On 12/20/2012 03:49 AM, Will Deacon wrote:
> Hi Stephen,
>
> On Wed, Dec 19, 2012 at 11:44:41PM +0000, Stephen Warren wrote:
>> From: Stephen Warren <swarren at nvidia.com>
>>
>> Both the regular reboot and shutdown paths do this. It seems reasonable
>> for kexec to work the same way.
>>
>> On the Tegra ARM SoC at least, this change allows kexec to work with SMP
>> enabled. ARM's machine_shutdown() simply puts all CPUs into a loop. If
>> the code of that loop is over-written, the CPUs may crash, and cause the
>> kexec'd kernel not to be able to initialize them. In practice, this
>> causes the kexec'd kernel to hang and/or crash. The intended way to
>> solve this is for ARM machines to provide a cpu_kill SMP operation to
>> e.g. power down the CPUs, or place them in reset. However, at least on
>> Tegra, the implementation of that function would simply be duplicating
>> the hotplug code that already exists, so it seems simpler to just call
>> disable_nonboot_cpus() for the kexec path, just like reboot/shutdown.
>
> If you do manage to get this merged, please can you follow up with a patch
> to remove the smp_kill_cpus bits from arch/arm/kernel/smp.c please? It only
> exists as a hook to do exactly this and currently nobody is using it afaict.
I originally implemented this in
arch/arm/kernel/process.c:machine_shutdown(), which currently is:
void machine_shutdown(void)
{
#ifdef CONFIG_SMP
smp_send_stop();
#endif
}
and I changed it to something like:
void machine_shutdown(void)
{
#ifdef CONFIG_HOTPLUG_CPU
disable_nonboot_cpus();
#elifdef CONFIG_SMP
smp_send_stop();
#endif
}
... but then figured that moving it up into the core kexec code would be
better, so that everything always worked the same way.
Anyway, the change above addresses Eric's concern about isolating the
change to ARM. Does that seem like a reasonable thing for the ARM code
to do?
