RE: [PATCH V1 1/2] Xen acpi memory hotplug driver

["Liu, Jinsong" <jinsong.liu@intel.com>]

Konrad Rzeszutek Wilk wrote:
> On Thu, Dec 06, 2012 at 04:27:36AM +0000, Liu, Jinsong wrote:
>>>>>> diff --git a/drivers/xen/Kconfig b/drivers/xen/Kconfig index
>>>>>> 126d8ce..abd0396 100644 --- a/drivers/xen/Kconfig
>>>>>> +++ b/drivers/xen/Kconfig
>>>>>> @@ -206,4 +206,15 @@ config XEN_MCE_LOG
>>>>>>  	  Allow kernel fetching MCE error from Xen platform and
>>>>>>  	  converting it into Linux mcelog format for mcelog tools
>>>>>> 
>>>>>> +config XEN_ACPI_MEMORY_HOTPLUG
>>>>>> +	bool "Xen ACPI memory hotplug"
>>>>> 
>>>>> There should be a way to make this a module.
>>>> 
>>>> I have some concerns to make it a module:
>>>> 1. xen and native memhotplug driver both work as module, while we
>>>> need early load xen driver. 
>>>> 2. if possible, a xen stub driver may solve load sequence issue,
>>>>   but it may involve other issues * if xen driver load then unload,
>>>> native driver may have chance to load successfully;
>>> 
>>> The stub driver would still "occupy" the ACPI bus for the memory
>>> hotplug PnP, so I think this would not be a problem.
>>> 
>> 
>> I'm not quite clear your mean here, do you mean it has
>> 1. xen_stub driver + xen_memhoplug driver, then xen_strub driver
>> unload and entirely replaced by xen_memhotplug driver, or 
>> 2. xen_stub driver (w/ stub ops) + xen_memhotplug ops (not driver),
>> then xen_stub driver keep occupying but its stub ops later replaced
>> by xen_memhotplug ops?  
> 
> #2
>> 
>> If in way #1, it has risk that native driver may load (if xen driver
>> unload). 
>> If in way #2, xen_memhotplug ops lose the chance to probe/add/bind
>> existed memory devices (since it's done when driver registerred). 
> 
> Could the stub driver have a queue of events?

If so, why not do 'real' add ops (like our patch did, to build-in xen memory hotplug logic)?
I'm not quite clear your purpose of insisting module -- what's advantage of module you prefer?

> 
>> 
>>>>   * if xen driver load --> unload --> load again, then it will lose
>>>> hotplug notification during unload period;
>>> 
>>> Sure. But I think we can do it with this driver? After all the
>>> function of it is to just tell the firmware to turn on/off sockets
>>> - and if we miss one notification we won't take advantage of the
>>> power savings - but we can do that later on. 
>>> 
>> 
>> Not only inform firmware.
>> Hotplug notify callback will invoke acpi_bus_add -> ... ->
>> implicitly invoke drv->ops.add method to add the hotadded memory
>> device.  
> 
> Gotcha.

? So it will lose the notification and no way to add the new memory device in the future.

Xen memory hotplug logic consist of 2 parts:
1) driver logic (.add/.remove etc)
2) notification install/callback logic
If you want to use 'xen_stub driver + .add/.remove ops', then notification install/callback logic would implement with xen_stub driver (means in build-in part, otherwise it would lose notification when the ops unload) --> but that would make xen_stub in big build-in size.

>> 
>>> 
>>>>   * if xen driver load --> unload --> load again, then it will
>>>> re-add all memory devices, but the handle for 'booting memory
>>>> device' and 'hotplug memory device' are different while we have no
>>>> way to distinguish these 2 kind of devices.
>>> 
>>> Wouldn't the stub driver hold onto that?
>>> 
>> 
>> Same question as comment #1. Do you mean it has a xen_stub driver
>> (w/ stub ops) and a xen_memhotplug ops? 
> 
> Correct.
>> 
>>>> 
>>>> IMHO I think to make xen hotplug logic as module may involves
>>>> unexpected result. Is there any obvious advantages of doing so?
>>>> after all we have provided config choice to user. Thoughts?
>>> 
>>> Yes, it becomes a module - which is what we want.
>>> 
>> 
>> What I meant here is, module will bring some unexpected issues for
>> xen hotplug. 
>> We can provide user 'bool' config choice, let them decide to
>> build-in or not, but not 'tristate' choice. 
> 
> What would be involved in making it an tristate choice?
>> 
>> Thanks
>> Jinsong