Re: [PATCH 7/7] ACPI / scan: Make memory hotplug driver use struct acpi_scan_handler

[Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com>]

Hi Vasilis,

2013/02/20 19:42, Vasilis Liaskovitis wrote:
> Hi Yasuaki,
>
> On Wed, Feb 20, 2013 at 12:35:48PM +0900, Yasuaki Ishimatsu wrote:
>> Hi Vasilis,
>>
>> 2013/02/20 3:11, Vasilis Liaskovitis wrote:
>>> Hi,
>>>
>>> On Sun, Feb 17, 2013 at 04:27:18PM +0100, Rafael J. Wysocki wrote:
>>>> From: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
>>>>
>>>> Make the ACPI memory hotplug driver use struct acpi_scan_handler
>>>> for representing the object used to set up ACPI memory hotplug
>>>> functionality and to remove hotplug memory ranges and data
>>>> structures used by the driver before unregistering ACPI device
>>>> nodes representing memory.  Register the new struct acpi_scan_handler
>>>> object with the help of acpi_scan_add_handler_with_hotplug() to allow
>>>> user space to manipulate the attributes of the memory hotplug
>>>> profile.
>>>
>>> Let's consider an example where we want acpi memory device ejection to be safely
>>> handled by userspace. We do the following:
>>>
>>> echo 0 > /sys/firmware/acpi/hotplug/memory/autoeject
>>> echo 1 > /sys/firmware/acpi/hotplug/memory/uevents
>>>
>>> We succesfully hotplug acpi device:
>>> /sys/devices/LNXSYSTM:00/LNXSYSBUS:00/PNP0C80:00
>>> and its corresponding memblocks /sys/devices/system/memory/memoryXX are
>>> also successfully onlined.
>>>
>>> On an eject request, since uevents == 1, the kernel will emit KOBJ_OFFLINE for:
>>> /sys/devices/LNXSYSTM:00/LNXSYSBUS:00/PNP0C80:00
>>>
>>> Can userspace know which memblocks in /sys/devices/system/memory/memoryXX/
>>> correspond to the acpi device /sys/devices/LNXSYSTM:00/LNXSYSBUS:00/PNP0C80:00 ?
>>> This will be needed so that userspace tries to offline the memblocks (and only
>>> if successful, issue the eject operation on the acpi device). As far as I see,
>>> we don't create any sysfs links or files for this scenario - can userspace get
>>> this info somehow?
>>
>>>
>>> /sys/devices/system/memory/memoryXX/phys_device needs to be properly implemented
>>> for this to work I think, see Documentation/ABI/testing/sysfs-memory
>>>
>>> The following test patch works toward that direction. Let me know if it's of
>>> interest or if there are better ideas /comments.
>>
>> How about use ../PNP0C80:00/physical_node/resources file?
>> In my system, the file shows following information.
>>
>> $ cat /sys/bus/acpi/devices/PNP0C80\:00/physical_node/resources
>> state = active
>> mem 0x0-0x80000000
>> mem 0x100000000-0x800000000
>>
>> It means PNP0C80:00's memory ranges are "0x0-0x7fffffff" and
>> "0x100000000-0x7ffffffff". In x86 architecture, memory section size is
>> 128MiB. So, if these memory range is divided by 128MiB, you can
>> calculate memory section number as follow:
>>
>> 0x0-0x7fffffff => 0x0-0x10
>> 0x100000000-0x7ffffffff => 0x20-0xff
>>
>> But there is one problem. The problem is that resources file of added memory
>> is not created. If the problem is fixed, I think you can use the way.
>

> thanks for your suggestion. Is this resources file a property of the
> physical_node or of each acpi devices?
>
> If it's a node specific file could there be a chance that adjacent memory
> ranges get merged? We 'd like these to not get merged.

This information is created by pnppacpi_init().
It seems that:
   - resources file is created to each acpi_devices.
   - the memory range does not get merged.

Thanks,
Yasuaki Ishimatsu
>
> I will look more into this property. I don't see it currently in my system
> (probably because initial memory is not backed by acpi devices in my
>   seabios/virtual machine).
>
>>
> [...]
>>> +int acpi_memory_phys_device(unsigned long start_pfn)
>>> +{
>>> +	struct acpi_memory_device *mem_dev;
>>> +	struct acpi_memory_info *info;
>>> +	unsigned long start_addr = start_pfn << PAGE_SHIFT;
>>> +	int id = 0;
>>> +
>>> +	list_for_each_entry(mem_dev, &acpi_mem_device_list, mem_device_list) {
>>> +		list_for_each_entry(info, &mem_dev->res_list, list) {
>>> +			if ((info->start_addr <= start_addr) &&
>>> +				(info->start_addr + info->length > start_addr))
>>> +				return id;
>>> +		}
>>> +		id++;
>>> +	}
>>
>> I don't think this solve your problem.
>>
>> When hot adding memory device in my system, consecutive index number is
>> applied to PNP0C80 as follows:
>>
>> $ ls /sys/bus/acpi/devices/ |grep PNP0C80
>> PNP0C80:00
>> PNP0C80:01  => hot added memory device
>> PNP0C80:02  => hot added memory device
>>
>> In this case, we can know PNP0C80:YY by memoryXX/phys_device file.
>> But if hot removing and adding the same device, index number is changed
>> as follows:
>>
>> $ ls /sys/bus/acpi/devices/
>> PNP0C80:00
>> PNP0C80:03  => hot added memory device
>> PNP0C80:04  => hot added memory device
>>
>> In this case, we cannot know PNP0C80:YY by memoryXX/phys_device file.
>>
>
> thanks, yes you are right. I forgot each new hotplug event will create a new
> PNP0C80:XX device where XX always increases. So the hot-add/hot-remove/hot-add
> scenario would have a problem.
> Then it would be enough to be able to return this monotonically increasing
> counter from phys_device instead of the current list iterator. Is this counter
> available somehwere in drivers/acpi/scan.c or bus.c? I 'll take a look.
>
> thanks,
>
> - Vasilis
>