[PATCH] kexec: disable cpu hotplug until the rebooting cpu is stable

[Pingfan Liu]

On Wed, Jan 26, 2022 at 11:06 PM Valentin Schneider
<valentin.schneider at arm.com> wrote:
>
> On 26/01/22 10:45, Pingfan Liu wrote:
> > On Wed, Jan 26, 2022 at 12:29 AM Valentin Schneider
> > <valentin.schneider at arm.com> wrote:
> >>
> >> On 25/01/22 11:39, Pingfan Liu wrote:
> >> > The following identical code piece appears in both
> >> > migrate_to_reboot_cpu() and smp_shutdown_nonboot_cpus():
> >> >
> >> >       if (!cpu_online(primary_cpu))
> >> >               primary_cpu = cpumask_first(cpu_online_mask);
> >> >
> >> > Although the kexec-reboot task can get through a cpu_down() on its cpu,
> >> > this code looks a little confusing.
> >> >
> >> > Make things straight forward by keep cpu hotplug disabled until
> >> > smp_shutdown_nonboot_cpus() holds cpu_add_remove_lock. By this way, the
> >> > rebooting cpu can keep unchanged.
> >> >
> >>
> >> So is this supposed to be a refactor with no change in behaviour? AFAICT it
> >> actually does change things (and isn't necessarily clearer).
> >>
> > Yes, as you have seen, it does change behavior. Before this patch,
> > there is a breakage:
> >   migrate_to_reboot_cpu();
> >   cpu_hotplug_enable();
> >                                      ----------> technical, here can
> > comes a cpu_down(this_cpu)
> >   machine_shutdown();
> >
> > And this patch squeezes out this breakage.
> >
>
> Ok, that's worth pointing out in the changelog.
>
Sure, I will update it.

> >> > diff --git a/kernel/kexec_core.c b/kernel/kexec_core.c
> >> > index 68480f731192..db4fa6b174e3 100644
> >> > --- a/kernel/kexec_core.c
> >> > +++ b/kernel/kexec_core.c
> >> > @@ -1168,14 +1168,12 @@ int kernel_kexec(void)
> >> >               kexec_in_progress = true;
> >> >               kernel_restart_prepare("kexec reboot");
> >> >               migrate_to_reboot_cpu();
> >> > -
> >> >               /*
> >> > -              * migrate_to_reboot_cpu() disables CPU hotplug assuming that
> >> > -              * no further code needs to use CPU hotplug (which is true in
> >> > -              * the reboot case). However, the kexec path depends on using
> >> > -              * CPU hotplug again; so re-enable it here.
> >> > +              * migrate_to_reboot_cpu() disables CPU hotplug. If an arch
> >> > +              * relies on the cpu teardown to achieve reboot, it needs to
> >> > +              * re-enable CPU hotplug there.
> >> >                */
> >> > -             cpu_hotplug_enable();
> >> > +
> >>
> >> Not all archs map machine_shutdown() to smp_shutdown_nonboot_cpus(), other
> >> archs will now be missing a cpu_hotplug_enable() prior to a kexec
> >> machine_shutdown(). That said, AFAICT none of those archs rely on the
> >> hotplug machinery in machine_shutdown(), so it might be OK, but that's not
> >> obvious at all.
> >>
> > At the first glance, it may be not obvious, but tracing down
> > cpu_hotplug_enable() to the variable cpu_hotplug_disabled, you can
> > find out the limited involved functions are all related to
> > cpu_up/cpu_down.
> >
> > IOW, if no code path connects with the interface of cpu_up/cpu_down,
> > then kexec-reboot will not be affected.
> >
>
> That's my point, this only works if the other archs truly do not rely on
> hotplug for machine_shutdown(), which seems to be the case but it wouldn't
> hurt for you to double-check that and explicitely call it out in the
> changelog.
>
OK, I will update the change log. BTW, besides x86, I have just
finished the test on powerpc, both of them works fine with this patch

> > And after this patch, it is more clear how to handle the following cases:
> > arch/arm/kernel/reboot.c:94:    smp_shutdown_nonboot_cpus(reboot_cpu);
> > arch/arm64/kernel/process.c:88: smp_shutdown_nonboot_cpus(reboot_cpu);
> > arch/ia64/kernel/process.c:578: smp_shutdown_nonboot_cpus(reboot_cpu);
> >
>
> FWIW riscv is also concerned.

I think its current statement is right.
arch/riscv/kernel/machine_kexec.c:135:
smp_shutdown_nonboot_cpus(smp_processor_id());

Thanks,

Pingfan
>
> > Thanks,
> > Pingfan