On Mon, Apr 11, 2016 at 5:57 PM, Shaohua Li <s...@fb.com> wrote: > Calvin found 'perf record -a --call-graph dwarf -- sleep 5' making clocksource > switching to hpet. We found similar symptom in another machine. Here is an > example: > > [8224517.520885] timekeeping watchdog: Marking clocksource 'tsc' as unstable, > because the skew is too large: > [8224517.540032] 'hpet' wd_now: ffffffff wd_last: b39c0bd mask: > ffffffff > [8224517.553092] 'tsc' cs_now: 48ceac7013714e cs_last: 48ceac25be34ac > mask: ffffffffffffffff > [8224517.569849] Switched to clocksource hpet > > In both machines, wd_now is 0xffffffff. The tsc time looks correct, the cpu > is 2.5G > (0x48ceac7013714e - 0x48ceac25be34ac)/2500000 = 0.4988s > 0.4988s matches WATCHDOG_INTERVAL. Since hpet reads to 0xffffffff in both > machines, this sounds not coincidence, hept is crappy. > > This patch tries to workaround this issue. We do retry if hpet has 0xffffff > value. > In the relevant machine, the hpet counter doesn't read to 0xffffffff later. > The chance hpet has 0xffffffff counter is very small, this patch should have > no > impact for good hpet. > > I'm open if there is better solution.
Hrm.. So can you characterize this bad behavior a bit more for us? Does every read of the HPET return 0xFFFFFFFF ? Or does it just occasionally start returning -1 values? Or once it trips and starts returning -1 does it always return -1? I'm trying to understand if there is a way to catch and disqualify that clocksource earlier then in the watchdog logic. thanks -john
