On Tue, May 03, 2011 at 12:49:55AM +0930, Arthur Marsh wrote: > > > Don Zickus wrote, on 02/05/11 22:29: > >On Sat, Apr 30, 2011 at 01:06:07PM +0930, Arthur Marsh wrote: > >> > >> > >> > >>Yes, with plain 2.6.39-rc5-git2 I had problems with shutting down > >>and with "BUG: soft lockup - CPU#0 stuck for 63s!" errors. > >> > >>After adding 'nmi_watchdog=0' I could cleanly shutdown and have not > >>seen any more "BUG: soft lockup - CPU#0 stuck for 63s!" errors. > > > >Ok. That's good. Can I see the dmesg log when you boot with out > >'nmi_watchdog=0'. I want to see the failure nmi watchdog spits out on > >boot and what subsystem perf thinks it found (should be none). > > > >This sounds like the softlock timer started and is hanging for some reason > >because the nmi watchdog failed to start. I thought those issues were > >fixed a while ago, maybe not. > > > >Cheers, > >Don > > > > I hope that it is alright providing the dmesg log for > 2.6.39-rc5-git4, which I've attached.
Interesting, I had just assumed a PII machine had no lapic and thus no perf counters. Guess I was wrong. It uses the p6 perf event subsystem and the lockup detector happily registers with it. This is odd because the commit you bisected down to, dealt with failure cases if the nmi watchdog couldn't register and initialize itself. But your dmesg clearly shows success there, so I am not sure why your machine would react negatively towards that commit. I do have to say, I have never really tested the lockup detector code using the p6 perf stuff, so I just assumed it worked correctly. I am wondering if the nmi or timer logic takes a long time on your machine that it makes the userspace tasks seem like the are running for a very long time. Could you boot with nmi_watchdog=0 and when you get a shell prompt, run something like perf top OR perf record grep -r foo /* I am curious to know if this is lockup detector related or the perf subystem related. Cheers, Don -- To UNSUBSCRIBE, email to [email protected] with a subject of "unsubscribe". Trouble? Contact [email protected]
