Re: [PATCH] ACPI, cpu hotplug: move try_offline_node() after acpi_unmap_lsapic()

[Toshi Kani <toshi.kani@hp.com>]

On Sat, 2013-08-10 at 01:29 +0200, Rafael J. Wysocki wrote:
> On Friday, August 09, 2013 04:16:56 PM Toshi Kani wrote:
> > On Fri, 2013-08-09 at 15:28 +0800, Tang Chen wrote:
> > > On 08/07/2013 12:56 AM, Toshi Kani wrote:
> > > > On Tue, 2013-08-06 at 19:11 +0900, Yasuaki Ishimatsu wrote:
> > > >> try_offline_node() checks that all cpus related with removed node have been
> > > >> removed by using cpu_present_bits. If all cpus related with removed node have
> > > >> been removed, try_offline_node() clears the node information.
> > > >>
> > > >> But try_offline_node() called from acpi_processor_remove() never clears
> > > >> the node information. For disabling cpu_present_bits, acpi_unmap_lsapic()
> > > >> need be called. But acpi_unmap_lsapic() is called after try_offline_node()
> > > >> runs. So when try_offline_node() runs, the cpu's cpu_present_bits is always
> > > >> set.
> > > >>
> > > >> This patch moves try_offline_node() after acpi_unmap_lsapic().
> > > >>
> > > >> Signed-off-by: Yasuaki Ishimatsu<isimatu.yasuaki@jp.fujitsu.com>
> > > >
> > > > The change looks good to me.
> > > >
> > > > Acked-by: Toshi Kani<toshi.kani@hp.com>
> > > >
> > > > BTW, do you know why try_offline_node() has to use stop_machine()?
> > > 
> > > try_offline_node() is used to check if the node could be hot-removed
> > > after each memory or cpu hot-remove operation.
> > > 
> > > In memory hot-remove path, we have lock_memory_hotplug() to series all
> > > the memory hot-remove options.
> > > 
> > > But when doing cpu hot-remove,
> > > 
> > > acpi_processor_remove()
> > >   |->try_offline_node()
> > > 
> > > There is no lock to protect it. I think, when we are going to hot-remove
> > > a node, others should not do any memory or cpu hotplug operation. In memory
> > > hotplug path, we have lock_memory_hotplug(). But in cpu hotplug path, I
> > > didn't find any lock. So we used stop_machine() to call check_cpu_on_node().
> > > If we find any cpu still present, we return and do not remove the node.
> > 
> > CPU/Memory hotplug operations and sysfs eject are serialized with
> > acpi_os_hotplug_execute().  CPU online/offline is protected by
> > cpu_hotplug_[begin|done]() and [get|put]_online_cpus().  But, yes,
> > online/offline and hotplug operations are not serialized.  I tried to
> > serialize them before, but that framework was not well received.
> 
> What about lock_device_hotplug()?  It is taken by both online/offline and
> the ACPI hotplug code, isn't it?

Oh, that's right!  I forgot about this one.  Yes, lock_device_hotplug()
nicely protects online/offline and hotplug operations. :-)

> > Anyway, it looks to me that cpu_up()->mem_online_node() path can race
> > with try_offline_node().
> 
> It can in principle, but I'm not sure if there's a way to trigger that
> race.  Do you have an example?

With lock_device_hotplug(), I agree that we do not have this race
condition -- cpu_up() may not run while other hotplug is running.
store_online() will be blocked at lock_device_hotplug() in such case.
When store_online() acquired the lock, this CPU may have been deleted.
So, we still need to make sure that this case is handled properly.  I
suppose sysfs keeps *dev valid with ref_count (Is that right?).  I think
cpu_up() needs to check with cpu_present(), not cpu_possible(), at the
top.  Otherwise, cpu_to_node(cpu) may return NUMA_NO_NODE (-1), which is
probably not a good value for node_online(nid).

Thanks,
-Toshi