Hi Thomas/Daniel et al,
This isn't about the problem I reported earlier, where you advised to add ONESHOT_STOPPED mode: https://lkml.org/lkml/2014/5/9/508. Above problem was about stopping the clock-event device when its not used anymore.
This ($subject) problem was initially spotted on Ivybrdge V2, 12 core X86 server by Santosh. And then I reproduced it on Dual core ARM Exynos (isn't that frequent as it was on x86 though).
Problem: Getting spurious ticks where hrtimer_interrupt() returns without servicing any hrtimers.
Kernel hack to catch this: http://pastebin.com/bTM7nqDc (Over 3.16-rc3) X86 boot logs: http://pastebin.com/E6axDnsa (search: hrtimer_interrupt) /proc/cpuinfo: http://pastebin.com/uQx9TmsA
The last I could debug it to is:
- Clockevent device is programmed for time 'x' seconds (Verified this by storing next-event from within lapic_next_event()). - Tick fires ~300 us before 'x' - Traversing through the list of hrtimers doesn't result in any pending hrtimer and we simply return. And so *spurious* interrupt.
- Happens when ticks are active or stopped (search for "tick-stopped" in logs)
Driver monitored for x86: arch/x86/kernel/apic/apic.c Similar behavior observed on exynos with arm_arch_timer.c
I couldn't get any deeper into it to see what's going on. From the behavior It looks lik the calculations we are doing with dev->mult/shift gives timeout <= next-event, whereas it should be >= ? Not at all sure though.
Reported-by: Santosh Shukla santosh.shukla@linaro.org
Note: Even the Hacky patchset that tried to to disable clockevent device when not used anymore, isn't able to fix it: https://lkml.org/lkml/2014/5/9/99..
-- viresh