.6.36-rc3 at a glance appears to be working correctly. |