Re: [PATCH] Do not use cpu_to_node() to find an offlined cpu's node.

[Wen Congyang <wency@cn.fujitsu.com>]

At 10/09/2012 06:04 PM, David Rientjes Wrote:
> On Tue, 9 Oct 2012, Tang Chen wrote:
> 
>>>> Eek, the nid shouldn't be -1 yet, though, for cpu hotplug since this
>>>> should be called at CPU_DYING level and migrate_tasks() still sees a valid
>>>> cpu.
>>
>> As Wen said below, nid is now set to -1 when cpu is hotremoved.
>> I reproduce this problem in this situation:
>>
>> all cpus are online, and hot remove a system board directorily, without
>> offlining any cpu.
>>
>> As a result, the removed cpu's nid is set to -1, and this causes
>> problems.
>>
> 
> Let's add Andrew to the cc list then, because I'm nacking 
> cpu_hotplug-unmap-cpu2node-when-the-cpu-is-hotremoved.patch in the -mm 
> tree for this reason.
> 
> We can only clear a cpu-to-node mapping when the cpu is completely 
> offline, not before or during the CPU_DYING stage.  Kernel code, such as 

I clear cpu-to-node mapping when the cpu is hotremoved. If the cpu is onlined,
it will be offlined before clearing cpu-to-node mapping.

Here is the code in driver/acpi/processor_driver.c:
=============
static int acpi_processor_handle_eject(struct acpi_processor *pr)
{
	if (cpu_online(pr->id))
		cpu_down(pr->id);      <========== cpu is offlined here.

	arch_unregister_cpu(pr->id);
	acpi_unmap_lsapic(pr->id);     <========== I clear the mapping here.
	return (0);
}
=============

The real problem is that: we don't migrate this task to another cpu when
the cpu is offlined. I guess this task is not in running state, and it
is not in the cpu's runqueue.

Thanks
Wen Congyang

> the sched code that you are now trying to "fix", depends on this mapping 
> to work correctly; obviously no audit was done of cpu hotplug code 
> depending on it before the patch was proposed.
> 
> I say "fix" because even this workaround isn't a good solution since it 
> would be much better to pick another cpu on the same node as the offlining 
> cpu for the runqueue before falling back to the set of all allowed nodes.  
> We lose all NUMA affinity information with that patch.  There's no reason 
> why we shouldn't know the node of a cpu that is being offlined.
> 
> So nack to cpu_hotplug-unmap-cpu2node-when-the-cpu-is-hotremoved.patch.  
> After it's removed because it's buggy, this "fix" will no longer be 
> necessary.
> 

This patch is try to fix a bug: the kernel will be panicked after removing a
node.