Re: acpi_bus_trim does not detach devices in post order

[Rafael J. Wysocki]

On Wednesday, August 07, 2013 09:35:51 AM Yasuaki Ishimatsu wrote:
> (2013/08/06 23:26), Rafael J. Wysocki wrote:
> > On Tuesday, August 06, 2013 07:06:37 PM Yasuaki Ishimatsu wrote:
> >> (2013/08/06 12:07), Yasuaki Ishimatsu wrote:
> >>>
> >>> I acked the following commit but I hit a problem by the commit.
> >>> So I report it.
> >>>
> >>> commit cecdb193c8d91a42d9489d00618cc3dfff92e55a
> >>> Author: Rafael J. Wysocki <rafael.j.wysocki@xxxxxxxxx>
> >>> Date:   Tue Jan 15 13:24:02 2013 +0100
> >>>
> >>>       ACPI / scan: Change the implementation of acpi_bus_trim()
> >>>
> >>> Before applying the commit, acpi_bus_trim() detachs devices in post order.
> >>>
> >>> When I hot add memory devices and processor devices by container device
> >>> in my x86 box, memory devices are added first and processor devices are added
> >>> second. So I expect that processor devices are removed first and memory
> >>> devices are removed second when I remove them. And before applying the
> >>> commit, acpi_bus_trim() did so.
> >>>
> >>> But after appling the commit, acpi_bus_trim() does not detach devices in
> >>> post order. So when I remove them, memory devices are removed first and
> >>> processor devices are removed second.
> >>>
> >>> By this, I hit a problem.
> >>>
> >>> In Linux on x86 arch, NUMA node is depend on memory devices. So new NUMA
> >>> node is created at memory hot adding. Thus when I hot add memory devices and
> >>> processor devices, we must hot add memory device first. Otherwise, processor
> >>> devices are not set to correct NUMA node number.
> >>>
> >>> And Linux expects that when removing them, processor devices are removed
> >>> first before removing memory devices. But acpi_bus_trim() does not do so.
> >>> By this, NUMA node is not cleared in my x86 box when hot removing memory device
> >>> and processor devices. When removing memory devices, NUMA node is cleared.
> >>> But if there are processor devices related with the NUMA node, NUMA node is
> >>> not be cleared at memory hot removing.
> >>>
> >>> So when I remove them, NUMA node's sysfs file remained as follows:
> >>
> >> I had little mistake. CPU also tries to clear NUMA node.
> >> But current implementation has bug. So I'll fix it.
> >
> 
> > Do I understand correctly that with your fix at
> >
> > https://patchwork.kernel.org/patch/2839298/
> >
> > the current acpi_bus_trim() implementation will be sufficient?
> 
> No. The patch just fixed implementation of CPU hotplug.

My question was not sufficiently precise. :-)

I wanted to ask if your patch was sufficient to address the specific breakage
you were seeing without modifying acpi_bus_trim().

> A problem I think is that acpi_bus_trim() does not detach devices in
> post-order.

That is not exactly post-order, but children are guaranteed to be processed
before their parents.  If that guarantee is sufficient, there's no problem.
Otherwise, acpi_bus_trim() may need to be modified, but first I'd like to
see a real life example where that really matters.

> And my patch does not fix it. So if some device has dependency
> of other device, similar problem will occur.

If there is a dependency that is not a parent-child one, we'll have a problem,
but in that case relying on ordering will not be robust enough anyway in my
opinion.

Thanks,
Rafael


-- 
I speak only for myself.
Rafael J. Wysocki, Intel Open Source Technology Center.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/