TSC sched_clock overflow workaround
Hey Thomas, Ingo,
I thought Peter had sent this on, but apparently its not landed yet, so
I wanted to make sure you got it.
After ~208 days, the sched_clock multiplication overflows and causes some false positives in the softlockup watchdog. I don't think there are any actual panics going on, as an earlier related div by zero issue was already fixed. However, the false positives do continue to worry folks.
We still need a proper fix, which will probably be in part making sched_clock do periodic accumulation, as well as possibly changing users of sched_clock (like the softlockup watchdog) use timekeeping timestamps rather then sched_clock.
thanks
-john
Peter's workaround is available in the git repository at:
git://git.linaro.org/people/jstultz/linux.git fortglx/3.2/tip/sched/clock
Peter Zijlstra (1):
x86, intel: Don't mark sched_clock() as stable
arch/x86/kernel/cpu/intel.c | 7 +++++++ 1 files changed, 7 insertions(+), 0 deletions(-)
Quoted 10 lines · John Stultz
Hey Thomas, Ingo,
I thought Peter had sent this on, but apparently its not landed yet, so
I wanted to make sure you got it.After ~208 days, the
sched_clockmultiplication overflows and causes some false positives in the softlockup watchdog. I don't think there are any actual panics going on, as an earlier related div by zero issue was already fixed. However, the false positives do continue to worry folks.
Quoted 4 lines
We still need a proper fix, which will probably be in part making
sched_clockdo periodic accumulation, as well as possibly changing users ofsched_clock(like the softlockup watchdog) use timekeeping timestamps rather thensched_clock.
I'd *much* rather see a real fix for this than an ugly workaround that disables a very real speedup.
Thanks,
Ingo