Messages in this thread | | | Subject | Re: [tip:sched/urgent] sched: Fix cross-cpu clock sync on remote wakeups | From | Peter Zijlstra <> | Date | Tue, 31 May 2011 15:11:56 +0200 |
| |
On Tue, 2011-05-31 at 14:56 +0200, Borislav Petkov wrote: > Hey Peter, > > On Tue, May 31, 2011 at 12:31:20PM +0000, tip-bot for Peter Zijlstra wrote: > > Commit-ID: f01114cb59d670e9b4f2c335930dd57db96e9360 > > Gitweb: http://git.kernel.org/tip/f01114cb59d670e9b4f2c335930dd57db96e9360 > > Author: Peter Zijlstra <a.p.zijlstra@chello.nl> > > AuthorDate: Tue, 31 May 2011 12:26:55 +0200 > > Committer: Ingo Molnar <mingo@elte.hu> > > CommitDate: Tue, 31 May 2011 14:19:56 +0200 > > > > sched: Fix cross-cpu clock sync on remote wakeups > > > > Markus reported that commit 317f394160e ("sched: Move the second half > > of ttwu() to the remote cpu") caused some accounting funnies on his AMD > > Phenom II X4, such as weird 'top' results. > > > > It turns out that this is due to non-synced TSC > > this would mean that his machine doesn't pass the TSC sync check at boot > but that's a F10h and they usu. have synchronized TSCs? > > I'm confused.
Well, I don't have a modern AMD system to verify on, but the only explanation is sched_clock weirdness (different code from the GTOD tsc stuff). I could not reproduce on an Intel Westmere machine, but could on a Core2.
The sched_clock_cpu stuff basically takes a GTOD timestamp every tick and uses sched_clock() (tsc + cyc2ns) to provide delta increments, when TSCs are synced every cpu should return the same value and the patch is a nop.
If they aren't synced the per-cpu sched_clock_cpu() values can drift up to about 2 jiffies (when the ticks drift about 1 and the slower of the two has a stuck tsc while the faster of the two does progress at the normal rate). In that case doing a clock update cross-cpu will ensure time monotonicity between those two cpus.
So by this patch making a difference we prove that sched_clock_cpu()'s aren't synced on the affected machines. That leaves two options, either the TSC are screwy or there's a bug somewhere in the sched_clock* code.
I'd be more than happy if you could take a look at the relevant code since all code can use more eyes :-)
| |