Re: tick/sched: iowait_sleeptime resides in tick_cpu_sched structure is a negative value

[luanshi]

Hi Frederi,

在 2022/1/6 21:40, Frederic Weisbecker 写道:
> On Thu, Jan 06, 2022 at 05:14:02PM +0800, luanshi wrote:
> > Hi All,
> >
> > We encounted a hardlockup problem on ARM server, after some debug found
> > that:
> >
> > PID: 0      TASK: ffff0400064de300  CPU: 126  COMMAND: "swapper/126"
> >   #0 [ffff8000250f3c90] __crash_kexec at ffff80001013a064
> >   #1 [ffff8000250f3e30] panic at ffff800010afd028
> >   #2 [ffff8000250f3f10] nmi_panic at ffff80001004a9e0
> >   #3 [ffff8000250f3f20] watchdog_hardlockup_check at ffff80001017b7f0
> >   #4 [ffff8000250f3f40] sdei_watchdog_callback at ffff80001003c9a4
> >   #5 [ffff8000250f3f50] sdei_event_handler at ffff8000108816b8
> >   #6 [ffff8000250f3f80] _sdei_handler at ffff800010b1c2e4
> >   #7 [ffff8000250f3fd0] __sdei_handler at ffff800010b139b0
> >   #8 [ffff8000250f3ff0] __sdei_asm_handler at ffff800010014c18
> > --- <IRQ stack> ---
> >   #9 [ffff800013973f10] __cpu_do_idle at ffff800010b13764
> > #10 [ffff800013973f20] arch_cpu_idle at ffff800010b137ec
> > #11 [ffff800013973f30] default_idle_call at ffff800010b1b5cc
> > #12 [ffff800013973f50] cpuidle_idle_call at ffff80001009648c
> > #13 [ffff800013973f90] do_idle at ffff8000100965b0
> > #14 [ffff800013973fc0] cpu_startup_entry at ffff8000100967d4
> > #15 [ffff800013973fe0] secondary_start_kernel at ffff800010026bb0
> >
> >
> > per_cpu(tick_cpu_sched, 126) = $1 = {
> >    sched_timer = {
> >      node = {
> >        node = {
> >          __rb_parent_color = 18446603337117384112,
> >          rb_right = 0x0,
> >          rb_left = 0x0
> >        },
> >        expires = 6108564000000
> >      },
> >      _softexpires = 6108564000000,
> >      function = 0xffff800010122ec0 <tick_sched_timer>,
> >      base = 0xffff04473bbcc780,
> >      state = 1 '\001',
> >      is_rel = 0 '\000',
> >      is_soft = 0 '\000',
> >      is_hard = 1 '\001'
> >    },
> >    check_clocks = 0,
> >    nohz_mode = NOHZ_MODE_INACTIVE,
> >    inidle = 1,
> >    tick_stopped = 0,
> >    idle_active = 1,
> >    do_timer_last = 0,
> >    got_idle_tick = 1,
> >    last_tick = 0,
> >    next_tick = 0,
> >    idle_jiffies = 0,
> >    idle_calls = 0,
> >    idle_sleeps = 0,
> >    idle_entrytime = 5012087709249,
> >    idle_waketime = 0,
> >    idle_exittime = 0,
> >    idle_sleeptime = 4936136669951,
> >    iowait_sleeptime = -1942739704,
> >    last_jiffies = 0,
> >    timer_expires = 0,
> >    timer_expires_base = 0,
> >    next_timer = 0,
> >    idle_expires = 0,
> >    tick_dep_mask = {
> >      counter = 0
> >    }
> > }
> >
> > iowait_sleeptime = -1942739704,
> > ----------
> > iowait_sleeptime is monotonically increasing, under what circumstances iowait_sleeptime
> > can be a negative value?
> >
> > For detailed information:
> > https://bugzilla.kernel.org/show_bug.cgi?id=215458
> >
> >
> > Can you give me some suggestions for debugging.
> Some racy updates can happen if cpufreq or "cat /proc/stat" do concurrent
> updates. But for that nohz needs to be running and I see your CPU clock has
> NOHZ_MODE_INACTIVE. Perhaps it's only for that CPU though.
>
> What is the value of tick_nohz_active in your dump?

nohz=off was configured in the startup parameters

crash> tick_nohz_enabled
tick_nohz_enabled = $1 = false
crash> tick_nohz_active
tick_nohz_active = $2 = 0


and current no cpufreq driver was used:

crash> cpufreq_driver
cpufreq_driver = $4 = (struct cpufreq_driver *) 0x0


Thanks,

Liguang

> > Thanks,
> > Liguang