On Thursday 12 March 2009, John Stultz wrote: > * Reverting both 6c9bacb41c10 and 5cd1c9c5cf30 against 2.6-git did not > seem to help the issue, as originally reported. I also reverted > 49b5cf34727a as well just in case, and it didn't help either. Hmm, interesting. I think you may have reproduced a similar failure case, but not exactly the same as mine. I wonder if what you have reproduced is the same as Debian bug report http://bugs.debian.org/511334. That is with 2.6.26 which would match the fact that the reverts that work reliably for me did not work for you. I've never had any problems with Debian's 2.6.26 in my configuration. For me 2.6.28.7 is rock solid with the two reverts. I've attached my kernel config and my Hercules config. I think the critical thing in my config may be the "TIMERINT 500" setting for Hercules, which is somewhat similar to your "goes away if HZ is anything except 250" observation. > That's about all I can get for today. I'm out of town until Monday, so > I'll start digging back into it then. From your other replies I've seen that I have been way off track at some points. It looks as if I have been looking at events too early in the boot and been confused by the use of clock->xtime_nsec. My apologies for any confusion caused by that. I have a few more tests I think I can usefully do, but will try harder not to confuse the issue with my inexpert ramblings anymore :-P My main problem in trying to nail this down has been that any printks that are added tend to flood the system to such an extend that you cannot actually observe anything. Thanks a lot for your replies and work on this.