Re: [PATCH v2 1/2] tick: Remove unreasonable detached state set in tick_shutdown()

[Bibo Mao]

On 2025/9/4 下午11:57, Thomas Gleixner wrote:
> On Thu, Sep 04 2025 at 15:17, Bibo Mao wrote:
> > Function clockevents_switch_state() will check whether it has already
> > switched to specified state, do nothing if it has.
> >
> > In function tick_shutdown(), it will set detached state at first and
> > call clockevents_switch_state() in clockevents_exchange_device(). The
> > function clockevents_switch_state() will do nothing since it is already
> > detached state. So the tick timer device will not be shutdown when CPU
> > is offline. In guest VM system, timer interrupt will prevent vCPU to
> > sleep if vCPU is hot removed.
> >
> > Here remove state set before calling clockevents_exchange_device(),
> > its state will be set in function clockevents_switch_state() if it
> > succeeds to do so.
>
> This explanation is incomplete. tick_shutdown() did this because it was
> originally invoked on a life CPU and not on the outgoing CPU.
>
> That got changed in
>
>    3b1596a21fbf ("clockevents: Shutdown and unregister current clockevents at CPUHP_AP_TICK_DYING")
>
> which is the actual root cause.
>
> The pile of 'Fixes:' below is just enumerating the subsequent problems.
>
> > Fixes: bf9a001fb8e4 ("clocksource/drivers/timer-tegra: Remove clockevents shutdown call on offlining")
> > Fixes: cd165ce8314f ("clocksource/drivers/qcom: Remove clockevents shutdown call on offlining")
> > Fixes: 30f8c70a85bc ("clocksource/drivers/armada-370-xp: Remove clockevents shutdown call on offlining")
> > Fixes: ba23b6c7f974 ("clocksource/drivers/exynos_mct: Remove clockevents shutdown call on offlining")
> > Fixes: 15b810e0496e ("clocksource/drivers/arm_global_timer: Remove clockevents shutdown call on offlining")
> > Fixes: 78b5c2ca5f27 ("clocksource/drivers/arm_arch_timer: Remove clockevents shutdown call on offlining")
> > Fixes: 900053d9eedf ("ARM: smp_twd: Remove clockevents shutdown call on offlining")
> >
> > Signed-off-by: Bibo Mao <maobibo@xxxxxxxxxxx>
> > Reviewed-by: Frederic Weisbecker <frederic@xxxxxxxxxx>
> > ---
> >   kernel/time/tick-common.c | 5 -----
> >   1 file changed, 5 deletions(-)
> >
> > diff --git a/kernel/time/tick-common.c b/kernel/time/tick-common.c
> > index 9a3859443c04..eb9b777f5492 100644
> > --- a/kernel/time/tick-common.c
> > +++ b/kernel/time/tick-common.c
> > @@ -424,11 +424,6 @@ void tick_shutdown(unsigned int cpu)
> >   
> >   	td->mode = TICKDEV_MODE_PERIODIC;
> >   	if (dev) {
> > -		/*
> > -		 * Prevent that the clock events layer tries to call
> > -		 * the set mode function!
> > -		 */
> > -		clockevent_set_state(dev, CLOCK_EVT_STATE_DETACHED);
> >   		clockevents_exchange_device(dev, NULL);
> >   		dev->event_handler = clockevents_handle_noop;
> >   		td->evtdev = NULL;
>
> Can this pretty please cleanup the misleading comment above
> tick_shutdown() as well?
>
>   * Shutdown an event device on a given cpu:
>   *
>   * This is called on a life CPU, when a CPU is dead. So we cannot
>   * access the hardware device itself.
>   * We just set the mode and remove it from the lists.
>
> That should have been removed or updated with 3b1596a21fbf too, no?
>
> With that the cpu argument is not longer useful either, because this is
> now guaranteed to be invoked on the outgoing CPU, no?

It is not easy with my poor English to spell out the comments :(
How about the patch like this?

Function clockevents_switch_state() will check whether it has already
switched to specified state, do nothing if it has.

In function tick_shutdown(), it will set detached state at first and
call clockevents_switch_state() in clockevents_exchange_device(). The
function clockevents_switch_state() will do nothing since it is already
detached state. So the tick timer device will not be shutdown when CPU
is offline.

Function tick_shutdown() did this because it was originally invoked
on a life CPU and not on the outgoing CPU. Now this function is called
on the outgoing CPU, the hardware device can be accessed.

Here remove state set before calling clockevents_exchange_device(), its
state will be set in function clockevents_switch_state() if it succeeds
to do so.

Fixes: 3b1596a21fbf ("clockevents: Shutdown and unregister current 
clockevents at CPUHP_AP_TICK_DYING")


 /*
- * Shutdown an event device on a given cpu:
+ * Shutdown an event device on the outgoing CPU:
  *
- * This is called on a life CPU, when a CPU is dead. So we cannot
- * access the hardware device itself.
- * We just set the mode and remove it from the lists.
+ * Called by the dying CPU during teardown, with clockevents_lock held
+ * and interrupts disabled.
  */
-void tick_shutdown(unsigned int cpu)
+void tick_shutdown(void)
 {
-       struct tick_device *td = &per_cpu(tick_cpu_device, cpu);
+       struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
        struct clock_event_device *dev = td->evtdev;

        td->mode = TICKDEV_MODE_PERIODIC;
        if (dev) {
-               /*
-                * Prevent that the clock events layer tries to call
-                * the set mode function!
-                */
-               clockevent_set_state(dev, CLOCK_EVT_STATE_DETACHED);
                clockevents_exchange_device(dev, NULL);
                dev->event_handler = clockevents_handle_noop;
                td->evtdev = NULL;

Regards
Bibo Mao
> Thanks,
>
>          tglx